blob: 0d6e5bfeaef86fbdc83bbbe368d4d91ee8fd1f5c [file] [log] [blame]
Frank Barchard1a953052020-11-16 18:44:58 -08001# Copyright 2020 Google LLC
Marat Dukhan08c4a432019-10-03 09:29:21 -07002#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
Marat Dukhan13b68f22020-11-12 11:55:19 -080062 "src/operators/depth-to-space-nchw2nhwc.c",
Marat Dukhan0e521172020-11-25 13:10:04 -080063 "src/operators/depth-to-space-nhwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070064 "src/operators/fully-connected-nc.c",
65 "src/operators/global-average-pooling-ncw.c",
66 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070067 "src/operators/leaky-relu-nc.c",
68 "src/operators/max-pooling-nhwc.c",
69 "src/operators/prelu-nc.c",
Artsiom Ablavatski97918102020-10-27 15:52:59 -070070 "src/operators/resize-bilinear-nchw.c",
Marat Dukhane8265432020-04-28 18:42:59 -070071 "src/operators/resize-bilinear-nhwc.c",
72 "src/operators/sigmoid-nc.c",
73 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070074 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070075 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070076]
77
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070078SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070079 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070080 "src/subgraph/add2.c",
81 "src/subgraph/argmax-pooling-2d.c",
82 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070083 "src/subgraph/bankers-rounding.c",
84 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070085 "src/subgraph/clamp.c",
86 "src/subgraph/convolution-2d.c",
87 "src/subgraph/deconvolution-2d.c",
Artsiom Ablavatskibbe85062020-11-05 14:07:37 -080088 "src/subgraph/depth-to-space.c",
Frank Barchard9cef5ea2020-11-18 14:52:08 -080089 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070090 "src/subgraph/divide.c",
Marat Dukhana1600202020-12-01 22:17:16 -080091 "src/subgraph/elu.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070092 "src/subgraph/floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -070093 "src/subgraph/fully-connected.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070094 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070096 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070098 "src/subgraph/maximum2.c",
99 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700100 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700101 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700102 "src/subgraph/prelu.c",
103 "src/subgraph/sigmoid.c",
104 "src/subgraph/softmax.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700105 "src/subgraph/square-root.c",
106 "src/subgraph/square.c",
107 "src/subgraph/squared-difference.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700108 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700109 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700110 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700111 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700112 "src/subgraph/unpooling-2d.c",
113]
114
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800115TABLE_SRCS = [
116 "src/tables/exp2-k-over-64.c",
117 "src/tables/exp2-k-over-2048.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800118 "src/tables/exp2minus-k-over-4.c",
119 "src/tables/exp2minus-k-over-8.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800120 "src/tables/exp2minus-k-over-16.c",
Marat Dukhan1f256fc2020-09-24 21:27:14 -0700121 "src/tables/exp2minus-k-over-64.c",
122 "src/tables/exp2minus-k-over-2048.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800123]
124
Marat Dukhan08c4a432019-10-03 09:29:21 -0700125SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800126 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800127 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800128 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700129 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
130 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700131 "src/f32-clamp/gen/scalar-x1.c",
132 "src/f32-clamp/gen/scalar-x2.c",
133 "src/f32-clamp/gen/scalar-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700135 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700136 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700137 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700138 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
139 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
140 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700141 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700142 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
143 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
144 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700145 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700146 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
147 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
148 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
151 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
152 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700153 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700154 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
155 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
156 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700157 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700158 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
159 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
160 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700161 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
162 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
163 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700164 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700165 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
Marat Dukhan91249d22020-10-24 12:02:51 -0700166 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
167 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
168 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
169 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
170 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700171 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
173 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700174 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
Marat Dukhancf5b3c32020-10-25 19:21:10 -0700175 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700176 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
177 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
178 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700179 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
180 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
181 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
182 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700183 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700184 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
185 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700186 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
Marat Dukhanc4efb002020-10-25 23:14:47 -0700187 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700188 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
Marat Dukhan29c0c332020-10-28 22:11:00 -0700189 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
190 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
191 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
192 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
193 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
194 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
195 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
196 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
197 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
198 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700199 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700200 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
201 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700202 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
203 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
204 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700205 "src/f32-gemm/gen/1x4-minmax-scalar.c",
206 "src/f32-gemm/gen/1x4-relu-scalar.c",
207 "src/f32-gemm/gen/1x4-scalar.c",
208 "src/f32-gemm/gen/2x4-minmax-scalar.c",
209 "src/f32-gemm/gen/2x4-relu-scalar.c",
210 "src/f32-gemm/gen/2x4-scalar.c",
211 "src/f32-gemm/gen/4x2-minmax-scalar.c",
212 "src/f32-gemm/gen/4x2-relu-scalar.c",
213 "src/f32-gemm/gen/4x2-scalar.c",
214 "src/f32-gemm/gen/4x4-minmax-scalar.c",
215 "src/f32-gemm/gen/4x4-relu-scalar.c",
216 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700217 "src/f32-hswish/gen/hswish-scalar-x1.c",
218 "src/f32-hswish/gen/hswish-scalar-x2.c",
219 "src/f32-hswish/gen/hswish-scalar-x4.c",
XNNPACK Team21432672020-10-19 19:58:48 -0700220 "src/f32-ibilinear-chw/gen/scalar-p1.c",
221 "src/f32-ibilinear-chw/gen/scalar-p2.c",
222 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700223 "src/f32-ibilinear/gen/scalar-c1.c",
224 "src/f32-ibilinear/gen/scalar-c2.c",
225 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700226 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700227 "src/f32-igemm/gen/1x4-relu-scalar.c",
228 "src/f32-igemm/gen/1x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700229 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700230 "src/f32-igemm/gen/2x4-relu-scalar.c",
231 "src/f32-igemm/gen/2x4-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700232 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700233 "src/f32-igemm/gen/4x2-relu-scalar.c",
234 "src/f32-igemm/gen/4x2-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700235 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700236 "src/f32-igemm/gen/4x4-relu-scalar.c",
237 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700238 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
239 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
240 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700241 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
242 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
243 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
244 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800245 "src/f32-prelu/gen/scalar-2x1.c",
246 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800247 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800248 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700249 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800250 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
251 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700252 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800253 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800254 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700255 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800256 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
257 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700258 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700259 "src/f32-relu/gen/scalar-x1.c",
260 "src/f32-relu/gen/scalar-x2.c",
261 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700262 "src/f32-relu/gen/scalar-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700263 "src/f32-rmax/scalar.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800264 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
265 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
266 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700267 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
268 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
269 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800270 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
271 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
272 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700273 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
274 "src/f32-spmm/gen/1x1-minmax-scalar.c",
275 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
276 "src/f32-spmm/gen/2x1-minmax-scalar.c",
277 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
278 "src/f32-spmm/gen/4x1-minmax-scalar.c",
279 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
280 "src/f32-spmm/gen/8x1-minmax-scalar.c",
281 "src/f32-spmm/gen/8x2-minmax-scalar.c",
282 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700283 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
284 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
285 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700287 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
288 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
289 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700291 "src/f32-vbinary/gen/vadd-scalar-x1.c",
292 "src/f32-vbinary/gen/vadd-scalar-x2.c",
293 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700303 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
304 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
305 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700307 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
308 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
309 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700311 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
312 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
313 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700315 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
316 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
317 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700319 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
320 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
321 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700323 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
324 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
325 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700327 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
328 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
329 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800331 "src/f32-vbinary/gen/vmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800335 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
336 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
337 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800339 "src/f32-vbinary/gen/vmin-scalar-x1.c",
340 "src/f32-vbinary/gen/vmin-scalar-x2.c",
341 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800343 "src/f32-vbinary/gen/vminc-scalar-x1.c",
344 "src/f32-vbinary/gen/vminc-scalar-x2.c",
345 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700347 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
348 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
349 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700351 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
352 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
353 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700355 "src/f32-vbinary/gen/vmul-scalar-x1.c",
356 "src/f32-vbinary/gen/vmul-scalar-x2.c",
357 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
368 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
369 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
380 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
381 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700383 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
384 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
385 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700386 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700387 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
388 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
389 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700390 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700391 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
392 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
393 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700394 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700395 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
396 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
397 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700398 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700399 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
400 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
401 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700402 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700403 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
404 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
405 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700406 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700407 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
408 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
409 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700410 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700411 "src/f32-vbinary/gen/vsub-scalar-x1.c",
412 "src/f32-vbinary/gen/vsub-scalar-x2.c",
413 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700414 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700415 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
416 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
417 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700418 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700419 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
420 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
421 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700422 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700423 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
424 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
425 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700426 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800427 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
428 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
429 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
430 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
431 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
432 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
433 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
434 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
435 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
436 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
437 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
438 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700439 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
440 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
441 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700442 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
443 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
444 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700445 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
446 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
447 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700448 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
449 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
450 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
451 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
452 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
453 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
454 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
455 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
456 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700457 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
458 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
459 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700460 "src/f32-vunary/gen/vabs-scalar-x1.c",
461 "src/f32-vunary/gen/vabs-scalar-x2.c",
462 "src/f32-vunary/gen/vabs-scalar-x4.c",
463 "src/f32-vunary/gen/vneg-scalar-x1.c",
464 "src/f32-vunary/gen/vneg-scalar-x2.c",
465 "src/f32-vunary/gen/vneg-scalar-x4.c",
466 "src/f32-vunary/gen/vsqr-scalar-x1.c",
467 "src/f32-vunary/gen/vsqr-scalar-x2.c",
468 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhande390d42020-11-29 19:32:18 -0800469 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
470 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
471 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
Marat Dukhanc60742b2020-11-23 12:33:27 -0800472 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
473 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
474 "src/math/expm1minus-scalar-rr2-p5.c",
475 "src/math/expm1minus-scalar-rr2-p6.c",
Frank Barchard22136062020-11-24 18:44:46 -0800476 "src/math/expminus-scalar-rr2-lut64-p2.c",
477 "src/math/expminus-scalar-rr2-lut2048-p1.c",
478 "src/math/expminus-scalar-rr2-p5.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700479 "src/math/roundd-scalar-addsub.c",
480 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700481 "src/math/roundd-scalar-floor.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700482 "src/math/roundne-scalar-addsub.c",
483 "src/math/roundne-scalar-nearbyint.c",
484 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700485 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700486 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700487 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700488 "src/math/roundz-scalar-addsub.c",
489 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700490 "src/math/roundz-scalar-trunc.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700491 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700492 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -0700493 "src/math/sigmoid-scalar-rr2-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700494 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
495 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700496 "src/qs8-requantization/fp32-scalar-lrintf.c",
497 "src/qs8-requantization/fp32-scalar-magic.c",
498 "src/qs8-requantization/precise-scalar-signed64.c",
499 "src/qs8-requantization/precise-scalar-unsigned32.c",
500 "src/qs8-requantization/precise-scalar-unsigned64.c",
501 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700502 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
503 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
504 "src/qu8-dwconv/up1x9-minmax-scalar.c",
505 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
506 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700507 "src/qu8-gemm/2x2-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700508 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
509 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700510 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700511 "src/qu8-requantization/fp32-scalar-lrintf.c",
512 "src/qu8-requantization/fp32-scalar-magic.c",
513 "src/qu8-requantization/precise-scalar-signed64.c",
514 "src/qu8-requantization/precise-scalar-unsigned32.c",
515 "src/qu8-requantization/precise-scalar-unsigned64.c",
516 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700517 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700518 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700519 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700520 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700521 "src/u8-rmax/scalar.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700522 "src/x8-lut/scalar.c",
523 "src/x8-zip/x2-scalar.c",
524 "src/x8-zip/x3-scalar.c",
525 "src/x8-zip/x4-scalar.c",
526 "src/x8-zip/xm-scalar.c",
Marat Dukhanad71b9a2020-11-20 00:01:51 -0800527 "src/x32-depthtospace2d-chw2hwc/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700528 "src/x32-fill/scalar-float.c",
529 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700530 "src/x32-packx/x2-scalar.c",
531 "src/x32-packx/x3-scalar.c",
532 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700533 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700534 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700535 "src/x32-unpool/scalar.c",
536 "src/x32-zip/x2-scalar.c",
537 "src/x32-zip/x3-scalar.c",
538 "src/x32-zip/x4-scalar.c",
539 "src/x32-zip/xm-scalar.c",
Marat Dukhan048931b2020-11-24 20:53:54 -0800540 "src/xx-copy/memcpy.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700541]
542
Marat Dukhan436ebe62019-12-04 15:10:12 -0800543WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700544 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
545 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700546 "src/f32-clamp/gen/wasm-x1.c",
547 "src/f32-clamp/gen/wasm-x2.c",
548 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700549 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
550 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700551 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
552 "src/f32-dwconv/gen/up1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700553 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
554 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700555 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
556 "src/f32-dwconv/gen/up1x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700557 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
558 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700559 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
560 "src/f32-dwconv/gen/up1x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700561 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
562 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700563 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
564 "src/f32-dwconv/gen/up2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700565 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
566 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700567 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
568 "src/f32-dwconv/gen/up2x9-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700569 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
570 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700571 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
572 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700573 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
574 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700575 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
576 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
577 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
578 "src/f32-gemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700579 "src/f32-gemm/gen/1x4-relu-wasm.c",
580 "src/f32-gemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700581 "src/f32-gemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700582 "src/f32-gemm/gen/2x4-relu-wasm.c",
583 "src/f32-gemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700584 "src/f32-gemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700585 "src/f32-gemm/gen/4x2-relu-wasm.c",
586 "src/f32-gemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700587 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700588 "src/f32-gemm/gen/4x4-relu-wasm.c",
589 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700590 "src/f32-hswish/gen/hswish-wasm-x1.c",
591 "src/f32-hswish/gen/hswish-wasm-x2.c",
592 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700593 "src/f32-igemm/gen/1x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700594 "src/f32-igemm/gen/1x4-relu-wasm.c",
595 "src/f32-igemm/gen/1x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700596 "src/f32-igemm/gen/2x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700597 "src/f32-igemm/gen/2x4-relu-wasm.c",
598 "src/f32-igemm/gen/2x4-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700599 "src/f32-igemm/gen/4x2-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700600 "src/f32-igemm/gen/4x2-relu-wasm.c",
601 "src/f32-igemm/gen/4x2-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700602 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700603 "src/f32-igemm/gen/4x4-relu-wasm.c",
604 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700605 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
606 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
607 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700608 "src/f32-prelu/gen/wasm-2x1.c",
609 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700610 "src/f32-relu/gen/wasm-x1.c",
611 "src/f32-relu/gen/wasm-x2.c",
612 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700613 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700622 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
623 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
624 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
625 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700630 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
631 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
632 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
633 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700638 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
639 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
640 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
641 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700642 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
643 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
644 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700645 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800646 "src/f32-vbinary/gen/vmax-wasm-x1.c",
647 "src/f32-vbinary/gen/vmax-wasm-x2.c",
648 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700649 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800650 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
651 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
652 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700653 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800654 "src/f32-vbinary/gen/vmin-wasm-x1.c",
655 "src/f32-vbinary/gen/vmin-wasm-x2.c",
656 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700657 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800658 "src/f32-vbinary/gen/vminc-wasm-x1.c",
659 "src/f32-vbinary/gen/vminc-wasm-x2.c",
660 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700661 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700662 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
663 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
664 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700665 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700666 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
667 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
668 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700669 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700670 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
671 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
672 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
673 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700674 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
675 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
676 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700677 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700678 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
679 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
680 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
681 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700682 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
683 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
684 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700685 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700686 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
687 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
688 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
689 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700690 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
691 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
692 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700693 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700694 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
695 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
696 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
697 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700698 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
699 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
700 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700701 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700702 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
703 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
704 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
705 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700706 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
707 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
708 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700709 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -0800710 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x1.c",
711 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x2.c",
712 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x3.c",
713 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x4.c",
714 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x5.c",
715 "src/f32-velu/gen/velu-wasm-rr2-lut16-p3-x6.c",
716 "src/f32-velu/gen/velu-wasm-rr2-p6-x1.c",
717 "src/f32-velu/gen/velu-wasm-rr2-p6-x2.c",
718 "src/f32-velu/gen/velu-wasm-rr2-p6-x3.c",
719 "src/f32-velu/gen/velu-wasm-rr2-p6-x4.c",
720 "src/f32-velu/gen/velu-wasm-rr2-p6-x5.c",
721 "src/f32-velu/gen/velu-wasm-rr2-p6-x6.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700722 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
723 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
724 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700725 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
726 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
727 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800728]
729
Marat Dukhan290055c2020-06-09 12:24:29 -0700730WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700731 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
732 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
733 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700734 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
735 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
736 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
737 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700738 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
739 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
740 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
741 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Frank Barchard22136062020-11-24 18:44:46 -0800742 "src/f32-conv-hwc2chw/3x3s2p1c3x4-wasmsimd-2x2.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700743 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700744 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700745 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700746 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700747 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
748 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700749 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700750 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700751 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700752 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
753 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
754 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
755 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
756 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
757 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
758 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700759 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700760 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700761 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700762 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
763 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700764 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700765 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700766 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700767 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
768 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
769 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
770 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
771 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
772 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
Frank Barchard3b800452020-11-22 12:12:35 -0800773 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc2.c",
774 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc3.c",
775 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4-acc4.c",
776 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-1x4.c",
777 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4-acc2.c",
778 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-2x4.c",
779 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-3x4.c",
780 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-4x4.c",
781 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-5x4.c",
782 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-arm-6x4.c",
783 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc2.c",
784 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc3.c",
785 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4-acc4.c",
786 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-1x4.c",
787 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4-acc2.c",
788 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-2x4.c",
789 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-3x4.c",
790 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-4x4.c",
791 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-5x4.c",
792 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-wasmsimd-x86-6x4.c",
Frank Barchardb88d0112020-12-04 09:07:13 -0800793 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc2.c",
794 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc3.c",
795 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4-acc4.c",
796 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-1x4.c",
797 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4-acc2.c",
798 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-2x4.c",
799 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-3x4.c",
800 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-arm-4x4.c",
801 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc2.c",
802 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc3.c",
803 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4-acc4.c",
804 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-1x4.c",
805 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4-acc2.c",
806 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-2x4.c",
807 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-3x4.c",
808 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-wasmsimd-x86-4x4.c",
Frank Barchard20a07412020-11-30 23:30:00 -0800809 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc2.c",
810 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc3.c",
811 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc4.c",
812 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4-acc5.c",
813 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-1x4.c",
814 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc2.c",
815 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4-acc3.c",
816 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-2x4.c",
817 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4-acc2.c",
818 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-3x4.c",
819 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4-acc2.c",
820 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-4x4.c",
821 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-arm-5x4.c",
822 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc2.c",
823 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc3.c",
824 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc4.c",
825 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4-acc5.c",
826 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-1x4.c",
827 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc2.c",
828 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4-acc3.c",
829 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-2x4.c",
830 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4-acc2.c",
831 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-3x4.c",
832 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4-acc2.c",
833 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-4x4.c",
834 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-wasmsimd-x86-5x4.c",
Frank Barchard729f07b2020-12-03 16:14:33 -0800835 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc2.c",
836 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc3.c",
837 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc4.c",
838 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4-acc5.c",
839 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-1x4.c",
840 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc2.c",
841 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4-acc3.c",
842 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-2x4.c",
843 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4-acc2.c",
844 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-arm-3x4.c",
845 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc2.c",
846 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc3.c",
847 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc4.c",
848 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4-acc5.c",
849 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-1x4.c",
850 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc2.c",
851 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4-acc3.c",
852 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-2x4.c",
853 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4-acc2.c",
854 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-wasmsimd-x86-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700855 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
856 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700857 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
858 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
859 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
860 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700861 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
862 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
863 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
864 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
865 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
866 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
867 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
868 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
869 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
870 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
871 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
872 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
873 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
874 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
875 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
876 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
877 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
878 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
879 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
880 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
881 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
882 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
883 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
884 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
885 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
886 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
887 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
888 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
889 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
890 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
891 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
892 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
893 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
894 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
895 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
896 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
897 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
898 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
899 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
900 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
901 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
902 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
903 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
904 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
905 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
906 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
907 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
908 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
909 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
910 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
911 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
912 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
913 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
914 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
915 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
916 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
917 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
918 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
919 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
920 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
921 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
922 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
923 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
924 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
925 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
926 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
927 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
928 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
929 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
930 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700931 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
932 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
933 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
XNNPACK Team965272b2020-10-23 21:10:15 -0700934 "src/f32-ibilinear-chw/gen/wasmsimd-p4.c",
935 "src/f32-ibilinear-chw/gen/wasmsimd-p8.c",
Frank Barchard35db7d02020-10-26 13:37:34 -0700936 "src/f32-ibilinear/gen/wasmsimd-c4.c",
937 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -0700938 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
939 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
940 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
941 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
942 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
943 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
944 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
945 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
946 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
947 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
948 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
949 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
950 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
951 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
952 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
953 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
954 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
955 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
956 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
957 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
958 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
959 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
960 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
961 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
962 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
963 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
964 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
965 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
966 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
967 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
968 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
969 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
970 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
971 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
972 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
973 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
974 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
975 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
976 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
977 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700978 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
979 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700980 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
981 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
982 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
983 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700984 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
985 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700986 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
987 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
988 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700989 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
990 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700991 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
992 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
993 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
994 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
995 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
996 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
997 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700998 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
999 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001000 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
1001 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
1002 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
1003 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001004 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001005 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001006 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001007 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
1008 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001009 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001010 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
1011 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001012 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -07001013 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
1014 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001015 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001016 "src/f32-relu/gen/wasmsimd-x4.c",
1017 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -07001018 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -07001019 "src/f32-rmax/wasmsimd-arm.c",
1020 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -07001021 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
1022 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
1023 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
1024 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
1025 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
1026 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001027 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
1028 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
1029 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
1030 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
1031 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
1032 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001033 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined-x2.c",
1034 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001035 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x2.c",
1036 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001037 "src/f32-spmm/gen/4x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001038 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined-x2.c",
1039 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001040 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x2.c",
1041 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001042 "src/f32-spmm/gen/4x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001043 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined-x2.c",
1044 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001045 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x2.c",
1046 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001047 "src/f32-spmm/gen/8x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001048 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined-x2.c",
1049 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001050 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x2.c",
1051 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001052 "src/f32-spmm/gen/8x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001053 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined-x2.c",
