blob: 496b6f5782c826f2d67913cfe6d00a7bfc146ccc [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",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800219 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
225 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700228 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/1x1-minmax-scalar.c",
230 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar.c",
232 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar.c",
234 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar.c",
236 "src/f32-spmm/gen/8x2-minmax-scalar.c",
237 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700238 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700241 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vadd-scalar-x1.c",
245 "src/f32-vbinary/gen/vadd-scalar-x2.c",
246 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700250 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
251 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
252 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700253 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700259 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
260 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
261 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700262 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
263 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
264 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
272 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
273 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800274 "src/f32-vbinary/gen/vmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vmax-scalar-x4.c",
277 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
278 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
279 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
280 "src/f32-vbinary/gen/vmin-scalar-x1.c",
281 "src/f32-vbinary/gen/vmin-scalar-x2.c",
282 "src/f32-vbinary/gen/vmin-scalar-x4.c",
283 "src/f32-vbinary/gen/vminc-scalar-x1.c",
284 "src/f32-vbinary/gen/vminc-scalar-x2.c",
285 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700289 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
290 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
291 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vmul-scalar-x1.c",
293 "src/f32-vbinary/gen/vmul-scalar-x2.c",
294 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700298 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
299 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
300 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700301 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
302 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
303 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
311 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
312 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700313 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
314 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
315 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700316 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
317 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
318 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700319 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
320 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
321 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700322 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
323 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
324 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
325 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
326 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
327 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700334 "src/f32-vbinary/gen/vsub-scalar-x1.c",
335 "src/f32-vbinary/gen/vsub-scalar-x2.c",
336 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700337 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
338 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
339 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700340 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
341 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
342 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700343 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
344 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
345 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700346 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
347 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
348 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700349 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
350 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
351 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700352 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
353 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
354 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
355 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
356 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
357 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
358 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
359 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
360 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700361 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
362 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
363 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700364 "src/f32-vunary/gen/vabs-scalar-x1.c",
365 "src/f32-vunary/gen/vabs-scalar-x2.c",
366 "src/f32-vunary/gen/vabs-scalar-x4.c",
367 "src/f32-vunary/gen/vneg-scalar-x1.c",
368 "src/f32-vunary/gen/vneg-scalar-x2.c",
369 "src/f32-vunary/gen/vneg-scalar-x4.c",
370 "src/f32-vunary/gen/vsqr-scalar-x1.c",
371 "src/f32-vunary/gen/vsqr-scalar-x2.c",
372 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800373 "src/math/expminus-scalar-lut2048-p1.c",
374 "src/math/expminus-scalar-lut64-p2.c",
375 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700376 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700377 "src/math/roundne-scalar-nearbyint.c",
378 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700379 "src/math/roundd-scalar-addsub.c",
380 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700381 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700382 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700383 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700384 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700385 "src/math/roundz-scalar-addsub.c",
386 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700387 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800388 "src/math/sigmoid-scalar-lut2048-p1-div.c",
389 "src/math/sigmoid-scalar-lut64-p2-div.c",
390 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700391 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
392 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700393 "src/qs8-requantization/fp32-scalar-lrintf.c",
394 "src/qs8-requantization/fp32-scalar-magic.c",
395 "src/qs8-requantization/precise-scalar-signed64.c",
396 "src/qs8-requantization/precise-scalar-unsigned32.c",
397 "src/qs8-requantization/precise-scalar-unsigned64.c",
398 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700399 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
400 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
401 "src/qu8-dwconv/up1x9-minmax-scalar.c",
402 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
403 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700404 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
405 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700406 "src/qu8-gemm/2x2-minmax-scalar.c",
407 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700408 "src/qu8-requantization/fp32-scalar-lrintf.c",
409 "src/qu8-requantization/fp32-scalar-magic.c",
410 "src/qu8-requantization/precise-scalar-signed64.c",
411 "src/qu8-requantization/precise-scalar-unsigned32.c",
412 "src/qu8-requantization/precise-scalar-unsigned64.c",
413 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700414 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700415 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700416 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700417 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700419 "src/x32-fill/scalar-float.c",
420 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x2-scalar.c",
422 "src/x32-packx/x3-scalar.c",
423 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700424 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700425 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700426 "src/x32-unpool/scalar.c",
427 "src/x32-zip/x2-scalar.c",
428 "src/x32-zip/x3-scalar.c",
429 "src/x32-zip/x4-scalar.c",
430 "src/x32-zip/xm-scalar.c",
431 "src/x8-lut/scalar.c",
432 "src/x8-zip/x2-scalar.c",
433 "src/x8-zip/x3-scalar.c",
434 "src/x8-zip/x4-scalar.c",
435 "src/x8-zip/xm-scalar.c",
436]
437
Marat Dukhan436ebe62019-12-04 15:10:12 -0800438WASM_UKERNELS = [
Frank Barchardd91624e2020-08-12 13:45:56 -0700439 "src/f32-relu/wasm.S",
Marat Dukhan99936602020-04-11 16:47:01 -0700440 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
441 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700442 "src/f32-clamp/gen/wasm-x1.c",
443 "src/f32-clamp/gen/wasm-x2.c",
444 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700445 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
446 "src/f32-dwconv/gen/up1x4-wasm.c",
447 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
448 "src/f32-dwconv/gen/up1x9-wasm.c",
449 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
450 "src/f32-dwconv/gen/up1x25-wasm.c",
451 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
452 "src/f32-dwconv/gen/up2x4-wasm.c",
453 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
454 "src/f32-dwconv/gen/up2x9-wasm.c",
455 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
456 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700457 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
458 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
459 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
460 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700461 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
462 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700463 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
464 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
465 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
466 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700467 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
468 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700469 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
470 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700471 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
472 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
473 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700474 "src/f32-gemm/gen/1x4-wasm.c",
475 "src/f32-gemm/gen/2x4-wasm.c",
476 "src/f32-gemm/gen/4x2-wasm.c",
477 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700478 "src/f32-gemm/gen/1x4-relu-wasm.c",
479 "src/f32-gemm/gen/2x4-relu-wasm.c",
480 "src/f32-gemm/gen/4x2-relu-wasm.c",
481 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700482 "src/f32-gemm/gen/1x4-minmax-wasm.c",
483 "src/f32-gemm/gen/2x4-minmax-wasm.c",
484 "src/f32-gemm/gen/4x2-minmax-wasm.c",
485 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700486 "src/f32-hswish/gen/hswish-wasm-x1.c",
487 "src/f32-hswish/gen/hswish-wasm-x2.c",
488 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700489 "src/f32-igemm/gen/1x4-wasm.c",
490 "src/f32-igemm/gen/2x4-wasm.c",
491 "src/f32-igemm/gen/4x2-wasm.c",
492 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700493 "src/f32-igemm/gen/1x4-relu-wasm.c",
494 "src/f32-igemm/gen/2x4-relu-wasm.c",
495 "src/f32-igemm/gen/4x2-relu-wasm.c",
496 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700497 "src/f32-igemm/gen/1x4-minmax-wasm.c",
498 "src/f32-igemm/gen/2x4-minmax-wasm.c",
499 "src/f32-igemm/gen/4x2-minmax-wasm.c",
500 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700501 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
502 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
503 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700504 "src/f32-prelu/gen/wasm-2x1.c",
505 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700506 "src/f32-relu/gen/wasm-x1.c",
507 "src/f32-relu/gen/wasm-x2.c",
508 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700509 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
510 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
511 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
512 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
513 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
