blob: 2fc11cee1ac15b076ed9d1b2a33db6d53eeecb7a [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
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",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800425 "src/math/sigmoid-scalar-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001074 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001075 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001076 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001077 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001078 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001079 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001080 "src/x32-zip/x2-wasmsimd.c",
1081 "src/x32-zip/x3-wasmsimd.c",
1082 "src/x32-zip/x4-wasmsimd.c",
1083 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001084 "src/math/roundne-wasmsimd-addsub.c",
1085 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001086 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001087 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001088 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001089 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001090 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001091 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1092 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001093]
1094
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001095PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001096 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001097 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001098 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001099 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001100 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001101 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1102 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1103 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001104]
1105
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001106PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001107 "src/qs8-requantization/precise-psimd.c",
1108 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001109 "src/qu8-requantization/precise-psimd.c",
1110 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001111]
1112
Marat Dukhan08c4a432019-10-03 09:29:21 -07001113# ISA-specific micro-kernels
1114NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001115 "src/f32-argmaxpool/4x-neon-c4.c",
1116 "src/f32-argmaxpool/9p8x-neon-c4.c",
1117 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001118 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1119 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001120 "src/f32-clamp/gen/neon-x4.c",
1121 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001122 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1123 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001124 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1125 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001126 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1127 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001128 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1129 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001130 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1131 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1132 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1133 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001134 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1135 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1136 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1137 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001142 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001143 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1144 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001145 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1146 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1147 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1148 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1149 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1150 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1151 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1152 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1153 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1154 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1155 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1156 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1157 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1158 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1159 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1160 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1161 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1162 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1163 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1164 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1165 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1166 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1167 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1168 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1169 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1170 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1171 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1172 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1173 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1174 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1175 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001176 "src/f32-hswish/gen/hswish-neon-x4.c",
1177 "src/f32-hswish/gen/hswish-neon-x8.c",
1178 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001179 "src/f32-ibilinear/gen/neon-c4.c",
1180 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001181 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1182 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1183 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1184 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1185 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1186 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1187 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1188 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1189 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1190 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1191 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1192 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1193 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1194 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1195 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1196 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001197 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1198 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1199 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001200 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1201 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001202 "src/f32-prelu/gen/neon-1x4.c",
1203 "src/f32-prelu/gen/neon-1x8.c",
1204 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001205 "src/f32-prelu/gen/neon-2x4.c",
1206 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001207 "src/f32-prelu/gen/neon-2x16.c",
1208 "src/f32-prelu/gen/neon-4x4.c",
1209 "src/f32-prelu/gen/neon-4x8.c",
1210 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001211 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1212 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1213 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1214 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1215 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1216 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1217 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1218 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001235 "src/f32-relu/gen/neon-x4.c",
1236 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001237 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001238 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001239 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1240 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1241 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1242 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1243 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1244 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1245 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1246 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1247 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1248 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1249 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1250 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1251 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1252 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001257 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1258 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1259 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1260 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001261 "src/f32-vbinary/gen/vmax-neon-x4.c",
1262 "src/f32-vbinary/gen/vmax-neon-x8.c",
1263 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1264 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1265 "src/f32-vbinary/gen/vmin-neon-x4.c",
1266 "src/f32-vbinary/gen/vmin-neon-x8.c",
1267 "src/f32-vbinary/gen/vminc-neon-x4.c",
1268 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001269 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1272 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1273 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1274 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001275 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1276 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1277 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1278 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001279 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001283 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1284 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001285 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1286 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001287 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1288 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1289 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1290 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1291 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1292 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1293 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1294 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001295 "src/f32-vunary/gen/vabs-neon-x4.c",
1296 "src/f32-vunary/gen/vabs-neon-x8.c",
1297 "src/f32-vunary/gen/vneg-neon-x4.c",
1298 "src/f32-vunary/gen/vneg-neon-x8.c",
1299 "src/f32-vunary/gen/vsqr-neon-x4.c",
1300 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001301 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1302 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1303 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1304 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001305 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1306 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1307 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1308 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1309 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1310 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1311 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1312 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001313 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1314 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1315 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1316 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1317 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1318 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1319 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1320 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001321 "src/qs8-requantization/precise-neon.c",
1322 "src/qs8-requantization/fp32-neon.c",
1323 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001324 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1325 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1326 "src/qu8-dwconv/up8x9-minmax-neon.c",
1327 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1328 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1329 "src/qu8-gemm/4x8-minmax-neon.c",
1330 "src/qu8-gemm/8x8-minmax-neon.c",
1331 "src/qu8-igemm/4x8-minmax-neon.c",
1332 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001333 "src/qu8-requantization/precise-neon.c",
1334 "src/qu8-requantization/fp32-neon.c",
1335 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001336 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001337 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001338 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001339 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001340 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001341 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001342 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001343 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001344 "src/x32-zip/x2-neon.c",
1345 "src/x32-zip/x3-neon.c",
1346 "src/x32-zip/x4-neon.c",
1347 "src/x32-zip/xm-neon.c",
1348 "src/x8-zip/x2-neon.c",
1349 "src/x8-zip/x3-neon.c",
1350 "src/x8-zip/x4-neon.c",
1351 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001352 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001353 "src/math/roundd-neon-addsub.c",
1354 "src/math/roundd-neon-cvt.c",
1355 "src/math/roundu-neon-addsub.c",
1356 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001357 "src/math/roundz-neon-addsub.c",
1358 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001359 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001360 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1361 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1362 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1363 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1364 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1365 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001366 "src/math/sqrt-neon-nr1rsqrts.c",
1367 "src/math/sqrt-neon-nr2rsqrts.c",
1368 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001369]
1370
1371NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001372 "src/f32-ibilinear/gen/neonfma-c4.c",
1373 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001374 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1375 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1376 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1377 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1378 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1379 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1380 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1381 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1382 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001383 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1384 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1385 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1386 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001387 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1388 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1389 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1390 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001391 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1392 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1393 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1394 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001395 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1396 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1397 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1398 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1399 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1400 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1401 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1402 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1403 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1404 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1405 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1406 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1407 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1408 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1409 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1410 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1411 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1412 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001413 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1414 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001415 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1416 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1417 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1418 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1419 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1420 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1421 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1422 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1423 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1424 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1425 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1426 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1427 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1428 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1429 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1430 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1431 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001439 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1440 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1441 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1442 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1443 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1444 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1445 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1446 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1447 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1448 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1449 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1450 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1451 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1452 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1453 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1454 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001493 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1494 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001495 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1496 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1497 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1498 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1499 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1500 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1501 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1502 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1503 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1504 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1505 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1506 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1507 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1508 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1509 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1510 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1511 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1512 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1513 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1514 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001515 "src/math/exp-neonfma-lut64-p2.c",
1516 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001517 "src/math/expminus-neonfma-lut2048-p1.c",
1518 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001519 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001520 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1521 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1522 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1523 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1524 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1525 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1526 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1527 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1528 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1529 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1530 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1531 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1532 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1533 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1534 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1535 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1536 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1537 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001538 "src/math/sqrt-neonfma-nr1fma.c",
1539 "src/math/sqrt-neonfma-nr2fma.c",
1540 "src/math/sqrt-neonfma-nr3fma.c",
1541 "src/math/sqrt-neonfma-nr2fma1adj.c",
1542 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001543]
1544
1545AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001546 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1547 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1548 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1549 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1550 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1551 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1553 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1554 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1555 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1556 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1557 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1558 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1559 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1560 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1561 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1562 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1563 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1564 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1565 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1566 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1567 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1568 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1569 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1570 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1571 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001572 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1573 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001574 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1575 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001576 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1577 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001578 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1579 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001580 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1581 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1582 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1583 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1584 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001585 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1586 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1587 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1588 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1589 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1590 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1591 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1592 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1593 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1594 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1595 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1596 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1597 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1598 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1599 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1600 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1601 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001603 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1604 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1605 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1606 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1607 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1608 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1609 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1610 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1611 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1612 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1613 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1614 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1615 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1616 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1617 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1618 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1619 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1620 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001621 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1622 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001623 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1624 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1625 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1626 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1627 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1628 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001629]
1630
Marat Dukhan8853b822020-05-07 12:19:01 -07001631NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001632 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1633 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1634 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1635 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1636 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1637 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1638 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1639 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001640 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001641 "src/math/roundd-neonv8.c",
1642 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001643 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001644]
1645
Marat Dukhan08c4a432019-10-03 09:29:21 -07001646AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001647 "src/f16-clamp/gen/neonfp16arith-x16.c",
1648 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001649 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1650 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1651 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1652 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1653 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1654 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1655 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1656 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1657 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1658 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1659 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1660 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001661 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1662 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001663 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1664 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001665 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001666 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001667 "src/f16-relu/gen/neonfp16arith-x16.c",
1668 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001669 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1670 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1671 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1672 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1673 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1674 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1675 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1676 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1677 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1678 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1679 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1680 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1681 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1682 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1683 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1684 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1685 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001697 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1698 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001699 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1700 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1701 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1702 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1703 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1704 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1705 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1706 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001707 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1708 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1709 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1710 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001711 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1712 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1713 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1714 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1715 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1719 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1720 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1721 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1722 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001723 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1724 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1725 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1726 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1727 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1728 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1729 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1730 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731]
1732
Benoit Jacoba9644732020-08-13 12:48:55 -07001733NEONDOT_UKERNELS = [
1734 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1735 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1736 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001737 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1738 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1739 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001740]
1741
Marat Dukhan08c4a432019-10-03 09:29:21 -07001742SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001743 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1744 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001745 "src/f32-clamp/gen/sse-x4.c",
1746 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001747 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1748 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001749 "src/f32-dwconv-chw/3x3p1-sse.c",
1750 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001751 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1752 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1753 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1754 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1755 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1756 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1757 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1758 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1759 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1760 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1761 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1762 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001763 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001764 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1765 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001766 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1767 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1768 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1769 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1770 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1771 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1772 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1773 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1774 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1775 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1776 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1777 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1778 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001779 "src/f32-hswish/gen/hswish-sse-x4.c",
1780 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001781 "src/f32-ibilinear/gen/sse-c4.c",
1782 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001783 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1784 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1785 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1786 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1787 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1788 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1789 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001790 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1791 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1792 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001793 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001794 "src/f32-prelu/gen/sse-2x4.c",
1795 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001796 "src/f32-relu/gen/sse-x4.c",
1797 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001798 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001799 "src/f32-spmm/gen/4x1-minmax-sse.c",
1800 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001801 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001802 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1803 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1804 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1805 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1806 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1807 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1808 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1809 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001810 "src/f32-vbinary/gen/vmax-sse-x4.c",
1811 "src/f32-vbinary/gen/vmax-sse-x8.c",
1812 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1813 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1814 "src/f32-vbinary/gen/vmin-sse-x4.c",
1815 "src/f32-vbinary/gen/vmin-sse-x8.c",
1816 "src/f32-vbinary/gen/vminc-sse-x4.c",
1817 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001818 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1819 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1820 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1821 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1822 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001826 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1827 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1828 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1829 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001830 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1832 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1833 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001834 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1835 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001836 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1837 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001838 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1839 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001840 "src/f32-vunary/gen/vabs-sse-x4.c",
1841 "src/f32-vunary/gen/vabs-sse-x8.c",
1842 "src/f32-vunary/gen/vneg-sse-x4.c",
1843 "src/f32-vunary/gen/vneg-sse-x8.c",
1844 "src/f32-vunary/gen/vsqr-sse-x4.c",
1845 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001846 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001847 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001848 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001849 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001850 "src/math/roundd-sse-addsub.c",