1054 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001055 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x2.c",
1056 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001057 "src/f32-spmm/gen/16x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001058 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined-x2.c",
1059 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001060 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x2.c",
1061 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001062 "src/f32-spmm/gen/16x1-minmax-wasmsimd-x86.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001063 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined-x2.c",
1064 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001065 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x2.c",
1066 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001067 "src/f32-spmm/gen/32x1-minmax-wasmsimd-arm.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08001068 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined-x2.c",
1069 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-pipelined.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001070 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x2.c",
1071 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86-x4.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001072 "src/f32-spmm/gen/32x1-minmax-wasmsimd-x86.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001073 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
1074 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001075 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001076 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
1077 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001078 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001079 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
1080 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001081 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001082 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
1083 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001084 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001085 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
1086 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001087 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001088 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
1089 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001090 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001091 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
1092 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001093 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001094 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
1095 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001096 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001097 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
1098 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001099 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001100 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
1101 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001102 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001103 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
1104 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001105 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001106 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
1107 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001108 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001109 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
1110 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001111 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001112 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
1113 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001114 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001115 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
1116 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001117 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001118 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
1119 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001120 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001121 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
1122 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001123 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001124 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
1125 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001126 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001127 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
1128 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001129 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001130 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
1131 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001132 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001133 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
1134 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001135 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001136 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
1137 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001138 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001139 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
1140 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001141 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001142 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
1143 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001144 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001145 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
1146 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001147 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001148 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
1149 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001150 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001151 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
1152 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001153 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001154 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
1155 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001156 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001157 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
1158 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001159 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001160 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
1161 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001162 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001163 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
1164 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001165 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001166 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
1167 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001168 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001169 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
1170 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001171 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001172 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
1173 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001174 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001175 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
1176 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001177 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001178 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
1179 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001180 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001181 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
1182 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001183 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001184 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
1185 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001186 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001187 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
1188 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001189 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001190 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
1191 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001192 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001193 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
1194 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001195 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001196 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
1197 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001198 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001199 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
1200 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001201 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -07001202 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
1203 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001204 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001205 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1206 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001207 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001208 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1209 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001210 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001211 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1212 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001213 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001214 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1215 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001216 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001217 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1218 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001219 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001220 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1221 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001222 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001223 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x4.c",
1224 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x8.c",
1225 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x12.c",
1226 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x16.c",
1227 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x20.c",
1228 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001229 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x4.c",
1230 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x8.c",
1231 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x12.c",
1232 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x16.c",
1233 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x20.c",
1234 "src/f32-velu/gen/velu-wasmsimd-arm-rr2-p6-x24.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001235 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x4.c",
1236 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x8.c",
1237 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x12.c",
1238 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x16.c",
1239 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x20.c",
1240 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-lut16-p3-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001241 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x4.c",
1242 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x8.c",
1243 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x12.c",
1244 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x16.c",
1245 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x20.c",
1246 "src/f32-velu/gen/velu-wasmsimd-x86-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001247 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1248 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1249 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1250 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001251 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001252 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001253 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001254 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001255 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1256 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1257 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1258 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001259 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1260 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1261 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1262 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1263 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1264 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1265 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1266 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1267 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1268 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001269 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1270 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001271 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1272 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1273 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1274 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1275 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1276 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001277 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-andnot.c",
1278 "src/math/expm1minus-wasmsimd-rr2-lut16-p3-max.c",
1279 "src/math/expm1minus-wasmsimd-rr2-p6-andnot.c",
1280 "src/math/expm1minus-wasmsimd-rr2-p6-max.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001281 "src/math/roundd-wasmsimd-addsub.c",
1282 "src/math/roundd-wasmsimd-cvt.c",
1283 "src/math/roundne-wasmsimd-addsub.c",
1284 "src/math/roundu-wasmsimd-addsub.c",
1285 "src/math/roundu-wasmsimd-cvt.c",
1286 "src/math/roundz-wasmsimd-addsub.c",
1287 "src/math/roundz-wasmsimd-cvt.c",
1288 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1289 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001290 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1291 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1292 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001293 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1294 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1295 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001296 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1297 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1298 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001299 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001300 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001301 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001302 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1303 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001304 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001305 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1306 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001307 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1308 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001309 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001310 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001311 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001312 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001313 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001314 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001315 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001316 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1317 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1318 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1319 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1320 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1321 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1322 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1323 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001324 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001325 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001326 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001327 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001328 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001329 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001330 "src/x32-zip/x2-wasmsimd.c",
1331 "src/x32-zip/x3-wasmsimd.c",
1332 "src/x32-zip/x4-wasmsimd.c",
1333 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001334]
1335
Marat Dukhan102e9ea2020-12-05 22:35:18 -08001336PSIMD_ACCMATH_UKERNELS = [
1337 "src/qs8-requantization/fp32-psimd.c",
1338 "src/qs8-requantization/precise-psimd.c",
1339 "src/qu8-requantization/fp32-psimd.c",
1340 "src/qu8-requantization/precise-psimd.c",
1341]
1342
Marat Dukhan08c4a432019-10-03 09:29:21 -07001343# ISA-specific micro-kernels
1344NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001345 "src/f32-argmaxpool/4x-neon-c4.c",
1346 "src/f32-argmaxpool/9p8x-neon-c4.c",
1347 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001348 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1349 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001350 "src/f32-clamp/gen/neon-x4.c",
1351 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001352 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001353 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001354 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001355 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001356 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001357 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001358 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001359 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001360 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001361 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001362 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001363 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001364 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001365 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1366 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
1367 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1368 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
1369 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001370 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001371 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001372 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
1373 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
1374 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001375 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001376 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
Marat Dukhanc581e482020-10-24 01:28:11 -07001377 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
1378 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
1379 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
1380 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
1381 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001382 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
1383 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
1384 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001385 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001386 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001387 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
1388 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
1389 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001390 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
1391 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
1392 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
1393 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001394 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001395 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
1396 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001397 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001398 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001399 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001400 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001401 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
1402 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001403 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
1404 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
1405 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
1406 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
1407 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
1408 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
1409 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
1410 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001411 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001412 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001413 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001414 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1415 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001416 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001417 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1418 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001419 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001420 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1423 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1424 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001425 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1426 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001427 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1428 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1430 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001431 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1432 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1433 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1434 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1435 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1436 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1437 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1438 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1439 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1440 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1441 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1442 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1443 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1444 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1445 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1446 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001447 "src/f32-hswish/gen/hswish-neon-x4.c",
1448 "src/f32-hswish/gen/hswish-neon-x8.c",
1449 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001450 "src/f32-ibilinear/gen/neon-c4.c",
1451 "src/f32-ibilinear/gen/neon-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001452 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001453 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001454 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001455 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1456 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001457 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001458 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1459 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1460 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1461 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001462 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1463 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001464 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1465 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001466 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1467 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001468 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1469 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1470 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001471 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1472 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001473 "src/f32-prelu/gen/neon-1x4.c",
1474 "src/f32-prelu/gen/neon-1x8.c",
1475 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001476 "src/f32-prelu/gen/neon-2x4.c",
1477 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001478 "src/f32-prelu/gen/neon-2x16.c",
1479 "src/f32-prelu/gen/neon-4x4.c",
1480 "src/f32-prelu/gen/neon-4x8.c",
1481 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001482 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001483 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001484 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001485 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1486 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001487 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001488 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1489 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001490 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001491 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1492 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001493 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1494 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1495 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1496 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1497 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1498 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1499 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1500 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1501 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1502 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1503 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1504 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1505 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001506 "src/f32-relu/gen/neon-x4.c",
1507 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001508 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001509 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001510 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1511 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1512 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1513 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1514 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1515 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1516 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1517 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1518 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1519 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1520 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1521 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001522 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1523 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1524 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1525 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1526 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1527 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001528 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1529 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1530 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1531 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001532 "src/f32-vbinary/gen/vmax-neon-x4.c",
1533 "src/f32-vbinary/gen/vmax-neon-x8.c",
1534 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1535 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1536 "src/f32-vbinary/gen/vmin-neon-x4.c",
1537 "src/f32-vbinary/gen/vmin-neon-x8.c",
1538 "src/f32-vbinary/gen/vminc-neon-x4.c",
1539 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001540 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1541 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1542 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1543 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1544 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1545 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001546 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1547 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1548 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1549 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001550 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1551 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1552 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1553 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001554 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
1555 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
1556 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
1557 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
1558 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
1559 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
1560 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
1561 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
1562 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
1563 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
1564 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
1565 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001566 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1567 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001568 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1569 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001570 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1571 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001572 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1573 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1574 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1575 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1576 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1577 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001578 "src/f32-vunary/gen/vabs-neon-x4.c",
1579 "src/f32-vunary/gen/vabs-neon-x8.c",
1580 "src/f32-vunary/gen/vneg-neon-x4.c",
1581 "src/f32-vunary/gen/vneg-neon-x8.c",
1582 "src/f32-vunary/gen/vsqr-neon-x4.c",
1583 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhande390d42020-11-29 19:32:18 -08001584 "src/math/expm1minus-neon-rr2-lut16-p3.c",
1585 "src/math/expm1minus-neon-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001586 "src/math/roundd-neon-addsub.c",
1587 "src/math/roundd-neon-cvt.c",
1588 "src/math/roundne-neon-addsub.c",
1589 "src/math/roundu-neon-addsub.c",
1590 "src/math/roundu-neon-cvt.c",
1591 "src/math/roundz-neon-addsub.c",
1592 "src/math/roundz-neon-cvt.c",
1593 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
1594 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1595 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1596 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1597 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1598 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1599 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
1600 "src/math/sqrt-neon-nr1rsqrts.c",
1601 "src/math/sqrt-neon-nr2rsqrts.c",
1602 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001603 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1604 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1605 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1606 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001607 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1608 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1609 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1610 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001611 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1612 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1613 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1614 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001615 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001616 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001617 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001618 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1619 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001620 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001621 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001622 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001623 "src/qs8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001624 "src/qs8-requantization/precise-neon.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001625 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001626 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1627 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1628 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1629 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1630 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1631 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1632 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1633 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001634 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1635 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1636 "src/qu8-dwconv/up8x9-minmax-neon.c",
1637 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1638 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1639 "src/qu8-gemm/4x8-minmax-neon.c",
1640 "src/qu8-gemm/8x8-minmax-neon.c",
1641 "src/qu8-igemm/4x8-minmax-neon.c",
1642 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001643 "src/qu8-requantization/fp32-neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001644 "src/qu8-requantization/precise-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001645 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001646 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001647 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001648 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001649 "src/u8-rmax/neon.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001650 "src/x8-zip/x2-neon.c",
1651 "src/x8-zip/x3-neon.c",
1652 "src/x8-zip/x4-neon.c",
1653 "src/x8-zip/xm-neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001654 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001655 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001656 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001657 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001658 "src/x32-zip/x2-neon.c",
1659 "src/x32-zip/x3-neon.c",
1660 "src/x32-zip/x4-neon.c",
1661 "src/x32-zip/xm-neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662]
1663
1664NEONFMA_UKERNELS = [
Frank Barchard04336c12020-10-22 16:48:55 -07001665 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1666 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1667 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1668 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1669 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1670 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1671 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
1672 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1673 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
1674 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1675 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
1676 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1677 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1678 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1679 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1680 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1681 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1682 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1683 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1684 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1685 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
1686 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1687 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1688 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1689 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1690 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1691 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1692 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1693 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1694 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001695 "src/f32-ibilinear/gen/neonfma-c4.c",
1696 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001697 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001698 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001699 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001700 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1701 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001702 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1703 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001704 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1705 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001706 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1707 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001708 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001709 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001710 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001711 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1712 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001713 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001714 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1715 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001716 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001717 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1718 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001719 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1720 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1721 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1722 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1723 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1724 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1725 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1726 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1727 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1728 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1729 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1730 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1731 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001732 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1733 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1734 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1735 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1736 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1737 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001738 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1739 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1740 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1741 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1742 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1743 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001744 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1745 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1746 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1747 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1748 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1749 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001750 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1751 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1752 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1753 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1754 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1755 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001756 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1757 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1758 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1759 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1760 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1761 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001762 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1763 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1764 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1765 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1766 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1767 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001768 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1769 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1770 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1771 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1772 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1773 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1774 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1775 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1776 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1777 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1778 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1779 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1780 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1781 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1782 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1783 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1784 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1785 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Marat Dukhan2fa7a0c2020-12-06 19:09:02 -08001786 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1787 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1788 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1789 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1790 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1791 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1792 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1793 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1794 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1795 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1796 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1797 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1798 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08001799 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1800 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1801 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1802 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1803 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1804 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1805 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1806 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1807 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1808 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1809 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1810 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001811 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1812 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001813 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1814 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1815 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1816 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1817 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1818 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1819 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1820 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1821 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1822 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1823 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1824 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1825 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1826 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1827 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1828 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1829 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1830 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1831 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1832 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08001833 "src/math/exp-neonfma-rr2-lut64-p2.c",
1834 "src/math/exp-neonfma-rr2-p5.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08001835 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1836 "src/math/expm1minus-neonfma-rr1-p6.c",
Marat Dukhan9dd119a2020-11-20 18:20:04 -08001837 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1838 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1839 "src/math/expminus-neonfma-rr2-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001840 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1841 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1842 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001843 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1844 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1845 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001846 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1847 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1848 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001849 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1850 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1851 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001852 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1853 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1854 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001855 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1856 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1857 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001858 "src/math/sqrt-neonfma-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001859 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001860 "src/math/sqrt-neonfma-nr2fma.c",
1861 "src/math/sqrt-neonfma-nr2fma1adj.c",
1862 "src/math/sqrt-neonfma-nr3fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863]
1864
1865AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001866 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001867 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001868 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001869 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001870 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001871 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001872 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001873 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001874 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001875 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1876 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1877 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001878 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001879 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
Marat Dukhan1268a242020-10-24 00:36:32 -07001880 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1881 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1882 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1883 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1884 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001885 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1886 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1887 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001888 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
Marat Dukhan82f0c322020-10-25 19:17:35 -07001889 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001890 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1891 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1892 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001893 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1894 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1895 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1896 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001897 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001898 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1899 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001900 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001901 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001902 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
Marat Dukhan149f0ea2020-10-26 12:50:33 -07001903 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07001904 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1905 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001906 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1907 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1908 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1909 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1910 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1911 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1912 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1913 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
Marat Dukhan30d4b252020-10-29 16:33:22 -07001914 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07001915 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001916 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1917 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1918 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1919 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1920 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1921 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1922 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1923 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1924 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1925 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1926 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1927 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1928 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1929 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1930 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1931 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1932 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1933 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1934 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1935 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001936 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1937 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1938 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1939 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1940 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1941 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1942 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1943 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1944 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1945 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1946 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1947 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001948 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1949 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1950 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1951 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1952 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1953 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001954 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1955 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001956 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1957 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001958 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1959 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001960 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1961 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07001962 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1963 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001964 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1965 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1966 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1967 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1968 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1969 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001970 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1971 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001972 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001973 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001974 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001975 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001976 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001977 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001978]
1979
Marat Dukhan8853b822020-05-07 12:19:01 -07001980NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001981 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1982 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001983 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1984 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1985 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1986 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1987 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1988 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001989 "src/math/roundd-neonv8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001990 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001991 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001992 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001993]