514 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
515 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
516 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
517 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
518 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
519 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
520 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700521 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
522 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
523 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
524 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
525 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
526 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
527 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
528 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
529 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
530 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
531 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
532 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800533 "src/f32-vbinary/gen/vmax-wasm-x1.c",
534 "src/f32-vbinary/gen/vmax-wasm-x2.c",
535 "src/f32-vbinary/gen/vmax-wasm-x4.c",
536 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
537 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
538 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
539 "src/f32-vbinary/gen/vmin-wasm-x1.c",
540 "src/f32-vbinary/gen/vmin-wasm-x2.c",
541 "src/f32-vbinary/gen/vmin-wasm-x4.c",
542 "src/f32-vbinary/gen/vminc-wasm-x1.c",
543 "src/f32-vbinary/gen/vminc-wasm-x2.c",
544 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700545 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
546 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
547 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
548 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
549 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
550 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
551 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
552 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
553 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
554 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
557 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
558 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
559 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
560 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
561 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
562 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700563 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
564 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
565 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
566 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
569 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
570 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
571 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
572 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
573 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
574 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
575 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
576 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
577 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
578 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
579 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
580 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700581 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
582 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
583 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700584 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
585 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
586 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800587]
588
Marat Dukhan290055c2020-06-09 12:24:29 -0700589WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700590 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
591 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
592 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700593 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
594 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
595 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
596 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700597 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
598 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
599 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
600 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700601 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
602 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
603 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
604 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
605 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
606 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700607 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
608 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
609 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
610 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
611 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
612 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
613 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
614 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
615 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
616 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
617 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
618 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
619 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
620 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
621 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
622 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
623 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
624 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
625 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
626 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
627 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
628 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
629 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
630 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700631 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
632 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
633 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700634 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700635 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
636 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
637 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
638 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700639 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
640 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
641 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
642 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
643 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
644 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
645 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
646 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
647 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
648 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
649 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
650 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
651 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
652 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
653 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
654 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
655 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
656 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
657 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
658 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
659 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
660 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
661 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
662 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
663 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
664 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
665 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
666 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
667 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
668 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700669 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
670 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700671 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
672 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
673 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
674 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
675 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
676 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
677 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
678 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
679 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
680 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
681 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
682 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
683 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
684 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
685 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
686 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
687 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
688 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
689 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
690 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
691 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
692 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
693 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
694 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
695 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
696 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
697 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
698 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
699 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
700 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700701 "src/f32-ibilinear/gen/wasmsimd-c4.c",
702 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700703 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
704 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
705 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700706 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700707 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
708 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
709 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
710 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700711 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
712 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
713 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
714 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
715 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
716 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
717 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
718 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
719 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
720 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
721 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
722 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
723 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
724 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
725 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
726 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
727 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
728 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
729 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
730 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
731 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
732 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
733 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
734 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
735 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
736 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
737 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
738 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
739 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
740 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700741 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
742 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700743 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
744 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