1851 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001852 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001853 "src/math/sqrt-sse-hh1mac.c",
1854 "src/math/sqrt-sse-nr1mac.c",
1855 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001856]
1857
1858SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001859 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1860 "src/f32-argmaxpool/4x-sse2-c4.c",
1861 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001862 "src/f32-prelu/gen/sse2-2x4.c",
1863 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001864 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1865 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1866 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1867 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1868 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1869 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1870 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1871 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1872 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1873 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1874 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1875 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001876 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001877 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001878 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001879 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001880 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1881 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001882 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1883 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001884 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1885 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1886 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1887 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1888 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1889 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1890 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1891 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001892 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1893 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1894 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001895 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1896 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1897 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001898 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1899 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1900 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001901 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1902 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001903 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1904 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001905 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1906 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001907 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1908 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001909 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001910 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1911 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001912 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001913 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1914 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1915 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001916 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1917 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001918 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1919 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001920 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1921 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001922 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001923 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1924 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001925 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001926 "src/qs8-requantization/precise-sse2.c",
1927 "src/qs8-requantization/fp32-sse2.c",
1928 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001929 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1930 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1931 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1932 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001933 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1934 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1935 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1936 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001937 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1938 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1939 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1940 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1941 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1942 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1943 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1944 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001945 "src/qu8-requantization/precise-sse2.c",
1946 "src/qu8-requantization/fp32-sse2.c",
1947 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001948 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001949 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001950 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001951 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001952 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953 "src/x32-zip/x2-sse2.c",
1954 "src/x32-zip/x3-sse2.c",
1955 "src/x32-zip/x4-sse2.c",
1956 "src/x32-zip/xm-sse2.c",
1957 "src/x8-zip/x2-sse2.c",
1958 "src/x8-zip/x3-sse2.c",
1959 "src/x8-zip/x4-sse2.c",
1960 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001961 "src/math/exp-sse2-p5.c",
1962 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001963 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001964 "src/math/roundd-sse2-cvt.c",
1965 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001966 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001967 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001968]
1969
1970SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001971 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1972 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1973 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001974 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1975 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1976 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001977 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1978 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1979 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001980 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1981 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001982 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1983 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001984 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1985 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001986 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1987 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001988 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001989 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1990 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001991 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001992 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1993 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1994 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001995 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1996 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001997 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1998 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001999 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2000 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002001 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002002 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2003 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002004 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002005 "src/qs8-requantization/precise-ssse3.c",
2006 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002007 "src/qu8-requantization/precise-ssse3.c",
2008 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002009]
2010
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002011SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002012 "src/f32-prelu/gen/sse41-2x4.c",
2013 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002014 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002015 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002016 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002017 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002018 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2019 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002020 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2021 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002022 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2023 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2024 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2025 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2026 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2027 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2028 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2029 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002030 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2031 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2032 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002033 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2034 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2035 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002036 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2037 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2038 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002039 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2040 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002041 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2042 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002043 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2044 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002045 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2046 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002047 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002048 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2049 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002050 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002051 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2052 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2053 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002054 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2055 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002056 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2057 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002058 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2059 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002060 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002061 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2062 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002063 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002064 "src/qs8-requantization/fp32-sse4.c",
2065 "src/qs8-requantization/precise-sse4.c",
2066 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002067 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2068 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2069 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2070 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002071 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2072 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2073 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2074 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002075 "src/qu8-requantization/precise-sse4.c",
2076 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002077 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002078 "src/math/roundd-sse41.c",
2079 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002080 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002081]
2082
Marat Dukhan08c4a432019-10-03 09:29:21 -07002083AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002084 "src/f32-clamp/gen/avx-x8.c",
2085 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002086 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2087 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2088 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2089 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2090 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2091 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2092 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2093 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2094 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2095 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2096 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2097 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2098 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2099 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2100 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2101 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2102 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2103 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2104 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2105 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2106 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2107 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2108 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2109 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2110 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2111 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2112 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2113 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2114 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2115 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002116 "src/f32-hswish/gen/hswish-avx-x8.c",
2117 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002118 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2119 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2120 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2121 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2122 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2123 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2124 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2125 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2126 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002127 "src/f32-prelu/gen/avx-2x8.c",
2128 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002129 "src/f32-relu/gen/avx-x16.c",
2130 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002131 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002132 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2133 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2134 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2135 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2136 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2137 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2138 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2139 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002140 "src/f32-vbinary/gen/vmax-avx-x8.c",
2141 "src/f32-vbinary/gen/vmax-avx-x16.c",
2142 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2143 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2144 "src/f32-vbinary/gen/vmin-avx-x8.c",
2145 "src/f32-vbinary/gen/vmin-avx-x16.c",
2146 "src/f32-vbinary/gen/vminc-avx-x8.c",
2147 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002148 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2149 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2150 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2151 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2152 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2153 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2154 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2155 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002156 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2157 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2158 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2159 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002160 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2161 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2162 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2163 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002164 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2165 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002166 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002167 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2168 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2169 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2170 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2171 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2172 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2173 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2174 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002175 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2176 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002177 "src/f32-vunary/gen/vabs-avx-x8.c",
2178 "src/f32-vunary/gen/vabs-avx-x16.c",
2179 "src/f32-vunary/gen/vneg-avx-x8.c",
2180 "src/f32-vunary/gen/vneg-avx-x16.c",
2181 "src/f32-vunary/gen/vsqr-avx-x8.c",
2182 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002183]
2184
Marat Dukhan1566fee2020-08-02 21:55:41 -07002185XOP_UKERNELS = [
2186 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2187 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2188 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2189 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002190 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2191 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002192 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2193 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002194 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002195 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2196 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002197 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002198 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2199 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2200 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002201 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2202 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2203 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2204 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2205 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2206 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002207 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002208 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2209 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002210 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002211]
2212
Marat Dukhanfda12b82019-11-21 12:27:59 -08002213FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002214 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2215 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2216 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2217 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2218 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2219 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2220 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2221 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2222 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2223 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2224 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2225 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2226 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2227 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2228 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2229 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2230 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2231 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2232 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2233 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2234 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2235 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2236 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2237 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2238 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2239 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2240 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2241 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2242 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2243 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2244 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2245 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2246 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2247 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2248 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2249 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2250 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2251 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2252 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2253 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002254 "src/f32-hswish/gen/hswish-fma3-x8.c",
2255 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002256 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2257 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2258 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2259 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2260 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2261 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2262 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2263 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2264 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2265 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2266 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2267 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2268 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2269 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002270 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2271 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2272 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2273 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2274 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2275 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2276 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2277 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002278 "src/math/sqrt-fma3-nr1fma.c",
2279 "src/math/sqrt-fma3-nr2fma.c",
2280 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002281]
2282
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002283AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002284 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2285 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2286 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2287 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2288 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2289 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2290 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2291 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2292 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2293 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2294 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2295 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2296 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2297 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2298 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2299 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2300 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2301 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2302 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2303 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2304 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2305 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2306 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2307 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2308 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2309 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2310 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2311 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2312 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2313 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2314 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2315 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2316 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2317 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2318 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2319 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002320 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2321 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2322 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2323 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2324 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2325 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2326 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2327 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2328 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2329 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2330 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2331 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2332 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2333 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2334 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2335 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2336 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2337 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2338 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2339 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2340 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2341 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2342 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2343 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2344 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2345 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002350 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2351 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2352 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2353 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2354 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2355 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2356 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2357 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2358 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2359 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2360 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2361 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2362 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2363 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2364 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2365 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2366 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2367 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2368 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2369 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2370 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2371 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2372 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2373 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002374 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2375 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002376 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2377 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2378 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2379 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002380 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2381 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2382 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002383 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2384 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2385 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002386 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2387 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2388 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002389 "src/math/exp-avx2-p5.c",
2390 "src/math/exp-avx2-perm-p3.c",
2391 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002392 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002393 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002394 "src/math/sigmoid-avx2-rr2-p5-div.c",
2395 "src/math/sigmoid-avx2-rr1-p5-div.c",
2396 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2397 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2398 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2399 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002400]
2401
Marat Dukhan08c4a432019-10-03 09:29:21 -07002402AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002403 "src/f32-clamp/gen/avx512f-x16.c",
2404 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002405 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2406 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2407 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2408 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2409 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2410 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2411 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2412 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2413 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2414 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2415 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2416 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2417 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2418 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2419 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2420 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2421 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2422 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2423 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2424 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2425 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2426 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2427 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2428 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002429 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2430 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002431 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2432 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2433 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2434 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2435 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2436 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002437 "src/f32-prelu/gen/avx512f-2x16.c",
2438 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002439 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2440 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2441 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2442 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2443 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2444 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2445 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2446 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2447 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2448 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2449 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2450 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2451 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2452 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2453 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2454 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2455 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2456 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2457 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2458 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2459 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2460 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2461 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2462 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2463 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2464 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2465 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2466 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2467 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2468 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2469 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2470 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2471 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2472 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2473 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2474 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002475 "src/f32-relu/gen/avx512f-x16.c",
2476 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002477 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002478 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2479 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2480 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2481 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2482 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2483 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2484 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2485 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002486 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2487 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2488 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2489 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2490 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2491 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2492 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2493 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002494 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2495 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2496 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2497 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2498 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2499 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2500 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2501 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002502 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2503 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2504 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2505 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002506 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2507 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2508 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2509 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002510 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2511 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002512 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002513 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2514 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2515 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2516 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2517 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2518 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2519 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2520 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2521 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2522 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2523 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2524 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2525 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2526 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2527 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2528 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2529 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2530 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2531 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2532 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2533 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2534 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2535 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2536 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002537 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2538 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2539 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2540 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2541 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2542 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2543 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2544 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002545 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2546 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2547 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2548 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2549 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2550 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2551 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2552 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002553 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2554 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2555 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2556 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2557 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2558 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002559 "src/math/exp-avx512f-p5-scalef.c",