1994
Marat Dukhan08c4a432019-10-03 09:29:21 -07001995AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001996 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07001997 "src/f16-clamp/gen/neonfp16arith-x16.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001998 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1999 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
2000 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
2001 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002002 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
2003 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
2004 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
2005 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
2006 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
2007 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
2008 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
2009 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07002010 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
2011 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002012 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
2013 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
2014 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
2015 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
2016 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
2017 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
2018 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
2019 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
2020 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2021 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2022 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2023 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2024 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2025 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2026 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2027 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002028 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002029 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
2030 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
2031 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
2032 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
2033 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
2034 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
2035 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
2036 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
2037 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Frank Barchardb1966592020-05-12 13:47:06 -07002038 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07002039 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002040 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002041 "src/f16-relu/gen/neonfp16arith-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002042 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002043 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002044 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002045 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002046 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002047 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002048 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002049 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
2050 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
2051 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
2052 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
2053 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
2054 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
2055 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
2056 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
2057 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
2058 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
2059 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
2060 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
2061 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
2062 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
2063 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
2064 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
2065 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
2066 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
2067 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
2068 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
2069 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
2070 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
2071 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
2072 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
2073 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
2074 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
2075 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
2076 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
2077 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
2078 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
2079 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002080]
2081
Benoit Jacoba9644732020-08-13 12:48:55 -07002082NEONDOT_UKERNELS = [
Marat Dukhand4c83032020-09-09 23:20:12 -07002083 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002084 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002085 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002086 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002087 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002088 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002089 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002090 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002091 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002092 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07002093 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002094 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07002095 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002096 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
Frank Barchard2fa17452020-09-22 09:51:46 -07002097 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002098 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
Frank Barcharda48848f2020-09-24 00:59:24 -07002099 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002100 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07002101]
2102
Marat Dukhan08c4a432019-10-03 09:29:21 -07002103SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07002104 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
2105 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002106 "src/f32-clamp/gen/sse-x4.c",
2107 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07002108 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
2109 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002110 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
2111 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
2112 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
2113 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002114 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
2115 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002116 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
2117 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
2118 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
2119 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002120 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
2121 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002122 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
2123 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
2124 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002125 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002126 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
Marat Dukhan470078a2020-10-23 22:36:52 -07002127 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
2128 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
2129 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
2130 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
2131 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002132 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
2133 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
2134 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002135 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
Marat Dukhan0ff97182020-10-25 19:14:03 -07002136 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002137 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
2138 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
2139 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
Marat Dukhand0503892020-10-30 08:22:04 -07002140 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
2141 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
2142 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
2143 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
2144 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
2145 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
2146 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
2147 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
2148 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
2149 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
2150 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
2151 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
2152 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002153 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
2154 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
2155 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
2156 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
2157 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
2158 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
2159 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
2160 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
Marat Dukhanccca2142020-10-30 17:32:45 -07002161 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
Frank Barchard8ef44cd2020-11-03 12:30:23 -08002162 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07002163 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002164 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
2165 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002166 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
2167 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
2168 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
2169 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
2170 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
2171 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002172 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
2173 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
2174 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
2175 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
2176 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
2177 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
2178 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002179 "src/f32-hswish/gen/hswish-sse-x4.c",
2180 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07002181 "src/f32-ibilinear/gen/sse-c4.c",
2182 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002183 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
2184 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
2185 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
2186 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
2187 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
2188 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
2189 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002190 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
2191 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
2192 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002193 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07002194 "src/f32-prelu/gen/sse-2x4.c",
2195 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002196 "src/f32-relu/gen/sse-x4.c",
2197 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002198 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07002199 "src/f32-spmm/gen/4x1-minmax-sse.c",
2200 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07002201 "src/f32-spmm/gen/16x1-minmax-sse.c",
Frank Barchard846c0c62020-10-26 15:01:39 -07002202 "src/f32-spmm/gen/32x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002203 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
2204 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
2205 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
2206 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
2207 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
2208 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
2209 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
2210 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08002211 "src/f32-vbinary/gen/vmax-sse-x4.c",
2212 "src/f32-vbinary/gen/vmax-sse-x8.c",
2213 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
2214 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
2215 "src/f32-vbinary/gen/vmin-sse-x4.c",
2216 "src/f32-vbinary/gen/vmin-sse-x8.c",
2217 "src/f32-vbinary/gen/vminc-sse-x4.c",
2218 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002219 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
2220 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
2221 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
2222 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
2223 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
2224 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
2225 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
2226 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002227 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
2228 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
2229 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
2230 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002231 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
2232 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
2233 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
2234 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002235 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
2236 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002237 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
2238 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002239 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
2240 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002241 "src/f32-vunary/gen/vabs-sse-x4.c",
2242 "src/f32-vunary/gen/vabs-sse-x8.c",
2243 "src/f32-vunary/gen/vneg-sse-x4.c",
2244 "src/f32-vunary/gen/vneg-sse-x8.c",
2245 "src/f32-vunary/gen/vsqr-sse-x4.c",
2246 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002247 "src/math/roundd-sse-addsub.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002248 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002249 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002250 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002251 "src/math/sqrt-sse-hh1mac.c",
2252 "src/math/sqrt-sse-nr1mac.c",
2253 "src/math/sqrt-sse-nr2mac.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002254 "src/x32-fill/sse.c",
2255 "src/x32-packx/x4-sse.c",
2256 "src/x32-pad/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002257]
2258
2259SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08002260 "src/f32-argmaxpool/4x-sse2-c4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002261 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Marat Dukhan329da642019-11-19 21:44:39 -08002262 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002263 "src/f32-prelu/gen/sse2-2x4.c",
2264 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002265 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002266 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002267 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002268 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
2269 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002270 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002271 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
2272 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002273 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08002274 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
2275 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002276 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002277 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
2278 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
2279 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
2280 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
2281 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
2282 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002283 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002284 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002285 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002286 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002287 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
2288 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002289 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
2290 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
2291 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
2292 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
2293 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
2294 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
2295 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
2296 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
2297 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
2298 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
2299 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
2300 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002301 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
2302 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002303 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
2304 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002305 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
2306 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
2307 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
2308 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
2309 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
2310 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002311 "src/math/exp-sse2-rr2-lut64-p2.c",
2312 "src/math/exp-sse2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002313 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
Marat Dukhana438aca2020-11-20 15:45:01 -08002314 "src/math/expm1minus-sse2-rr2-p6.c",
Frank Barchard3b800452020-11-22 12:12:35 -08002315 "src/math/expminus-sse2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002316 "src/math/roundd-sse2-cvt.c",
2317 "src/math/roundne-sse2-cvt.c",
2318 "src/math/roundu-sse2-cvt.c",
2319 "src/math/roundz-sse2-cvt.c",
2320 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
2321 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
2322 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
2323 "src/math/sigmoid-sse2-rr2-p5-div.c",
2324 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
2325 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002326 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
2327 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
2328 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002329 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
2330 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
2331 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002332 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
2333 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
2334 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002335 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002336 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002337 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002338 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002339 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002340 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002341 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
2342 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002343 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002344 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
2345 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002346 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002347 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
2348 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
2349 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002350 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002351 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002352 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002353 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002354 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002355 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002356 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002357 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002358 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
2359 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002360 "src/qs8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002361 "src/qs8-requantization/precise-sse2.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002362 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002363 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
2364 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
2365 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
2366 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002367 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
2368 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
2369 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
2370 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002371 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
2372 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002373 "src/qu8-dwconv/up8x9-minmax-sse2.c",
2374 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
2375 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
2376 "src/qu8-gemm/2x4c8-minmax-sse2.c",
2377 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002378 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002379 "src/qu8-requantization/fp32-sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002380 "src/qu8-requantization/precise-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002381 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07002382 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002383 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07002384 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002385 "src/u8-rmax/sse2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002386 "src/x8-zip/x2-sse2.c",
2387 "src/x8-zip/x3-sse2.c",
2388 "src/x8-zip/x4-sse2.c",
2389 "src/x8-zip/xm-sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07002390 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002391 "src/x32-zip/x2-sse2.c",
2392 "src/x32-zip/x3-sse2.c",
2393 "src/x32-zip/x4-sse2.c",
2394 "src/x32-zip/xm-sse2.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002395]
2396
2397SSSE3_UKERNELS = [
Frank Barchard35db7d02020-10-26 13:37:34 -07002398 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
2399 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
2400 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002401 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
Frank Barchard35db7d02020-10-26 13:37:34 -07002402 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
Marat Dukhan98f2eeb2020-10-23 23:13:41 -07002403 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
2404 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
2405 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
2406 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
2407 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002408 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
2409 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2410 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002411 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2412 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2413 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002414 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2415 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2416 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002417 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002418 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002419 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002420 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002421 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002422 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002423 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
2424 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002425 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002426 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
2427 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002428 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002429 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
2430 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
2431 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002432 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002433 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002434 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002435 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002436 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002437 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002438 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002439 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002440 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
2441 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002442 "src/qs8-requantization/precise-ssse3.c",
2443 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002444 "src/qu8-requantization/precise-ssse3.c",
2445 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446]
2447
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002448SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002449 "src/f32-prelu/gen/sse41-2x4.c",
2450 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002451 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2452 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2453 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2454 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2455 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2456 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002457 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002458 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002459 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002460 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002461 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2462 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002463 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
2464 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
2465 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
2466 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
2467 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
2468 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
2469 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
2470 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
2471 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
2472 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
2473 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
2474 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002475 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2476 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002477 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2478 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002479 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2480 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2481 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2482 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2483 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2484 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2485 "src/math/roundd-sse41.c",
2486 "src/math/roundne-sse41.c",
2487 "src/math/roundu-sse41.c",
2488 "src/math/roundz-sse41.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002489 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2490 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2491 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002492 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2493 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2494 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002495 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2496 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2497 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002498 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002499 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002500 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002501 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002502 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002503 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002504 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
2505 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002506 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002507 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
2508 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002509 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002510 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
2511 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
2512 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002513 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002514 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002515 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002516 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002517 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002518 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002519 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002520 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002521 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
2522 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002523 "src/qs8-requantization/fp32-sse4.c",
2524 "src/qs8-requantization/precise-sse4.c",
2525 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002526 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2527 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2528 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2529 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002530 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2531 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2532 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2533 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002534 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2535 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2536 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2537 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002538 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2539 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2540 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2541 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002542 "src/qu8-requantization/precise-sse4.c",
2543 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002544]
2545
Marat Dukhan08c4a432019-10-03 09:29:21 -07002546AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002547 "src/f32-clamp/gen/avx-x8.c",
2548 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002549 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2550 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002551 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2552 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002553 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2554 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002555 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2556 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2557 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2558 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2559 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2560 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002561 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002562 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2563 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002564 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002565 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002566 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002567 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002568 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2569 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2570 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2571 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2572 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2573 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2574 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2575 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2576 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2577 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2578 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002579 "src/f32-hswish/gen/hswish-avx-x8.c",
2580 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002581 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002582 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2583 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002584 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002585 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002586 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002587 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002588 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2589 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002590 "src/f32-prelu/gen/avx-2x8.c",
2591 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002592 "src/f32-relu/gen/avx-x8.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002593 "src/f32-relu/gen/avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002594 "src/f32-rmax/avx.c",
T.J. Alumbaughdc2b29c2020-10-14 13:56:08 -07002595 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
2596 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
2597 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
2598 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
2599 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
2600 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
2601 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
2602 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
2603 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
2604 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
2605 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
2606 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
2607 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
2608 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
2609 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
2610 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
2611 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
2612 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
2613 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
2614 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002615 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2616 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2617 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2618 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2619 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2620 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2621 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2622 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002623 "src/f32-vbinary/gen/vmax-avx-x8.c",
2624 "src/f32-vbinary/gen/vmax-avx-x16.c",
2625 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2626 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2627 "src/f32-vbinary/gen/vmin-avx-x8.c",
2628 "src/f32-vbinary/gen/vmin-avx-x16.c",
2629 "src/f32-vbinary/gen/vminc-avx-x8.c",
2630 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002631 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2632 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2633 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2634 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2635 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2636 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2637 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2638 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002639 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2640 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2641 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2642 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002643 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2644 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2645 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2646 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002647 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
2648 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
2649 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
2650 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
2651 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
2652 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
2653 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
2654 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
2655 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
2656 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
2657 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
2658 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
2659 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
2660 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
2661 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
2662 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
2663 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
2664 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002665 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2666 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002667 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2668 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002669 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2670 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2671 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2672 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2673 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2674 "src/f32-vrnd/gen/vrndz-avx-x16.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07002675 "src/f32-vscale/avx-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002676 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2677 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002678 "src/f32-vunary/gen/vabs-avx-x8.c",
2679 "src/f32-vunary/gen/vabs-avx-x16.c",
2680 "src/f32-vunary/gen/vneg-avx-x8.c",
2681 "src/f32-vunary/gen/vneg-avx-x16.c",
2682 "src/f32-vunary/gen/vsqr-avx-x8.c",
2683 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002684 "src/math/exp-avx-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002685 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2686 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2687 "src/math/expm1minus-avx-rr2-p6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002688 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
Marat Dukhan2e4727b2020-09-24 00:48:39 -07002689 "src/math/sigmoid-avx-rr2-p5-div.c",
Marat Dukhan39ab16f2020-10-16 14:30:35 -07002690 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2691 "src/math/sigmoid-avx-rr2-p5-nr2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692]
2693
Marat Dukhan1566fee2020-08-02 21:55:41 -07002694XOP_UKERNELS = [
2695 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002696 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002697 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002698 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002699 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002700 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002701 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2702 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002703 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002704 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2705 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002706 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002707 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2708 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2709 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002710 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002711 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002712 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002713 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002714 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002715 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002716 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002717 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002718 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2719 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002720 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2721 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2722 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2723 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2724 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2725 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2726 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2727 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002728]
2729
Marat Dukhanfda12b82019-11-21 12:27:59 -08002730FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002731 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2732 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002733 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2734 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002735 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2736 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002737 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2738 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2739 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2740 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2741 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2742 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002743 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002744 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2745 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2746 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2747 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002748 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002749 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2750 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002751 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002752 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2753 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002754 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2755 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2756 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002757 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2758 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2759 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2760 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2761 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2762 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2763 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2764 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2765 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2766 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2767 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2768 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2769 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2770 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002771 "src/f32-hswish/gen/hswish-fma3-x8.c",
2772 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002773 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002774 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2775 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2776 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2777 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002778 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002779 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2780 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002781 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002782 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2783 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002784 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2785 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2786 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002787 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2788 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2789 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2790 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2791 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2792 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2793 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2794 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002795 "src/math/sqrt-fma3-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002796 "src/math/sqrt-fma3-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002797 "src/math/sqrt-fma3-nr2fma.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002798]
2799
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002800AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002801 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2802 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002803 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002804 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002805 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002806 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2807 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002808 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002809 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2810 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2811 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002812 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002813 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2814 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002815 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002816 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002817 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002818 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2819 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002820 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002821 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2822 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2823 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002824 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002825 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2826 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002827 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002828 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002829 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002830 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2831 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002832 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002833 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2834 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2835 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002836 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002837 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2838 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2839 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2840 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2841 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2842 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2843 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2844 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2845 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2846 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2847 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2848 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2849 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2850 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2851 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2852 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2853 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2854 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2855 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2856 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2857 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2858 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2859 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2860 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2861 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2862 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2863 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2864 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2865 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2866 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08002867 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2868 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2869 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2870 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2871 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2872 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2873 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2874 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2875 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2876 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2877 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2878 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2879 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2880 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2881 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2882 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2883 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2884 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2885 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2886 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2887 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2888 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2889 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2890 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2891 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2892 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2893 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2894 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2895 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2896 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2897 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2898 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2899 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2900 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2901 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2902 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2903 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2904 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2905 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2906 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002907 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2908 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2909 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2910 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2911 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2912 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2913 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2914 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2915 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2916 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2917 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2918 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2919 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2920 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2921 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2922 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2923 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2924 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2925 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2926 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2927 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2928 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2929 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2930 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08002931 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2932 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2933 "src/math/exp-avx2-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08002934 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2935 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2936 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2937 "src/math/expm1minus-avx2-rr1-p6.c",