745 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
746 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700747 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
748 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700749 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
750 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
751 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700752 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
753 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700754 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
755 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
756 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
757 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700758 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
759 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700760 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
761 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
762 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700763 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
764 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700765 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
766 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
767 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
768 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
769 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
770 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
771 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700772 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
773 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700774 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
775 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
776 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
777 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700778 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
779 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
780 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
781 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
782 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
783 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
784 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
785 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
786 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
787 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
788 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
789 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700790 "src/f32-relu/gen/wasmsimd-x4.c",
791 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700792 "src/f32-rmax/wasmsimd-arm.c",
793 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700794 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
795 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
796 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
797 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
798 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
799 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700800 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
801 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
802 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
803 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
804 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
805 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700806 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
807 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
808 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
809 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700810 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
811 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700812 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
813 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700814 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
815 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
816 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
817 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700818 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
819 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700820 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
821 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700822 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
823 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
824 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
825 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700826 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
827 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700828 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
829 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700830 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
831 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
832 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
833 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700834 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
835 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700836 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
837 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700838 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
839 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
840 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
841 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700842 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
843 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
844 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
845 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700846 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
847 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
848 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
849 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700850 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
851 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
852 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
853 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700854 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
855 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
856 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
857 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700858 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
859 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700860 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
861 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700862 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
863 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
864 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
865 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700866 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
867 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700868 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
869 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700870 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
871 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
872 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
873 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700876 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
877 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700878 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
879 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
880 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
881 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700882 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
883 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700884 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
885 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700886 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700888 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
889 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700890 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
891 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
892 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
893 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700894 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
895 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700896 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
897 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700898 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
899 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
900 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
901 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700902 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
903 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700904 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
905 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700906 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
907 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
908 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
909 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700910 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
911 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
912 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
913 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700914 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
915 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
916 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
917 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
918 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
919 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
920 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
921 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
922 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
923 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
924 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
925 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
926 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
927 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700928 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
929 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700930 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
931 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
932 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
933 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
934 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
935 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700936 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
937 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
938 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700939 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
940 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
941 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700942 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
943 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
944 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700945 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
946 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
947 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
948 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
949 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
950 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
951 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
952 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
953 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
954 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
955 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
956 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
957 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
958 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
959 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700960 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700961 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700962 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700963 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700964 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700965 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700966 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700967 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700968 "src/x32-zip/x2-wasmsimd.c",
969 "src/x32-zip/x3-wasmsimd.c",
970 "src/x32-zip/x4-wasmsimd.c",