2560 "src/math/exp-avx512f-p5.c",
2561 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002562 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002563 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002564 "src/math/sqrt-avx512f-nr1fma.c",
2565 "src/math/sqrt-avx512f-nr2fma.c",
2566 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002567]
2568
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002569AVX512SKX_UKERNELS = [
2570 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2571 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2572 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2573 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2574 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2575 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2576 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2577 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2578]
2579
Frank Barchardbcedc082020-08-17 18:00:51 -07002580WASM32_ASM_UKERNELS = [
2581 "src/f32-relu/wasm_shr_x1.S",
2582 "src/f32-relu/wasm_shr_x2.S",
2583 "src/f32-relu/wasm_shr_x4.S",
2584]
2585
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002587 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002588 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002589 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2590 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002591 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002592 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002593 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002594 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002595 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2596 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002597 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2598 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2599 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2600 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601]
2602
2603AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002604 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002605 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2606 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002607 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002608 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2609 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002610 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2611 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2612 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002613 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002614 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2615 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2616 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002617 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002618 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2619 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002620 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002621 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2622 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2623 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2624 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2625 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2626 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2627 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2628 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2629 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2630 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2631 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2632 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2633 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2634 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2635 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2636 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2637 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2638 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002639 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2640 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002641 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002642 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2643 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2644 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2645 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2646 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2647 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2648 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2649 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2650 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2651 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2652 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2653 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2654 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2655 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2656 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2657 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2658 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2659 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002660 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2661 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2662 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2663 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2664 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2665 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2666 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2667 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2668 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2669 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2670 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2671 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2672 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2673 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2674 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2675 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2676 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2677 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002678]
2679
Marat Dukhan1b354632020-03-23 12:50:22 -07002680INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002681 "src/xnnpack/argmaxpool.h",
2682 "src/xnnpack/avgpool.h",
2683 "src/xnnpack/clamp.h",
2684 "src/xnnpack/common.h",
2685 "src/xnnpack/conv.h",
2686 "src/xnnpack/dwconv.h",
2687 "src/xnnpack/gavgpool.h",
2688 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002689 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002690 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002691 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002692 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002693 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002694 "src/xnnpack/lut.h",
2695 "src/xnnpack/math.h",
2696 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002697 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002698 "src/xnnpack/packx.h",
2699 "src/xnnpack/pad.h",
2700 "src/xnnpack/params.h",
2701 "src/xnnpack/pavgpool.h",
2702 "src/xnnpack/ppmm.h",
2703 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002704 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002705 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002706 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002707 "src/xnnpack/rmax.h",
2708 "src/xnnpack/scalar-utils.h",
2709 "src/xnnpack/spmm.h",
2710 "src/xnnpack/unpool.h",
2711 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002712 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002713 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002714 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002715 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002716 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002717 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002718 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002719]
2720
2721INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722 "include/xnnpack.h",
2723 "src/xnnpack/allocator.h",
2724 "src/xnnpack/compute.h",
2725 "src/xnnpack/im2col.h",
2726 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002727 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002728 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729 "src/xnnpack/operator.h",
2730 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002731 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002732 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002733 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002734 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002735]
2736
Marat Dukhan1b354632020-03-23 12:50:22 -07002737ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002738 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002739]
2740
Marat Dukhan1b354632020-03-23 12:50:22 -07002741MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002742 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 "include/xnnpack.h",
2744]
2745
Marat Dukhan1b354632020-03-23 12:50:22 -07002746MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002747 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002748 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749 "src/xnnpack/requantization.h",
2750 "include/xnnpack.h",
2751]
2752
2753OPERATOR_TEST_PARAMS_HDRS = [
2754 "src/xnnpack/params.h",
2755 "src/xnnpack/common.h",
2756]
2757
2758WEIGHTS_PACK_HDRS = [
2759 "src/xnnpack/pack.h",
2760 "src/xnnpack/operator.h",
2761 "src/xnnpack/compute.h",
2762]
2763
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002764LOGGING_COPTS = select({
2765 # No logging in optimized mode
2766 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2767 # Full logging in debug mode
2768 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2769 # Error-only logging in default (fastbuild) mode
2770 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2771})
2772
Marat Dukhan3b59de22020-06-03 20:15:19 -07002773LOGGING_SRCS = select({
2774 # No logging in optimized mode
2775 ":optimized_build": [],
2776 "//conditions:default": [
2777 "src/operator-strings.c",
2778 "src/subgraph-strings.c",
2779 ],
2780})
2781
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002782LOGGING_HDRS = [
2783 "src/xnnpack/log.h",
2784]
2785
Marat Dukhan08c4a432019-10-03 09:29:21 -07002786xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002787 name = "tables",
2788 srcs = TABLE_SRCS,
2789 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002790 gcc_copts = xnnpack_gcc_std_copts(),
2791 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002792)
2793
2794xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002795 name = "scalar_ukernels",
2796 srcs = SCALAR_UKERNELS,
2797 hdrs = INTERNAL_HDRS,
2798 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002799 gcc_copts = xnnpack_gcc_std_copts(),
2800 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002801 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002802 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002803 "@FP16",
2804 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002805 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 ],
2807)
2808
2809xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002810 name = "scalar_ukernels_test_mode",
2811 srcs = SCALAR_UKERNELS,
2812 hdrs = INTERNAL_HDRS,
2813 aarch32_copts = ["-marm"],
2814 copts = [
2815 "-UNDEBUG",
2816 "-DXNN_TEST_MODE=1",
2817 ],
2818 gcc_copts = xnnpack_gcc_std_copts(),
2819 msvc_copts = xnnpack_msvc_std_copts(),
2820 deps = [
2821 ":tables",
2822 "@FP16",
2823 "@FXdiv",
2824 "@pthreadpool",
2825 ],
2826)
2827
2828xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002829 name = "wasm_ukernels",
2830 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002831 gcc_copts = xnnpack_gcc_std_copts(),
2832 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002833 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002834 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002835 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002836 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002837 "@FP16",
2838 "@FXdiv",
2839 "@pthreadpool",
2840 ],
2841)
2842
2843xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002844 name = "wasm_ukernels_test_mode",
2845 hdrs = INTERNAL_HDRS,
2846 copts = [
2847 "-UNDEBUG",
2848 "-DXNN_TEST_MODE=1",
2849 ],
2850 gcc_copts = xnnpack_gcc_std_copts(),
2851 msvc_copts = xnnpack_msvc_std_copts(),
2852 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002853 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002854 deps = [
2855 ":tables",
2856 "@FP16",
2857 "@FXdiv",
2858 "@pthreadpool",
2859 ],
2860)
2861
2862xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002863 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002864 hdrs = INTERNAL_HDRS,
2865 aarch32_copts = [
2866 "-marm",
2867 "-mfpu=neon",
2868 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002869 gcc_copts = xnnpack_gcc_std_copts(),
2870 gcc_x86_copts = ["-msse2"],
2871 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002872 optimized_copts = [
2873 "-O3",
2874 "-ffast-math",
2875 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002876 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002877 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002878 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002879 "@FP16",
2880 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002881 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 ],
2883)
2884
2885xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002886 name = "psimd_fastmath_ukernels_test_mode",
2887 hdrs = INTERNAL_HDRS,
2888 aarch32_copts = [
2889 "-marm",
2890 "-mfpu=neon",
2891 ],
2892 copts = [
2893 "-UNDEBUG",
2894 "-DXNN_TEST_MODE=1",
2895 ],
2896 gcc_copts = xnnpack_gcc_std_copts(),
2897 gcc_x86_copts = ["-msse2"],
2898 msvc_copts = xnnpack_msvc_std_copts(),
2899 optimized_copts = [
2900 "-O3",
2901 "-ffast-math",
2902 ],
2903 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2904 deps = [
2905 ":tables",
2906 "@FP16",
2907 "@psimd",
2908 "@pthreadpool",
2909 ],
2910)
2911
2912xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002913 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002914 hdrs = INTERNAL_HDRS,
2915 aarch32_copts = [
2916 "-marm",
2917 "-mfpu=neon",
2918 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002919 gcc_copts = xnnpack_gcc_std_copts(),
2920 gcc_x86_copts = ["-msse2"],
2921 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002922 optimized_copts = [
2923 "-O3",
2924 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002925 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002926 deps = [
2927 ":tables",
2928 "@FP16",
2929 "@psimd",
2930 "@pthreadpool",
2931 ],
2932)
2933
2934xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002935 name = "psimd_accmath_ukernels_test_mode",
2936 hdrs = INTERNAL_HDRS,
2937 aarch32_copts = [
2938 "-marm",
2939 "-mfpu=neon",
2940 ],
2941 copts = [
2942 "-UNDEBUG",
2943 "-DXNN_TEST_MODE=1",
2944 ],
2945 gcc_copts = xnnpack_gcc_std_copts(),
2946 gcc_x86_copts = ["-msse2"],
2947 msvc_copts = xnnpack_msvc_std_copts(),
2948 optimized_copts = [
2949 "-O3",
2950 ],
2951 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2952 deps = [
2953 ":tables",
2954 "@FP16",
2955 "@psimd",
2956 "@pthreadpool",
2957 ],
2958)
2959
2960xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002961 name = "neon_ukernels",
2962 hdrs = INTERNAL_HDRS,
2963 aarch32_copts = [
2964 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002965 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002966 "-mfpu=neon",
2967 ],
2968 aarch32_srcs = NEON_UKERNELS,
2969 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002970 gcc_copts = xnnpack_gcc_std_copts(),
2971 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002972 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002973 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002974 "@FP16",
2975 "@pthreadpool",
2976 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977)
2978
2979xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002980 name = "neon_ukernels_test_mode",
2981 hdrs = INTERNAL_HDRS,
2982 aarch32_copts = [
2983 "-marm",
2984 "-march=armv7-a",
2985 "-mfpu=neon",
2986 ],
2987 aarch32_srcs = NEON_UKERNELS,
2988 aarch64_srcs = NEON_UKERNELS,
2989 copts = [
2990 "-UNDEBUG",
2991 "-DXNN_TEST_MODE=1",
2992 ],
2993 gcc_copts = xnnpack_gcc_std_copts(),
2994 msvc_copts = xnnpack_msvc_std_copts(),
2995 deps = [
2996 ":tables",
2997 "@FP16",
2998 "@pthreadpool",
2999 ],
3000)
3001
3002xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003003 name = "neonfma_ukernels",
3004 hdrs = INTERNAL_HDRS,
3005 aarch32_copts = [
3006 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003007 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003008 "-mfpu=neon-vfpv4",
3009 ],
3010 aarch32_srcs = NEONFMA_UKERNELS,
3011 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003012 apple_aarch32_copts = [
3013 "-mcpu=swift",
3014 "-mtune=generic",
3015 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003016 gcc_copts = xnnpack_gcc_std_copts(),
3017 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003018 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003019 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003020 "@FP16",
3021 "@pthreadpool",
3022 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003023)
3024
3025xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003026 name = "neonfma_ukernels_test_mode",
3027 hdrs = INTERNAL_HDRS,
3028 aarch32_copts = [
3029 "-marm",
3030 "-march=armv7-a",
3031 "-mfpu=neon-vfpv4",
3032 ],
3033 aarch32_srcs = NEONFMA_UKERNELS,
3034 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003035 apple_aarch32_copts = [
3036 "-mcpu=swift",
3037 "-mtune=generic",
3038 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003039 copts = [
3040 "-UNDEBUG",
3041 "-DXNN_TEST_MODE=1",
3042 ],
3043 gcc_copts = xnnpack_gcc_std_copts(),
3044 msvc_copts = xnnpack_msvc_std_copts(),
3045 deps = [
3046 ":tables",
3047 "@FP16",
3048 "@pthreadpool",
3049 ],
3050)
3051
3052xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003053 name = "neonv8_ukernels",
3054 hdrs = INTERNAL_HDRS,
3055 aarch32_copts = [
3056 "-marm",
3057 "-march=armv8-a",
3058 "-mfpu=neon-fp-armv8",
3059 ],
3060 aarch32_srcs = NEONV8_UKERNELS,
3061 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003062 apple_aarch32_copts = [
3063 "-mcpu=cyclone",
3064 "-mtune=generic",
3065 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003066 gcc_copts = xnnpack_gcc_std_copts(),
3067 msvc_copts = xnnpack_msvc_std_copts(),
3068 deps = [
3069 ":tables",
3070 "@FP16",
3071 "@pthreadpool",
3072 ],
3073)
3074
3075xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003076 name = "neonv8_ukernels_test_mode",
3077 hdrs = INTERNAL_HDRS,
3078 aarch32_copts = [
3079 "-marm",
3080 "-march=armv8-a",
3081 "-mfpu=neon-fp-armv8",
3082 ],
3083 aarch32_srcs = NEONV8_UKERNELS,
3084 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003085 apple_aarch32_copts = [
3086 "-mcpu=cyclone",
3087 "-mtune=generic",
3088 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003089 copts = [
3090 "-UNDEBUG",
3091 "-DXNN_TEST_MODE=1",
3092 ],
3093 gcc_copts = xnnpack_gcc_std_copts(),
3094 msvc_copts = xnnpack_msvc_std_copts(),
3095 deps = [
3096 ":tables",
3097 "@FP16",
3098 "@pthreadpool",
3099 ],
3100)
3101
3102xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003103 name = "neonfp16arith_ukernels",
3104 hdrs = INTERNAL_HDRS,
3105 aarch64_copts = ["-march=armv8.2-a+fp16"],
3106 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003107 gcc_copts = xnnpack_gcc_std_copts(),
3108 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003109 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003110 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003111 "@FP16",
3112 "@pthreadpool",
3113 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003114)
3115
3116xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003117 name = "neonfp16arith_ukernels_test_mode",
3118 hdrs = INTERNAL_HDRS,
3119 aarch64_copts = ["-march=armv8.2-a+fp16"],
3120 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3121 copts = [
3122 "-UNDEBUG",
3123 "-DXNN_TEST_MODE=1",
3124 ],
3125 gcc_copts = xnnpack_gcc_std_copts(),
3126 msvc_copts = xnnpack_msvc_std_copts(),
3127 deps = [
3128 ":tables",
3129 "@FP16",
3130 "@pthreadpool",
3131 ],
3132)
3133
3134xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003135 name = "neondot_ukernels",
3136 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003137 aarch32_copts = [
3138 "-marm",
3139 "-march=armv8.2-a+dotprod",
3140 "-mfpu=neon-fp-armv8",
3141 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003142 aarch32_srcs = NEONDOT_UKERNELS,
3143 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3144 aarch64_srcs = NEONDOT_UKERNELS,
3145 gcc_copts = xnnpack_gcc_std_copts(),
3146 msvc_copts = xnnpack_msvc_std_copts(),
3147 deps = [
3148 ":tables",
3149 "@FP16",
3150 "@pthreadpool",
3151 ],
3152)
3153
3154xnnpack_cc_library(
3155 name = "neondot_ukernels_test_mode",
3156 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003157 aarch32_copts = [
3158 "-marm",
3159 "-march=armv8.2-a+dotprod",
3160 "-mfpu=neon-fp-armv8",
3161 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003162 aarch32_srcs = NEONDOT_UKERNELS,
3163 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3164 aarch64_srcs = NEONDOT_UKERNELS,
3165 copts = [
3166 "-UNDEBUG",
3167 "-DXNN_TEST_MODE=1",
3168 ],
3169 gcc_copts = xnnpack_gcc_std_copts(),
3170 msvc_copts = xnnpack_msvc_std_copts(),
3171 deps = [
3172 ":tables",
3173 "@FP16",
3174 "@pthreadpool",
3175 ],
3176)
3177
3178xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003179 name = "sse2_ukernels",
3180 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003181 gcc_copts = xnnpack_gcc_std_copts(),
3182 gcc_x86_copts = ["-msse2"],
3183 msvc_copts = xnnpack_msvc_std_copts(),
3184 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003185 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003186 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003187 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003188 "@FP16",
3189 "@pthreadpool",
3190 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003191)
3192
3193xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003194 name = "sse2_ukernels_test_mode",
3195 hdrs = INTERNAL_HDRS,
3196 copts = [
3197 "-UNDEBUG",
3198 "-DXNN_TEST_MODE=1",
3199 ],
3200 gcc_copts = xnnpack_gcc_std_copts(),
3201 gcc_x86_copts = ["-msse2"],
3202 msvc_copts = xnnpack_msvc_std_copts(),
3203 msvc_x86_32_copts = ["/arch:SSE2"],
3204 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3205 deps = [
3206 ":tables",
3207 "@FP16",
3208 "@pthreadpool",
3209 ],
3210)
3211
3212xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003213 name = "ssse3_ukernels",
3214 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003215 gcc_copts = xnnpack_gcc_std_copts(),
3216 gcc_x86_copts = ["-mssse3"],
3217 msvc_copts = xnnpack_msvc_std_copts(),
3218 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003219 x86_srcs = SSSE3_UKERNELS,
3220 deps = [
3221 ":tables",
3222 "@FP16",
3223 "@pthreadpool",
3224 ],
3225)
3226
3227xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003228 name = "ssse3_ukernels_test_mode",
3229 hdrs = INTERNAL_HDRS,
3230 copts = [
3231 "-UNDEBUG",
3232 "-DXNN_TEST_MODE=1",
3233 ],
3234 gcc_copts = xnnpack_gcc_std_copts(),
3235 gcc_x86_copts = ["-mssse3"],
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 msvc_x86_32_copts = ["/arch:SSE2"],
3238 x86_srcs = SSSE3_UKERNELS,
3239 deps = [
3240 ":tables",
3241 "@FP16",
3242 "@pthreadpool",
3243 ],
3244)
3245
3246xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003247 name = "sse41_ukernels",
3248 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003249 gcc_copts = xnnpack_gcc_std_copts(),
3250 gcc_x86_copts = ["-msse4.1"],
3251 msvc_copts = xnnpack_msvc_std_copts(),
3252 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003253 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003254 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003255 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003256 "@FP16",
3257 "@pthreadpool",
3258 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003259)
3260
3261xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003262 name = "sse41_ukernels_test_mode",
3263 hdrs = INTERNAL_HDRS,
3264 copts = [
3265 "-UNDEBUG",
3266 "-DXNN_TEST_MODE=1",
3267 ],
3268 gcc_copts = xnnpack_gcc_std_copts(),
3269 gcc_x86_copts = ["-msse4.1"],
3270 msvc_copts = xnnpack_msvc_std_copts(),
3271 msvc_x86_32_copts = ["/arch:SSE2"],
3272 x86_srcs = SSE41_UKERNELS,
3273 deps = [
3274 ":tables",
3275 "@FP16",
3276 "@pthreadpool",
3277 ],
3278)
3279
3280xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003281 name = "avx_ukernels",
3282 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003283 gcc_copts = xnnpack_gcc_std_copts(),
3284 gcc_x86_copts = ["-mavx"],
3285 msvc_copts = xnnpack_msvc_std_copts(),
3286 msvc_x86_32_copts = ["/arch:AVX"],
3287 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003288 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003289 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003290 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003291 "@FP16",
3292 "@pthreadpool",
3293 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003294)
3295
3296xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003297 name = "avx_ukernels_test_mode",
3298 hdrs = INTERNAL_HDRS,
3299 copts = [
3300 "-UNDEBUG",
3301 "-DXNN_TEST_MODE=1",
3302 ],
3303 gcc_copts = xnnpack_gcc_std_copts(),
3304 gcc_x86_copts = ["-mavx"],
3305 msvc_copts = xnnpack_msvc_std_copts(),
3306 msvc_x86_32_copts = ["/arch:AVX"],
3307 msvc_x86_64_copts = ["/arch:AVX"],
3308 x86_srcs = AVX_UKERNELS,
3309 deps = [
3310 ":tables",
3311 "@FP16",
3312 "@pthreadpool",
3313 ],
3314)
3315
3316xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003317 name = "xop_ukernels",
3318 hdrs = INTERNAL_HDRS,
3319 gcc_copts = xnnpack_gcc_std_copts(),
3320 gcc_x86_copts = ["-mxop"],
3321 msvc_copts = xnnpack_msvc_std_copts(),
3322 msvc_x86_32_copts = ["/arch:AVX"],
3323 msvc_x86_64_copts = ["/arch:AVX"],
3324 x86_srcs = XOP_UKERNELS,
3325 deps = [
3326 ":tables",
3327 "@FP16",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
3333 name = "xop_ukernels_test_mode",
3334 hdrs = INTERNAL_HDRS,
3335 copts = [
3336 "-UNDEBUG",
3337 "-DXNN_TEST_MODE=1",
3338 ],
3339 gcc_copts = xnnpack_gcc_std_copts(),
3340 gcc_x86_copts = ["-mxop"],
3341 msvc_copts = xnnpack_msvc_std_copts(),
3342 msvc_x86_32_copts = ["/arch:AVX"],
3343 msvc_x86_64_copts = ["/arch:AVX"],
3344 x86_srcs = XOP_UKERNELS,
3345 deps = [
3346 ":tables",
3347 "@FP16",
3348 "@pthreadpool",
3349 ],
3350)
3351
3352xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003353 name = "fma3_ukernels",
3354 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003355 gcc_copts = xnnpack_gcc_std_copts(),
3356 gcc_x86_copts = ["-mfma"],
3357 msvc_copts = xnnpack_msvc_std_copts(),
3358 msvc_x86_32_copts = ["/arch:AVX"],
3359 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003360 x86_srcs = FMA3_UKERNELS,
3361 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003362 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003363 "@FP16",
3364 "@pthreadpool",
3365 ],
3366)
3367
3368xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003369 name = "fma3_ukernels_test_mode",
3370 hdrs = INTERNAL_HDRS,
3371 copts = [
3372 "-UNDEBUG",
3373 "-DXNN_TEST_MODE=1",
3374 ],
3375 gcc_copts = xnnpack_gcc_std_copts(),
3376 gcc_x86_copts = ["-mfma"],
3377 msvc_copts = xnnpack_msvc_std_copts(),
3378 msvc_x86_32_copts = ["/arch:AVX"],
3379 msvc_x86_64_copts = ["/arch:AVX"],
3380 x86_srcs = FMA3_UKERNELS,
3381 deps = [
3382 ":tables",
3383 "@FP16",
3384 "@pthreadpool",
3385 ],
3386)
3387
3388xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003389 name = "avx2_ukernels",
3390 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003391 gcc_copts = xnnpack_gcc_std_copts(),
3392 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003393 "-mfma",
3394 "-mavx2",
3395 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003396 msvc_copts = xnnpack_msvc_std_copts(),
3397 msvc_x86_32_copts = ["/arch:AVX2"],
3398 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003399 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003400 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003401 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003402 "@FP16",
3403 "@pthreadpool",
3404 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003405)
3406
3407xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003408 name = "avx2_ukernels_test_mode",
3409 hdrs = INTERNAL_HDRS,
3410 copts = [
3411 "-UNDEBUG",
3412 "-DXNN_TEST_MODE=1",
3413 ],
3414 gcc_copts = xnnpack_gcc_std_copts(),
3415 gcc_x86_copts = [
3416 "-mfma",
3417 "-mavx2",
3418 ],
3419 msvc_copts = xnnpack_msvc_std_copts(),
3420 msvc_x86_32_copts = ["/arch:AVX2"],
3421 msvc_x86_64_copts = ["/arch:AVX2"],
3422 x86_srcs = AVX2_UKERNELS,
3423 deps = [
3424 ":tables",
3425 "@FP16",
3426 "@pthreadpool",
3427 ],
3428)
3429
3430xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003431 name = "avx512f_ukernels",
3432 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003433 gcc_copts = xnnpack_gcc_std_copts(),
3434 gcc_x86_copts = ["-mavx512f"],
3435 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3436 msvc_copts = xnnpack_msvc_std_copts(),
3437 msvc_x86_32_copts = ["/arch:AVX512"],
3438 msvc_x86_64_copts = ["/arch:AVX512"],
3439 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003441 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003442 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003443 "@FP16",
3444 "@pthreadpool",
3445 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446)
3447
3448xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003449 name = "avx512f_ukernels_test_mode",
3450 hdrs = INTERNAL_HDRS,
3451 copts = [
3452 "-UNDEBUG",
3453 "-DXNN_TEST_MODE=1",
3454 ],
3455 gcc_copts = xnnpack_gcc_std_copts(),
3456 gcc_x86_copts = ["-mavx512f"],
3457 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3458 msvc_copts = xnnpack_msvc_std_copts(),
3459 msvc_x86_32_copts = ["/arch:AVX512"],
3460 msvc_x86_64_copts = ["/arch:AVX512"],
3461 msys_copts = ["-fno-asynchronous-unwind-tables"],
3462 x86_srcs = AVX512F_UKERNELS,
3463 deps = [
3464 ":tables",
3465 "@FP16",
3466 "@pthreadpool",
3467 ],
3468)
3469
3470xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003471 name = "avx512skx_ukernels",
3472 hdrs = INTERNAL_HDRS,
3473 gcc_copts = xnnpack_gcc_std_copts(),
3474 gcc_x86_copts = [
3475 "-mavx512f",
3476 "-mavx512cd",
3477 "-mavx512bw",
3478 "-mavx512dq",
3479 "-mavx512vl",
3480 ],
3481 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3482 msvc_copts = xnnpack_msvc_std_copts(),
3483 msvc_x86_32_copts = ["/arch:AVX512"],
3484 msvc_x86_64_copts = ["/arch:AVX512"],
3485 msys_copts = ["-fno-asynchronous-unwind-tables"],
3486 x86_srcs = AVX512SKX_UKERNELS,
3487 deps = [
3488 ":tables",
3489 "@FP16",
3490 "@pthreadpool",
3491 ],
3492)
3493
3494xnnpack_cc_library(
3495 name = "avx512skx_ukernels_test_mode",
3496 hdrs = INTERNAL_HDRS,
3497 copts = [
3498 "-UNDEBUG",
3499 "-DXNN_TEST_MODE=1",
3500 ],
3501 gcc_copts = xnnpack_gcc_std_copts(),
3502 gcc_x86_copts = [
3503 "-mavx512f",
3504 "-mavx512cd",
3505 "-mavx512bw",
3506 "-mavx512dq",
3507 "-mavx512vl",
3508 ],
3509 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3510 msvc_copts = xnnpack_msvc_std_copts(),
3511 msvc_x86_32_copts = ["/arch:AVX512"],
3512 msvc_x86_64_copts = ["/arch:AVX512"],
3513 msys_copts = ["-fno-asynchronous-unwind-tables"],
3514 x86_srcs = AVX512SKX_UKERNELS,
3515 deps = [
3516 ":tables",
3517 "@FP16",
3518 "@pthreadpool",
3519 ],
3520)
3521
3522xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003523 name = "asm_ukernels",
3524 hdrs = ["src/xnnpack/assembly.h"],
3525 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003526 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003528 wasm_srcs = WASM32_ASM_UKERNELS,
3529 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003530)
3531
Marat Dukhan3b59de22020-06-03 20:15:19 -07003532xnnpack_cc_library(
3533 name = "logging_utils",
3534 srcs = LOGGING_SRCS,
3535 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3536 copts = LOGGING_COPTS + [
3537 "-Isrc",
3538 "-Iinclude",
3539 ] + select({
3540 ":debug_build": [],
3541 "//conditions:default": xnnpack_min_size_copts(),
3542 }),
3543 gcc_copts = xnnpack_gcc_std_copts(),
3544 msvc_copts = xnnpack_msvc_std_copts(),
3545 visibility = xnnpack_visibility(),
3546 deps = [
3547 "@FP16",
3548 "@clog",
3549 "@pthreadpool",
3550 ],
3551)
3552
Marat Dukhan08c4a432019-10-03 09:29:21 -07003553xnnpack_aggregate_library(
3554 name = "ukernels",
3555 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003556 ":neon_ukernels",
3557 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003558 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003559 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003560 ":asm_ukernels",
3561 ],
3562 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003563 ":neon_ukernels",
3564 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003565 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003566 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003567 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568 ":asm_ukernels",
3569 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003570 generic_deps = [
3571 ":scalar_ukernels",
3572 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003573 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003574 ":psimd_fastmath_ukernels",
3575 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003577 wasm_deps = [
3578 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003579 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003580 ],
3581 wasmsimd_deps = [
3582 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003583 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003584 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003587 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003588 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003590 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003591 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003592 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003594 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003595 ],
3596)
3597
Marat Dukhan33fcf782020-05-24 14:27:15 -07003598xnnpack_aggregate_library(
3599 name = "ukernels_test_mode",
3600 aarch32_deps = [
3601 ":neon_ukernels_test_mode",
3602 ":neonfma_ukernels_test_mode",
3603 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003604 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003605 ":asm_ukernels",
3606 ],
3607 aarch64_deps = [
3608 ":neon_ukernels_test_mode",
3609 ":neonfma_ukernels_test_mode",
3610 ":neonv8_ukernels_test_mode",
3611 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003612 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003613 ":asm_ukernels",
3614 ],
3615 generic_deps = [
3616 ":scalar_ukernels_test_mode",
3617 ],
3618 psimd_deps = [
3619 ":psimd_fastmath_ukernels_test_mode",
3620 ":psimd_accmath_ukernels_test_mode",
3621 ],
3622 wasm_deps = [
3623 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003624 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003625 ],
3626 wasmsimd_deps = [
3627 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003628 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003629 ],
3630 x86_deps = [
3631 ":sse2_ukernels_test_mode",
3632 ":ssse3_ukernels_test_mode",
3633 ":sse41_ukernels_test_mode",
3634 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003635 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003636 ":fma3_ukernels_test_mode",
3637 ":avx2_ukernels_test_mode",
3638 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003639 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003640 ],
3641)
3642
Marat Dukhan86df3842020-06-17 14:03:31 -07003643# This target must be compatible with Android NDK r17 and compile with Clang 6.