Frank Barcharde7223ee2020-12-04 19:04:01 -08002938 "src/math/expminus-avx2-rr2-p5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002939 "src/math/extexp-avx2-p5.c",
2940 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2941 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2942 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2943 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
2944 "src/math/sigmoid-avx2-rr1-p5-div.c",
2945 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
2946 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2947 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2948 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2949 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2950 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2951 "src/math/sigmoid-avx2-rr2-p5-div.c",
2952 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2953 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002954 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002955 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002956 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2957 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002958 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002959 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002960 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002961 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002962 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002963 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002964 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002965 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002966 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2967 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2968 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002969 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2970 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2971 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2972 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2973 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2974 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2975 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2976 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002977]
2978
Marat Dukhan08c4a432019-10-03 09:29:21 -07002979AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002980 "src/f32-clamp/gen/avx512f-x16.c",
2981 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002982 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2983 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002984 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2985 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002986 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2987 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Frank Barchard04336c12020-10-22 16:48:55 -07002988 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2989 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2990 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2991 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2992 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2993 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002994 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2995 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2996 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2997 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2998 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2999 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003000 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
3001 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
3002 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
3003 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
3004 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
3005 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07003006 "src/f32-hswish/gen/hswish-avx512f-x16.c",
3007 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07003008 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
3009 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
3010 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
3011 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
3012 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
3013 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07003014 "src/f32-prelu/gen/avx512f-2x16.c",
3015 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003016 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3017 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003018 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003019 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003020 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003021 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3022 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003023 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003024 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3025 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3026 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003027 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003028 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
3029 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003030 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003031 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003032 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003033 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
3034 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003035 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003036 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
3037 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
3038 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003039 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003040 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
3041 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003042 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003043 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003044 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003045 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
3046 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003047 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003048 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
3049 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
3050 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003051 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07003052 "src/f32-relu/gen/avx512f-x16.c",
3053 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003054 "src/f32-rmax/avx512f.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003055 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
3056 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
3057 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
3058 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
3059 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
3060 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
3061 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
3062 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
3063 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
3064 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
3065 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
3066 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
3067 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
3068 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
3069 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
3070 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003071 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
3072 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
3073 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
3074 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
3075 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
3076 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
3077 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
3078 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
3079 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
3080 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
3081 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
3082 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
3083 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
3084 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
3085 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
3086 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
Marat Dukhand9ca7e62020-09-23 23:45:29 -07003087 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
3088 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
3089 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
3090 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
3091 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
3092 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
3093 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
3094 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
3095 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
3096 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
3097 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
3098 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
3099 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
3100 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
3101 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
3102 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003103 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
3104 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
3105 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
3106 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
3107 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
3108 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
3109 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
3110 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08003111 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
3112 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
3113 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
3114 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
3115 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
3116 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
3117 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
3118 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003119 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
3120 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
3121 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
3122 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
3123 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
3124 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
3125 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
3126 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07003127 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
3128 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
3129 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
3130 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07003131 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
3132 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
3133 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
3134 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhaned6baaf2020-12-01 15:07:08 -08003135 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
3136 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
3137 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
3138 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
3139 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
3140 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
3141 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
3142 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
3143 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
3144 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
3145 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
3146 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
3147 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
3148 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
3149 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
3150 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07003151 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
3152 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003153 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
3154 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
3155 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
3156 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
3157 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
3158 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
3159 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
3160 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
Frank Barchardbeca6522020-10-30 22:34:35 -07003161 "src/f32-vscale/avx512f-x64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003162 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
3163 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
3164 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
3165 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
3166 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
3167 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
3168 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
3169 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
3170 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
3171 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
3172 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
3173 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
3174 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
3175 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
3176 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
3177 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
3178 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
3179 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
3180 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
3181 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
3182 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
3183 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
3184 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
3185 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07003186 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
3187 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
3188 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
3189 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
3190 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
3191 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
3192 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
3193 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07003194 "src/f32-vunary/gen/vabs-avx512f-x16.c",
3195 "src/f32-vunary/gen/vabs-avx512f-x32.c",
3196 "src/f32-vunary/gen/vneg-avx512f-x16.c",
3197 "src/f32-vunary/gen/vneg-avx512f-x32.c",
3198 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
3199 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhanb7633f22020-11-20 16:34:56 -08003200 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
3201 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
3202 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
3203 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
3204 "src/math/exp-avx512f-rr2-p5-scalef.c",
3205 "src/math/exp-avx512f-rr2-p5.c",
Marat Dukhande390d42020-11-29 19:32:18 -08003206 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
3207 "src/math/expm1minus-avx512f-rr1-p6.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07003208 "src/math/extexp-avx512f-p5.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003209 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003210 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003211 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003212 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
Marat Dukhan6a34c5f2020-09-22 21:44:15 -07003213 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003214 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003215 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003216 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003217 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
3218 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
3219 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
3220 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
3221 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
3222 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
3223 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
3224 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
3225 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
3226 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003227 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
Marat Dukhan36173d22020-10-15 17:14:26 -07003228 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003229 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
3230 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
3231 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
3232 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003233 "src/math/sqrt-avx512f-nr1fma.c",
Marat Dukhan84000762020-06-29 18:38:43 -07003234 "src/math/sqrt-avx512f-nr1fma1adj.c",
Frank Barchard04336c12020-10-22 16:48:55 -07003235 "src/math/sqrt-avx512f-nr2fma.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003236]
3237
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003238AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07003239 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
3240 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003241 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
3242 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
3243 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
3244 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
3245 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
3246 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
3247 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
3248 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
3249]
3250
Frank Barchardbcedc082020-08-17 18:00:51 -07003251WASM32_ASM_UKERNELS = [
3252 "src/f32-relu/wasm_shr_x1.S",
3253 "src/f32-relu/wasm_shr_x2.S",
3254 "src/f32-relu/wasm_shr_x4.S",
3255]
3256
Marat Dukhan08c4a432019-10-03 09:29:21 -07003257AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07003258 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07003259 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003260 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3261 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003262 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003263 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07003264 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003265 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003266 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
3267 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07003268 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
3269 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
3270 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
3271 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003272]
3273
3274AARCH64_ASM_UKERNELS = [
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003275 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003276 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003277 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003278 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07003279 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003280 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07003281 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003282 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
3283 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
3284 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
3285 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
3286 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
3287 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
3288 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003289 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
3290 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003291 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3292 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3293 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003294 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
3295 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003296 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3297 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
3298 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3299 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003300 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003301 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
3302 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003303 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
3304 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
3305 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
3306 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003307 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003308 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003309 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003310 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003311 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
3312 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
3313 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3314 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
3315 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
3316 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
3317 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3318 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
3319 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3320 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3321 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
3322 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
3323 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3324 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3325 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
3326 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3327 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3328 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3329 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3330 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
3331 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
3332 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003333 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003334 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003335 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
3336 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Frank Barchard04336c12020-10-22 16:48:55 -07003337 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
3338 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
3339 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
3340 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
3341 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
3342 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003343 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
3344 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
3345 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
3346 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07003347 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
3348 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Frank Barchardf1fd89e2020-10-06 23:27:11 -07003349 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Frank Barchardf2742c42020-10-12 21:16:46 -07003350 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Frank Barcharda5237a52020-10-06 16:18:46 -07003351 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003352]
3353
Marat Dukhan1b354632020-03-23 12:50:22 -07003354INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003355 "src/xnnpack/argmaxpool.h",
3356 "src/xnnpack/avgpool.h",
3357 "src/xnnpack/clamp.h",
3358 "src/xnnpack/common.h",
3359 "src/xnnpack/conv.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08003360 "src/xnnpack/depthtospace.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003361 "src/xnnpack/dwconv.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003362 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003363 "src/xnnpack/gavgpool.h",
3364 "src/xnnpack/gemm.h",
3365 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07003366 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003367 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08003368 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003369 "src/xnnpack/lut.h",
3370 "src/xnnpack/math.h",
3371 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003372 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003373 "src/xnnpack/packx.h",
3374 "src/xnnpack/pad.h",
3375 "src/xnnpack/params.h",
3376 "src/xnnpack/pavgpool.h",
3377 "src/xnnpack/ppmm.h",
3378 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003379 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003380 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003381 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003382 "src/xnnpack/rmax.h",
3383 "src/xnnpack/scalar-utils.h",
3384 "src/xnnpack/spmm.h",
3385 "src/xnnpack/unpool.h",
3386 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003387 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003388 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07003389 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07003390 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07003391 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08003392 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003393 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07003394]
3395
3396INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003397 "include/xnnpack.h",
3398 "src/xnnpack/allocator.h",
3399 "src/xnnpack/compute.h",
3400 "src/xnnpack/im2col.h",
3401 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003402 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07003403 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003404 "src/xnnpack/operator.h",
3405 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003406 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003407 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003408 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08003409 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003410]
3411
Marat Dukhan1b354632020-03-23 12:50:22 -07003412ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003413 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414]
3415
Marat Dukhan1b354632020-03-23 12:50:22 -07003416MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003417 "include/xnnpack.h",
Frank Barchard35db7d02020-10-26 13:37:34 -07003418 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003419]
3420
Marat Dukhan1b354632020-03-23 12:50:22 -07003421MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Frank Barchard35db7d02020-10-26 13:37:34 -07003422 "include/xnnpack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003423 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07003424 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003425 "src/xnnpack/requantization.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003426]
3427
3428OPERATOR_TEST_PARAMS_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003429 "src/xnnpack/common.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003430 "src/xnnpack/params.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431]
3432
3433WEIGHTS_PACK_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003434 "src/xnnpack/compute.h",
Frank Barchard04336c12020-10-22 16:48:55 -07003435 "src/xnnpack/operator.h",
3436 "src/xnnpack/pack.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437]
3438
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003439LOGGING_COPTS = select({
3440 # No logging in optimized mode
3441 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
3442 # Full logging in debug mode
3443 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
3444 # Error-only logging in default (fastbuild) mode
3445 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
3446})
3447
Marat Dukhan3b59de22020-06-03 20:15:19 -07003448LOGGING_SRCS = select({
3449 # No logging in optimized mode
3450 ":optimized_build": [],
3451 "//conditions:default": [
3452 "src/operator-strings.c",
3453 "src/subgraph-strings.c",
3454 ],
3455})
3456
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003457LOGGING_HDRS = [
3458 "src/xnnpack/log.h",
3459]
3460
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003462 name = "tables",
3463 srcs = TABLE_SRCS,
3464 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003465 gcc_copts = xnnpack_gcc_std_copts(),
3466 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003467)
3468
3469xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003470 name = "scalar_ukernels",
3471 srcs = SCALAR_UKERNELS,
3472 hdrs = INTERNAL_HDRS,
3473 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003474 gcc_copts = xnnpack_gcc_std_copts(),
3475 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003477 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003478 "@FP16",
3479 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003480 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003481 ],
3482)
3483
3484xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003485 name = "scalar_ukernels_test_mode",
3486 srcs = SCALAR_UKERNELS,
3487 hdrs = INTERNAL_HDRS,
3488 aarch32_copts = ["-marm"],
3489 copts = [
3490 "-UNDEBUG",
3491 "-DXNN_TEST_MODE=1",
3492 ],
3493 gcc_copts = xnnpack_gcc_std_copts(),
3494 msvc_copts = xnnpack_msvc_std_copts(),
3495 deps = [
3496 ":tables",
3497 "@FP16",
3498 "@FXdiv",
3499 "@pthreadpool",
3500 ],
3501)
3502
3503xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08003504 name = "wasm_ukernels",
3505 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003506 gcc_copts = xnnpack_gcc_std_copts(),
3507 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08003508 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07003509 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08003510 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003511 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08003512 "@FP16",
3513 "@FXdiv",
3514 "@pthreadpool",
3515 ],
3516)
3517
3518xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003519 name = "wasm_ukernels_test_mode",
3520 hdrs = INTERNAL_HDRS,
3521 copts = [
3522 "-UNDEBUG",
3523 "-DXNN_TEST_MODE=1",
3524 ],
3525 gcc_copts = xnnpack_gcc_std_copts(),
3526 msvc_copts = xnnpack_msvc_std_copts(),
3527 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07003528 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003529 deps = [
3530 ":tables",
3531 "@FP16",
3532 "@FXdiv",
3533 "@pthreadpool",
3534 ],
3535)
3536
3537xnnpack_cc_library(
Marat Dukhan102e9ea2020-12-05 22:35:18 -08003538 name = "psimd_accmath_ukernels",
3539 hdrs = INTERNAL_HDRS,
3540 aarch32_copts = [
3541 "-marm",
3542 "-mfpu=neon",
3543 ],
3544 gcc_copts = xnnpack_gcc_std_copts(),
3545 gcc_x86_copts = ["-msse2"],
3546 msvc_copts = xnnpack_msvc_std_copts(),
3547 optimized_copts = [
3548 "-O3",
3549 ],
3550 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3551 deps = [
3552 ":tables",
3553 "@FP16",
3554 "@psimd",
3555 "@pthreadpool",
3556 ],
3557)
3558
3559xnnpack_cc_library(
3560 name = "psimd_accmath_ukernels_test_mode",
3561 hdrs = INTERNAL_HDRS,
3562 aarch32_copts = [
3563 "-marm",
3564 "-mfpu=neon",
3565 ],
3566 copts = [
3567 "-UNDEBUG",
3568 "-DXNN_TEST_MODE=1",
3569 ],
3570 gcc_copts = xnnpack_gcc_std_copts(),
3571 gcc_x86_copts = ["-msse2"],
3572 msvc_copts = xnnpack_msvc_std_copts(),
3573 optimized_copts = [
3574 "-O3",
3575 ],
3576 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3577 deps = [
3578 ":tables",
3579 "@FP16",
3580 "@psimd",
3581 "@pthreadpool",
3582 ],
3583)
3584
3585xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 name = "neon_ukernels",
3587 hdrs = INTERNAL_HDRS,
3588 aarch32_copts = [
3589 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003590 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003591 "-mfpu=neon",
3592 ],
3593 aarch32_srcs = NEON_UKERNELS,
3594 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003595 gcc_copts = xnnpack_gcc_std_copts(),
3596 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003597 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003598 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003599 "@FP16",
3600 "@pthreadpool",
3601 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602)
3603
3604xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003605 name = "neon_ukernels_test_mode",
3606 hdrs = INTERNAL_HDRS,
3607 aarch32_copts = [
3608 "-marm",
3609 "-march=armv7-a",
3610 "-mfpu=neon",
3611 ],
3612 aarch32_srcs = NEON_UKERNELS,
3613 aarch64_srcs = NEON_UKERNELS,
3614 copts = [
3615 "-UNDEBUG",
3616 "-DXNN_TEST_MODE=1",
3617 ],
3618 gcc_copts = xnnpack_gcc_std_copts(),
3619 msvc_copts = xnnpack_msvc_std_copts(),
3620 deps = [
3621 ":tables",
3622 "@FP16",
3623 "@pthreadpool",
3624 ],
3625)
3626
3627xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003628 name = "neonfma_ukernels",
3629 hdrs = INTERNAL_HDRS,
3630 aarch32_copts = [
3631 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003632 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003633 "-mfpu=neon-vfpv4",
3634 ],
3635 aarch32_srcs = NEONFMA_UKERNELS,
3636 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003637 apple_aarch32_copts = [
3638 "-mcpu=swift",
3639 "-mtune=generic",
3640 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003641 gcc_copts = xnnpack_gcc_std_copts(),
3642 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003643 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003644 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003645 "@FP16",
3646 "@pthreadpool",
3647 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648)
3649
3650xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003651 name = "neonfma_ukernels_test_mode",
3652 hdrs = INTERNAL_HDRS,
3653 aarch32_copts = [
3654 "-marm",
3655 "-march=armv7-a",
3656 "-mfpu=neon-vfpv4",
3657 ],
3658 aarch32_srcs = NEONFMA_UKERNELS,
3659 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003660 apple_aarch32_copts = [
3661 "-mcpu=swift",
3662 "-mtune=generic",
3663 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003664 copts = [
3665 "-UNDEBUG",
3666 "-DXNN_TEST_MODE=1",
3667 ],
3668 gcc_copts = xnnpack_gcc_std_copts(),
3669 msvc_copts = xnnpack_msvc_std_copts(),
3670 deps = [
3671 ":tables",
3672 "@FP16",
3673 "@pthreadpool",
3674 ],
3675)
3676
3677xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003678 name = "neonv8_ukernels",
3679 hdrs = INTERNAL_HDRS,
3680 aarch32_copts = [
3681 "-marm",
3682 "-march=armv8-a",
3683 "-mfpu=neon-fp-armv8",
3684 ],
3685 aarch32_srcs = NEONV8_UKERNELS,
3686 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003687 apple_aarch32_copts = [
3688 "-mcpu=cyclone",
3689 "-mtune=generic",
3690 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003691 gcc_copts = xnnpack_gcc_std_copts(),
3692 msvc_copts = xnnpack_msvc_std_copts(),
3693 deps = [
3694 ":tables",
3695 "@FP16",
3696 "@pthreadpool",
3697 ],
3698)
3699
3700xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003701 name = "neonv8_ukernels_test_mode",
3702 hdrs = INTERNAL_HDRS,
3703 aarch32_copts = [
3704 "-marm",
3705 "-march=armv8-a",
3706 "-mfpu=neon-fp-armv8",
3707 ],
3708 aarch32_srcs = NEONV8_UKERNELS,
3709 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003710 apple_aarch32_copts = [
3711 "-mcpu=cyclone",
3712 "-mtune=generic",
3713 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003714 copts = [
3715 "-UNDEBUG",
3716 "-DXNN_TEST_MODE=1",
3717 ],
3718 gcc_copts = xnnpack_gcc_std_copts(),
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 deps = [
3721 ":tables",
3722 "@FP16",
3723 "@pthreadpool",
3724 ],
3725)
3726
3727xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 name = "neonfp16arith_ukernels",
3729 hdrs = INTERNAL_HDRS,
3730 aarch64_copts = ["-march=armv8.2-a+fp16"],
3731 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003732 gcc_copts = xnnpack_gcc_std_copts(),
3733 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003734 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003735 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003736 "@FP16",
3737 "@pthreadpool",
3738 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003739)
3740
3741xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003742 name = "neonfp16arith_ukernels_test_mode",
3743 hdrs = INTERNAL_HDRS,
3744 aarch64_copts = ["-march=armv8.2-a+fp16"],
3745 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
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 deps = [
3753 ":tables",
3754 "@FP16",
3755 "@pthreadpool",
3756 ],
3757)
3758
3759xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003760 name = "neondot_ukernels",
3761 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003762 aarch32_copts = [
3763 "-marm",
3764 "-march=armv8.2-a+dotprod",
3765 "-mfpu=neon-fp-armv8",
3766 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003767 aarch32_srcs = NEONDOT_UKERNELS,
3768 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3769 aarch64_srcs = NEONDOT_UKERNELS,
3770 gcc_copts = xnnpack_gcc_std_copts(),
3771 msvc_copts = xnnpack_msvc_std_copts(),
3772 deps = [
3773 ":tables",
3774 "@FP16",
3775 "@pthreadpool",
3776 ],
3777)
3778
3779xnnpack_cc_library(
3780 name = "neondot_ukernels_test_mode",
3781 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003782 aarch32_copts = [
3783 "-marm",
3784 "-march=armv8.2-a+dotprod",
3785 "-mfpu=neon-fp-armv8",
3786 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003787 aarch32_srcs = NEONDOT_UKERNELS,
3788 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3789 aarch64_srcs = NEONDOT_UKERNELS,
3790 copts = [
3791 "-UNDEBUG",
3792 "-DXNN_TEST_MODE=1",
3793 ],
3794 gcc_copts = xnnpack_gcc_std_copts(),
3795 msvc_copts = xnnpack_msvc_std_copts(),
3796 deps = [
3797 ":tables",
3798 "@FP16",
3799 "@pthreadpool",
3800 ],
3801)
3802
3803xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003804 name = "sse2_ukernels",
3805 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003806 gcc_copts = xnnpack_gcc_std_copts(),
3807 gcc_x86_copts = ["-msse2"],
3808 msvc_copts = xnnpack_msvc_std_copts(),
3809 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003810 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003811 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003812 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003813 "@FP16",
3814 "@pthreadpool",
3815 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003816)
3817
3818xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003819 name = "sse2_ukernels_test_mode",
3820 hdrs = INTERNAL_HDRS,
3821 copts = [
3822 "-UNDEBUG",
3823 "-DXNN_TEST_MODE=1",
3824 ],
3825 gcc_copts = xnnpack_gcc_std_copts(),
3826 gcc_x86_copts = ["-msse2"],
3827 msvc_copts = xnnpack_msvc_std_copts(),
3828 msvc_x86_32_copts = ["/arch:SSE2"],
3829 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3830 deps = [
3831 ":tables",
3832 "@FP16",
3833 "@pthreadpool",
3834 ],
3835)
3836
3837xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003838 name = "ssse3_ukernels",
3839 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003840 gcc_copts = xnnpack_gcc_std_copts(),
3841 gcc_x86_copts = ["-mssse3"],
3842 msvc_copts = xnnpack_msvc_std_copts(),
3843 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003844 x86_srcs = SSSE3_UKERNELS,
3845 deps = [
3846 ":tables",
3847 "@FP16",
3848 "@pthreadpool",
3849 ],
3850)
3851
3852xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003853 name = "ssse3_ukernels_test_mode",
3854 hdrs = INTERNAL_HDRS,
3855 copts = [
3856 "-UNDEBUG",
3857 "-DXNN_TEST_MODE=1",
3858 ],
3859 gcc_copts = xnnpack_gcc_std_copts(),
3860 gcc_x86_copts = ["-mssse3"],
3861 msvc_copts = xnnpack_msvc_std_copts(),
3862 msvc_x86_32_copts = ["/arch:SSE2"],
3863 x86_srcs = SSSE3_UKERNELS,
3864 deps = [
3865 ":tables",
3866 "@FP16",
3867 "@pthreadpool",
3868 ],
3869)
3870
3871xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003872 name = "sse41_ukernels",
3873 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003874 gcc_copts = xnnpack_gcc_std_copts(),
3875 gcc_x86_copts = ["-msse4.1"],
3876 msvc_copts = xnnpack_msvc_std_copts(),
3877 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003878 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003879 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003880 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003881 "@FP16",
3882 "@pthreadpool",
3883 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003884)
3885
3886xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003887 name = "sse41_ukernels_test_mode",
3888 hdrs = INTERNAL_HDRS,
3889 copts = [
3890 "-UNDEBUG",
3891 "-DXNN_TEST_MODE=1",
3892 ],
3893 gcc_copts = xnnpack_gcc_std_copts(),
3894 gcc_x86_copts = ["-msse4.1"],
3895 msvc_copts = xnnpack_msvc_std_copts(),
3896 msvc_x86_32_copts = ["/arch:SSE2"],
3897 x86_srcs = SSE41_UKERNELS,
3898 deps = [
3899 ":tables",
3900 "@FP16",
3901 "@pthreadpool",
3902 ],
3903)
3904
3905xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003906 name = "avx_ukernels",
3907 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003908 gcc_copts = xnnpack_gcc_std_copts(),
3909 gcc_x86_copts = ["-mavx"],
3910 msvc_copts = xnnpack_msvc_std_copts(),
3911 msvc_x86_32_copts = ["/arch:AVX"],
3912 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003913 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003914 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003915 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003916 "@FP16",
3917 "@pthreadpool",
3918 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003919)
3920
3921xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003922 name = "avx_ukernels_test_mode",
3923 hdrs = INTERNAL_HDRS,
3924 copts = [
3925 "-UNDEBUG",
3926 "-DXNN_TEST_MODE=1",
3927 ],
3928 gcc_copts = xnnpack_gcc_std_copts(),
3929 gcc_x86_copts = ["-mavx"],
3930 msvc_copts = xnnpack_msvc_std_copts(),
3931 msvc_x86_32_copts = ["/arch:AVX"],
3932 msvc_x86_64_copts = ["/arch:AVX"],
3933 x86_srcs = AVX_UKERNELS,
3934 deps = [
3935 ":tables",
3936 "@FP16",
3937 "@pthreadpool",
3938 ],
3939)
3940
3941xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003942 name = "xop_ukernels",
3943 hdrs = INTERNAL_HDRS,
3944 gcc_copts = xnnpack_gcc_std_copts(),
3945 gcc_x86_copts = ["-mxop"],
3946 msvc_copts = xnnpack_msvc_std_copts(),
3947 msvc_x86_32_copts = ["/arch:AVX"],
3948 msvc_x86_64_copts = ["/arch:AVX"],
3949 x86_srcs = XOP_UKERNELS,
3950 deps = [
3951 ":tables",
3952 "@FP16",
3953 "@pthreadpool",
3954 ],
3955)
3956
3957xnnpack_cc_library(
3958 name = "xop_ukernels_test_mode",
3959 hdrs = INTERNAL_HDRS,
3960 copts = [
3961 "-UNDEBUG",
3962 "-DXNN_TEST_MODE=1",
3963 ],
3964 gcc_copts = xnnpack_gcc_std_copts(),
3965 gcc_x86_copts = ["-mxop"],
3966 msvc_copts = xnnpack_msvc_std_copts(),
3967 msvc_x86_32_copts = ["/arch:AVX"],
3968 msvc_x86_64_copts = ["/arch:AVX"],
3969 x86_srcs = XOP_UKERNELS,
3970 deps = [
3971 ":tables",
3972 "@FP16",
3973 "@pthreadpool",
3974 ],
3975)
3976
3977xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003978 name = "fma3_ukernels",
3979 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003980 gcc_copts = xnnpack_gcc_std_copts(),
3981 gcc_x86_copts = ["-mfma"],
3982 msvc_copts = xnnpack_msvc_std_copts(),
3983 msvc_x86_32_copts = ["/arch:AVX"],
3984 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003985 x86_srcs = FMA3_UKERNELS,
3986 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003987 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003988 "@FP16",
3989 "@pthreadpool",
3990 ],
3991)
3992
3993xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003994 name = "fma3_ukernels_test_mode",
3995 hdrs = INTERNAL_HDRS,
3996 copts = [
3997 "-UNDEBUG",
3998 "-DXNN_TEST_MODE=1",
3999 ],
4000 gcc_copts = xnnpack_gcc_std_copts(),
4001 gcc_x86_copts = ["-mfma"],
4002 msvc_copts = xnnpack_msvc_std_copts(),
4003 msvc_x86_32_copts = ["/arch:AVX"],
4004 msvc_x86_64_copts = ["/arch:AVX"],
4005 x86_srcs = FMA3_UKERNELS,
4006 deps = [
4007 ":tables",
4008 "@FP16",
4009 "@pthreadpool",
4010 ],
4011)
4012
4013xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004014 name = "avx2_ukernels",
4015 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004016 gcc_copts = xnnpack_gcc_std_copts(),
4017 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004018 "-mfma",
4019 "-mavx2",
4020 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004021 msvc_copts = xnnpack_msvc_std_copts(),
4022 msvc_x86_32_copts = ["/arch:AVX2"],
4023 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004024 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004025 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004026 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004027 "@FP16",
4028 "@pthreadpool",
4029 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004030)
4031
4032xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004033 name = "avx2_ukernels_test_mode",
4034 hdrs = INTERNAL_HDRS,
4035 copts = [
4036 "-UNDEBUG",
4037 "-DXNN_TEST_MODE=1",
4038 ],
4039 gcc_copts = xnnpack_gcc_std_copts(),
4040 gcc_x86_copts = [
4041 "-mfma",
4042 "-mavx2",
4043 ],
4044 msvc_copts = xnnpack_msvc_std_copts(),
4045 msvc_x86_32_copts = ["/arch:AVX2"],
4046 msvc_x86_64_copts = ["/arch:AVX2"],
4047 x86_srcs = AVX2_UKERNELS,
4048 deps = [
4049 ":tables",
4050 "@FP16",
4051 "@pthreadpool",
4052 ],
4053)
4054
4055xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004056 name = "avx512f_ukernels",
4057 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004058 gcc_copts = xnnpack_gcc_std_copts(),
4059 gcc_x86_copts = ["-mavx512f"],
4060 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4061 msvc_copts = xnnpack_msvc_std_copts(),
4062 msvc_x86_32_copts = ["/arch:AVX512"],
4063 msvc_x86_64_copts = ["/arch:AVX512"],
4064 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004065 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08004066 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08004067 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08004068 "@FP16",
4069 "@pthreadpool",
4070 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004071)
4072
4073xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004074 name = "avx512f_ukernels_test_mode",
4075 hdrs = INTERNAL_HDRS,
4076 copts = [
4077 "-UNDEBUG",
4078 "-DXNN_TEST_MODE=1",
4079 ],
4080 gcc_copts = xnnpack_gcc_std_copts(),
4081 gcc_x86_copts = ["-mavx512f"],
4082 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4083 msvc_copts = xnnpack_msvc_std_copts(),
4084 msvc_x86_32_copts = ["/arch:AVX512"],
4085 msvc_x86_64_copts = ["/arch:AVX512"],
4086 msys_copts = ["-fno-asynchronous-unwind-tables"],
4087 x86_srcs = AVX512F_UKERNELS,
4088 deps = [
4089 ":tables",
4090 "@FP16",
4091 "@pthreadpool",
4092 ],
4093)
4094
4095xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004096 name = "avx512skx_ukernels",
4097 hdrs = INTERNAL_HDRS,
4098 gcc_copts = xnnpack_gcc_std_copts(),
4099 gcc_x86_copts = [
4100 "-mavx512f",
4101 "-mavx512cd",
4102 "-mavx512bw",
4103 "-mavx512dq",
4104 "-mavx512vl",
4105 ],
4106 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4107 msvc_copts = xnnpack_msvc_std_copts(),
4108 msvc_x86_32_copts = ["/arch:AVX512"],
4109 msvc_x86_64_copts = ["/arch:AVX512"],
4110 msys_copts = ["-fno-asynchronous-unwind-tables"],
4111 x86_srcs = AVX512SKX_UKERNELS,
4112 deps = [
4113 ":tables",
4114 "@FP16",
4115 "@pthreadpool",
4116 ],
4117)
4118
4119xnnpack_cc_library(
4120 name = "avx512skx_ukernels_test_mode",
4121 hdrs = INTERNAL_HDRS,
4122 copts = [
4123 "-UNDEBUG",
4124 "-DXNN_TEST_MODE=1",
4125 ],
4126 gcc_copts = xnnpack_gcc_std_copts(),
4127 gcc_x86_copts = [
4128 "-mavx512f",
4129 "-mavx512cd",
4130 "-mavx512bw",
4131 "-mavx512dq",
4132 "-mavx512vl",
4133 ],
4134 mingw_copts = ["-fno-asynchronous-unwind-tables"],
4135 msvc_copts = xnnpack_msvc_std_copts(),
4136 msvc_x86_32_copts = ["/arch:AVX512"],
4137 msvc_x86_64_copts = ["/arch:AVX512"],
4138 msys_copts = ["-fno-asynchronous-unwind-tables"],
4139 x86_srcs = AVX512SKX_UKERNELS,
4140 deps = [
4141 ":tables",
4142 "@FP16",
4143 "@pthreadpool",
4144 ],
4145)
4146
4147xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004148 name = "asm_ukernels",
4149 hdrs = ["src/xnnpack/assembly.h"],
4150 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard31bb45b2020-10-06 00:26:33 -07004151 aarch64_copts = ["-march=armv8.2-a+fp16+dotprod"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004152 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07004153 wasm_srcs = WASM32_ASM_UKERNELS,
4154 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004155)
4156
Marat Dukhan3b59de22020-06-03 20:15:19 -07004157xnnpack_cc_library(
4158 name = "logging_utils",
4159 srcs = LOGGING_SRCS,
4160 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4161 copts = LOGGING_COPTS + [
4162 "-Isrc",
4163 "-Iinclude",
4164 ] + select({
4165 ":debug_build": [],
4166 "//conditions:default": xnnpack_min_size_copts(),
4167 }),
4168 gcc_copts = xnnpack_gcc_std_copts(),
4169 msvc_copts = xnnpack_msvc_std_copts(),
4170 visibility = xnnpack_visibility(),
4171 deps = [
4172 "@FP16",
4173 "@clog",
4174 "@pthreadpool",
4175 ],
4176)
4177
Marat Dukhan08c4a432019-10-03 09:29:21 -07004178xnnpack_aggregate_library(
4179 name = "ukernels",
4180 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004181 ":neon_ukernels",
4182 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004183 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004184 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004185 ":asm_ukernels",
4186 ],
4187 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004188 ":neon_ukernels",
4189 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07004190 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004191 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07004192 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 ":asm_ukernels",
4194 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004195 generic_deps = [
4196 ":scalar_ukernels",
4197 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004198 psimd_deps = [
4199 ":psimd_accmath_ukernels",
4200 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004201 wasm_deps = [
4202 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004203 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004204 ],
4205 wasmsimd_deps = [
4206 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07004207 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004208 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004210 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004211 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08004212 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004214 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08004215 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004216 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004217 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004218 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 ],
4220)
4221
Marat Dukhan33fcf782020-05-24 14:27:15 -07004222xnnpack_aggregate_library(
4223 name = "ukernels_test_mode",
4224 aarch32_deps = [
4225 ":neon_ukernels_test_mode",
4226 ":neonfma_ukernels_test_mode",
4227 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004228 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004229 ":asm_ukernels",
4230 ],
4231 aarch64_deps = [
4232 ":neon_ukernels_test_mode",
4233 ":neonfma_ukernels_test_mode",
4234 ":neonv8_ukernels_test_mode",
4235 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07004236 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004237 ":asm_ukernels",
4238 ],
4239 generic_deps = [
4240 ":scalar_ukernels_test_mode",
4241 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004242 psimd_deps = [
4243 ":psimd_accmath_ukernels_test_mode",
4244 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07004245 wasm_deps = [
4246 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004247 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004248 ],
4249 wasmsimd_deps = [
4250 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07004251 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004252 ],
4253 x86_deps = [
4254 ":sse2_ukernels_test_mode",
4255 ":ssse3_ukernels_test_mode",
4256 ":sse41_ukernels_test_mode",
4257 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004258 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004259 ":fma3_ukernels_test_mode",
4260 ":avx2_ukernels_test_mode",
4261 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004262 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004263 ],
4264)
4265
Marat Dukhan86df3842020-06-17 14:03:31 -07004266# This target must be compatible with Android NDK r17 and compile with Clang 6.