971 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700972 "src/math/roundne-wasmsimd-addsub.c",
973 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700974 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700975 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700976 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700977 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700978 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700979 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
980 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700981]
982
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800983PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700984 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700985 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700986 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700987 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700988 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700989 "src/f32-spmm/gen/4x1-minmax-psimd.c",
990 "src/f32-spmm/gen/8x1-minmax-psimd.c",
991 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700992]
993
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800994PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700995 "src/qs8-requantization/precise-psimd.c",
996 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700997 "src/qu8-requantization/precise-psimd.c",
998 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800999]
1000
Marat Dukhan08c4a432019-10-03 09:29:21 -07001001# ISA-specific micro-kernels
1002NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001003 "src/f32-argmaxpool/4x-neon-c4.c",
1004 "src/f32-argmaxpool/9p8x-neon-c4.c",
1005 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001006 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1007 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001008 "src/f32-clamp/gen/neon-x4.c",
1009 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001010 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1011 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001012 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1013 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001014 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1015 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001016 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1017 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001018 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1019 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1020 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1021 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001022 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1023 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1024 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1025 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001026 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1027 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1028 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1029 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001030 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001031 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1032 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001033 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1034 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1035 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1036 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1038 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1039 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1040 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1041 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1042 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1044 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1045 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1046 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1047 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1048 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1049 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1050 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1051 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1052 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1053 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1054 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1055 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1056 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1057 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1058 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1059 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1060 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1061 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1062 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1063 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001064 "src/f32-hswish/gen/hswish-neon-x4.c",
1065 "src/f32-hswish/gen/hswish-neon-x8.c",
1066 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001067 "src/f32-ibilinear/gen/neon-c4.c",
1068 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001069 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1070 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1071 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1072 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1073 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1074 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1075 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1076 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1077 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1078 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1079 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1080 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1081 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1082 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1083 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1084 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001085 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1086 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1087 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001088 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1089 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001090 "src/f32-prelu/gen/neon-1x4.c",
1091 "src/f32-prelu/gen/neon-1x8.c",
1092 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001093 "src/f32-prelu/gen/neon-2x4.c",
1094 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001095 "src/f32-prelu/gen/neon-2x16.c",
1096 "src/f32-prelu/gen/neon-4x4.c",
1097 "src/f32-prelu/gen/neon-4x8.c",
1098 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001099 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1100 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1101 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1103 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1104 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1105 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1106 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1107 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1108 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1109 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1110 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1111 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1112 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1113 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1114 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1115 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1116 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1117 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1118 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1119 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1120 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1121 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1122 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001123 "src/f32-relu/gen/neon-x4.c",
1124 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001126 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001127 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1128 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1129 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1130 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1131 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1132 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1133 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1134 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1135 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1136 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1137 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1138 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1139 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1140 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1141 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1142 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1143 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1144 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001145 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1146 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1147 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1148 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001149 "src/f32-vbinary/gen/vmax-neon-x4.c",
1150 "src/f32-vbinary/gen/vmax-neon-x8.c",
1151 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1152 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1153 "src/f32-vbinary/gen/vmin-neon-x4.c",
1154 "src/f32-vbinary/gen/vmin-neon-x8.c",
1155 "src/f32-vbinary/gen/vminc-neon-x4.c",
1156 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001157 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1158 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1159 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1160 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1161 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1162 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001163 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1164 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1165 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1166 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001167 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1168 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1169 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1170 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001171 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1172 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001173 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1174 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001175 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1176 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1177 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1178 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1179 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1180 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1181 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1182 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001183 "src/f32-vunary/gen/vabs-neon-x4.c",
1184 "src/f32-vunary/gen/vabs-neon-x8.c",
1185 "src/f32-vunary/gen/vneg-neon-x4.c",
1186 "src/f32-vunary/gen/vneg-neon-x8.c",
1187 "src/f32-vunary/gen/vsqr-neon-x4.c",
1188 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001189 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1190 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1191 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1192 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001193 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1194 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1195 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1196 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1197 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1198 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1199 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1200 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001201 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1202 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1203 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1204 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1205 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1206 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1207 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1208 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001209 "src/qs8-requantization/precise-neon.c",