3644# It is similar to :ukernels target, but excludes microkernels which require
3645# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003646xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003647 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003648 aarch32_deps = [
3649 ":neon_ukernels",
3650 ":neonfma_ukernels",
3651 ":neonv8_ukernels",
3652 ":asm_ukernels",
3653 ],
3654 aarch64_deps = [
3655 ":neon_ukernels",
3656 ":neonfma_ukernels",
3657 ":neonv8_ukernels",
3658 ":asm_ukernels",
3659 ],
3660 generic_deps = [
3661 ":scalar_ukernels",
3662 ],
3663 psimd_deps = [
3664 ":psimd_fastmath_ukernels",
3665 ":psimd_accmath_ukernels",
3666 ],
3667 wasm_deps = [
3668 ":wasm_ukernels",
3669 ],
3670 wasmsimd_deps = [
3671 ":wasm_ukernels",
3672 ],
3673 x86_deps = [
3674 ":sse2_ukernels",
3675 ":ssse3_ukernels",
3676 ":sse41_ukernels",
3677 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003678 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003679 ":fma3_ukernels",
3680 ":avx2_ukernels",
3681 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003682 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003683 ],
3684)
3685
Marat Dukhan08c4a432019-10-03 09:29:21 -07003686xnnpack_cc_library(
3687 name = "im2col",
3688 srcs = ["src/im2col.c"],
3689 hdrs = [
3690 "src/xnnpack/common.h",
3691 "src/xnnpack/im2col.h",
3692 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003693 gcc_copts = xnnpack_gcc_std_copts(),
3694 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003695)
3696
3697xnnpack_cc_library(
3698 name = "indirection",
3699 srcs = ["src/indirection.c"],
3700 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003701 gcc_copts = xnnpack_gcc_std_copts(),
3702 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003703 deps = [
3704 "@FP16",
3705 "@FXdiv",
3706 "@pthreadpool",
3707 ],
3708)
3709
3710xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003711 name = "indirection_test_mode",
3712 srcs = ["src/indirection.c"],
3713 hdrs = INTERNAL_HDRS,
3714 copts = [
3715 "-UNDEBUG",
3716 "-DXNN_TEST_MODE=1",
3717 ],
3718 gcc_copts = xnnpack_gcc_std_copts(),
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 deps = [
3721 "@FP16",
3722 "@FXdiv",
3723 "@pthreadpool",
3724 ],
3725)
3726
3727xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003728 name = "packing",
3729 srcs = ["src/packing.c"],
3730 hdrs = INTERNAL_HDRS,
3731 gcc_copts = xnnpack_gcc_std_copts(),
3732 msvc_copts = xnnpack_msvc_std_copts(),
3733 deps = [
3734 "@FP16",
3735 "@FXdiv",
3736 "@pthreadpool",
3737 ],
3738)
3739
3740xnnpack_cc_library(
3741 name = "packing_test_mode",
3742 srcs = ["src/packing.c"],
3743 hdrs = INTERNAL_HDRS,
3744 copts = [
3745 "-UNDEBUG",
3746 "-DXNN_TEST_MODE=1",
3747 ],
3748 gcc_copts = xnnpack_gcc_std_copts(),
3749 msvc_copts = xnnpack_msvc_std_copts(),
3750 deps = [
3751 "@FP16",
3752 "@FXdiv",
3753 "@pthreadpool",
3754 ],
3755)
3756
3757xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003758 name = "operator_run",
3759 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003760 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003761 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003762 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3763 "//conditions:default": [],
3764 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003765 gcc_copts = xnnpack_gcc_std_copts(),
3766 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003768 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003769 "@FP16",
3770 "@FXdiv",
3771 "@clog",
3772 "@pthreadpool",
3773 ],
3774)
3775
Chao Mei6ddfc602020-05-13 22:29:36 -07003776xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003777 name = "operator_run_test_mode",
3778 srcs = ["src/operator-run.c"],
3779 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3780 copts = LOGGING_COPTS + [
3781 "-UNDEBUG",
3782 "-DXNN_TEST_MODE=1",
3783 ] + select({
3784 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3785 "//conditions:default": [],
3786 }),
3787 gcc_copts = xnnpack_gcc_std_copts(),
3788 msvc_copts = xnnpack_msvc_std_copts(),
3789 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003790 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003791 "@FP16",
3792 "@FXdiv",
3793 "@clog",
3794 "@pthreadpool",
3795 ],
3796)
3797
3798xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003799 name = "memory_planner",
3800 srcs = ["src/memory-planner.c"],
3801 hdrs = INTERNAL_HDRS,
3802 defines = select({
3803 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3804 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3805 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3806 }),
3807 gcc_copts = xnnpack_gcc_std_copts(),
3808 msvc_copts = xnnpack_msvc_std_copts(),
3809 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003810 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003811 "@pthreadpool",
3812 ],
3813)
3814
Marat Dukhan33fcf782020-05-24 14:27:15 -07003815xnnpack_cc_library(
3816 name = "memory_planner_test_mode",
3817 srcs = ["src/memory-planner.c"],
3818 hdrs = INTERNAL_HDRS,
3819 copts = [
3820 "-UNDEBUG",
3821 "-DXNN_TEST_MODE=1",
3822 ],
3823 defines = select({
3824 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3825 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3826 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3827 }),
3828 gcc_copts = xnnpack_gcc_std_copts(),
3829 msvc_copts = xnnpack_msvc_std_copts(),
3830 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003831 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003832 "@pthreadpool",
3833 ],
3834)
3835
Marat Dukhan08c4a432019-10-03 09:29:21 -07003836cc_library(
3837 name = "enable_assembly",
3838 defines = select({
3839 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3840 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003841 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003842 }),
3843)
3844
Marat Dukhan9de90e02020-06-18 16:04:12 -07003845cc_library(
3846 name = "enable_sparse",
3847 defines = select({
3848 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3849 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003850 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003851 }),
3852)
3853
Marat Dukhancf056b22019-10-07 10:26:29 -07003854xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855 name = "operators",
3856 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003857 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003859 ],
3860 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003861 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003862 "-Isrc",
3863 "-Iinclude",
3864 ] + select({
3865 ":debug_build": [],
3866 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003867 }) + select({
3868 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3869 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003870 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003871 gcc_copts = xnnpack_gcc_std_copts(),
3872 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003873 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003875 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003876 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003877 "@FP16",
3878 "@FXdiv",
3879 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003880 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003881 ],
3882)
3883
Marat Dukhan10a38082020-04-17 03:58:35 -07003884xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003885 name = "operators_test_mode",
3886 srcs = OPERATOR_SRCS + [
3887 "src/memory.c",
3888 "src/operator-delete.c",
3889 ],
3890 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3891 copts = LOGGING_COPTS + [
3892 "-Isrc",
3893 "-Iinclude",
3894 "-UNDEBUG",
3895 "-DXNN_TEST_MODE=1",
3896 ] + select({
3897 ":debug_build": [],
3898 "//conditions:default": xnnpack_min_size_copts(),
3899 }) + select({
3900 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3901 "//conditions:default": [],
3902 }),
3903 gcc_copts = xnnpack_gcc_std_copts(),
3904 msvc_copts = xnnpack_msvc_std_copts(),
3905 deps = [
3906 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003907 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003908 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003909 "@FP16",
3910 "@FXdiv",
3911 "@clog",
3912 "@pthreadpool",
3913 ],
3914)
3915
3916xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003917 name = "XNNPACK",
3918 srcs = [
3919 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003920 "src/runtime.c",
3921 "src/subgraph.c",
3922 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003923 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003924 hdrs = ["include/xnnpack.h"],
3925 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003926 "-Isrc",
3927 "-Iinclude",
3928 ] + select({
3929 ":debug_build": [],
3930 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003931 }) + select({
3932 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3933 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003934 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003935 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003936 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003937 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003938 visibility = xnnpack_visibility(),
3939 deps = [
3940 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003941 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003942 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003943 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003944 ":operator_run",
3945 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003946 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003947 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003948 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003949 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003950 ] + select({
3951 ":emscripten": [],
3952 "//conditions:default": ["@cpuinfo"],
3953 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003954)
3955
Marat Dukhan10a38082020-04-17 03:58:35 -07003956xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003957 name = "XNNPACK_test_mode",
3958 srcs = [
3959 "src/init.c",
3960 "src/runtime.c",
3961 "src/subgraph.c",
3962 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003963 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003964 hdrs = ["include/xnnpack.h"],
3965 copts = LOGGING_COPTS + [
3966 "-Isrc",
3967 "-Iinclude",
3968 "-UNDEBUG",
3969 "-DXNN_TEST_MODE=1",
3970 ] + select({
3971 ":debug_build": [],
3972 "//conditions:default": xnnpack_min_size_copts(),
3973 }) + select({
3974 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3975 "//conditions:default": [],
3976 }),
3977 gcc_copts = xnnpack_gcc_std_copts(),
3978 includes = ["include"],
3979 msvc_copts = xnnpack_msvc_std_copts(),
3980 visibility = xnnpack_visibility(),
3981 deps = [
3982 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003983 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003984 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003985 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003986 ":operator_run_test_mode",
3987 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003988 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003989 "@clog",
3990 "@FP16",
3991 "@pthreadpool",
3992 ] + select({
3993 ":emscripten": [],
3994 "//conditions:default": ["@cpuinfo"],
3995 }),
3996)
3997
3998xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003999 name = "xnnpack_f32",
4000 srcs = [
4001 "src/init.c",
4002 "src/runtime.c",
4003 "src/subgraph.c",
4004 "src/tensor.c",
4005 ] + SUBGRAPH_SRCS,
4006 hdrs = ["include/xnnpack.h"],
4007 copts = LOGGING_COPTS + [
4008 "-Isrc",
4009 "-Iinclude",
4010 ] + select({
4011 ":debug_build": [],
4012 "//conditions:default": xnnpack_min_size_copts(),
4013 }) + select({
4014 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4015 "//conditions:default": [],
4016 }),
4017 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004018 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004019 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004020 "XNN_NO_U8_OPERATORS",
4021 "XNN_NO_X8_OPERATORS",
4022 "XNN_NO_F16_OPERATORS",
4023 "XNN_NO_X16_OPERATORS",
4024 ],
4025 gcc_copts = xnnpack_gcc_std_copts(),
4026 includes = ["include"],
4027 msvc_copts = xnnpack_msvc_std_copts(),
4028 visibility = xnnpack_visibility(),
4029 deps = [
4030 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004031 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004032 ":logging_utils",
4033 ":memory_planner",
4034 ":operator_run",
4035 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004036 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004037 "@clog",
4038 "@FP16",
4039 "@pthreadpool",
4040 ] + select({
4041 ":emscripten": [],
4042 "//conditions:default": ["@cpuinfo"],
4043 }),
4044)
4045
4046xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004047 name = "xnnpack_operators_nhwc_f32",
4048 srcs = [
4049 "src/init.c",
4050 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004051 hdrs = ["include/xnnpack.h"],
4052 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004053 "-Isrc",
4054 "-Iinclude",
4055 ] + select({
4056 ":debug_build": [],
4057 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004058 }) + select({
4059 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4060 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004061 }),
4062 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004063 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004064 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004065 "XNN_NO_U8_OPERATORS",
4066 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004067 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004068 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004069 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004070 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004071 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004072 visibility = xnnpack_visibility(),
4073 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004074 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004075 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004076 ":operator_run",
4077 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004078 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004079 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004080 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004081 ] + select({
4082 ":emscripten": [],
4083 "//conditions:default": ["@cpuinfo"],
4084 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004085)
4086
Marat Dukhancf056b22019-10-07 10:26:29 -07004087xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004088 name = "bench_utils",
4089 srcs = ["bench/utils.cc"],
4090 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004091 deps = [
4092 "@com_google_benchmark//:benchmark",
4093 "@cpuinfo",
4094 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004095)
4096
Frank Barchard7e955972019-10-11 10:34:25 -07004097######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098
4099xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004100 name = "qs8_gemm_bench",
4101 srcs = [
4102 "bench/gemm.h",
4103 "bench/qs8-gemm.cc",
4104 "src/xnnpack/AlignedAllocator.h",
4105 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4106 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4107)
4108
4109xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004110 name = "qs8_requantization_bench",
4111 srcs = [
4112 "bench/qs8-requantization.cc",
4113 "src/xnnpack/requantization-stubs.h",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + MICROKERNEL_BENCHMARK_HDRS,
4116 deps = MICROKERNEL_BENCHMARK_DEPS,
4117)
4118
4119xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004120 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004121 srcs = [
4122 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004123 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004124 "src/xnnpack/AlignedAllocator.h",
4125 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004126 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004127 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004128)
4129
4130xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004131 name = "qu8_requantization_bench",
4132 srcs = [
4133 "bench/qu8-requantization.cc",
4134 "src/xnnpack/requantization-stubs.h",
4135 "src/xnnpack/AlignedAllocator.h",
4136 ] + MICROKERNEL_BENCHMARK_HDRS,
4137 deps = MICROKERNEL_BENCHMARK_DEPS,
4138)
4139
4140xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004141 name = "f16_igemm_bench",
4142 srcs = [
4143 "bench/f16-igemm.cc",
4144 "bench/conv.h",
4145 "bench/google/conv.h",
4146 "src/xnnpack/AlignedAllocator.h",
4147 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004148 deps = MICROKERNEL_BENCHMARK_DEPS + [
4149 ":indirection",
4150 ":packing",
4151 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004152)
4153
4154xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004155 name = "f16_gemm_bench",
4156 srcs = [
4157 "bench/f16-gemm.cc",
4158 "bench/gemm.h",
4159 "src/xnnpack/AlignedAllocator.h",
4160 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004161 deps = MICROKERNEL_BENCHMARK_DEPS + [
4162 ":packing",
4163 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004164)
4165
4166xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004167 name = "f16_spmm_bench",
4168 srcs = [
4169 "bench/f16-spmm.cc",
4170 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004171 "src/xnnpack/AlignedAllocator.h",
4172 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004173 deps = MICROKERNEL_BENCHMARK_DEPS,
4174)
4175
4176xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004177 name = "f32_igemm_bench",
4178 srcs = [
4179 "bench/f32-igemm.cc",
4180 "bench/conv.h",
4181 "src/xnnpack/AlignedAllocator.h",
4182 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004183 deps = MICROKERNEL_BENCHMARK_DEPS + [
4184 ":indirection",
4185 ":packing",
4186 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187)
4188
4189xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004190 name = "f16_relu_bench",
4191 srcs = [
4192 "bench/f16-relu.cc",
4193 "src/xnnpack/AlignedAllocator.h",
4194 ] + MICROKERNEL_BENCHMARK_HDRS,
4195 deps = MICROKERNEL_BENCHMARK_DEPS,
4196)
4197
4198xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004199 name = "f32_conv_hwc_bench",
4200 srcs = [
4201 "bench/f32-conv-hwc.cc",