4267# It is similar to :ukernels target, but excludes microkernels which require
4268# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07004269xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07004270 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004271 aarch32_deps = [
4272 ":neon_ukernels",
4273 ":neonfma_ukernels",
4274 ":neonv8_ukernels",
4275 ":asm_ukernels",
4276 ],
4277 aarch64_deps = [
4278 ":neon_ukernels",
4279 ":neonfma_ukernels",
4280 ":neonv8_ukernels",
4281 ":asm_ukernels",
4282 ],
4283 generic_deps = [
4284 ":scalar_ukernels",
4285 ],
Marat Dukhan102e9ea2020-12-05 22:35:18 -08004286 psimd_deps = [
4287 ":psimd_accmath_ukernels",
4288 ],
Marat Dukhanae046f52020-06-15 13:16:14 -07004289 wasm_deps = [
4290 ":wasm_ukernels",
4291 ],
4292 wasmsimd_deps = [
4293 ":wasm_ukernels",
4294 ],
4295 x86_deps = [
4296 ":sse2_ukernels",
4297 ":ssse3_ukernels",
4298 ":sse41_ukernels",
4299 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07004300 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004301 ":fma3_ukernels",
4302 ":avx2_ukernels",
4303 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07004304 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07004305 ],
4306)
4307
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308xnnpack_cc_library(
4309 name = "im2col",
4310 srcs = ["src/im2col.c"],
4311 hdrs = [
4312 "src/xnnpack/common.h",
4313 "src/xnnpack/im2col.h",
4314 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004315 gcc_copts = xnnpack_gcc_std_copts(),
4316 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004317)
4318
4319xnnpack_cc_library(
4320 name = "indirection",
4321 srcs = ["src/indirection.c"],
4322 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004323 gcc_copts = xnnpack_gcc_std_copts(),
4324 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325 deps = [
4326 "@FP16",
4327 "@FXdiv",
4328 "@pthreadpool",
4329 ],
4330)
4331
4332xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004333 name = "indirection_test_mode",
4334 srcs = ["src/indirection.c"],
4335 hdrs = INTERNAL_HDRS,
4336 copts = [
4337 "-UNDEBUG",
4338 "-DXNN_TEST_MODE=1",
4339 ],
4340 gcc_copts = xnnpack_gcc_std_copts(),
4341 msvc_copts = xnnpack_msvc_std_copts(),
4342 deps = [
4343 "@FP16",
4344 "@FXdiv",
4345 "@pthreadpool",
4346 ],
4347)
4348
4349xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07004350 name = "packing",
4351 srcs = ["src/packing.c"],
4352 hdrs = INTERNAL_HDRS,
4353 gcc_copts = xnnpack_gcc_std_copts(),
4354 msvc_copts = xnnpack_msvc_std_copts(),
4355 deps = [
4356 "@FP16",
4357 "@FXdiv",
4358 "@pthreadpool",
4359 ],
4360)
4361
4362xnnpack_cc_library(
4363 name = "packing_test_mode",
4364 srcs = ["src/packing.c"],
4365 hdrs = INTERNAL_HDRS,
4366 copts = [
4367 "-UNDEBUG",
4368 "-DXNN_TEST_MODE=1",
4369 ],
4370 gcc_copts = xnnpack_gcc_std_copts(),
4371 msvc_copts = xnnpack_msvc_std_copts(),
4372 deps = [
4373 "@FP16",
4374 "@FXdiv",
4375 "@pthreadpool",
4376 ],
4377)
4378
4379xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004380 name = "operator_run",
4381 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07004382 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004383 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07004384 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4385 "//conditions:default": [],
4386 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004387 gcc_copts = xnnpack_gcc_std_copts(),
4388 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004389 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004390 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004391 "@FP16",
4392 "@FXdiv",
4393 "@clog",
4394 "@pthreadpool",
4395 ],
4396)
4397
Chao Mei6ddfc602020-05-13 22:29:36 -07004398xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004399 name = "operator_run_test_mode",
4400 srcs = ["src/operator-run.c"],
4401 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4402 copts = LOGGING_COPTS + [
4403 "-UNDEBUG",
4404 "-DXNN_TEST_MODE=1",
4405 ] + select({
4406 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4407 "//conditions:default": [],
4408 }),
4409 gcc_copts = xnnpack_gcc_std_copts(),
4410 msvc_copts = xnnpack_msvc_std_copts(),
4411 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004412 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004413 "@FP16",
4414 "@FXdiv",
4415 "@clog",
4416 "@pthreadpool",
4417 ],
4418)
4419
4420xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07004421 name = "memory_planner",
4422 srcs = ["src/memory-planner.c"],
4423 hdrs = INTERNAL_HDRS,
4424 defines = select({
4425 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4426 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4427 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4428 }),
4429 gcc_copts = xnnpack_gcc_std_copts(),
4430 msvc_copts = xnnpack_msvc_std_copts(),
4431 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004432 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004433 "@pthreadpool",
4434 ],
4435)
4436
Marat Dukhan33fcf782020-05-24 14:27:15 -07004437xnnpack_cc_library(
4438 name = "memory_planner_test_mode",
4439 srcs = ["src/memory-planner.c"],
4440 hdrs = INTERNAL_HDRS,
4441 copts = [
4442 "-UNDEBUG",
4443 "-DXNN_TEST_MODE=1",
4444 ],
4445 defines = select({
4446 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
4447 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
4448 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
4449 }),
4450 gcc_copts = xnnpack_gcc_std_copts(),
4451 msvc_copts = xnnpack_msvc_std_copts(),
4452 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07004453 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004454 "@pthreadpool",
4455 ],
4456)
4457
Marat Dukhan08c4a432019-10-03 09:29:21 -07004458cc_library(
4459 name = "enable_assembly",
4460 defines = select({
4461 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
4462 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07004463 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464 }),
4465)
4466
Marat Dukhan9de90e02020-06-18 16:04:12 -07004467cc_library(
4468 name = "enable_sparse",
4469 defines = select({
4470 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
4471 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07004472 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07004473 }),
4474)
4475
Marat Dukhancf056b22019-10-07 10:26:29 -07004476xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004477 name = "operators",
4478 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08004479 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004480 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07004481 ],
4482 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004483 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004484 "-Isrc",
4485 "-Iinclude",
4486 ] + select({
4487 ":debug_build": [],
4488 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004489 }) + select({
4490 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4491 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004493 gcc_copts = xnnpack_gcc_std_copts(),
4494 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07004496 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004497 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004498 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004499 "@FP16",
4500 "@FXdiv",
4501 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004503 ],
4504)
4505
Marat Dukhan10a38082020-04-17 03:58:35 -07004506xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004507 name = "operators_test_mode",
4508 srcs = OPERATOR_SRCS + [
4509 "src/memory.c",
4510 "src/operator-delete.c",
4511 ],
4512 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
4513 copts = LOGGING_COPTS + [
4514 "-Isrc",
4515 "-Iinclude",
4516 "-UNDEBUG",
4517 "-DXNN_TEST_MODE=1",
4518 ] + select({
4519 ":debug_build": [],
4520 "//conditions:default": xnnpack_min_size_copts(),
4521 }) + select({
4522 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4523 "//conditions:default": [],
4524 }),
4525 gcc_copts = xnnpack_gcc_std_copts(),
4526 msvc_copts = xnnpack_msvc_std_copts(),
4527 deps = [
4528 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004529 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07004530 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004531 "@FP16",
4532 "@FXdiv",
4533 "@clog",
4534 "@pthreadpool",
4535 ],
4536)
4537
4538xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004539 name = "XNNPACK",
4540 srcs = [
4541 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08004542 "src/runtime.c",
4543 "src/subgraph.c",
4544 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004545 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004546 hdrs = ["include/xnnpack.h"],
4547 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004548 "-Isrc",
4549 "-Iinclude",
4550 ] + select({
4551 ":debug_build": [],
4552 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004553 }) + select({
4554 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4555 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004556 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07004557 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004558 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004559 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004560 visibility = xnnpack_visibility(),
4561 deps = [
4562 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004563 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004564 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004565 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004566 ":operator_run",
4567 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004568 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004569 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004570 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004571 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004572 ] + select({
4573 ":emscripten": [],
4574 "//conditions:default": ["@cpuinfo"],
4575 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576)
4577
Marat Dukhan10a38082020-04-17 03:58:35 -07004578xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004579 name = "XNNPACK_test_mode",
4580 srcs = [
4581 "src/init.c",
4582 "src/runtime.c",
4583 "src/subgraph.c",
4584 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004585 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004586 hdrs = ["include/xnnpack.h"],
4587 copts = LOGGING_COPTS + [
4588 "-Isrc",
4589 "-Iinclude",
4590 "-UNDEBUG",
4591 "-DXNN_TEST_MODE=1",
4592 ] + select({
4593 ":debug_build": [],
4594 "//conditions:default": xnnpack_min_size_copts(),
4595 }) + select({
4596 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4597 "//conditions:default": [],
4598 }),
4599 gcc_copts = xnnpack_gcc_std_copts(),
4600 includes = ["include"],
4601 msvc_copts = xnnpack_msvc_std_copts(),
4602 visibility = xnnpack_visibility(),
4603 deps = [
4604 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004605 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004606 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004607 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004608 ":operator_run_test_mode",
4609 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004610 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004611 "@clog",
4612 "@FP16",
4613 "@pthreadpool",
4614 ] + select({
4615 ":emscripten": [],
4616 "//conditions:default": ["@cpuinfo"],
4617 }),
4618)
4619
4620xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004621 name = "xnnpack_f32",
4622 srcs = [
4623 "src/init.c",
4624 "src/runtime.c",
4625 "src/subgraph.c",
4626 "src/tensor.c",
4627 ] + SUBGRAPH_SRCS,
4628 hdrs = ["include/xnnpack.h"],
4629 copts = LOGGING_COPTS + [
4630 "-Isrc",
4631 "-Iinclude",
4632 ] + select({
4633 ":debug_build": [],
4634 "//conditions:default": xnnpack_min_size_copts(),
4635 }) + select({
4636 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4637 "//conditions:default": [],
4638 }),
4639 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004640 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004641 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004642 "XNN_NO_U8_OPERATORS",
4643 "XNN_NO_X8_OPERATORS",
4644 "XNN_NO_F16_OPERATORS",
4645 "XNN_NO_X16_OPERATORS",
4646 ],
4647 gcc_copts = xnnpack_gcc_std_copts(),
4648 includes = ["include"],
4649 msvc_copts = xnnpack_msvc_std_copts(),
4650 visibility = xnnpack_visibility(),
4651 deps = [
4652 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004653 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004654 ":logging_utils",
4655 ":memory_planner",
4656 ":operator_run",
4657 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004658 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004659 "@clog",
4660 "@FP16",
4661 "@pthreadpool",
4662 ] + select({
4663 ":emscripten": [],
4664 "//conditions:default": ["@cpuinfo"],
4665 }),
4666)
4667
4668xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004669 name = "xnnpack_operators_nhwc_f32",
4670 srcs = [
4671 "src/init.c",
4672 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004673 hdrs = ["include/xnnpack.h"],
4674 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004675 "-Isrc",
4676 "-Iinclude",
4677 ] + select({
4678 ":debug_build": [],
4679 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004680 }) + select({
4681 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4682 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004683 }),
4684 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004685 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004686 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004687 "XNN_NO_U8_OPERATORS",
4688 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004689 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004690 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004691 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004693 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004694 visibility = xnnpack_visibility(),
4695 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004696 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004697 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004698 ":operator_run",
4699 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004700 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004701 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004702 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004703 ] + select({
4704 ":emscripten": [],
4705 "//conditions:default": ["@cpuinfo"],
4706 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004707)
4708
Marat Dukhancf056b22019-10-07 10:26:29 -07004709xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004710 name = "bench_utils",
4711 srcs = ["bench/utils.cc"],
4712 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004713 deps = [
4714 "@com_google_benchmark//:benchmark",
4715 "@cpuinfo",
4716 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004717)
4718
Frank Barchard7e955972019-10-11 10:34:25 -07004719######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004720
4721xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004722 name = "qs8_gemm_bench",
4723 srcs = [
4724 "bench/gemm.h",
4725 "bench/qs8-gemm.cc",
4726 "src/xnnpack/AlignedAllocator.h",
4727 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Frank Barchard31328cb2020-10-12 11:55:18 -07004728 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
4729 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan595e1702020-07-31 10:12:52 -07004730)
4731
4732xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004733 name = "qs8_requantization_bench",
4734 srcs = [
4735 "bench/qs8-requantization.cc",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004736 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004737 "src/xnnpack/requantization-stubs.h",
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004738 ] + MICROKERNEL_BENCHMARK_HDRS,
4739 deps = MICROKERNEL_BENCHMARK_DEPS,
4740)
4741
4742xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004743 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004744 srcs = [
4745 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004746 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747 "src/xnnpack/AlignedAllocator.h",
4748 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004749 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004750 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004751)
4752
4753xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004754 name = "qu8_requantization_bench",
4755 srcs = [
4756 "bench/qu8-requantization.cc",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004757 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07004758 "src/xnnpack/requantization-stubs.h",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004759 ] + MICROKERNEL_BENCHMARK_HDRS,
4760 deps = MICROKERNEL_BENCHMARK_DEPS,
4761)
4762
4763xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004764 name = "f16_igemm_bench",
4765 srcs = [
4766 "bench/f16-igemm.cc",
4767 "bench/conv.h",
4768 "bench/google/conv.h",
4769 "src/xnnpack/AlignedAllocator.h",
4770 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004771 deps = MICROKERNEL_BENCHMARK_DEPS + [
4772 ":indirection",
4773 ":packing",
4774 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004775)
4776
4777xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778 name = "f16_gemm_bench",
4779 srcs = [
4780 "bench/f16-gemm.cc",
4781 "bench/gemm.h",
4782 "src/xnnpack/AlignedAllocator.h",
4783 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004784 deps = MICROKERNEL_BENCHMARK_DEPS + [
4785 ":packing",
4786 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004787)
4788
4789xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004790 name = "f16_spmm_bench",
4791 srcs = [
4792 "bench/f16-spmm.cc",
4793 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004794 "src/xnnpack/AlignedAllocator.h",
4795 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004796 deps = MICROKERNEL_BENCHMARK_DEPS,
4797)
4798
4799xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004800 name = "f32_igemm_bench",
4801 srcs = [
4802 "bench/f32-igemm.cc",
4803 "bench/conv.h",
4804 "src/xnnpack/AlignedAllocator.h",
4805 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004806 deps = MICROKERNEL_BENCHMARK_DEPS + [
4807 ":indirection",
4808 ":packing",
4809 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004810)
4811
4812xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004813 name = "f16_relu_bench",
4814 srcs = [
4815 "bench/f16-relu.cc",
4816 "src/xnnpack/AlignedAllocator.h",
4817 ] + MICROKERNEL_BENCHMARK_HDRS,
4818 deps = MICROKERNEL_BENCHMARK_DEPS,
4819)
4820
4821xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004822 name = "f32_conv_hwc_bench",
4823 srcs = [
4824 "bench/f32-conv-hwc.cc",
4825 "bench/dconv.h",
4826 "src/xnnpack/AlignedAllocator.h",
4827 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004828 deps = MICROKERNEL_BENCHMARK_DEPS + [
4829 ":packing",
4830 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831)
4832
4833xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004834 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004835 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004836 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004837 "bench/dconv.h",
4838 "src/xnnpack/AlignedAllocator.h",
4839 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004840 deps = MICROKERNEL_BENCHMARK_DEPS + [
4841 ":packing",
4842 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004843)
4844
4845xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004846 name = "f16_dwconv_bench",
4847 srcs = [
4848 "bench/f16-dwconv.cc",
4849 "bench/dwconv.h",
4850 "bench/google/dwconv.h",
4851 "src/xnnpack/AlignedAllocator.h",
4852 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004853 deps = MICROKERNEL_BENCHMARK_DEPS + [
4854 ":indirection",
4855 ":packing",
4856 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004857)
4858
4859xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 name = "f32_dwconv_bench",
4861 srcs = [
4862 "bench/f32-dwconv.cc",
4863 "bench/dwconv.h",
4864 "src/xnnpack/AlignedAllocator.h",
4865 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004866 deps = MICROKERNEL_BENCHMARK_DEPS + [
4867 ":indirection",
4868 ":packing",
4869 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870)
4871
4872xnnpack_benchmark(
Marat Dukhanbf715f92020-10-23 20:17:00 -07004873 name = "f32_dwconv2d_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004874 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07004875 "bench/f32-dwconv2d-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004876 "bench/dwconv.h",
4877 "src/xnnpack/AlignedAllocator.h",
4878 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004879 deps = MICROKERNEL_BENCHMARK_DEPS + [
4880 ":indirection",
4881 ":packing",
4882 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883)
4884
4885xnnpack_benchmark(
4886 name = "f32_gemm_bench",
4887 srcs = [
4888 "bench/f32-gemm.cc",
4889 "bench/gemm.h",
4890 "src/xnnpack/AlignedAllocator.h",
4891 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004892 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004893 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004894)
4895
4896xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004897 name = "f32_hswish_bench",
4898 srcs = [
4899 "bench/f32-hswish.cc",
4900 "src/xnnpack/AlignedAllocator.h",
4901 ] + MICROKERNEL_BENCHMARK_HDRS,
4902 deps = MICROKERNEL_BENCHMARK_DEPS,
4903)
4904
4905xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004906 name = "f32_raddexpminusmax_bench",
4907 srcs = [
4908 "bench/f32-raddexpminusmax.cc",
4909 "src/xnnpack/AlignedAllocator.h",
4910 ] + MICROKERNEL_BENCHMARK_HDRS,
4911 deps = MICROKERNEL_BENCHMARK_DEPS,
4912)
4913
4914xnnpack_benchmark(
4915 name = "f32_raddextexp_bench",
4916 srcs = [
4917 "bench/f32-raddextexp.cc",
4918 "src/xnnpack/AlignedAllocator.h",
4919 ] + MICROKERNEL_BENCHMARK_HDRS,
4920 deps = MICROKERNEL_BENCHMARK_DEPS,
4921)
4922
4923xnnpack_benchmark(
4924 name = "f32_raddstoreexpminusmax_bench",
4925 srcs = [
4926 "bench/f32-raddstoreexpminusmax.cc",
4927 "src/xnnpack/AlignedAllocator.h",
4928 ] + MICROKERNEL_BENCHMARK_HDRS,
4929 deps = MICROKERNEL_BENCHMARK_DEPS,
4930)
4931
4932xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004933 name = "f32_relu_bench",
4934 srcs = [
4935 "bench/f32-relu.cc",
4936 "src/xnnpack/AlignedAllocator.h",
4937 ] + MICROKERNEL_BENCHMARK_HDRS,
4938 deps = MICROKERNEL_BENCHMARK_DEPS,
4939)
4940
4941xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942 name = "f32_rmax_bench",
4943 srcs = [
4944 "bench/f32-rmax.cc",
4945 "src/xnnpack/AlignedAllocator.h",
4946 ] + MICROKERNEL_BENCHMARK_HDRS,
4947 deps = MICROKERNEL_BENCHMARK_DEPS,
4948)
4949
4950xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004951 name = "f32_sigmoid_bench",
4952 srcs = [
4953 "bench/f32-sigmoid.cc",
4954 "src/xnnpack/AlignedAllocator.h",
4955 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004956 deps = MICROKERNEL_BENCHMARK_DEPS,
4957)
4958
4959xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 name = "f32_spmm_bench",
4961 srcs = [
4962 "bench/f32-spmm.cc",
4963 "bench/gemm.h",
4964 "src/xnnpack/AlignedAllocator.h",
4965 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004966 deps = MICROKERNEL_BENCHMARK_DEPS,
4967)
4968
4969xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004970 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004971 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004972 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004973 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004974 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004975 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004976)
4977
4978xnnpack_benchmark(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08004979 name = "f32_velu_bench",
4980 srcs = [
4981 "bench/f32-velu.cc",
4982 "src/xnnpack/AlignedAllocator.h",
4983 ] + MICROKERNEL_BENCHMARK_HDRS,
4984 deps = MICROKERNEL_BENCHMARK_DEPS,
4985)
4986
4987xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004988 name = "f32_vscaleexpminusmax_bench",
4989 srcs = [
4990 "bench/f32-vscaleexpminusmax.cc",
4991 "src/xnnpack/AlignedAllocator.h",
4992 ] + MICROKERNEL_BENCHMARK_HDRS,
4993 deps = MICROKERNEL_BENCHMARK_DEPS,
4994)
4995
4996xnnpack_benchmark(
4997 name = "f32_vscaleextexp_bench",
4998 srcs = [
4999 "bench/f32-vscaleextexp.cc",
5000 "src/xnnpack/AlignedAllocator.h",
5001 ] + MICROKERNEL_BENCHMARK_HDRS,
5002 deps = MICROKERNEL_BENCHMARK_DEPS,
5003)
5004
5005xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005006 name = "f32_vsqrt_bench",