1210 "src/qs8-requantization/fp32-neon.c",
1211 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001212 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1213 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1214 "src/qu8-dwconv/up8x9-minmax-neon.c",
1215 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1216 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1217 "src/qu8-gemm/4x8-minmax-neon.c",
1218 "src/qu8-gemm/8x8-minmax-neon.c",
1219 "src/qu8-igemm/4x8-minmax-neon.c",
1220 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001221 "src/qu8-requantization/precise-neon.c",
1222 "src/qu8-requantization/fp32-neon.c",
1223 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001224 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001225 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001226 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001227 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001228 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001229 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001230 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001231 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001232 "src/x32-zip/x2-neon.c",
1233 "src/x32-zip/x3-neon.c",
1234 "src/x32-zip/x4-neon.c",
1235 "src/x32-zip/xm-neon.c",
1236 "src/x8-zip/x2-neon.c",
1237 "src/x8-zip/x3-neon.c",
1238 "src/x8-zip/x4-neon.c",
1239 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001240 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001241 "src/math/roundd-neon-addsub.c",
1242 "src/math/roundd-neon-cvt.c",
1243 "src/math/roundu-neon-addsub.c",
1244 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001245 "src/math/roundz-neon-addsub.c",
1246 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001247 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001248 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1249 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1250 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1251 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1252 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1253 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001254 "src/math/sqrt-neon-nr1rsqrts.c",
1255 "src/math/sqrt-neon-nr2rsqrts.c",
1256 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001257]
1258
1259NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001260 "src/f32-ibilinear/gen/neonfma-c4.c",
1261 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001262 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1263 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1264 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1265 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1266 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1267 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1268 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1269 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1270 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001271 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1272 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1273 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1274 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001275 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1276 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1277 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1278 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001279 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1280 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1281 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1282 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001283 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1284 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1285 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1286 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1287 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1288 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1289 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1290 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1291 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1292 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1293 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1294 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1295 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1296 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1297 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1298 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1299 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1300 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001301 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1302 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001303 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1304 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1305 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1306 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1307 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1308 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1309 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1310 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1311 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1312 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1313 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1314 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1315 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1316 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1317 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1318 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1319 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1320 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1321 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1322 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1323 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1324 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1325 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1326 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001327 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001381 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1382 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001383 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1384 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1385 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1386 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1387 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1388 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1389 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1390 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1391 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1392 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1393 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1394 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1395 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1396 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1397 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1398 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1399 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1400 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1401 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1402 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001403 "src/math/exp-neonfma-lut64-p2.c",
1404 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001405 "src/math/expminus-neonfma-lut2048-p1.c",
1406 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001407 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001408 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1409 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1410 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1411 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1412 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1413 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1414 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1415 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1416 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1417 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1418 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1419 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1420 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1421 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1422 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1423 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1424 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1425 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001426 "src/math/sqrt-neonfma-nr1fma.c",
1427 "src/math/sqrt-neonfma-nr2fma.c",
1428 "src/math/sqrt-neonfma-nr3fma.c",
1429 "src/math/sqrt-neonfma-nr2fma1adj.c",
1430 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431]
1432
1433AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001434 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1435 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1436 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1437 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1438 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1439 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001440 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1441 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1442 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1443 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1444 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1445 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1446 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1447 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1448 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1449 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1450 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1451 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1452 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1453 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1454 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1455 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1456 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1457 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1458 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1459 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001460 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1461 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001462 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1463 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001464 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1465 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001466 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1467 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001468 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1469 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1470 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1471 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1472 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001473 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001491 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1492 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1493 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1494 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1495 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1496 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1497 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1498 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1499 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1500 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1501 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1502 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1503 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1504 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1505 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1506 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1507 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1508 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001509 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1510 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001511 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1512 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1513 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1514 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1515 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1516 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001517]