4202 "bench/dconv.h",
4203 "src/xnnpack/AlignedAllocator.h",
4204 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004205 deps = MICROKERNEL_BENCHMARK_DEPS + [
4206 ":packing",
4207 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004208)
4209
4210xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004211 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004212 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004213 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004214 "bench/dconv.h",
4215 "src/xnnpack/AlignedAllocator.h",
4216 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004217 deps = MICROKERNEL_BENCHMARK_DEPS + [
4218 ":packing",
4219 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004220)
4221
4222xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004223 name = "f16_dwconv_bench",
4224 srcs = [
4225 "bench/f16-dwconv.cc",
4226 "bench/dwconv.h",
4227 "bench/google/dwconv.h",
4228 "src/xnnpack/AlignedAllocator.h",
4229 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004230 deps = MICROKERNEL_BENCHMARK_DEPS + [
4231 ":indirection",
4232 ":packing",
4233 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004234)
4235
4236xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004237 name = "f32_dwconv_bench",
4238 srcs = [
4239 "bench/f32-dwconv.cc",
4240 "bench/dwconv.h",
4241 "src/xnnpack/AlignedAllocator.h",
4242 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004243 deps = MICROKERNEL_BENCHMARK_DEPS + [
4244 ":indirection",
4245 ":packing",
4246 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004247)
4248
4249xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004250 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004251 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004252 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 "bench/dwconv.h",
4254 "src/xnnpack/AlignedAllocator.h",
4255 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004256 deps = MICROKERNEL_BENCHMARK_DEPS + [
4257 ":indirection",
4258 ":packing",
4259 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004260)
4261
4262xnnpack_benchmark(
4263 name = "f32_gemm_bench",
4264 srcs = [
4265 "bench/f32-gemm.cc",
4266 "bench/gemm.h",
4267 "src/xnnpack/AlignedAllocator.h",
4268 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004269 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004270 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004271)
4272
4273xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004274 name = "f32_hswish_bench",
4275 srcs = [
4276 "bench/f32-hswish.cc",
4277 "src/xnnpack/AlignedAllocator.h",
4278 ] + MICROKERNEL_BENCHMARK_HDRS,
4279 deps = MICROKERNEL_BENCHMARK_DEPS,
4280)
4281
4282xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004283 name = "f32_raddexpminusmax_bench",
4284 srcs = [
4285 "bench/f32-raddexpminusmax.cc",
4286 "src/xnnpack/AlignedAllocator.h",
4287 ] + MICROKERNEL_BENCHMARK_HDRS,
4288 deps = MICROKERNEL_BENCHMARK_DEPS,
4289)
4290
4291xnnpack_benchmark(
4292 name = "f32_raddextexp_bench",
4293 srcs = [
4294 "bench/f32-raddextexp.cc",
4295 "src/xnnpack/AlignedAllocator.h",
4296 ] + MICROKERNEL_BENCHMARK_HDRS,
4297 deps = MICROKERNEL_BENCHMARK_DEPS,
4298)
4299
4300xnnpack_benchmark(
4301 name = "f32_raddstoreexpminusmax_bench",
4302 srcs = [
4303 "bench/f32-raddstoreexpminusmax.cc",
4304 "src/xnnpack/AlignedAllocator.h",
4305 ] + MICROKERNEL_BENCHMARK_HDRS,
4306 deps = MICROKERNEL_BENCHMARK_DEPS,
4307)
4308
4309xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004310 name = "f32_relu_bench",
4311 srcs = [
4312 "bench/f32-relu.cc",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + MICROKERNEL_BENCHMARK_HDRS,
4315 deps = MICROKERNEL_BENCHMARK_DEPS,
4316)
4317
4318xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319 name = "f32_rmax_bench",
4320 srcs = [
4321 "bench/f32-rmax.cc",
4322 "src/xnnpack/AlignedAllocator.h",
4323 ] + MICROKERNEL_BENCHMARK_HDRS,
4324 deps = MICROKERNEL_BENCHMARK_DEPS,
4325)
4326
4327xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004328 name = "f32_sigmoid_bench",
4329 srcs = [
4330 "bench/f32-sigmoid.cc",
4331 "src/xnnpack/AlignedAllocator.h",
4332 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004333 deps = MICROKERNEL_BENCHMARK_DEPS,
4334)
4335
4336xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004337 name = "f32_spmm_bench",
4338 srcs = [
4339 "bench/f32-spmm.cc",
4340 "bench/gemm.h",
4341 "src/xnnpack/AlignedAllocator.h",
4342 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004343 deps = MICROKERNEL_BENCHMARK_DEPS,
4344)
4345
4346xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004347 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004348 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004349 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004350 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004351 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004352 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004353)
4354
4355xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004356 name = "f32_vscaleexpminusmax_bench",
4357 srcs = [
4358 "bench/f32-vscaleexpminusmax.cc",
4359 "src/xnnpack/AlignedAllocator.h",
4360 ] + MICROKERNEL_BENCHMARK_HDRS,
4361 deps = MICROKERNEL_BENCHMARK_DEPS,
4362)
4363
4364xnnpack_benchmark(
4365 name = "f32_vscaleextexp_bench",
4366 srcs = [
4367 "bench/f32-vscaleextexp.cc",
4368 "src/xnnpack/AlignedAllocator.h",
4369 ] + MICROKERNEL_BENCHMARK_HDRS,
4370 deps = MICROKERNEL_BENCHMARK_DEPS,
4371)
4372
4373xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004374 name = "f32_vsqrt_bench",
4375 srcs = [
4376 "bench/f32-vsqrt.cc",
4377 "src/xnnpack/AlignedAllocator.h",
4378 ] + MICROKERNEL_BENCHMARK_HDRS,
4379 deps = MICROKERNEL_BENCHMARK_DEPS,
4380)
4381
4382xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004383 name = "f32_im2col_gemm_bench",
4384 srcs = [
4385 "bench/f32-im2col-gemm.cc",
4386 "bench/conv.h",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004389 deps = MICROKERNEL_BENCHMARK_DEPS + [
4390 ":im2col",
4391 ":packing",
4392 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393)
4394
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004395xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004396 name = "rounding_bench",
4397 srcs = [
4398 "bench/rounding.cc",
4399 "src/xnnpack/math-stubs.h",
4400 "src/xnnpack/AlignedAllocator.h",
4401 ] + MICROKERNEL_BENCHMARK_HDRS,
4402 deps = MICROKERNEL_BENCHMARK_DEPS,
4403)
4404
Marat Dukhan08c4a432019-10-03 09:29:21 -07004405########################### Benchmarks for operators ###########################
4406
4407xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004408 name = "average_pooling_bench",
4409 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004410 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004411 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004412 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004413)
4414
4415xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004416 name = "bankers_rounding_bench",
4417 srcs = ["bench/bankers-rounding.cc"],
4418 copts = xnnpack_optional_tflite_copts(),
4419 tags = ["nowin32"],
4420 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4421)
4422
4423xnnpack_benchmark(
4424 name = "ceiling_bench",
4425 srcs = ["bench/ceiling.cc"],
4426 copts = xnnpack_optional_tflite_copts(),
4427 tags = ["nowin32"],
4428 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4429)
4430
4431xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004432 name = "channel_shuffle_bench",
4433 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004434 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004435)
4436
4437xnnpack_benchmark(
4438 name = "convolution_bench",
4439 srcs = ["bench/convolution.cc"],
4440 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004441 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004442 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004443)
4444
4445xnnpack_benchmark(
4446 name = "deconvolution_bench",
4447 srcs = ["bench/deconvolution.cc"],
4448 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004449 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004450 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451)
4452
4453xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004454 name = "floor_bench",
4455 srcs = ["bench/floor.cc"],
4456 copts = xnnpack_optional_tflite_copts(),
4457 tags = ["nowin32"],
4458 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4459)
4460
4461xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004462 name = "global_average_pooling_bench",
4463 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004464 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004465)
4466
4467xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004468 name = "hardswish_bench",
4469 srcs = ["bench/hardswish.cc"],
4470 copts = xnnpack_optional_tflite_copts(),
4471 tags = ["nowin32"],
4472 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4473)
4474
4475xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476 name = "max_pooling_bench",
4477 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004478 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004479)
4480
4481xnnpack_benchmark(
4482 name = "sigmoid_bench",
4483 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004484 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004485 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004486 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004487)
4488
4489xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004490 name = "prelu_bench",
4491 srcs = ["bench/prelu.cc"],
4492 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004493 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004494 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004495)
4496
4497xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004498 name = "softmax_bench",
4499 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004500 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004501 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004502 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503)
4504
Marat Dukhan87727142020-06-24 15:24:10 -07004505xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004506 name = "square_root_bench",
4507 srcs = ["bench/square-root.cc"],
4508 copts = xnnpack_optional_tflite_copts(),
4509 tags = ["nowin32"],
4510 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4511)
4512
4513xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004514 name = "truncation_bench",
4515 srcs = ["bench/truncation.cc"],
4516 deps = OPERATOR_BENCHMARK_DEPS,
4517)
4518
Marat Dukhanc068bb62019-10-04 13:24:39 -07004519############################# End-to-end benchmarks ############################
4520
4521cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004522 name = "fp32_mobilenet_v1",
4523 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004524 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004525 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004526 linkstatic = True,
4527 deps = [
4528 ":XNNPACK",
4529 "@pthreadpool",
4530 ],
4531)
4532
4533cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004534 name = "fp16_mobilenet_v1",
4535 srcs = ["models/fp16-mobilenet-v1.cc"],
4536 hdrs = ["models/models.h"],
4537 copts = xnnpack_std_cxxopts(),
4538 linkstatic = True,
4539 deps = [
4540 ":XNNPACK",
4541 "@FP16",
4542 "@pthreadpool",
4543 ],
4544)
4545
4546cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004547 name = "qs8_mobilenet_v1",
4548 srcs = ["models/qs8-mobilenet-v1.cc"],
4549 hdrs = ["models/models.h"],
4550 copts = xnnpack_std_cxxopts(),
4551 linkstatic = True,
4552 deps = [
4553 ":XNNPACK",
4554 "@pthreadpool",
4555 ],
4556)
4557
4558cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004559 name = "fp32_mobilenet_v2",
4560 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004561 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004562 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004563 linkstatic = True,
4564 deps = [
4565 ":XNNPACK",
4566 "@pthreadpool",
4567 ],
4568)
4569
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004570cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004571 name = "fp16_mobilenet_v2",
4572 srcs = ["models/fp16-mobilenet-v2.cc"],
4573 hdrs = ["models/models.h"],
4574 copts = xnnpack_std_cxxopts(),
4575 linkstatic = True,
4576 deps = [
4577 ":XNNPACK",
4578 "@FP16",
4579 "@pthreadpool",
4580 ],
4581)
4582
4583cc_library(
4584 name = "fp32_mobilenet_v3_large",
4585 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004586 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004587 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004588 linkstatic = True,
4589 deps = [
4590 ":XNNPACK",
4591 "@pthreadpool",
4592 ],
4593)
4594
4595cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004596 name = "fp32_mobilenet_v3_small",
4597 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004598 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004599 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004600 linkstatic = True,
4601 deps = [
4602 ":XNNPACK",
4603 "@pthreadpool",
4604 ],
4605)
4606
Marat Dukhanc068bb62019-10-04 13:24:39 -07004607xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004608 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004609 srcs = [
4610 "bench/f32-dwconv-e2e.cc",
4611 "bench/end2end.h",
4612 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004613 deps = MICROKERNEL_BENCHMARK_DEPS + [
4614 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004615 ":fp32_mobilenet_v1",
4616 ":fp32_mobilenet_v2",
4617 ":fp32_mobilenet_v3_large",
4618 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004619 ],
4620)
4621
4622xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004623 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004624 srcs = [
4625 "bench/f32-gemm-e2e.cc",
4626 "bench/end2end.h",
4627 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004628 deps = MICROKERNEL_BENCHMARK_DEPS + [
4629 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004630 ":fp32_mobilenet_v1",
4631 ":fp32_mobilenet_v2",
4632 ":fp32_mobilenet_v3_large",
4633 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004634 ],
4635)
4636
4637xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004638 name = "end2end_bench",
4639 srcs = ["bench/end2end.cc"],
4640 deps = [
4641 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004642 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004643 ":fp16_mobilenet_v1",
4644 ":fp16_mobilenet_v2",
4645 ":fp32_mobilenet_v1",
4646 ":fp32_mobilenet_v2",
4647 ":fp32_mobilenet_v3_large",
4648 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004649 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004650 "@pthreadpool",
4651 ],
4652)
4653
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004654#################### Accuracy evaluation for math functions ####################
4655
4656xnnpack_benchmark(
4657 name = "f32_exp_eval",
4658 srcs = [
4659 "eval/f32-exp.cc",
4660 "src/xnnpack/AlignedAllocator.h",
4661 ] + ACCURACY_EVAL_HDRS,
4662 deps = ACCURACY_EVAL_DEPS,
4663)
4664
Marat Dukhan515c9772019-10-17 18:07:57 -07004665xnnpack_benchmark(
4666 name = "f32_expminus_eval",
4667 srcs = [
4668 "eval/f32-expminus.cc",
4669 "src/xnnpack/AlignedAllocator.h",
4670 ] + ACCURACY_EVAL_HDRS,
4671 deps = ACCURACY_EVAL_DEPS,
4672)
4673
Marat Dukhan98ba4412019-10-23 02:14:28 -07004674xnnpack_benchmark(
4675 name = "f32_extexp_eval",
4676 srcs = [
4677 "eval/f32-extexp.cc",
4678 "src/xnnpack/AlignedAllocator.h",
4679 ] + ACCURACY_EVAL_HDRS,
4680 deps = ACCURACY_EVAL_DEPS,
4681)
4682
Marat Dukhan8853b822020-05-07 12:19:01 -07004683xnnpack_unit_test(
4684 name = "f32_roundne_eval",
4685 srcs = [
4686 "eval/f32-roundne.cc",
4687 "src/xnnpack/AlignedAllocator.h",
4688 "src/xnnpack/math-stubs.h",
4689 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004690 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004691 deps = MICROKERNEL_TEST_DEPS,
4692)
4693
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004694xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004695 name = "f32_roundd_eval",
4696 srcs = [
4697 "eval/f32-roundd.cc",
4698 "src/xnnpack/AlignedAllocator.h",
4699 "src/xnnpack/math-stubs.h",
4700 ] + MICROKERNEL_TEST_HDRS,
4701 automatic = False,
4702 deps = MICROKERNEL_TEST_DEPS,
4703)
4704
4705xnnpack_unit_test(
4706 name = "f32_roundu_eval",
4707 srcs = [
4708 "eval/f32-roundu.cc",
4709 "src/xnnpack/AlignedAllocator.h",
4710 "src/xnnpack/math-stubs.h",
4711 ] + MICROKERNEL_TEST_HDRS,
4712 automatic = False,
4713 deps = MICROKERNEL_TEST_DEPS,
4714)
4715
4716xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004717 name = "f32_roundz_eval",
4718 srcs = [
4719 "eval/f32-roundz.cc",
4720 "src/xnnpack/AlignedAllocator.h",
4721 "src/xnnpack/math-stubs.h",
4722 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004723 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004724 deps = MICROKERNEL_TEST_DEPS,
4725)
4726
Marat Dukhan346a9e52019-11-15 09:06:30 -08004727xnnpack_benchmark(
4728 name = "f32_sigmoid_eval",
4729 srcs = [
4730 "eval/f32-sigmoid.cc",
4731 "src/xnnpack/AlignedAllocator.h",
4732 ] + ACCURACY_EVAL_HDRS,
4733 deps = ACCURACY_EVAL_DEPS,
4734)
4735
Marat Dukhan84000762020-06-29 18:38:43 -07004736xnnpack_benchmark(
4737 name = "f32_sqrt_eval",
4738 srcs = [
4739 "eval/f32-sqrt.cc",
4740 "src/xnnpack/AlignedAllocator.h",
4741 ] + ACCURACY_EVAL_HDRS,
4742 deps = ACCURACY_EVAL_DEPS + [
4743 ":bench_utils",
4744 ],
4745)
4746
Marat Dukhan08c4a432019-10-03 09:29:21 -07004747######################### Unit tests for micro-kernels #########################
4748
4749xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004750 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004751 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004752 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004753 "test/gemm-microkernel-tester.h",
4754 "src/xnnpack/AlignedAllocator.h",
4755 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004756 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004757)
4758
4759xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004760 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004761 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004762 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004763 "test/spmm-microkernel-tester.h",
4764 "src/xnnpack/AlignedAllocator.h",
4765 ] + MICROKERNEL_TEST_HDRS,