5007 srcs = [
5008 "bench/f32-vsqrt.cc",
5009 "src/xnnpack/AlignedAllocator.h",
5010 ] + MICROKERNEL_BENCHMARK_HDRS,
5011 deps = MICROKERNEL_BENCHMARK_DEPS,
5012)
5013
5014xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015 name = "f32_im2col_gemm_bench",
5016 srcs = [
5017 "bench/f32-im2col-gemm.cc",
5018 "bench/conv.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005021 deps = MICROKERNEL_BENCHMARK_DEPS + [
5022 ":im2col",
5023 ":packing",
5024 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005025)
5026
Marat Dukhanfe7acb62020-03-09 19:30:05 -07005027xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005028 name = "rounding_bench",
5029 srcs = [
5030 "bench/rounding.cc",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005031 "src/xnnpack/AlignedAllocator.h",
Frank Barchard04336c12020-10-22 16:48:55 -07005032 "src/xnnpack/math-stubs.h",
Marat Dukhanffbf96a2020-05-14 02:59:08 -07005033 ] + MICROKERNEL_BENCHMARK_HDRS,
5034 deps = MICROKERNEL_BENCHMARK_DEPS,
5035)
5036
Marat Dukhan08c4a432019-10-03 09:29:21 -07005037########################### Benchmarks for operators ###########################
5038
5039xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005040 name = "average_pooling_bench",
5041 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07005042 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005043 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005044 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045)
5046
5047xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005048 name = "bankers_rounding_bench",
5049 srcs = ["bench/bankers-rounding.cc"],
5050 copts = xnnpack_optional_tflite_copts(),
5051 tags = ["nowin32"],
5052 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5053)
5054
5055xnnpack_benchmark(
5056 name = "ceiling_bench",
5057 srcs = ["bench/ceiling.cc"],
5058 copts = xnnpack_optional_tflite_copts(),
5059 tags = ["nowin32"],
5060 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5061)
5062
5063xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005064 name = "channel_shuffle_bench",
5065 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005066 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005067)
5068
5069xnnpack_benchmark(
5070 name = "convolution_bench",
5071 srcs = ["bench/convolution.cc"],
5072 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005073 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005074 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075)
5076
5077xnnpack_benchmark(
5078 name = "deconvolution_bench",
5079 srcs = ["bench/deconvolution.cc"],
5080 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005081 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005082 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005083)
5084
5085xnnpack_benchmark(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08005086 name = "elu_bench",
5087 srcs = ["bench/elu.cc"],
5088 copts = xnnpack_optional_tflite_copts(),
5089 tags = ["nowin32"],
5090 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5091)
5092
5093xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005094 name = "floor_bench",
5095 srcs = ["bench/floor.cc"],
5096 copts = xnnpack_optional_tflite_copts(),
5097 tags = ["nowin32"],
5098 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5099)
5100
5101xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005102 name = "global_average_pooling_bench",
5103 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005104 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005105)
5106
5107xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07005108 name = "hardswish_bench",
5109 srcs = ["bench/hardswish.cc"],
5110 copts = xnnpack_optional_tflite_copts(),
5111 tags = ["nowin32"],
5112 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5113)
5114
5115xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005116 name = "max_pooling_bench",
5117 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005118 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005119)
5120
5121xnnpack_benchmark(
5122 name = "sigmoid_bench",
5123 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08005124 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005125 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005126 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005127)
5128
5129xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07005130 name = "prelu_bench",
5131 srcs = ["bench/prelu.cc"],
5132 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07005133 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005134 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07005135)
5136
5137xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005138 name = "softmax_bench",
5139 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08005140 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07005141 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07005142 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07005143)
5144
Marat Dukhan87727142020-06-24 15:24:10 -07005145xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005146 name = "square_root_bench",
5147 srcs = ["bench/square-root.cc"],
5148 copts = xnnpack_optional_tflite_copts(),
5149 tags = ["nowin32"],
5150 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
5151)
5152
5153xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07005154 name = "truncation_bench",
5155 srcs = ["bench/truncation.cc"],
5156 deps = OPERATOR_BENCHMARK_DEPS,
5157)
5158
Marat Dukhanc068bb62019-10-04 13:24:39 -07005159############################# End-to-end benchmarks ############################
5160
5161cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005162 name = "fp32_mobilenet_v1",
5163 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005164 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005165 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005166 linkstatic = True,
5167 deps = [
5168 ":XNNPACK",
5169 "@pthreadpool",
5170 ],
5171)
5172
5173cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005174 name = "fp16_mobilenet_v1",
5175 srcs = ["models/fp16-mobilenet-v1.cc"],
5176 hdrs = ["models/models.h"],
5177 copts = xnnpack_std_cxxopts(),
5178 linkstatic = True,
5179 deps = [
5180 ":XNNPACK",
5181 "@FP16",
5182 "@pthreadpool",
5183 ],
5184)
5185
5186cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005187 name = "qs8_mobilenet_v1",
5188 srcs = ["models/qs8-mobilenet-v1.cc"],
5189 hdrs = ["models/models.h"],
5190 copts = xnnpack_std_cxxopts(),
5191 linkstatic = True,
5192 deps = [
5193 ":XNNPACK",
5194 "@pthreadpool",
5195 ],
5196)
5197
5198cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07005199 name = "qs8_mobilenet_v2",
5200 srcs = ["models/qs8-mobilenet-v2.cc"],
5201 hdrs = ["models/models.h"],
5202 copts = xnnpack_std_cxxopts(),
5203 linkstatic = True,
5204 deps = [
5205 ":XNNPACK",
5206 "@pthreadpool",
5207 ],
5208)
5209
5210cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005211 name = "fp32_mobilenet_v2",
5212 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07005213 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005214 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07005215 linkstatic = True,
5216 deps = [
5217 ":XNNPACK",
5218 "@pthreadpool",
5219 ],
5220)
5221
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005222cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005223 name = "fp16_mobilenet_v2",
5224 srcs = ["models/fp16-mobilenet-v2.cc"],
5225 hdrs = ["models/models.h"],
5226 copts = xnnpack_std_cxxopts(),
5227 linkstatic = True,
5228 deps = [
5229 ":XNNPACK",
5230 "@FP16",
5231 "@pthreadpool",
5232 ],
5233)
5234
5235cc_library(
5236 name = "fp32_mobilenet_v3_large",
5237 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005238 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005239 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005240 linkstatic = True,
5241 deps = [
5242 ":XNNPACK",
5243 "@pthreadpool",
5244 ],
5245)
5246
5247cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005248 name = "fp16_mobilenet_v3_large",
5249 srcs = ["models/fp16-mobilenet-v3-large.cc"],
5250 hdrs = ["models/models.h"],
5251 copts = xnnpack_std_cxxopts(),
5252 linkstatic = True,
5253 deps = [
5254 ":XNNPACK",
5255 "@FP16",
5256 "@pthreadpool",
5257 ],
5258)
5259
5260cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07005261 name = "fp32_mobilenet_v3_small",
5262 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005263 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08005264 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005265 linkstatic = True,
5266 deps = [
5267 ":XNNPACK",
5268 "@pthreadpool",
5269 ],
5270)
5271
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005272cc_library(
5273 name = "fp16_mobilenet_v3_small",
5274 srcs = ["models/fp16-mobilenet-v3-small.cc"],
5275 hdrs = ["models/models.h"],
5276 copts = xnnpack_std_cxxopts(),
5277 linkstatic = True,
5278 deps = [
5279 ":XNNPACK",
5280 "@FP16",
5281 "@pthreadpool",
5282 ],
5283)
5284
Marat Dukhanc068bb62019-10-04 13:24:39 -07005285xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07005286 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005287 srcs = [
5288 "bench/f32-dwconv-e2e.cc",
5289 "bench/end2end.h",
5290 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07005291 deps = MICROKERNEL_BENCHMARK_DEPS + [
5292 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005293 ":fp32_mobilenet_v1",
5294 ":fp32_mobilenet_v2",
5295 ":fp32_mobilenet_v3_large",
5296 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07005297 ],
5298)
5299
5300xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07005301 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08005302 srcs = [
5303 "bench/f32-gemm-e2e.cc",
5304 "bench/end2end.h",
5305 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07005306 deps = MICROKERNEL_BENCHMARK_DEPS + [
5307 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005308 ":fp32_mobilenet_v1",
5309 ":fp32_mobilenet_v2",
5310 ":fp32_mobilenet_v3_large",
5311 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07005312 ],
5313)
5314
5315xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07005316 name = "end2end_bench",
5317 srcs = ["bench/end2end.cc"],
5318 deps = [
5319 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07005320 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005321 ":fp16_mobilenet_v1",
5322 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07005323 ":fp16_mobilenet_v3_large",
5324 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07005325 ":fp32_mobilenet_v1",
5326 ":fp32_mobilenet_v2",
5327 ":fp32_mobilenet_v3_large",
5328 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07005329 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07005330 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07005331 "@pthreadpool",
5332 ],
5333)
5334
Marat Dukhan6adff4e2019-10-14 18:32:07 -07005335#################### Accuracy evaluation for math functions ####################
5336
5337xnnpack_benchmark(
5338 name = "f32_exp_eval",
5339 srcs = [
5340 "eval/f32-exp.cc",
5341 "src/xnnpack/AlignedAllocator.h",
5342 ] + ACCURACY_EVAL_HDRS,
5343 deps = ACCURACY_EVAL_DEPS,
5344)
5345
Marat Dukhan515c9772019-10-17 18:07:57 -07005346xnnpack_benchmark(
5347 name = "f32_expminus_eval",
5348 srcs = [
5349 "eval/f32-expminus.cc",
5350 "src/xnnpack/AlignedAllocator.h",
5351 ] + ACCURACY_EVAL_HDRS,
5352 deps = ACCURACY_EVAL_DEPS,
5353)
5354
Marat Dukhan98ba4412019-10-23 02:14:28 -07005355xnnpack_benchmark(
Marat Dukhana438aca2020-11-20 15:45:01 -08005356 name = "f32_expm1minus_eval",
5357 srcs = [
5358 "eval/f32-expm1minus.cc",
5359 "src/xnnpack/AlignedAllocator.h",
5360 ] + ACCURACY_EVAL_HDRS,
Marat Dukhande390d42020-11-29 19:32:18 -08005361 deps = ACCURACY_EVAL_DEPS + [
5362 ":bench_utils",
5363 ],
Marat Dukhana438aca2020-11-20 15:45:01 -08005364)
5365
5366xnnpack_benchmark(
Marat Dukhan98ba4412019-10-23 02:14:28 -07005367 name = "f32_extexp_eval",
5368 srcs = [
5369 "eval/f32-extexp.cc",
5370 "src/xnnpack/AlignedAllocator.h",
5371 ] + ACCURACY_EVAL_HDRS,
5372 deps = ACCURACY_EVAL_DEPS,
5373)
5374
Marat Dukhan8853b822020-05-07 12:19:01 -07005375xnnpack_unit_test(
5376 name = "f32_roundne_eval",
5377 srcs = [
5378 "eval/f32-roundne.cc",
5379 "src/xnnpack/AlignedAllocator.h",
5380 "src/xnnpack/math-stubs.h",
5381 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07005382 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07005383 deps = MICROKERNEL_TEST_DEPS,
5384)
5385
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005386xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005387 name = "f32_roundd_eval",
5388 srcs = [
5389 "eval/f32-roundd.cc",
5390 "src/xnnpack/AlignedAllocator.h",
5391 "src/xnnpack/math-stubs.h",
5392 ] + MICROKERNEL_TEST_HDRS,
5393 automatic = False,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
5398 name = "f32_roundu_eval",
5399 srcs = [
5400 "eval/f32-roundu.cc",
5401 "src/xnnpack/AlignedAllocator.h",
5402 "src/xnnpack/math-stubs.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 automatic = False,
5405 deps = MICROKERNEL_TEST_DEPS,
5406)
5407
5408xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005409 name = "f32_roundz_eval",
5410 srcs = [
5411 "eval/f32-roundz.cc",
5412 "src/xnnpack/AlignedAllocator.h",
5413 "src/xnnpack/math-stubs.h",
5414 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07005415 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07005416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
Marat Dukhan346a9e52019-11-15 09:06:30 -08005419xnnpack_benchmark(
5420 name = "f32_sigmoid_eval",
5421 srcs = [
5422 "eval/f32-sigmoid.cc",
5423 "src/xnnpack/AlignedAllocator.h",
5424 ] + ACCURACY_EVAL_HDRS,
5425 deps = ACCURACY_EVAL_DEPS,
5426)
5427
Marat Dukhan84000762020-06-29 18:38:43 -07005428xnnpack_benchmark(
5429 name = "f32_sqrt_eval",
5430 srcs = [
5431 "eval/f32-sqrt.cc",
5432 "src/xnnpack/AlignedAllocator.h",
5433 ] + ACCURACY_EVAL_HDRS,
5434 deps = ACCURACY_EVAL_DEPS + [
5435 ":bench_utils",
5436 ],
5437)
5438
Marat Dukhan08c4a432019-10-03 09:29:21 -07005439######################### Unit tests for micro-kernels #########################
5440
5441xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07005442 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005443 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07005444 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005445 "test/gemm-microkernel-tester.h",
5446 "src/xnnpack/AlignedAllocator.h",
5447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005448 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005449)
5450
5451xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005452 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005453 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005454 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08005455 "test/spmm-microkernel-tester.h",
5456 "src/xnnpack/AlignedAllocator.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005462 name = "f32_argmaxpool_test",
5463 srcs = [
5464 "test/f32-argmaxpool.cc",
5465 "test/argmaxpool-microkernel-tester.h",
5466 "src/xnnpack/AlignedAllocator.h",
5467 ] + MICROKERNEL_TEST_HDRS,
5468 deps = MICROKERNEL_TEST_DEPS,
5469)
5470
5471xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005472 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005474 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005475 "test/avgpool-microkernel-tester.h",
5476 "src/xnnpack/AlignedAllocator.h",
5477 ] + MICROKERNEL_TEST_HDRS,
5478 deps = MICROKERNEL_TEST_DEPS,
5479)
5480
5481xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07005482 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005483 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07005484 "test/f32-ibilinear.cc",
5485 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08005486 "src/xnnpack/AlignedAllocator.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
XNNPACK Team21432672020-10-19 19:58:48 -07005492 name = "f32_ibilinear_chw_test",
5493 srcs = [
5494 "test/f32-ibilinear-chw.cc",
XNNPACK Team6be46b22020-10-22 23:34:54 -07005495 "test/ibilinear-microkernel-tester.h",
XNNPACK Team21432672020-10-19 19:58:48 -07005496 "src/xnnpack/AlignedAllocator.h",
5497 ] + MICROKERNEL_TEST_HDRS,
5498 deps = MICROKERNEL_TEST_DEPS,
5499)
5500
5501xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005502 name = "f16_clamp_test",
5503 srcs = [
5504 "test/f16-clamp.cc",
5505 "test/clamp-microkernel-tester.h",
5506 ] + MICROKERNEL_TEST_HDRS,
5507 deps = MICROKERNEL_TEST_DEPS,
5508)
5509
5510xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005511 name = "f32_clamp_test",
5512 srcs = [
5513 "test/f32-clamp.cc",
5514 "test/clamp-microkernel-tester.h",
5515 ] + MICROKERNEL_TEST_HDRS,
5516 deps = MICROKERNEL_TEST_DEPS,
5517)
5518
5519xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005520 name = "f32_igemm_test",
5521 srcs = [
5522 "test/f32-igemm.cc",
5523 "test/gemm-microkernel-tester.h",
5524 "src/xnnpack/AlignedAllocator.h",
5525 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005526 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005527)
5528
5529xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005530 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005531 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07005532 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005533 "test/gemm-microkernel-tester.h",
5534 "src/xnnpack/AlignedAllocator.h",
5535 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005536 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005537)
5538
5539xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07005540 name = "f32_igemm_minmax_test",
5541 srcs = [
5542 "test/f32-igemm-minmax.cc",
5543 "test/gemm-microkernel-tester.h",
5544 "src/xnnpack/AlignedAllocator.h",
5545 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005546 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07005547)
5548
5549xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005550 name = "f16_igemm_minmax_test",
5551 srcs = [
5552 "test/f16-igemm-minmax.cc",
5553 "test/gemm-microkernel-tester.h",
5554 "src/xnnpack/AlignedAllocator.h",
5555 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005556 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07005557)
5558
5559xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005560 name = "f32_conv_hwc_test",
5561 srcs = [
5562 "test/f32-conv-hwc.cc",
5563 "test/conv-hwc-microkernel-tester.h",
5564 "src/xnnpack/AlignedAllocator.h",
5565 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005566 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005567)
5568
5569xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005570 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005572 "test/f32-conv-hwc2chw.cc",
5573 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005574 "src/xnnpack/AlignedAllocator.h",
5575 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005576 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005577)
5578
5579xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005580 name = "f32_dwconv_test",
5581 srcs = [
5582 "test/f32-dwconv.cc",
5583 "test/dwconv-microkernel-tester.h",
5584 "src/xnnpack/AlignedAllocator.h",
5585 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005586 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005587)
5588
5589xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07005590 name = "f16_dwconv_minmax_test",
5591 srcs = [
5592 "test/f16-dwconv-minmax.cc",
5593 "test/dwconv-microkernel-tester.h",
5594 "src/xnnpack/AlignedAllocator.h",
5595 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005596 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07005597)
5598
5599xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005600 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005601 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005602 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005603 "test/dwconv-microkernel-tester.h",
5604 "src/xnnpack/AlignedAllocator.h",
5605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005606 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005607)
5608
5609xnnpack_unit_test(
Marat Dukhanbf715f92020-10-23 20:17:00 -07005610 name = "f32_dwconv2d_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005611 srcs = [
Marat Dukhanbf715f92020-10-23 20:17:00 -07005612 "test/f32-dwconv2d-chw.cc",
5613 "test/dwconv2d-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005614 "src/xnnpack/AlignedAllocator.h",
5615 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005616 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005617)
5618
5619xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005620 name = "f16_gavgpool_minmax_test",
5621 srcs = [
5622 "test/f16-gavgpool-minmax.cc",
5623 "test/gavgpool-microkernel-tester.h",
5624 "src/xnnpack/AlignedAllocator.h",
5625 ] + MICROKERNEL_TEST_HDRS,
5626 deps = MICROKERNEL_TEST_DEPS,
5627)
5628
5629xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005630 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005631 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005632 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005633 "test/gavgpool-microkernel-tester.h",
5634 "src/xnnpack/AlignedAllocator.h",
5635 ] + MICROKERNEL_TEST_HDRS,
5636 deps = MICROKERNEL_TEST_DEPS,
5637)
5638
5639xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005640 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005642 "test/f32-gavgpool-cw.cc",
5643 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005644 "src/xnnpack/AlignedAllocator.h",
5645 ] + MICROKERNEL_TEST_HDRS,
5646 deps = MICROKERNEL_TEST_DEPS,
5647)
5648
5649xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005650 name = "f32_gemm_test",
5651 srcs = [
5652 "test/f32-gemm.cc",
5653 "test/gemm-microkernel-tester.h",
5654 "src/xnnpack/AlignedAllocator.h",
5655 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005656 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005657)
5658
5659xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005660 name = "f32_gemm_relu_test",
5661 srcs = [
5662 "test/f32-gemm-relu.cc",
5663 "test/gemm-microkernel-tester.h",
5664 "src/xnnpack/AlignedAllocator.h",
5665 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005666 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005667)
5668
5669xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005670 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005671 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005672 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005673 "test/gemm-microkernel-tester.h",
5674 "src/xnnpack/AlignedAllocator.h",
5675 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005676 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005677)
5678
5679xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005680 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005681 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005682 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005683 "test/gemm-microkernel-tester.h",
5684 "src/xnnpack/AlignedAllocator.h",
5685 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005686 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005687)
5688
5689xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005690 name = "f16_hswish_test",
5691 srcs = [
5692 "test/f16-hswish.cc",
5693 "test/hswish-microkernel-tester.h",
5694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005699 name = "f32_hswish_test",
5700 srcs = [
5701 "test/f32-hswish.cc",
5702 "test/hswish-microkernel-tester.h",
5703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005708 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005709 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005710 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005711 "test/maxpool-microkernel-tester.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005717 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005718 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005719 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005720 "test/avgpool-microkernel-tester.h",
5721 "src/xnnpack/AlignedAllocator.h",
5722 ] + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS,
5724)
5725
5726xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005727 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005729 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005730 "test/gemm-microkernel-tester.h",
5731 "src/xnnpack/AlignedAllocator.h",
5732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005733 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005734)
5735
5736xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005737 name = "f16_prelu_test",
5738 srcs = [