1518
Marat Dukhan8853b822020-05-07 12:19:01 -07001519NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001520 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1521 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1522 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1523 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1524 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1525 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1526 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1527 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001528 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001529 "src/math/roundd-neonv8.c",
1530 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001531 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001532]
1533
Marat Dukhan08c4a432019-10-03 09:29:21 -07001534AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001535 "src/f16-clamp/gen/neonfp16arith-x16.c",
1536 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001537 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1538 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1539 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1540 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1541 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1542 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1543 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1544 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1545 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1546 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1547 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1548 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001549 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1550 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001551 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1552 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001553 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001554 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001555 "src/f16-relu/gen/neonfp16arith-x16.c",
1556 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001557 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1558 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1559 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1560 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1561 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1562 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1563 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1564 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1565 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1566 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1567 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1568 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1569 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1570 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1571 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1572 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1573 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1574 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1575 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1576 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1577 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1578 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1579 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1580 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1581 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1582 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1583 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1584 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001585 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1586 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001587 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1588 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1589 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1590 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1591 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1592 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1593 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1594 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001595 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1596 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1597 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1598 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001599 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1600 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1601 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1602 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1603 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1604 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1605 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1606 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1607 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1608 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1609 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1610 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001611 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1612 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1613 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1614 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1615 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1616 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1617 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1618 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001619]
1620
Benoit Jacoba9644732020-08-13 12:48:55 -07001621NEONDOT_UKERNELS = [
1622 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1623 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1624 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001625 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1626 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1627 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001628]
1629
Marat Dukhan08c4a432019-10-03 09:29:21 -07001630SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001631 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1632 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001633 "src/f32-clamp/gen/sse-x4.c",
1634 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001635 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1636 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001637 "src/f32-dwconv-chw/3x3p1-sse.c",
1638 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001639 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1640 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1641 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1642 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1643 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1644 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1645 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1646 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1647 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1648 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1649 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1650 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001651 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001652 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1653 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001654 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1655 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1656 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1657 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1658 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1659 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1660 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1661 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1662 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1663 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1664 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1665 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1666 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001667 "src/f32-hswish/gen/hswish-sse-x4.c",
1668 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001669 "src/f32-ibilinear/gen/sse-c4.c",
1670 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001671 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1672 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1673 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1674 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1675 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1676 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1677 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001678 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1679 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1680 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001681 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001682 "src/f32-prelu/gen/sse-2x4.c",
1683 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001684 "src/f32-relu/gen/sse-x4.c",
1685 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001686 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001687 "src/f32-spmm/gen/4x1-minmax-sse.c",
1688 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001689 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001690 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1691 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1692 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1693 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1694 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1695 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1696 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1697 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001698 "src/f32-vbinary/gen/vmax-sse-x4.c",
1699 "src/f32-vbinary/gen/vmax-sse-x8.c",
1700 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1701 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1702 "src/f32-vbinary/gen/vmin-sse-x4.c",
1703 "src/f32-vbinary/gen/vmin-sse-x8.c",
1704 "src/f32-vbinary/gen/vminc-sse-x4.c",
1705 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001706 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1707 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1708 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1709 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1710 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1711 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1712 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1713 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001714 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1715 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1716 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1717 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001718 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1719 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1720 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1721 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001722 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1723 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001724 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1725 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001726 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1727 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001728 "src/f32-vunary/gen/vabs-sse-x4.c",
1729 "src/f32-vunary/gen/vabs-sse-x8.c",
1730 "src/f32-vunary/gen/vneg-sse-x4.c",
1731 "src/f32-vunary/gen/vneg-sse-x8.c",
1732 "src/f32-vunary/gen/vsqr-sse-x4.c",
1733 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001734 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001735 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001736 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001737 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001738 "src/math/roundd-sse-addsub.c",
1739 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001740 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001741 "src/math/sqrt-sse-hh1mac.c",
1742 "src/math/sqrt-sse-nr1mac.c",
1743 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001744]
1745
1746SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001747 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1748 "src/f32-argmaxpool/4x-sse2-c4.c",
1749 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001750 "src/f32-prelu/gen/sse2-2x4.c",