4766 deps = MICROKERNEL_TEST_DEPS,
4767)
4768
4769xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004770 name = "f32_argmaxpool_test",
4771 srcs = [
4772 "test/f32-argmaxpool.cc",
4773 "test/argmaxpool-microkernel-tester.h",
4774 "src/xnnpack/AlignedAllocator.h",
4775 ] + MICROKERNEL_TEST_HDRS,
4776 deps = MICROKERNEL_TEST_DEPS,
4777)
4778
4779xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004780 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004782 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004783 "test/avgpool-microkernel-tester.h",
4784 "src/xnnpack/AlignedAllocator.h",
4785 ] + MICROKERNEL_TEST_HDRS,
4786 deps = MICROKERNEL_TEST_DEPS,
4787)
4788
4789xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004790 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004791 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004792 "test/f32-ibilinear.cc",
4793 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004794 "src/xnnpack/AlignedAllocator.h",
4795 ] + MICROKERNEL_TEST_HDRS,
4796 deps = MICROKERNEL_TEST_DEPS,
4797)
4798
4799xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004800 name = "f16_clamp_test",
4801 srcs = [
4802 "test/f16-clamp.cc",
4803 "test/clamp-microkernel-tester.h",
4804 ] + MICROKERNEL_TEST_HDRS,
4805 deps = MICROKERNEL_TEST_DEPS,
4806)
4807
4808xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004809 name = "f32_clamp_test",
4810 srcs = [
4811 "test/f32-clamp.cc",
4812 "test/clamp-microkernel-tester.h",
4813 ] + MICROKERNEL_TEST_HDRS,
4814 deps = MICROKERNEL_TEST_DEPS,
4815)
4816
4817xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004818 name = "f32_igemm_test",
4819 srcs = [
4820 "test/f32-igemm.cc",
4821 "test/gemm-microkernel-tester.h",
4822 "src/xnnpack/AlignedAllocator.h",
4823 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004824 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004825)
4826
4827xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004828 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004829 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004830 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004831 "test/gemm-microkernel-tester.h",
4832 "src/xnnpack/AlignedAllocator.h",
4833 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004834 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004835)
4836
4837xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004838 name = "f32_igemm_minmax_test",
4839 srcs = [
4840 "test/f32-igemm-minmax.cc",
4841 "test/gemm-microkernel-tester.h",
4842 "src/xnnpack/AlignedAllocator.h",
4843 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004844 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004845)
4846
4847xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004848 name = "f16_igemm_minmax_test",
4849 srcs = [
4850 "test/f16-igemm-minmax.cc",
4851 "test/gemm-microkernel-tester.h",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004854 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004855)
4856
4857xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858 name = "f32_conv_hwc_test",
4859 srcs = [
4860 "test/f32-conv-hwc.cc",
4861 "test/conv-hwc-microkernel-tester.h",
4862 "src/xnnpack/AlignedAllocator.h",
4863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004864 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004865)
4866
4867xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004868 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004870 "test/f32-conv-hwc2chw.cc",
4871 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004872 "src/xnnpack/AlignedAllocator.h",
4873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004874 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004878 name = "f32_dwconv_test",
4879 srcs = [
4880 "test/f32-dwconv.cc",
4881 "test/dwconv-microkernel-tester.h",
4882 "src/xnnpack/AlignedAllocator.h",
4883 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004884 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004885)
4886
4887xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004888 name = "f16_dwconv_minmax_test",
4889 srcs = [
4890 "test/f16-dwconv-minmax.cc",
4891 "test/dwconv-microkernel-tester.h",
4892 "src/xnnpack/AlignedAllocator.h",
4893 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004894 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004895)
4896
4897xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004898 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004900 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 "test/dwconv-microkernel-tester.h",
4902 "src/xnnpack/AlignedAllocator.h",
4903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004904 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004905)
4906
4907xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004908 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004909 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004910 "test/f32-dwconv-chw.cc",
4911 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004912 "src/xnnpack/AlignedAllocator.h",
4913 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004914 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004915)
4916
4917xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004918 name = "f16_gavgpool_minmax_test",
4919 srcs = [
4920 "test/f16-gavgpool-minmax.cc",
4921 "test/gavgpool-microkernel-tester.h",
4922 "src/xnnpack/AlignedAllocator.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004928 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004929 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004930 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931 "test/gavgpool-microkernel-tester.h",
4932 "src/xnnpack/AlignedAllocator.h",
4933 ] + MICROKERNEL_TEST_HDRS,
4934 deps = MICROKERNEL_TEST_DEPS,
4935)
4936
4937xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004938 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004939 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004940 "test/f32-gavgpool-cw.cc",
4941 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004942 "src/xnnpack/AlignedAllocator.h",
4943 ] + MICROKERNEL_TEST_HDRS,
4944 deps = MICROKERNEL_TEST_DEPS,
4945)
4946
4947xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004948 name = "f32_gemm_test",
4949 srcs = [
4950 "test/f32-gemm.cc",
4951 "test/gemm-microkernel-tester.h",
4952 "src/xnnpack/AlignedAllocator.h",
4953 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004954 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004955)
4956
4957xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004958 name = "f32_gemm_relu_test",
4959 srcs = [
4960 "test/f32-gemm-relu.cc",
4961 "test/gemm-microkernel-tester.h",
4962 "src/xnnpack/AlignedAllocator.h",
4963 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004964 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004965)
4966
4967xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004968 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004970 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004971 "test/gemm-microkernel-tester.h",
4972 "src/xnnpack/AlignedAllocator.h",
4973 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004974 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004975)
4976
4977xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004978 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004979 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004980 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004981 "test/gemm-microkernel-tester.h",
4982 "src/xnnpack/AlignedAllocator.h",
4983 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004984 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985)
4986
4987xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004988 name = "f16_hswish_test",
4989 srcs = [
4990 "test/f16-hswish.cc",
4991 "test/hswish-microkernel-tester.h",
4992 ] + MICROKERNEL_TEST_HDRS,
4993 deps = MICROKERNEL_TEST_DEPS,
4994)
4995
4996xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004997 name = "f32_hswish_test",
4998 srcs = [
4999 "test/f32-hswish.cc",
5000 "test/hswish-microkernel-tester.h",
5001 ] + MICROKERNEL_TEST_HDRS,
5002 deps = MICROKERNEL_TEST_DEPS,
5003)
5004
5005xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005006 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005007 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005008 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 "test/maxpool-microkernel-tester.h",
5010 ] + MICROKERNEL_TEST_HDRS,
5011 deps = MICROKERNEL_TEST_DEPS,
5012)
5013
5014xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005015 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005016 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005017 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005018 "test/avgpool-microkernel-tester.h",
5019 "src/xnnpack/AlignedAllocator.h",
5020 ] + MICROKERNEL_TEST_HDRS,
5021 deps = MICROKERNEL_TEST_DEPS,
5022)
5023
5024xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005025 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005027 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005028 "test/gemm-microkernel-tester.h",
5029 "src/xnnpack/AlignedAllocator.h",
5030 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005031 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005032)
5033
5034xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005035 name = "f16_prelu_test",
5036 srcs = [
5037 "test/f16-prelu.cc",
5038 "test/prelu-microkernel-tester.h",
5039 "src/xnnpack/AlignedAllocator.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045 name = "f32_prelu_test",
5046 srcs = [
5047 "test/f32-prelu.cc",
5048 "test/prelu-microkernel-tester.h",
5049 "src/xnnpack/AlignedAllocator.h",
5050 ] + MICROKERNEL_TEST_HDRS,
5051 deps = MICROKERNEL_TEST_DEPS,
5052)
5053
5054xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005055 name = "f32_raddexpminusmax_test",
5056 srcs = [
5057 "test/f32-raddexpminusmax.cc",
5058 "test/raddexpminusmax-microkernel-tester.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005064 name = "f32_raddextexp_test",
5065 srcs = [
5066 "test/f32-raddextexp.cc",
5067 "test/raddextexp-microkernel-tester.h",
5068 ] + MICROKERNEL_TEST_HDRS,
5069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
5072xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005073 name = "f32_raddstoreexpminusmax_test",
5074 srcs = [
5075 "test/f32-raddstoreexpminusmax.cc",
5076 "test/raddstoreexpminusmax-microkernel-tester.h",
5077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005082 name = "f32_relu_test",
5083 srcs = [
5084 "test/f32-relu.cc",
5085 "test/vunary-microkernel-tester.h",
5086 ] + MICROKERNEL_TEST_HDRS,
5087 deps = MICROKERNEL_TEST_DEPS,
5088)
5089
5090xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005091 name = "f32_rmax_test",
5092 srcs = [
5093 "test/f32-rmax.cc",
5094 "test/rmax-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005100 name = "f32_sigmoid_test",
5101 srcs = [
5102 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005103 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005109 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005110 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005111 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005112 "test/spmm-microkernel-tester.h",
5113 "src/xnnpack/AlignedAllocator.h",
5114 ] + MICROKERNEL_TEST_HDRS,
5115 deps = MICROKERNEL_TEST_DEPS,
5116)
5117
5118xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005119 name = "f16_vadd_minmax_test",
5120 srcs = [
5121 "test/f16-vadd-minmax.cc",
5122 "test/vbinary-microkernel-tester.h",
5123 ] + MICROKERNEL_TEST_HDRS,
5124 deps = MICROKERNEL_TEST_DEPS,
5125)
5126
5127xnnpack_unit_test(
5128 name = "f16_vaddc_minmax_test",
5129 srcs = [
5130 "test/f16-vaddc-minmax.cc",
5131 "test/vbinaryc-microkernel-tester.h",
5132 ] + MICROKERNEL_TEST_HDRS,
5133 deps = MICROKERNEL_TEST_DEPS,
5134)
5135
5136xnnpack_unit_test(
5137 name = "f16_vdiv_minmax_test",
5138 srcs = [
5139 "test/f16-vdiv-minmax.cc",
5140 "test/vbinary-microkernel-tester.h",
5141 ] + MICROKERNEL_TEST_HDRS,
5142 deps = MICROKERNEL_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
5146 name = "f16_vdivc_minmax_test",
5147 srcs = [
5148 "test/f16-vdivc-minmax.cc",
5149 "test/vbinaryc-microkernel-tester.h",
5150 ] + MICROKERNEL_TEST_HDRS,
5151 deps = MICROKERNEL_TEST_DEPS,
5152)
5153
5154xnnpack_unit_test(
5155 name = "f16_vrdivc_minmax_test",
5156 srcs = [
5157 "test/f16-vrdivc-minmax.cc",
5158 "test/vbinaryc-microkernel-tester.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
5164 name = "f16_vmax_test",
5165 srcs = [
5166 "test/f16-vmax.cc",
5167 "test/vbinary-microkernel-tester.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
5173 name = "f16_vmaxc_test",
5174 srcs = [
5175 "test/f16-vmaxc.cc",
5176 "test/vbinaryc-microkernel-tester.h",
5177 ] + MICROKERNEL_TEST_HDRS,
5178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
5182 name = "f16_vmin_test",
5183 srcs = [
5184 "test/f16-vmin.cc",
5185 "test/vbinary-microkernel-tester.h",
5186 ] + MICROKERNEL_TEST_HDRS,
5187 deps = MICROKERNEL_TEST_DEPS,
5188)
5189
5190xnnpack_unit_test(
5191 name = "f16_vminc_test",
5192 srcs = [
5193 "test/f16-vminc.cc",
5194 "test/vbinaryc-microkernel-tester.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
5200 name = "f16_vmul_minmax_test",
5201 srcs = [
5202 "test/f16-vmul-minmax.cc",
5203 "test/vbinary-microkernel-tester.h",
5204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
5209 name = "f16_vmulc_minmax_test",
5210 srcs = [
5211 "test/f16-vmulc-minmax.cc",
5212 "test/vbinaryc-microkernel-tester.h",
5213 ] + MICROKERNEL_TEST_HDRS,
5214 deps = MICROKERNEL_TEST_DEPS,
5215)
5216
5217xnnpack_unit_test(
5218 name = "f16_vsub_minmax_test",
5219 srcs = [
5220 "test/f16-vsub-minmax.cc",
5221 "test/vbinary-microkernel-tester.h",
5222 ] + MICROKERNEL_TEST_HDRS,
5223 deps = MICROKERNEL_TEST_DEPS,
5224)
5225
5226xnnpack_unit_test(
5227 name = "f16_vsubc_minmax_test",
5228 srcs = [
5229 "test/f16-vsubc-minmax.cc",
5230 "test/vbinaryc-microkernel-tester.h",
5231 ] + MICROKERNEL_TEST_HDRS,
5232 deps = MICROKERNEL_TEST_DEPS,
5233)
5234
5235xnnpack_unit_test(
5236 name = "f16_vrsubc_minmax_test",
5237 srcs = [
5238 "test/f16-vrsubc-minmax.cc",
5239 "test/vbinaryc-microkernel-tester.h",
5240 ] + MICROKERNEL_TEST_HDRS,
5241 deps = MICROKERNEL_TEST_DEPS,
5242)
5243
5244xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005245 name = "f32_vabs_test",
5246 srcs = [
5247 "test/f32-vabs.cc",
5248 "test/vunary-microkernel-tester.h",
5249 ] + MICROKERNEL_TEST_HDRS,
5250 deps = MICROKERNEL_TEST_DEPS,
5251)
5252
5253xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005254 name = "f32_vadd_test",
5255 srcs = [
5256 "test/f32-vadd.cc",
5257 "test/vbinary-microkernel-tester.h",
5258 ] + MICROKERNEL_TEST_HDRS,
5259 deps = MICROKERNEL_TEST_DEPS,
5260)
5261
5262xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005263 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005264 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005265 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005266 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005267 ] + MICROKERNEL_TEST_HDRS,
5268 deps = MICROKERNEL_TEST_DEPS,
5269)
5270
5271xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005272 name = "f32_vadd_relu_test",
5273 srcs = [
5274 "test/f32-vadd-relu.cc",
5275 "test/vbinary-microkernel-tester.h",
5276 ] + MICROKERNEL_TEST_HDRS,
5277 deps = MICROKERNEL_TEST_DEPS,
5278)
5279
5280xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005281 name = "f32_vaddc_test",
5282 srcs = [
5283 "test/f32-vaddc.cc",
5284 "test/vbinaryc-microkernel-tester.h",
5285 ] + MICROKERNEL_TEST_HDRS,
5286 deps = MICROKERNEL_TEST_DEPS,
5287)
5288
5289xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005290 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005291 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005292 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005293 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005294 ] + MICROKERNEL_TEST_HDRS,
5295 deps = MICROKERNEL_TEST_DEPS,
5296)
5297
5298xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005299 name = "f32_vaddc_relu_test",
5300 srcs = [
5301 "test/f32-vaddc-relu.cc",
5302 "test/vbinaryc-microkernel-tester.h",
5303 ] + MICROKERNEL_TEST_HDRS,
5304 deps = MICROKERNEL_TEST_DEPS,
5305)
5306
5307xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005308 name = "f32_vdiv_test",
5309 srcs = [
5310 "test/f32-vdiv.cc",
5311 "test/vbinary-microkernel-tester.h",
5312 ] + MICROKERNEL_TEST_HDRS,
5313 deps = MICROKERNEL_TEST_DEPS,
5314)
5315
5316xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005317 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005318 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005319 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005320 "test/vbinary-microkernel-tester.h",
5321 ] + MICROKERNEL_TEST_HDRS,
5322 deps = MICROKERNEL_TEST_DEPS,
5323)
5324
5325xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005326 name = "f32_vdiv_relu_test",
5327 srcs = [
5328 "test/f32-vdiv-relu.cc",
5329 "test/vbinary-microkernel-tester.h",
5330 ] + MICROKERNEL_TEST_HDRS,
5331 deps = MICROKERNEL_TEST_DEPS,
5332)
5333
5334xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005335 name = "f32_vdivc_test",
5336 srcs = [
5337 "test/f32-vdivc.cc",
5338 "test/vbinaryc-microkernel-tester.h",
5339 ] + MICROKERNEL_TEST_HDRS,
5340 deps = MICROKERNEL_TEST_DEPS,
5341)
5342
5343xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005344 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005345 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005346 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005347 "test/vbinaryc-microkernel-tester.h",
5348 ] + MICROKERNEL_TEST_HDRS,
5349 deps = MICROKERNEL_TEST_DEPS,
5350)
5351
5352xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005353 name = "f32_vdivc_relu_test",
5354 srcs = [
5355 "test/f32-vdivc-relu.cc",
5356 "test/vbinaryc-microkernel-tester.h",
5357 ] + MICROKERNEL_TEST_HDRS,
5358 deps = MICROKERNEL_TEST_DEPS,
5359)
5360
5361xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005362 name = "f32_vrdivc_test",
5363 srcs = [
5364 "test/f32-vrdivc.cc",
5365 "test/vbinaryc-microkernel-tester.h",
5366 ] + MICROKERNEL_TEST_HDRS,
5367 deps = MICROKERNEL_TEST_DEPS,
5368)
5369