5739 "test/f16-prelu.cc",
5740 "test/prelu-microkernel-tester.h",
5741 "src/xnnpack/AlignedAllocator.h",
5742 ] + MICROKERNEL_TEST_HDRS,
5743 deps = MICROKERNEL_TEST_DEPS,
5744)
5745
5746xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005747 name = "f32_prelu_test",
5748 srcs = [
5749 "test/f32-prelu.cc",
5750 "test/prelu-microkernel-tester.h",
5751 "src/xnnpack/AlignedAllocator.h",
5752 ] + MICROKERNEL_TEST_HDRS,
5753 deps = MICROKERNEL_TEST_DEPS,
5754)
5755
5756xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005757 name = "f32_raddexpminusmax_test",
5758 srcs = [
5759 "test/f32-raddexpminusmax.cc",
5760 "test/raddexpminusmax-microkernel-tester.h",
5761 ] + MICROKERNEL_TEST_HDRS,
5762 deps = MICROKERNEL_TEST_DEPS,
5763)
5764
5765xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005766 name = "f32_raddextexp_test",
5767 srcs = [
5768 "test/f32-raddextexp.cc",
5769 "test/raddextexp-microkernel-tester.h",
5770 ] + MICROKERNEL_TEST_HDRS,
5771 deps = MICROKERNEL_TEST_DEPS,
5772)
5773
5774xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005775 name = "f32_raddstoreexpminusmax_test",
5776 srcs = [
5777 "test/f32-raddstoreexpminusmax.cc",
5778 "test/raddstoreexpminusmax-microkernel-tester.h",
5779 ] + MICROKERNEL_TEST_HDRS,
5780 deps = MICROKERNEL_TEST_DEPS,
5781)
5782
5783xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005784 name = "f32_relu_test",
5785 srcs = [
5786 "test/f32-relu.cc",
5787 "test/vunary-microkernel-tester.h",
5788 ] + MICROKERNEL_TEST_HDRS,
5789 deps = MICROKERNEL_TEST_DEPS,
5790)
5791
5792xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005793 name = "f32_rmax_test",
5794 srcs = [
5795 "test/f32-rmax.cc",
5796 "test/rmax-microkernel-tester.h",
5797 ] + MICROKERNEL_TEST_HDRS,
5798 deps = MICROKERNEL_TEST_DEPS,
5799)
5800
5801xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005802 name = "f32_sigmoid_test",
5803 srcs = [
5804 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005805 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005806 ] + MICROKERNEL_TEST_HDRS,
5807 deps = MICROKERNEL_TEST_DEPS,
5808)
5809
5810xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005811 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005812 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005813 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005814 "test/spmm-microkernel-tester.h",
5815 "src/xnnpack/AlignedAllocator.h",
5816 ] + MICROKERNEL_TEST_HDRS,
5817 deps = MICROKERNEL_TEST_DEPS,
5818)
5819
5820xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005821 name = "f16_vadd_minmax_test",
5822 srcs = [
5823 "test/f16-vadd-minmax.cc",
5824 "test/vbinary-microkernel-tester.h",
5825 ] + MICROKERNEL_TEST_HDRS,
5826 deps = MICROKERNEL_TEST_DEPS,
5827)
5828
5829xnnpack_unit_test(
5830 name = "f16_vaddc_minmax_test",
5831 srcs = [
5832 "test/f16-vaddc-minmax.cc",
5833 "test/vbinaryc-microkernel-tester.h",
5834 ] + MICROKERNEL_TEST_HDRS,
5835 deps = MICROKERNEL_TEST_DEPS,
5836)
5837
5838xnnpack_unit_test(
5839 name = "f16_vdiv_minmax_test",
5840 srcs = [
5841 "test/f16-vdiv-minmax.cc",
5842 "test/vbinary-microkernel-tester.h",
5843 ] + MICROKERNEL_TEST_HDRS,
5844 deps = MICROKERNEL_TEST_DEPS,
5845)
5846
5847xnnpack_unit_test(
5848 name = "f16_vdivc_minmax_test",
5849 srcs = [
5850 "test/f16-vdivc-minmax.cc",
5851 "test/vbinaryc-microkernel-tester.h",
5852 ] + MICROKERNEL_TEST_HDRS,
5853 deps = MICROKERNEL_TEST_DEPS,
5854)
5855
5856xnnpack_unit_test(
5857 name = "f16_vrdivc_minmax_test",
5858 srcs = [
5859 "test/f16-vrdivc-minmax.cc",
5860 "test/vbinaryc-microkernel-tester.h",
5861 ] + MICROKERNEL_TEST_HDRS,
5862 deps = MICROKERNEL_TEST_DEPS,
5863)
5864
5865xnnpack_unit_test(
5866 name = "f16_vmax_test",
5867 srcs = [
5868 "test/f16-vmax.cc",
5869 "test/vbinary-microkernel-tester.h",
5870 ] + MICROKERNEL_TEST_HDRS,
5871 deps = MICROKERNEL_TEST_DEPS,
5872)
5873
5874xnnpack_unit_test(
5875 name = "f16_vmaxc_test",
5876 srcs = [
5877 "test/f16-vmaxc.cc",
5878 "test/vbinaryc-microkernel-tester.h",
5879 ] + MICROKERNEL_TEST_HDRS,
5880 deps = MICROKERNEL_TEST_DEPS,
5881)
5882
5883xnnpack_unit_test(
5884 name = "f16_vmin_test",
5885 srcs = [
5886 "test/f16-vmin.cc",
5887 "test/vbinary-microkernel-tester.h",
5888 ] + MICROKERNEL_TEST_HDRS,
5889 deps = MICROKERNEL_TEST_DEPS,
5890)
5891
5892xnnpack_unit_test(
5893 name = "f16_vminc_test",
5894 srcs = [
5895 "test/f16-vminc.cc",
5896 "test/vbinaryc-microkernel-tester.h",
5897 ] + MICROKERNEL_TEST_HDRS,
5898 deps = MICROKERNEL_TEST_DEPS,
5899)
5900
5901xnnpack_unit_test(
5902 name = "f16_vmul_minmax_test",
5903 srcs = [
5904 "test/f16-vmul-minmax.cc",
5905 "test/vbinary-microkernel-tester.h",
5906 ] + MICROKERNEL_TEST_HDRS,
5907 deps = MICROKERNEL_TEST_DEPS,
5908)
5909
5910xnnpack_unit_test(
5911 name = "f16_vmulc_minmax_test",
5912 srcs = [
5913 "test/f16-vmulc-minmax.cc",
5914 "test/vbinaryc-microkernel-tester.h",
5915 ] + MICROKERNEL_TEST_HDRS,
5916 deps = MICROKERNEL_TEST_DEPS,
5917)
5918
5919xnnpack_unit_test(
5920 name = "f16_vsub_minmax_test",
5921 srcs = [
5922 "test/f16-vsub-minmax.cc",
5923 "test/vbinary-microkernel-tester.h",
5924 ] + MICROKERNEL_TEST_HDRS,
5925 deps = MICROKERNEL_TEST_DEPS,
5926)
5927
5928xnnpack_unit_test(
5929 name = "f16_vsubc_minmax_test",
5930 srcs = [
5931 "test/f16-vsubc-minmax.cc",
5932 "test/vbinaryc-microkernel-tester.h",
5933 ] + MICROKERNEL_TEST_HDRS,
5934 deps = MICROKERNEL_TEST_DEPS,
5935)
5936
5937xnnpack_unit_test(
5938 name = "f16_vrsubc_minmax_test",
5939 srcs = [
5940 "test/f16-vrsubc-minmax.cc",
5941 "test/vbinaryc-microkernel-tester.h",
5942 ] + MICROKERNEL_TEST_HDRS,
5943 deps = MICROKERNEL_TEST_DEPS,
5944)
5945
5946xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005947 name = "f32_vabs_test",
5948 srcs = [
5949 "test/f32-vabs.cc",
5950 "test/vunary-microkernel-tester.h",
5951 ] + MICROKERNEL_TEST_HDRS,
5952 deps = MICROKERNEL_TEST_DEPS,
5953)
5954
5955xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005956 name = "f32_vadd_test",
5957 srcs = [
5958 "test/f32-vadd.cc",
5959 "test/vbinary-microkernel-tester.h",
5960 ] + MICROKERNEL_TEST_HDRS,
5961 deps = MICROKERNEL_TEST_DEPS,
5962)
5963
5964xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005965 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005966 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005967 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005968 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005969 ] + MICROKERNEL_TEST_HDRS,
5970 deps = MICROKERNEL_TEST_DEPS,
5971)
5972
5973xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005974 name = "f32_vadd_relu_test",
5975 srcs = [
5976 "test/f32-vadd-relu.cc",
5977 "test/vbinary-microkernel-tester.h",
5978 ] + MICROKERNEL_TEST_HDRS,
5979 deps = MICROKERNEL_TEST_DEPS,
5980)
5981
5982xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005983 name = "f32_vaddc_test",
5984 srcs = [
5985 "test/f32-vaddc.cc",
5986 "test/vbinaryc-microkernel-tester.h",
5987 ] + MICROKERNEL_TEST_HDRS,
5988 deps = MICROKERNEL_TEST_DEPS,
5989)
5990
5991xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005992 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005993 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005994 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005995 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005996 ] + MICROKERNEL_TEST_HDRS,
5997 deps = MICROKERNEL_TEST_DEPS,
5998)
5999
6000xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006001 name = "f32_vaddc_relu_test",
6002 srcs = [
6003 "test/f32-vaddc-relu.cc",
6004 "test/vbinaryc-microkernel-tester.h",
6005 ] + MICROKERNEL_TEST_HDRS,
6006 deps = MICROKERNEL_TEST_DEPS,
6007)
6008
6009xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006010 name = "f32_vdiv_test",
6011 srcs = [
6012 "test/f32-vdiv.cc",
6013 "test/vbinary-microkernel-tester.h",
6014 ] + MICROKERNEL_TEST_HDRS,
6015 deps = MICROKERNEL_TEST_DEPS,
6016)
6017
6018xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006019 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006020 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006021 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006022 "test/vbinary-microkernel-tester.h",
6023 ] + MICROKERNEL_TEST_HDRS,
6024 deps = MICROKERNEL_TEST_DEPS,
6025)
6026
6027xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006028 name = "f32_vdiv_relu_test",
6029 srcs = [
6030 "test/f32-vdiv-relu.cc",
6031 "test/vbinary-microkernel-tester.h",
6032 ] + MICROKERNEL_TEST_HDRS,
6033 deps = MICROKERNEL_TEST_DEPS,
6034)
6035
6036xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006037 name = "f32_vdivc_test",
6038 srcs = [
6039 "test/f32-vdivc.cc",
6040 "test/vbinaryc-microkernel-tester.h",
6041 ] + MICROKERNEL_TEST_HDRS,
6042 deps = MICROKERNEL_TEST_DEPS,
6043)
6044
6045xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006046 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006047 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006048 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006049 "test/vbinaryc-microkernel-tester.h",
6050 ] + MICROKERNEL_TEST_HDRS,
6051 deps = MICROKERNEL_TEST_DEPS,
6052)
6053
6054xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006055 name = "f32_vdivc_relu_test",
6056 srcs = [
6057 "test/f32-vdivc-relu.cc",
6058 "test/vbinaryc-microkernel-tester.h",
6059 ] + MICROKERNEL_TEST_HDRS,
6060 deps = MICROKERNEL_TEST_DEPS,
6061)
6062
6063xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006064 name = "f32_vrdivc_test",
6065 srcs = [
6066 "test/f32-vrdivc.cc",
6067 "test/vbinaryc-microkernel-tester.h",
6068 ] + MICROKERNEL_TEST_HDRS,
6069 deps = MICROKERNEL_TEST_DEPS,
6070)
6071
6072xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006073 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006074 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006075 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08006076 "test/vbinaryc-microkernel-tester.h",
6077 ] + MICROKERNEL_TEST_HDRS,
6078 deps = MICROKERNEL_TEST_DEPS,
6079)
6080
6081xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006082 name = "f32_vrdivc_relu_test",
6083 srcs = [
6084 "test/f32-vrdivc-relu.cc",
6085 "test/vbinaryc-microkernel-tester.h",
6086 ] + MICROKERNEL_TEST_HDRS,
6087 deps = MICROKERNEL_TEST_DEPS,
6088)
6089
6090xnnpack_unit_test(
Marat Dukhaned6baaf2020-12-01 15:07:08 -08006091 name = "f32_velu_test",
6092 srcs = [
6093 "test/f32-velu.cc",
6094 "test/vunary-microkernel-tester.h",
6095 ] + MICROKERNEL_TEST_HDRS,
6096 deps = MICROKERNEL_TEST_DEPS,
6097)
6098
6099xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08006100 name = "f32_vmax_test",
6101 srcs = [
6102 "test/f32-vmax.cc",
6103 "test/vbinary-microkernel-tester.h",
6104 ] + MICROKERNEL_TEST_HDRS,
6105 deps = MICROKERNEL_TEST_DEPS,
6106)
6107
6108xnnpack_unit_test(
6109 name = "f32_vmaxc_test",
6110 srcs = [
6111 "test/f32-vmaxc.cc",
6112 "test/vbinaryc-microkernel-tester.h",
6113 ] + MICROKERNEL_TEST_HDRS,
6114 deps = MICROKERNEL_TEST_DEPS,
6115)
6116
6117xnnpack_unit_test(
6118 name = "f32_vmin_test",
6119 srcs = [
6120 "test/f32-vmin.cc",
6121 "test/vbinary-microkernel-tester.h",
6122 ] + MICROKERNEL_TEST_HDRS,
6123 deps = MICROKERNEL_TEST_DEPS,
6124)
6125
6126xnnpack_unit_test(
6127 name = "f32_vminc_test",
6128 srcs = [
6129 "test/f32-vminc.cc",
6130 "test/vbinaryc-microkernel-tester.h",
6131 ] + MICROKERNEL_TEST_HDRS,
6132 deps = MICROKERNEL_TEST_DEPS,
6133)
6134
6135xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006136 name = "f32_vmul_test",
6137 srcs = [
6138 "test/f32-vmul.cc",
6139 "test/vbinary-microkernel-tester.h",
6140 ] + MICROKERNEL_TEST_HDRS,
6141 deps = MICROKERNEL_TEST_DEPS,
6142)
6143
6144xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006145 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006146 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006147 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006148 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006149 ] + MICROKERNEL_TEST_HDRS,
6150 deps = MICROKERNEL_TEST_DEPS,
6151)
6152
6153xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006154 name = "f32_vmul_relu_test",
6155 srcs = [
6156 "test/f32-vmul-relu.cc",
6157 "test/vbinary-microkernel-tester.h",
6158 ] + MICROKERNEL_TEST_HDRS,
6159 deps = MICROKERNEL_TEST_DEPS,
6160)
6161
6162xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006163 name = "f32_vmulc_test",
6164 srcs = [
6165 "test/f32-vmulc.cc",
6166 "test/vbinaryc-microkernel-tester.h",
6167 ] + MICROKERNEL_TEST_HDRS,
6168 deps = MICROKERNEL_TEST_DEPS,
6169)
6170
6171xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006172 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006173 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006174 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006175 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006176 ] + MICROKERNEL_TEST_HDRS,
6177 deps = MICROKERNEL_TEST_DEPS,
6178)
6179
6180xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006181 name = "f32_vmulc_relu_test",
6182 srcs = [
6183 "test/f32-vmulc-relu.cc",
6184 "test/vbinaryc-microkernel-tester.h",
6185 ] + MICROKERNEL_TEST_HDRS,
6186 deps = MICROKERNEL_TEST_DEPS,
6187)
6188
6189xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07006190 name = "f16_vmulcaddc_minmax_test",
6191 srcs = [
6192 "test/f16-vmulcaddc-minmax.cc",
6193 "test/vmulcaddc-microkernel-tester.h",
6194 "src/xnnpack/AlignedAllocator.h",
6195 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006196 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07006197)
6198
6199xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006200 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006201 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006202 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006203 "test/vmulcaddc-microkernel-tester.h",
6204 "src/xnnpack/AlignedAllocator.h",
6205 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006206 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207)
6208
6209xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07006210 name = "f32_vlrelu_test",
6211 srcs = [
6212 "test/f32-vlrelu.cc",
6213 "test/vunary-microkernel-tester.h",
6214 ] + MICROKERNEL_TEST_HDRS,
6215 deps = MICROKERNEL_TEST_DEPS,
6216)
6217
6218xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006219 name = "f32_vneg_test",
6220 srcs = [
6221 "test/f32-vneg.cc",
6222 "test/vunary-microkernel-tester.h",
6223 ] + MICROKERNEL_TEST_HDRS,
6224 deps = MICROKERNEL_TEST_DEPS,
6225)
6226
6227xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07006228 name = "f32_vrndne_test",
6229 srcs = [
6230 "test/f32-vrndne.cc",
6231 "test/vunary-microkernel-tester.h",
6232 ] + MICROKERNEL_TEST_HDRS,
6233 deps = MICROKERNEL_TEST_DEPS,
6234)
6235
6236xnnpack_unit_test(
6237 name = "f32_vrndz_test",
6238 srcs = [
6239 "test/f32-vrndz.cc",
6240 "test/vunary-microkernel-tester.h",
6241 ] + MICROKERNEL_TEST_HDRS,
6242 deps = MICROKERNEL_TEST_DEPS,
6243)
6244
6245xnnpack_unit_test(
6246 name = "f32_vrndu_test",
6247 srcs = [
6248 "test/f32-vrndu.cc",
6249 "test/vunary-microkernel-tester.h",
6250 ] + MICROKERNEL_TEST_HDRS,
6251 deps = MICROKERNEL_TEST_DEPS,
6252)
6253
6254xnnpack_unit_test(
6255 name = "f32_vrndd_test",
6256 srcs = [
6257 "test/f32-vrndd.cc",
6258 "test/vunary-microkernel-tester.h",
6259 ] + MICROKERNEL_TEST_HDRS,
6260 deps = MICROKERNEL_TEST_DEPS,
6261)
6262
6263xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07006264 name = "f32_vscale_test",
6265 srcs = [
6266 "test/f32-vscale.cc",
6267 "test/vscale-microkernel-tester.h",
6268 ] + MICROKERNEL_TEST_HDRS,
6269 deps = MICROKERNEL_TEST_DEPS,
6270)
6271
6272xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07006273 name = "f32_vscaleexpminusmax_test",
6274 srcs = [
6275 "test/f32-vscaleexpminusmax.cc",
6276 "test/vscaleexpminusmax-microkernel-tester.h",
6277 ] + MICROKERNEL_TEST_HDRS,
6278 deps = MICROKERNEL_TEST_DEPS,
6279)
6280
6281xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07006282 name = "f32_vscaleextexp_test",
6283 srcs = [
6284 "test/f32-vscaleextexp.cc",
6285 "test/vscaleextexp-microkernel-tester.h",
6286 ] + MICROKERNEL_TEST_HDRS,
6287 deps = MICROKERNEL_TEST_DEPS,
6288)
6289
6290xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006291 name = "f32_vsqr_test",
6292 srcs = [
6293 "test/f32-vsqr.cc",
6294 "test/vunary-microkernel-tester.h",
6295 ] + MICROKERNEL_TEST_HDRS,
6296 deps = MICROKERNEL_TEST_DEPS,
6297)
6298
6299xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07006300 name = "f32_vsqrdiff_test",
6301 srcs = [
6302 "test/f32-vsqrdiff.cc",
6303 "test/vbinary-microkernel-tester.h",
6304 ] + MICROKERNEL_TEST_HDRS,
6305 deps = MICROKERNEL_TEST_DEPS,
6306)
6307
6308xnnpack_unit_test(
6309 name = "f32_vsqrdiffc_test",
6310 srcs = [
6311 "test/f32-vsqrdiffc.cc",
6312 "test/vbinaryc-microkernel-tester.h",
6313 ] + MICROKERNEL_TEST_HDRS,
6314 deps = MICROKERNEL_TEST_DEPS,
6315)
6316
6317xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07006318 name = "f32_vsqrt_test",
6319 srcs = [
6320 "test/f32-vsqrt.cc",
6321 "test/vunary-microkernel-tester.h",
6322 ] + MICROKERNEL_TEST_HDRS,
6323 deps = MICROKERNEL_TEST_DEPS,
6324)
6325
6326xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006327 name = "f32_vsub_test",
6328 srcs = [
6329 "test/f32-vsub.cc",
6330 "test/vbinary-microkernel-tester.h",
6331 ] + MICROKERNEL_TEST_HDRS,
6332 deps = MICROKERNEL_TEST_DEPS,
6333)
6334
6335xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006336 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07006337 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006338 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006339 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006340 ] + MICROKERNEL_TEST_HDRS,
6341 deps = MICROKERNEL_TEST_DEPS,
6342)
6343
6344xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006345 name = "f32_vsub_relu_test",
6346 srcs = [
6347 "test/f32-vsub-relu.cc",
6348 "test/vbinary-microkernel-tester.h",
6349 ] + MICROKERNEL_TEST_HDRS,
6350 deps = MICROKERNEL_TEST_DEPS,
6351)
6352
6353xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006354 name = "f32_vsubc_test",
6355 srcs = [
6356 "test/f32-vsubc.cc",
6357 "test/vbinaryc-microkernel-tester.h",
6358 ] + MICROKERNEL_TEST_HDRS,
6359 deps = MICROKERNEL_TEST_DEPS,
6360)
6361
6362xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006363 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006364 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006365 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006366 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006367 ] + MICROKERNEL_TEST_HDRS,
6368 deps = MICROKERNEL_TEST_DEPS,
6369)
6370
6371xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006372 name = "f32_vsubc_relu_test",
6373 srcs = [
6374 "test/f32-vsubc-relu.cc",
6375 "test/vbinaryc-microkernel-tester.h",
6376 ] + MICROKERNEL_TEST_HDRS,
6377 deps = MICROKERNEL_TEST_DEPS,
6378)
6379
6380xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07006381 name = "f32_vrsubc_test",
6382 srcs = [
6383 "test/f32-vrsubc.cc",
6384 "test/vbinaryc-microkernel-tester.h",
6385 ] + MICROKERNEL_TEST_HDRS,
6386 deps = MICROKERNEL_TEST_DEPS,
6387)
6388
6389xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006390 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08006391 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07006392 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08006393 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07006394 ] + MICROKERNEL_TEST_HDRS,
6395 deps = MICROKERNEL_TEST_DEPS,
6396)
6397
6398xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07006399 name = "f32_vrsubc_relu_test",
6400 srcs = [
6401 "test/f32-vrsubc-relu.cc",
6402 "test/vbinaryc-microkernel-tester.h",
6403 ] + MICROKERNEL_TEST_HDRS,
6404 deps = MICROKERNEL_TEST_DEPS,
6405)
6406
6407xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006408 name = "qs8_dwconv_minmax_test",
6409 srcs = [
6410 "test/qs8-dwconv-minmax.cc",
6411 "test/dwconv-microkernel-tester.h",
6412 "src/xnnpack/AlignedAllocator.h",
6413 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6414 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6415)
6416
6417xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07006418 name = "qs8_gavgpool_minmax_test",
6419 srcs = [
6420 "test/qs8-gavgpool-minmax.cc",
6421 "test/gavgpool-microkernel-tester.h",
6422 "src/xnnpack/AlignedAllocator.h",
6423 ] + MICROKERNEL_TEST_HDRS,
6424 deps = MICROKERNEL_TEST_DEPS,
6425)
6426
6427xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07006428 name = "qs8_gemm_minmax_test",
6429 srcs = [
6430 "test/qs8-gemm-minmax.cc",
6431 "test/gemm-microkernel-tester.h",
6432 "src/xnnpack/AlignedAllocator.h",
6433 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6434 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6435)
6436
6437xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07006438 name = "qs8_igemm_minmax_test",
6439 srcs = [
6440 "test/qs8-igemm-minmax.cc",
6441 "test/gemm-microkernel-tester.h",
6442 "src/xnnpack/AlignedAllocator.h",
6443 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
6444 deps = MICROKERNEL_TEST_DEPS + [":packing"],
6445)
6446
6447xnnpack_unit_test(
6448 name = "qs8_requantization_test",
6449 srcs = [
6450 "src/xnnpack/requantization-stubs.h",
6451 "test/qs8-requantization.cc",
6452 "test/requantization-tester.h",
6453 ] + MICROKERNEL_TEST_HDRS,
6454 deps = MICROKERNEL_TEST_DEPS,
6455)
6456
6457xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07006458 name = "qs8_vadd_minmax_test",
6459 srcs = [
6460 "test/qs8-vadd-minmax.cc",
6461 "test/vadd-microkernel-tester.h",
6462 ] + MICROKERNEL_TEST_HDRS,
6463 deps = MICROKERNEL_TEST_DEPS,
6464)
6465
6466xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07006467 name = "qs8_vaddc_minmax_test",
6468 srcs = [
6469 "test/qs8-vaddc-minmax.cc",
6470 "test/vaddc-microkernel-tester.h",
6471 ] + MICROKERNEL_TEST_HDRS,
6472 deps = MICROKERNEL_TEST_DEPS,
6473)
6474
6475xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006476 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006477 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006478 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006479 "test/avgpool-microkernel-tester.h",
6480 "src/xnnpack/AlignedAllocator.h",
6481 ] + MICROKERNEL_TEST_HDRS,
6482 deps = MICROKERNEL_TEST_DEPS,
6483)
6484
6485xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006486 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006488 "test/qu8-dwconv-minmax.cc",
6489 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006490 "src/xnnpack/AlignedAllocator.h",
6491 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006492 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006493)
6494
6495xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006496 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006497 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07006498 "test/qu8-igemm-minmax.cc",
6499 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006500 "src/xnnpack/AlignedAllocator.h",
6501 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006502 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006503)
6504
6505xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006506 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006507 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006508 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006509 "test/gavgpool-microkernel-tester.h",
6510 "src/xnnpack/AlignedAllocator.h",
6511 ] + MICROKERNEL_TEST_HDRS,
6512 deps = MICROKERNEL_TEST_DEPS,
6513)
6514
6515xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006516 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006517 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006518 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006519 "test/gemm-microkernel-tester.h",
6520 "src/xnnpack/AlignedAllocator.h",
6521 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07006522 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006523)
6524
6525xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07006526 name = "qu8_requantization_test",
6527 srcs = [
6528 "src/xnnpack/requantization-stubs.h",
6529 "test/qu8-requantization.cc",
6530 "test/requantization-tester.h",
6531 ] + MICROKERNEL_TEST_HDRS,
6532 deps = MICROKERNEL_TEST_DEPS,
6533)
6534
6535xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07006536 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006537 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07006538 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006539 "test/vadd-microkernel-tester.h",
6540 ] + MICROKERNEL_TEST_HDRS,
6541 deps = MICROKERNEL_TEST_DEPS,
6542)
6543
6544xnnpack_unit_test(
6545 name = "u8_clamp_test",
6546 srcs = [
6547 "test/u8-clamp.cc",
6548 "test/clamp-microkernel-tester.h",
6549 ] + MICROKERNEL_TEST_HDRS,
6550 deps = MICROKERNEL_TEST_DEPS,
6551)
6552
6553xnnpack_unit_test(
6554 name = "u8_lut32norm_test",
6555 srcs = [
6556 "test/u8-lut32norm.cc",
6557 "test/lut-norm-microkernel-tester.h",
6558 ] + MICROKERNEL_TEST_HDRS,
6559 deps = MICROKERNEL_TEST_DEPS,
6560)
6561
6562xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07006563 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006564 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07006565 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006566 "test/maxpool-microkernel-tester.h",
6567 ] + MICROKERNEL_TEST_HDRS,
6568 deps = MICROKERNEL_TEST_DEPS,
6569)
6570
6571xnnpack_unit_test(
6572 name = "u8_rmax_test",
6573 srcs = [
6574 "test/u8-rmax.cc",
6575 "test/rmax-microkernel-tester.h",
6576 ] + MICROKERNEL_TEST_HDRS,
6577 deps = MICROKERNEL_TEST_DEPS,
6578)
6579
6580xnnpack_unit_test(
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006581 name = "x32_depthtospace2d_chw2hwc_test",
Yury Kartynnike7841862020-11-04 18:22:18 -08006582 srcs = [
Marat Dukhanad71b9a2020-11-20 00:01:51 -08006583 "test/x32-depthtospace2d-chw2hwc.cc",
6584 "test/depthtospace-microkernel-tester.h",
Yury Kartynnike7841862020-11-04 18:22:18 -08006585 ] + MICROKERNEL_TEST_HDRS,
6586 deps = MICROKERNEL_TEST_DEPS,
6587)
6588
6589xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07006590 name = "x32_fill_test",
6591 srcs = [
6592 "test/x32-fill.cc",
6593 "test/fill-microkernel-tester.h",
6594 ] + MICROKERNEL_TEST_HDRS,
6595 deps = MICROKERNEL_TEST_DEPS,
6596)
6597
6598xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006599 name = "x32_packx_test",
6600 srcs = [
6601 "test/x32-packx.cc",
6602 "test/pack-microkernel-tester.h",
6603 "src/xnnpack/AlignedAllocator.h",
6604 ] + MICROKERNEL_TEST_HDRS,
6605 deps = MICROKERNEL_TEST_DEPS,
6606)
6607
6608xnnpack_unit_test(
6609 name = "x32_pad_test",
6610 srcs = [
6611 "test/x32-pad.cc",
6612 "test/pad-microkernel-tester.h",
6613 ] + MICROKERNEL_TEST_HDRS,
6614 deps = MICROKERNEL_TEST_DEPS,
6615)
6616
6617xnnpack_unit_test(
6618 name = "x32_unpool_test",
6619 srcs = [
6620 "test/x32-unpool.cc",
6621 "test/unpool-microkernel-tester.h",
6622 ] + MICROKERNEL_TEST_HDRS,
6623 deps = MICROKERNEL_TEST_DEPS,
6624)
6625
6626xnnpack_unit_test(
6627 name = "x32_zip_test",
6628 srcs = [
6629 "test/x32-zip.cc",
6630 "test/zip-microkernel-tester.h",
6631 ] + MICROKERNEL_TEST_HDRS,
6632 deps = MICROKERNEL_TEST_DEPS,
6633)
6634
6635xnnpack_unit_test(
6636 name = "x8_lut_test",
6637 srcs = [
6638 "test/x8-lut.cc",
6639 "test/lut-microkernel-tester.h",
6640 ] + MICROKERNEL_TEST_HDRS,
6641 deps = MICROKERNEL_TEST_DEPS,
6642)
6643
6644xnnpack_unit_test(
6645 name = "x8_zip_test",
6646 srcs = [
6647 "test/x8-zip.cc",
6648 "test/zip-microkernel-tester.h",
6649 ] + MICROKERNEL_TEST_HDRS,
6650 deps = MICROKERNEL_TEST_DEPS,
6651)
6652
Marat Dukhan20c3b922020-03-10 03:45:06 -07006653########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006654
6655xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006656 name = "operator_size_test",
6657 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006658 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006659)
6660
Marat Dukhan20c3b922020-03-10 03:45:06 -07006661xnnpack_binary(
6662 name = "subgraph_size_test",
6663 srcs = ["test/subgraph-size.c"],
6664 deps = [":XNNPACK"],
6665)
6666
6667########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006668
6669xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006670 name = "abs_nc_test",
6671 srcs = [
6672 "test/abs-nc.cc",
6673 "test/abs-operator-tester.h",
6674 ],
6675 deps = OPERATOR_TEST_DEPS,
6676)
6677
6678xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006679 name = "add_nd_test",
6680 srcs = [
6681 "test/add-nd.cc",
6682 "test/binary-elementwise-operator-tester.h",
6683 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006684 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006685)
6686
6687xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006688 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006689 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006690 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006691 "test/argmax-pooling-operator-tester.h",
6692 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006693 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006694)
6695
6696xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006697 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006698 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006699 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006700 "test/average-pooling-operator-tester.h",
6701 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006702 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006703)
6704
6705xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006706 name = "bankers_rounding_nc_test",
6707 srcs = [
6708 "test/bankers-rounding-nc.cc",
6709 "test/bankers-rounding-operator-tester.h",
6710 ],
6711 deps = OPERATOR_TEST_DEPS,
6712)
6713
6714xnnpack_unit_test(
6715 name = "ceiling_nc_test",
6716 srcs = [
6717 "test/ceiling-nc.cc",
6718 "test/ceiling-operator-tester.h",
6719 ],
6720 deps = OPERATOR_TEST_DEPS,
6721)
6722
6723xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006724 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006725 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006726 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006727 "test/channel-shuffle-operator-tester.h",
6728 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006729 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006730)
6731
6732xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006733 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006734 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006735 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006736 "test/clamp-operator-tester.h",
6737 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006738 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006739)
6740
6741xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006742 name = "constant_pad_nd_test",
6743 srcs = [
6744 "test/constant-pad-nd.cc",
6745 "test/constant-pad-operator-tester.h",
6746 ],
6747 deps = OPERATOR_TEST_DEPS,
6748)
6749
6750xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006751 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006752 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006753 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006754 "test/convolution-operator-tester.h",
6755 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006756 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006757)
6758
6759xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006760 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006761 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006762 "test/convolution-nchw.cc",
6763 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006764 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006765 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006766)
6767
6768xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006769 name = "copy_nc_test",
6770 srcs = [
6771 "test/copy-nc.cc",
6772 "test/copy-operator-tester.h",
6773 ],
6774 deps = OPERATOR_TEST_DEPS,
6775)
6776
6777xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006778 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006779 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006780 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006781 "test/deconvolution-operator-tester.h",
6782 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006783 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006784)
6785
6786xnnpack_unit_test(
Marat Dukhan188d1042020-11-24 23:39:40 -08006787 name = "depth_to_space_nchw2nhwc_test",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006788 srcs = [
Marat Dukhan188d1042020-11-24 23:39:40 -08006789 "test/depth-to-space-nchw2nhwc.cc",
Artsiom Ablavatski0f1dc182020-11-05 19:21:50 -08006790 "test/depth-to-space-operator-tester.h",
6791 ] + OPERATOR_TEST_PARAMS_HDRS,
6792 deps = OPERATOR_TEST_DEPS,
6793)
6794
6795xnnpack_unit_test(
Marat Dukhan0e521172020-11-25 13:10:04 -08006796 name = "depth_to_space_nhwc_test",
6797 srcs = [
6798 "test/depth-to-space-nhwc.cc",
6799 "test/depth-to-space-operator-tester.h",
6800 ] + OPERATOR_TEST_PARAMS_HDRS,
6801 deps = OPERATOR_TEST_DEPS,
6802)
6803
6804xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006805 name = "divide_nd_test",
6806 srcs = [
6807 "test/binary-elementwise-operator-tester.h",
6808 "test/divide-nd.cc",
6809 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006810 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006811)
6812
6813xnnpack_unit_test(
Marat Dukhanb6bd4bc2020-12-01 17:01:40 -08006814 name = "elu_nc_test",
6815 srcs = [
6816 "test/elu-nc.cc",
6817 "test/elu-operator-tester.h",
6818 ],
6819 deps = OPERATOR_TEST_DEPS,
6820)
6821
6822xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006823 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006824 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006825 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006826 "test/fully-connected-operator-tester.h",
6827 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006828 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006829)
6830
6831xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006832 name = "floor_nc_test",
6833 srcs = [
6834 "test/floor-nc.cc",
6835 "test/floor-operator-tester.h",
6836 ],
6837 deps = OPERATOR_TEST_DEPS,
6838)
6839
6840xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006841 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006842 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006843 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006844 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006845 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006846 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006847)
6848
6849xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006850 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006851 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006852 "test/global-average-pooling-ncw.cc",
6853 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006854 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006855 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006856)
6857
6858xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006859 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006860 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006861 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006862 "test/hardswish-operator-tester.h",
6863 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006864 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006865)
6866
6867xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006868 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006869 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006870 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006871 "test/leaky-relu-operator-tester.h",
6872 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006873 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006874)
6875
6876xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006877 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006878 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006879 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006880 "test/max-pooling-operator-tester.h",
6881 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006882 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006883)
6884
6885xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006886 name = "maximum_nd_test",
6887 srcs = [
6888 "test/binary-elementwise-operator-tester.h",
6889 "test/maximum-nd.cc",
6890 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006891 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006892)
6893
6894xnnpack_unit_test(
6895 name = "minimum_nd_test",
6896 srcs = [
6897 "test/binary-elementwise-operator-tester.h",
6898 "test/minimum-nd.cc",
6899 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006900 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006901)
6902
6903xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006904 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006905 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006906 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006907 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006908 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006909 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006910)
6911
6912xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006913 name = "negate_nc_test",
6914 srcs = [
6915 "test/negate-nc.cc",
6916 "test/negate-operator-tester.h",
6917 ],
6918 deps = OPERATOR_TEST_DEPS,
6919)
6920
6921xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006922 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006923 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006924 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006925 "test/prelu-operator-tester.h",
6926 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006927 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006928)
6929
6930xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006931 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006932 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006933 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006934 "test/resize-bilinear-operator-tester.h",
6935 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006936 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006937)
6938
6939xnnpack_unit_test(
Artsiom Ablavatski97918102020-10-27 15:52:59 -07006940 name = "resize_bilinear_nchw_test",
6941 srcs = [
6942 "test/resize-bilinear-nchw.cc",
6943 "test/resize-bilinear-operator-tester.h",
6944 ] + OPERATOR_TEST_PARAMS_HDRS,
6945 deps = OPERATOR_TEST_DEPS,
6946)
6947
6948xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006949 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006950 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006951 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006952 "test/sigmoid-operator-tester.h",
6953 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006954 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006955)
6956
6957xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006958 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006959 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006960 "test/softmax-nc.cc",
6961 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006962 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006963 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006964)
6965
6966xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006967 name = "square_nc_test",
6968 srcs = [
6969 "test/square-nc.cc",
6970 "test/square-operator-tester.h",
6971 ],
6972 deps = OPERATOR_TEST_DEPS,
6973)
6974
6975xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006976 name = "square_root_nc_test",
6977 srcs = [
6978 "test/square-root-nc.cc",
6979 "test/square-root-operator-tester.h",
6980 ],
6981 deps = OPERATOR_TEST_DEPS,
6982)
6983
6984xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006985 name = "squared_difference_nd_test",
6986 srcs = [
6987 "test/binary-elementwise-operator-tester.h",
6988 "test/squared-difference-nd.cc",
6989 ],
6990 deps = OPERATOR_TEST_DEPS,
6991)
6992
6993xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006994 name = "subtract_nd_test",
6995 srcs = [
6996 "test/binary-elementwise-operator-tester.h",
6997 "test/subtract-nd.cc",
6998 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006999 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08007000)
7001
7002xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07007003 name = "truncation_nc_test",
7004 srcs = [
7005 "test/truncation-nc.cc",
7006 "test/truncation-operator-tester.h",
7007 ],
7008 deps = OPERATOR_TEST_DEPS,
7009)
7010
7011xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08007012 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007013 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08007014 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007015 "test/unpooling-operator-tester.h",
7016 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07007017 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07007018)
7019
Chao Mei6ddfc602020-05-13 22:29:36 -07007020############################### Misc unit tests ###############################
7021
7022xnnpack_unit_test(
7023 name = "memory_planner_test",
7024 srcs = [
7025 "test/memory-planner-test.cc",
7026 ],
7027 deps = [
7028 ":XNNPACK",
7029 ":memory_planner",
7030 ],
7031)
7032
XNNPACK Teamab8c4c82020-10-09 08:05:51 -07007033xnnpack_unit_test(
7034 name = "subgraph_nchw_test",
7035 srcs = [
7036 "src/xnnpack/subgraph.h",
7037 "test/subgraph-nchw.cc",
7038 "test/subgraph-tester.h",
7039 ],
7040 deps = [
7041 ":XNNPACK",
7042 ],
7043)
7044
Marat Dukhan08c4a432019-10-03 09:29:21 -07007045############################# Build configurations #############################
7046
Marat Dukhanb8642352019-10-30 15:43:02 -07007047# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07007048config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007049 name = "xnn_enable_assembly_explicit_true",
7050 define_values = {"xnn_enable_assembly": "true"},
7051)
7052
7053# Disables usage of assembly kernels.
7054config_setting(
7055 name = "xnn_enable_assembly_explicit_false",
7056 define_values = {"xnn_enable_assembly": "false"},
7057)
7058
Marat Dukhan9de90e02020-06-18 16:04:12 -07007059# Enables usage of sparse inference.
7060config_setting(
7061 name = "xnn_enable_sparse_explicit_true",
7062 define_values = {"xnn_enable_sparse": "true"},
7063)
7064
7065# Disables usage of sparse inference.
7066config_setting(
7067 name = "xnn_enable_sparse_explicit_false",
7068 define_values = {"xnn_enable_sparse": "false"},
7069)
7070
Marat Dukhan05702cf2020-03-26 15:41:33 -07007071# Disables usage of HMP-aware optimizations.
7072config_setting(
7073 name = "xnn_enable_hmp_explicit_false",
7074 define_values = {"xnn_enable_hmp": "false"},
7075)
7076
Chao Mei6ddfc602020-05-13 22:29:36 -07007077# Enable usage of optimized memory allocation
7078config_setting(
7079 name = "xnn_enable_memopt_explicit_true",
7080 define_values = {"xnnpack_enable_memopt": "true"},
7081)
7082
7083# Disable usage of optimized memory allocation
7084config_setting(
7085 name = "xnn_enable_memopt_explicit_false",
7086 define_values = {"xnnpack_enable_memopt": "false"},
7087)
7088
Marat Dukhanb8642352019-10-30 15:43:02 -07007089# Builds with -c dbg
7090config_setting(
7091 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007092 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07007093 "compilation_mode": "dbg",
7094 },
7095)
7096
7097# Builds with -c opt
7098config_setting(
7099 name = "optimized_build",
7100 values = {
7101 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007102 },
7103)
7104
7105config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07007106 name = "linux_k8",
7107 values = {"cpu": "k8"},
7108)
7109
7110config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007111 name = "linux_arm",
7112 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07007113)
7114
7115config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07007116 name = "linux_armeabi",
7117 values = {"cpu": "armeabi"},
7118)
7119
7120config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07007121 name = "linux_armhf",
7122 values = {"cpu": "armhf"},
7123)
7124
7125config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07007126 name = "linux_armv7a",
7127 values = {"cpu": "armv7a"},
7128)
7129
7130config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07007131 name = "linux_aarch64",
7132 values = {"cpu": "aarch64"},
7133)
7134
7135config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007136 name = "android",
7137 values = {"crosstool_top": "//external:android/crosstool"},
7138)
7139
7140config_setting(
7141 name = "android_armv7",
7142 values = {
7143 "crosstool_top": "//external:android/crosstool",
7144 "cpu": "armeabi-v7a",
7145 },
7146)
7147
7148config_setting(
7149 name = "android_arm64",
7150 values = {
7151 "crosstool_top": "//external:android/crosstool",
7152 "cpu": "arm64-v8a",
7153 },
7154)
7155
7156config_setting(
7157 name = "android_x86",
7158 values = {
7159 "crosstool_top": "//external:android/crosstool",
7160 "cpu": "x86",
7161 },
7162)
7163
7164config_setting(
7165 name = "android_x86_64",
7166 values = {
7167 "crosstool_top": "//external:android/crosstool",
7168 "cpu": "x86_64",
7169 },
7170)
7171
7172config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007173 name = "windows_x86_64",
7174 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007175)
7176
7177config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07007178 name = "windows_x86_64_clang",
7179 values = {
7180 "compiler": "clang-cl",
7181 "cpu": "x64_windows",
7182 },
7183)
7184
7185config_setting(
7186 name = "windows_x86_64_mingw",
7187 values = {
7188 "compiler": "mingw-gcc",
7189 "cpu": "x64_windows",
7190 },
7191)
7192
7193config_setting(
7194 name = "windows_x86_64_msys",
7195 values = {
7196 "compiler": "msys-gcc",
7197 "cpu": "x64_windows",
7198 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07007199)
7200
7201config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07007202 name = "macos_x86_64",
7203 values = {
7204 "apple_platform_type": "macos",
7205 "cpu": "darwin",
7206 },
7207)
7208
7209config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07007210 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007211 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07007212)
7213
7214config_setting(
7215 name = "emscripten_wasm",
7216 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007217 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007218 "cpu": "wasm",
7219 },
7220)
7221
7222config_setting(
7223 name = "emscripten_wasmsimd",
7224 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07007225 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007226 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07007227 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07007228 },
7229)
7230
7231config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007232 name = "ios_armv7",
7233 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007234 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007235 "cpu": "ios_armv7",
7236 },
7237)
7238
7239config_setting(
7240 name = "ios_arm64",
7241 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007242 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007243 "cpu": "ios_arm64",
7244 },
7245)
7246
7247config_setting(
7248 name = "ios_arm64e",
7249 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007250 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007251 "cpu": "ios_arm64e",
7252 },
7253)
7254
7255config_setting(
7256 name = "ios_x86",
7257 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007258 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007259 "cpu": "ios_i386",
7260 },
7261)
7262
7263config_setting(
7264 name = "ios_x86_64",
7265 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007266 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007267 "cpu": "ios_x86_64",
7268 },
7269)
7270
7271config_setting(
7272 name = "watchos_armv7k",
7273 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007274 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007275 "cpu": "watchos_armv7k",
7276 },
7277)
7278
7279config_setting(
7280 name = "watchos_arm64_32",
7281 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007282 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007283 "cpu": "watchos_arm64_32",
7284 },
7285)
7286
7287config_setting(
7288 name = "watchos_x86",
7289 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007290 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007291 "cpu": "watchos_i386",
7292 },
7293)
7294
7295config_setting(
7296 name = "watchos_x86_64",
7297 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007298 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007299 "cpu": "watchos_x86_64",
7300 },
7301)
7302
7303config_setting(
7304 name = "tvos_arm64",
7305 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007306 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007307 "cpu": "tvos_arm64",
7308 },
7309)
7310
7311config_setting(
7312 name = "tvos_x86_64",
7313 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08007314 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08007315 "cpu": "tvos_x86_64",
7316 },
7317)