5370xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005371 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005372 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005373 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005374 "test/vbinaryc-microkernel-tester.h",
5375 ] + MICROKERNEL_TEST_HDRS,
5376 deps = MICROKERNEL_TEST_DEPS,
5377)
5378
5379xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005380 name = "f32_vrdivc_relu_test",
5381 srcs = [
5382 "test/f32-vrdivc-relu.cc",
5383 "test/vbinaryc-microkernel-tester.h",
5384 ] + MICROKERNEL_TEST_HDRS,
5385 deps = MICROKERNEL_TEST_DEPS,
5386)
5387
5388xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005389 name = "f32_vmax_test",
5390 srcs = [
5391 "test/f32-vmax.cc",
5392 "test/vbinary-microkernel-tester.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
5398 name = "f32_vmaxc_test",
5399 srcs = [
5400 "test/f32-vmaxc.cc",
5401 "test/vbinaryc-microkernel-tester.h",
5402 ] + MICROKERNEL_TEST_HDRS,
5403 deps = MICROKERNEL_TEST_DEPS,
5404)
5405
5406xnnpack_unit_test(
5407 name = "f32_vmin_test",
5408 srcs = [
5409 "test/f32-vmin.cc",
5410 "test/vbinary-microkernel-tester.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
5416 name = "f32_vminc_test",
5417 srcs = [
5418 "test/f32-vminc.cc",
5419 "test/vbinaryc-microkernel-tester.h",
5420 ] + MICROKERNEL_TEST_HDRS,
5421 deps = MICROKERNEL_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005425 name = "f32_vmul_test",
5426 srcs = [
5427 "test/f32-vmul.cc",
5428 "test/vbinary-microkernel-tester.h",
5429 ] + MICROKERNEL_TEST_HDRS,
5430 deps = MICROKERNEL_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005434 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005435 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005436 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005437 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005438 ] + MICROKERNEL_TEST_HDRS,
5439 deps = MICROKERNEL_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005443 name = "f32_vmul_relu_test",
5444 srcs = [
5445 "test/f32-vmul-relu.cc",
5446 "test/vbinary-microkernel-tester.h",
5447 ] + MICROKERNEL_TEST_HDRS,
5448 deps = MICROKERNEL_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005452 name = "f32_vmulc_test",
5453 srcs = [
5454 "test/f32-vmulc.cc",
5455 "test/vbinaryc-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005461 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005462 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005463 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005464 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005470 name = "f32_vmulc_relu_test",
5471 srcs = [
5472 "test/f32-vmulc-relu.cc",
5473 "test/vbinaryc-microkernel-tester.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005479 name = "f16_vmulcaddc_minmax_test",
5480 srcs = [
5481 "test/f16-vmulcaddc-minmax.cc",
5482 "test/vmulcaddc-microkernel-tester.h",
5483 "src/xnnpack/AlignedAllocator.h",
5484 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005485 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005486)
5487
5488xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005489 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005490 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005491 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005492 "test/vmulcaddc-microkernel-tester.h",
5493 "src/xnnpack/AlignedAllocator.h",
5494 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005495 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005496)
5497
5498xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005499 name = "f32_vlrelu_test",
5500 srcs = [
5501 "test/f32-vlrelu.cc",
5502 "test/vunary-microkernel-tester.h",
5503 ] + MICROKERNEL_TEST_HDRS,
5504 deps = MICROKERNEL_TEST_DEPS,
5505)
5506
5507xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005508 name = "f32_vneg_test",
5509 srcs = [
5510 "test/f32-vneg.cc",
5511 "test/vunary-microkernel-tester.h",
5512 ] + MICROKERNEL_TEST_HDRS,
5513 deps = MICROKERNEL_TEST_DEPS,
5514)
5515
5516xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005517 name = "f32_vrndne_test",
5518 srcs = [
5519 "test/f32-vrndne.cc",
5520 "test/vunary-microkernel-tester.h",
5521 ] + MICROKERNEL_TEST_HDRS,
5522 deps = MICROKERNEL_TEST_DEPS,
5523)
5524
5525xnnpack_unit_test(
5526 name = "f32_vrndz_test",
5527 srcs = [
5528 "test/f32-vrndz.cc",
5529 "test/vunary-microkernel-tester.h",
5530 ] + MICROKERNEL_TEST_HDRS,
5531 deps = MICROKERNEL_TEST_DEPS,
5532)
5533
5534xnnpack_unit_test(
5535 name = "f32_vrndu_test",
5536 srcs = [
5537 "test/f32-vrndu.cc",
5538 "test/vunary-microkernel-tester.h",
5539 ] + MICROKERNEL_TEST_HDRS,
5540 deps = MICROKERNEL_TEST_DEPS,
5541)
5542
5543xnnpack_unit_test(
5544 name = "f32_vrndd_test",
5545 srcs = [
5546 "test/f32-vrndd.cc",
5547 "test/vunary-microkernel-tester.h",
5548 ] + MICROKERNEL_TEST_HDRS,
5549 deps = MICROKERNEL_TEST_DEPS,
5550)
5551
5552xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005553 name = "f32_vscale_test",
5554 srcs = [
5555 "test/f32-vscale.cc",
5556 "test/vscale-microkernel-tester.h",
5557 ] + MICROKERNEL_TEST_HDRS,
5558 deps = MICROKERNEL_TEST_DEPS,
5559)
5560
5561xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005562 name = "f32_vscaleexpminusmax_test",
5563 srcs = [
5564 "test/f32-vscaleexpminusmax.cc",
5565 "test/vscaleexpminusmax-microkernel-tester.h",
5566 ] + MICROKERNEL_TEST_HDRS,
5567 deps = MICROKERNEL_TEST_DEPS,
5568)
5569
5570xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005571 name = "f32_vscaleextexp_test",
5572 srcs = [
5573 "test/f32-vscaleextexp.cc",
5574 "test/vscaleextexp-microkernel-tester.h",
5575 ] + MICROKERNEL_TEST_HDRS,
5576 deps = MICROKERNEL_TEST_DEPS,
5577)
5578
5579xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005580 name = "f32_vsqr_test",
5581 srcs = [
5582 "test/f32-vsqr.cc",
5583 "test/vunary-microkernel-tester.h",
5584 ] + MICROKERNEL_TEST_HDRS,
5585 deps = MICROKERNEL_TEST_DEPS,
5586)
5587
5588xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005589 name = "f32_vsqrdiff_test",
5590 srcs = [
5591 "test/f32-vsqrdiff.cc",
5592 "test/vbinary-microkernel-tester.h",
5593 ] + MICROKERNEL_TEST_HDRS,
5594 deps = MICROKERNEL_TEST_DEPS,
5595)
5596
5597xnnpack_unit_test(
5598 name = "f32_vsqrdiffc_test",
5599 srcs = [
5600 "test/f32-vsqrdiffc.cc",
5601 "test/vbinaryc-microkernel-tester.h",
5602 ] + MICROKERNEL_TEST_HDRS,
5603 deps = MICROKERNEL_TEST_DEPS,
5604)
5605
5606xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005607 name = "f32_vsqrt_test",
5608 srcs = [
5609 "test/f32-vsqrt.cc",
5610 "test/vunary-microkernel-tester.h",
5611 ] + MICROKERNEL_TEST_HDRS,
5612 deps = MICROKERNEL_TEST_DEPS,
5613)
5614
5615xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005616 name = "f32_vsub_test",
5617 srcs = [
5618 "test/f32-vsub.cc",
5619 "test/vbinary-microkernel-tester.h",
5620 ] + MICROKERNEL_TEST_HDRS,
5621 deps = MICROKERNEL_TEST_DEPS,
5622)
5623
5624xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005625 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005626 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005627 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005628 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005629 ] + MICROKERNEL_TEST_HDRS,
5630 deps = MICROKERNEL_TEST_DEPS,
5631)
5632
5633xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005634 name = "f32_vsub_relu_test",
5635 srcs = [
5636 "test/f32-vsub-relu.cc",
5637 "test/vbinary-microkernel-tester.h",
5638 ] + MICROKERNEL_TEST_HDRS,
5639 deps = MICROKERNEL_TEST_DEPS,
5640)
5641
5642xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005643 name = "f32_vsubc_test",
5644 srcs = [
5645 "test/f32-vsubc.cc",
5646 "test/vbinaryc-microkernel-tester.h",
5647 ] + MICROKERNEL_TEST_HDRS,
5648 deps = MICROKERNEL_TEST_DEPS,
5649)
5650
5651xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005652 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005653 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005654 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005655 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005656 ] + MICROKERNEL_TEST_HDRS,
5657 deps = MICROKERNEL_TEST_DEPS,
5658)
5659
5660xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005661 name = "f32_vsubc_relu_test",
5662 srcs = [
5663 "test/f32-vsubc-relu.cc",
5664 "test/vbinaryc-microkernel-tester.h",
5665 ] + MICROKERNEL_TEST_HDRS,
5666 deps = MICROKERNEL_TEST_DEPS,
5667)
5668
5669xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005670 name = "f32_vrsubc_test",
5671 srcs = [
5672 "test/f32-vrsubc.cc",
5673 "test/vbinaryc-microkernel-tester.h",
5674 ] + MICROKERNEL_TEST_HDRS,
5675 deps = MICROKERNEL_TEST_DEPS,
5676)
5677
5678xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005679 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005680 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005681 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005682 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005683 ] + MICROKERNEL_TEST_HDRS,
5684 deps = MICROKERNEL_TEST_DEPS,
5685)
5686
5687xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005688 name = "f32_vrsubc_relu_test",
5689 srcs = [
5690 "test/f32-vrsubc-relu.cc",
5691 "test/vbinaryc-microkernel-tester.h",
5692 ] + MICROKERNEL_TEST_HDRS,
5693 deps = MICROKERNEL_TEST_DEPS,
5694)
5695
5696xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005697 name = "qs8_dwconv_minmax_test",
5698 srcs = [
5699 "test/qs8-dwconv-minmax.cc",
5700 "test/dwconv-microkernel-tester.h",
5701 "src/xnnpack/AlignedAllocator.h",
5702 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5703 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5704)
5705
5706xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005707 name = "qs8_gavgpool_minmax_test",
5708 srcs = [
5709 "test/qs8-gavgpool-minmax.cc",
5710 "test/gavgpool-microkernel-tester.h",
5711 "src/xnnpack/AlignedAllocator.h",
5712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005717 name = "qs8_gemm_minmax_test",
5718 srcs = [
5719 "test/qs8-gemm-minmax.cc",
5720 "test/gemm-microkernel-tester.h",
5721 "src/xnnpack/AlignedAllocator.h",
5722 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5723 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5724)
5725
5726xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005727 name = "qs8_igemm_minmax_test",
5728 srcs = [
5729 "test/qs8-igemm-minmax.cc",
5730 "test/gemm-microkernel-tester.h",
5731 "src/xnnpack/AlignedAllocator.h",
5732 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5733 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5734)
5735
5736xnnpack_unit_test(
5737 name = "qs8_requantization_test",
5738 srcs = [
5739 "src/xnnpack/requantization-stubs.h",
5740 "test/qs8-requantization.cc",
5741 "test/requantization-tester.h",
5742 ] + MICROKERNEL_TEST_HDRS,
5743 deps = MICROKERNEL_TEST_DEPS,
5744)
5745
5746xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005747 name = "qs8_vadd_minmax_test",
5748 srcs = [
5749 "test/qs8-vadd-minmax.cc",
5750 "test/vadd-microkernel-tester.h",
5751 ] + MICROKERNEL_TEST_HDRS,
5752 deps = MICROKERNEL_TEST_DEPS,
5753)
5754
5755xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005756 name = "qs8_vaddc_minmax_test",
5757 srcs = [
5758 "test/qs8-vaddc-minmax.cc",
5759 "test/vaddc-microkernel-tester.h",
5760 ] + MICROKERNEL_TEST_HDRS,
5761 deps = MICROKERNEL_TEST_DEPS,
5762)
5763
5764xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005765 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005766 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005767 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005768 "test/avgpool-microkernel-tester.h",
5769 "src/xnnpack/AlignedAllocator.h",
5770 ] + MICROKERNEL_TEST_HDRS,
5771 deps = MICROKERNEL_TEST_DEPS,
5772)
5773
5774xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005775 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005776 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005777 "test/qu8-dwconv-minmax.cc",
5778 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005779 "src/xnnpack/AlignedAllocator.h",
5780 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005781 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005782)
5783
5784xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005785 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005786 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005787 "test/qu8-igemm-minmax.cc",
5788 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005789 "src/xnnpack/AlignedAllocator.h",
5790 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005791 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005792)
5793
5794xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005795 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005796 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005797 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005798 "test/gavgpool-microkernel-tester.h",
5799 "src/xnnpack/AlignedAllocator.h",
5800 ] + MICROKERNEL_TEST_HDRS,
5801 deps = MICROKERNEL_TEST_DEPS,
5802)
5803
5804xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005805 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005806 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005807 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005808 "test/gemm-microkernel-tester.h",
5809 "src/xnnpack/AlignedAllocator.h",
5810 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005811 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005812)
5813
5814xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005815 name = "qu8_requantization_test",
5816 srcs = [
5817 "src/xnnpack/requantization-stubs.h",
5818 "test/qu8-requantization.cc",
5819 "test/requantization-tester.h",
5820 ] + MICROKERNEL_TEST_HDRS,
5821 deps = MICROKERNEL_TEST_DEPS,
5822)
5823
5824xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005825 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005826 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005827 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005828 "test/vadd-microkernel-tester.h",
5829 ] + MICROKERNEL_TEST_HDRS,
5830 deps = MICROKERNEL_TEST_DEPS,
5831)
5832
5833xnnpack_unit_test(
5834 name = "u8_clamp_test",
5835 srcs = [
5836 "test/u8-clamp.cc",
5837 "test/clamp-microkernel-tester.h",
5838 ] + MICROKERNEL_TEST_HDRS,
5839 deps = MICROKERNEL_TEST_DEPS,
5840)
5841
5842xnnpack_unit_test(
5843 name = "u8_lut32norm_test",
5844 srcs = [
5845 "test/u8-lut32norm.cc",
5846 "test/lut-norm-microkernel-tester.h",
5847 ] + MICROKERNEL_TEST_HDRS,
5848 deps = MICROKERNEL_TEST_DEPS,
5849)
5850
5851xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005852 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005853 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005854 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855 "test/maxpool-microkernel-tester.h",
5856 ] + MICROKERNEL_TEST_HDRS,
5857 deps = MICROKERNEL_TEST_DEPS,
5858)
5859
5860xnnpack_unit_test(
5861 name = "u8_rmax_test",
5862 srcs = [
5863 "test/u8-rmax.cc",
5864 "test/rmax-microkernel-tester.h",
5865 ] + MICROKERNEL_TEST_HDRS,
5866 deps = MICROKERNEL_TEST_DEPS,
5867)
5868
5869xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005870 name = "x32_fill_test",
5871 srcs = [
5872 "test/x32-fill.cc",
5873 "test/fill-microkernel-tester.h",
5874 ] + MICROKERNEL_TEST_HDRS,
5875 deps = MICROKERNEL_TEST_DEPS,
5876)
5877
5878xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005879 name = "x32_packx_test",
5880 srcs = [
5881 "test/x32-packx.cc",
5882 "test/pack-microkernel-tester.h",
5883 "src/xnnpack/AlignedAllocator.h",
5884 ] + MICROKERNEL_TEST_HDRS,
5885 deps = MICROKERNEL_TEST_DEPS,
5886)
5887
5888xnnpack_unit_test(
5889 name = "x32_pad_test",
5890 srcs = [
5891 "test/x32-pad.cc",
5892 "test/pad-microkernel-tester.h",
5893 ] + MICROKERNEL_TEST_HDRS,
5894 deps = MICROKERNEL_TEST_DEPS,
5895)
5896
5897xnnpack_unit_test(
5898 name = "x32_unpool_test",
5899 srcs = [
5900 "test/x32-unpool.cc",
5901 "test/unpool-microkernel-tester.h",
5902 ] + MICROKERNEL_TEST_HDRS,
5903 deps = MICROKERNEL_TEST_DEPS,
5904)
5905
5906xnnpack_unit_test(
5907 name = "x32_zip_test",
5908 srcs = [
5909 "test/x32-zip.cc",
5910 "test/zip-microkernel-tester.h",
5911 ] + MICROKERNEL_TEST_HDRS,
5912 deps = MICROKERNEL_TEST_DEPS,
5913)
5914
5915xnnpack_unit_test(
5916 name = "x8_lut_test",
5917 srcs = [
5918 "test/x8-lut.cc",
5919 "test/lut-microkernel-tester.h",
5920 ] + MICROKERNEL_TEST_HDRS,
5921 deps = MICROKERNEL_TEST_DEPS,
5922)
5923
5924xnnpack_unit_test(
5925 name = "x8_zip_test",
5926 srcs = [
5927 "test/x8-zip.cc",
5928 "test/zip-microkernel-tester.h",
5929 ] + MICROKERNEL_TEST_HDRS,
5930 deps = MICROKERNEL_TEST_DEPS,
5931)
5932
Marat Dukhan20c3b922020-03-10 03:45:06 -07005933########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934
5935xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005936 name = "operator_size_test",
5937 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005938 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005939)
5940
Marat Dukhan20c3b922020-03-10 03:45:06 -07005941xnnpack_binary(
5942 name = "subgraph_size_test",
5943 srcs = ["test/subgraph-size.c"],
5944 deps = [":XNNPACK"],
5945)
5946
5947########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005948
5949xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005950 name = "abs_nc_test",
5951 srcs = [
5952 "test/abs-nc.cc",
5953 "test/abs-operator-tester.h",
5954 ],
5955 deps = OPERATOR_TEST_DEPS,
5956)
5957
5958xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005959 name = "add_nd_test",
5960 srcs = [
5961 "test/add-nd.cc",
5962 "test/binary-elementwise-operator-tester.h",
5963 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005964 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005965)
5966
5967xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005968 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005969 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005970 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005971 "test/argmax-pooling-operator-tester.h",
5972 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005973 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005974)
5975
5976xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005977 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005978 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005979 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 "test/average-pooling-operator-tester.h",
5981 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005982 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005983)
5984
5985xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005986 name = "bankers_rounding_nc_test",
5987 srcs = [
5988 "test/bankers-rounding-nc.cc",
5989 "test/bankers-rounding-operator-tester.h",
5990 ],
5991 deps = OPERATOR_TEST_DEPS,
5992)
5993
5994xnnpack_unit_test(
5995 name = "ceiling_nc_test",
5996 srcs = [
5997 "test/ceiling-nc.cc",
5998 "test/ceiling-operator-tester.h",
5999 ],
6000 deps = OPERATOR_TEST_DEPS,
6001)
6002
6003xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006004 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006005 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006006 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006007 "test/channel-shuffle-operator-tester.h",
6008 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006009 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006010)
6011
6012xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006013 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006014 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006015 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006016 "test/clamp-operator-tester.h",
6017 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006018 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006019)
6020
6021xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006022 name = "constant_pad_nd_test",
6023 srcs = [
6024 "test/constant-pad-nd.cc",
6025 "test/constant-pad-operator-tester.h",
6026 ],
6027 deps = OPERATOR_TEST_DEPS,
6028)
6029
6030xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006031 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006032 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006033 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034 "test/convolution-operator-tester.h",
6035 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006036 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006037)
6038
6039xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006040 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006041 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006042 "test/convolution-nchw.cc",
6043 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006044 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006045 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006046)
6047
6048xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006049 name = "copy_nc_test",
6050 srcs = [
6051 "test/copy-nc.cc",
6052 "test/copy-operator-tester.h",
6053 ],
6054 deps = OPERATOR_TEST_DEPS,
6055)
6056
6057xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006058 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006059 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006060 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 "test/deconvolution-operator-tester.h",
6062 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006063 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006064)
6065
6066xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006067 name = "divide_nd_test",
6068 srcs = [
6069 "test/binary-elementwise-operator-tester.h",
6070 "test/divide-nd.cc",
6071 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006072 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006073)
6074
6075xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006076 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006077 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006078 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 "test/fully-connected-operator-tester.h",
6080 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006081 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082)
6083
6084xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006085 name = "floor_nc_test",
6086 srcs = [
6087 "test/floor-nc.cc",
6088 "test/floor-operator-tester.h",
6089 ],
6090 deps = OPERATOR_TEST_DEPS,
6091)
6092
6093xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006094 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006095 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006096 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006097 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006098 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006099 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006100)
6101
6102xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006103 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006104 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006105 "test/global-average-pooling-ncw.cc",
6106 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006107 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006108 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006109)
6110
6111xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006112 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006113 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006114 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006115 "test/hardswish-operator-tester.h",
6116 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006117 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006118)
6119
6120xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006121 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006122 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006123 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006124 "test/leaky-relu-operator-tester.h",
6125 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006126 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006127)
6128
6129xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006130 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006131 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006132 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006133 "test/max-pooling-operator-tester.h",
6134 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006135 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006136)
6137
6138xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006139 name = "maximum_nd_test",
6140 srcs = [
6141 "test/binary-elementwise-operator-tester.h",
6142 "test/maximum-nd.cc",
6143 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006144 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006145)
6146
6147xnnpack_unit_test(
6148 name = "minimum_nd_test",
6149 srcs = [
6150 "test/binary-elementwise-operator-tester.h",
6151 "test/minimum-nd.cc",
6152 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006153 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006154)
6155
6156xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006157 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006158 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006159 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006160 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006161 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006162 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006163)
6164
6165xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006166 name = "negate_nc_test",
6167 srcs = [
6168 "test/negate-nc.cc",
6169 "test/negate-operator-tester.h",
6170 ],
6171 deps = OPERATOR_TEST_DEPS,
6172)
6173
6174xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006175 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006176 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006177 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006178 "test/prelu-operator-tester.h",
6179 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006180 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006181)
6182
6183xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006184 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006185 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006186 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006187 "test/resize-bilinear-operator-tester.h",
6188 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006189 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006190)
6191
6192xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006193 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006194 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006195 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006196 "test/sigmoid-operator-tester.h",
6197 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006198 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006199)
6200
6201xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006202 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006203 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006204 "test/softmax-nc.cc",
6205 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006206 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006207 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006208)
6209
6210xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006211 name = "square_nc_test",
6212 srcs = [
6213 "test/square-nc.cc",
6214 "test/square-operator-tester.h",
6215 ],
6216 deps = OPERATOR_TEST_DEPS,
6217)
6218
6219xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006220 name = "square_root_nc_test",
6221 srcs = [
6222 "test/square-root-nc.cc",
6223 "test/square-root-operator-tester.h",
6224 ],
6225 deps = OPERATOR_TEST_DEPS,
6226)
6227
6228xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006229 name = "squared_difference_nd_test",
6230 srcs = [
6231 "test/binary-elementwise-operator-tester.h",
6232 "test/squared-difference-nd.cc",
6233 ],
6234 deps = OPERATOR_TEST_DEPS,
6235)
6236
6237xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006238 name = "subtract_nd_test",
6239 srcs = [
6240 "test/binary-elementwise-operator-tester.h",
6241 "test/subtract-nd.cc",
6242 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006243 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006244)
6245
6246xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006247 name = "truncation_nc_test",
6248 srcs = [
6249 "test/truncation-nc.cc",
6250 "test/truncation-operator-tester.h",
6251 ],
6252 deps = OPERATOR_TEST_DEPS,
6253)
6254
6255xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006256 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006257 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006258 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006259 "test/unpooling-operator-tester.h",
6260 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006261 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006262)
6263
Chao Mei6ddfc602020-05-13 22:29:36 -07006264############################### Misc unit tests ###############################
6265
6266xnnpack_unit_test(
6267 name = "memory_planner_test",
6268 srcs = [
6269 "test/memory-planner-test.cc",
6270 ],
6271 deps = [
6272 ":XNNPACK",
6273 ":memory_planner",
6274 ],
6275)
6276
Marat Dukhan08c4a432019-10-03 09:29:21 -07006277############################# Build configurations #############################
6278
Marat Dukhanb8642352019-10-30 15:43:02 -07006279# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006280config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006281 name = "xnn_enable_assembly_explicit_true",
6282 define_values = {"xnn_enable_assembly": "true"},
6283)
6284
6285# Disables usage of assembly kernels.
6286config_setting(
6287 name = "xnn_enable_assembly_explicit_false",
6288 define_values = {"xnn_enable_assembly": "false"},
6289)
6290
Marat Dukhan9de90e02020-06-18 16:04:12 -07006291# Enables usage of sparse inference.
6292config_setting(
6293 name = "xnn_enable_sparse_explicit_true",
6294 define_values = {"xnn_enable_sparse": "true"},
6295)
6296
6297# Disables usage of sparse inference.
6298config_setting(
6299 name = "xnn_enable_sparse_explicit_false",
6300 define_values = {"xnn_enable_sparse": "false"},
6301)
6302
Marat Dukhan05702cf2020-03-26 15:41:33 -07006303# Disables usage of HMP-aware optimizations.
6304config_setting(
6305 name = "xnn_enable_hmp_explicit_false",
6306 define_values = {"xnn_enable_hmp": "false"},
6307)
6308
Chao Mei6ddfc602020-05-13 22:29:36 -07006309# Enable usage of optimized memory allocation
6310config_setting(
6311 name = "xnn_enable_memopt_explicit_true",
6312 define_values = {"xnnpack_enable_memopt": "true"},
6313)
6314
6315# Disable usage of optimized memory allocation
6316config_setting(
6317 name = "xnn_enable_memopt_explicit_false",
6318 define_values = {"xnnpack_enable_memopt": "false"},
6319)
6320
Marat Dukhanb8642352019-10-30 15:43:02 -07006321# Builds with -c dbg
6322config_setting(
6323 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006324 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006325 "compilation_mode": "dbg",
6326 },
6327)
6328
6329# Builds with -c opt
6330config_setting(
6331 name = "optimized_build",
6332 values = {
6333 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006334 },
6335)
6336
6337config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006338 name = "linux_k8",
6339 values = {"cpu": "k8"},
6340)
6341
6342config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006343 name = "linux_arm",
6344 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006345)
6346
6347config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006348 name = "linux_armeabi",
6349 values = {"cpu": "armeabi"},
6350)
6351
6352config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006353 name = "linux_armhf",
6354 values = {"cpu": "armhf"},
6355)
6356
6357config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006358 name = "linux_armv7a",
6359 values = {"cpu": "armv7a"},
6360)
6361
6362config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006363 name = "linux_aarch64",
6364 values = {"cpu": "aarch64"},
6365)
6366
6367config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006368 name = "android",
6369 values = {"crosstool_top": "//external:android/crosstool"},
6370)
6371
6372config_setting(
6373 name = "android_armv7",
6374 values = {
6375 "crosstool_top": "//external:android/crosstool",
6376 "cpu": "armeabi-v7a",
6377 },
6378)
6379
6380config_setting(
6381 name = "android_arm64",
6382 values = {
6383 "crosstool_top": "//external:android/crosstool",
6384 "cpu": "arm64-v8a",
6385 },
6386)
6387
6388config_setting(
6389 name = "android_x86",
6390 values = {
6391 "crosstool_top": "//external:android/crosstool",
6392 "cpu": "x86",
6393 },
6394)
6395
6396config_setting(
6397 name = "android_x86_64",
6398 values = {
6399 "crosstool_top": "//external:android/crosstool",
6400 "cpu": "x86_64",
6401 },
6402)
6403
6404config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006405 name = "windows_x86_64",
6406 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006407)
6408
6409config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006410 name = "windows_x86_64_clang",
6411 values = {
6412 "compiler": "clang-cl",
6413 "cpu": "x64_windows",
6414 },
6415)
6416
6417config_setting(
6418 name = "windows_x86_64_mingw",
6419 values = {
6420 "compiler": "mingw-gcc",
6421 "cpu": "x64_windows",
6422 },
6423)
6424
6425config_setting(
6426 name = "windows_x86_64_msys",
6427 values = {
6428 "compiler": "msys-gcc",
6429 "cpu": "x64_windows",
6430 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006431)
6432
6433config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006434 name = "macos_x86_64",
6435 values = {
6436 "apple_platform_type": "macos",
6437 "cpu": "darwin",
6438 },
6439)
6440
6441config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006442 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006443 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006444)
6445
6446config_setting(
6447 name = "emscripten_wasm",
6448 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006449 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006450 "cpu": "wasm",
6451 },
6452)
6453
6454config_setting(
6455 name = "emscripten_wasmsimd",
6456 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006457 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006458 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006459 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006460 },
6461)
6462
6463config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006464 name = "ios_armv7",
6465 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006466 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006467 "cpu": "ios_armv7",
6468 },
6469)
6470
6471config_setting(
6472 name = "ios_arm64",
6473 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006474 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006475 "cpu": "ios_arm64",
6476 },
6477)
6478
6479config_setting(
6480 name = "ios_arm64e",
6481 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006482 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006483 "cpu": "ios_arm64e",
6484 },
6485)
6486
6487config_setting(
6488 name = "ios_x86",
6489 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006490 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006491 "cpu": "ios_i386",
6492 },
6493)
6494
6495config_setting(
6496 name = "ios_x86_64",
6497 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006498 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006499 "cpu": "ios_x86_64",
6500 },
6501)
6502
6503config_setting(
6504 name = "watchos_armv7k",
6505 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006506 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006507 "cpu": "watchos_armv7k",
6508 },
6509)
6510
6511config_setting(
6512 name = "watchos_arm64_32",
6513 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006514 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006515 "cpu": "watchos_arm64_32",
6516 },
6517)
6518
6519config_setting(
6520 name = "watchos_x86",
6521 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006522 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006523 "cpu": "watchos_i386",
6524 },
6525)
6526
6527config_setting(
6528 name = "watchos_x86_64",
6529 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006530 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006531 "cpu": "watchos_x86_64",
6532 },
6533)
6534
6535config_setting(
6536 name = "tvos_arm64",
6537 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006538 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006539 "cpu": "tvos_arm64",
6540 },
6541)
6542
6543config_setting(
6544 name = "tvos_x86_64",
6545 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006546 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006547 "cpu": "tvos_x86_64",
6548 },
6549)