1751 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001752 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1753 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1754 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1755 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1756 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1757 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1758 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1759 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1760 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1761 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1762 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1763 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001764 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001765 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001766 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001767 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001768 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1769 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001770 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1771 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001772 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1773 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1774 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1775 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1776 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1777 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1778 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1779 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001780 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1781 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1782 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001783 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1784 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1785 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001786 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1787 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1788 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001789 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1790 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001791 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1792 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001793 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1794 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001795 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1796 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001797 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001798 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1799 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001800 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001801 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1802 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1803 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001804 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1805 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001806 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1807 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001808 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1809 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001810 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001811 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1812 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001813 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001814 "src/qs8-requantization/precise-sse2.c",
1815 "src/qs8-requantization/fp32-sse2.c",
1816 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001817 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1818 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1819 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1820 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001821 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1822 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1823 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1824 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001825 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1826 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1827 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1828 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1829 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1830 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1831 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1832 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001833 "src/qu8-requantization/precise-sse2.c",
1834 "src/qu8-requantization/fp32-sse2.c",
1835 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001836 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001837 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001838 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001839 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001840 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/x32-zip/x2-sse2.c",
1842 "src/x32-zip/x3-sse2.c",
1843 "src/x32-zip/x4-sse2.c",
1844 "src/x32-zip/xm-sse2.c",
1845 "src/x8-zip/x2-sse2.c",
1846 "src/x8-zip/x3-sse2.c",
1847 "src/x8-zip/x4-sse2.c",
1848 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001849 "src/math/exp-sse2-p5.c",
1850 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001851 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001852 "src/math/roundd-sse2-cvt.c",
1853 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001854 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001855 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001856]
1857
1858SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001859 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1860 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1861 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001862 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1863 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1864 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001865 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1866 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1867 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001868 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1869 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001870 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1871 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001872 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1873 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001874 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1875 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001876 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001877 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1878 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001879 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001880 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1881 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1882 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001883 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1884 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001885 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1886 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001887 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1888 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001889 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001890 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1891 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001892 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001893 "src/qs8-requantization/precise-ssse3.c",
1894 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001895 "src/qu8-requantization/precise-ssse3.c",
1896 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001897]
1898
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001899SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001900 "src/f32-prelu/gen/sse41-2x4.c",
1901 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001902 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001903 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001904 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001905 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001906 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1907 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001908 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1909 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001910 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1911 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1912 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1913 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1914 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1915 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1916 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1917 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001918 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1919 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1920 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001921 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1922 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1923 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001924 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1925 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1926 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001927 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1928 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001929 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001931 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1932 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001933 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1934 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001935 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001936 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1937 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001938 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001939 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1940 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1941 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001942 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1943 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001944 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1945 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001946 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1947 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001948 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001949 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1950 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001951 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001952 "src/qs8-requantization/fp32-sse4.c",
1953 "src/qs8-requantization/precise-sse4.c",
1954 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001955 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1956 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1957 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1958 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001959 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
1960 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
1961 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
1962 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001963 "src/qu8-requantization/precise-sse4.c",
1964 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001965 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001966 "src/math/roundd-sse41.c",
1967 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001968 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001969]
1970
Marat Dukhan08c4a432019-10-03 09:29:21 -07001971AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001972 "src/f32-clamp/gen/avx-x8.c",
1973 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001974 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1975 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1976 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1977 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1978 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1979 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1980 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1981 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1982 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1983 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1984 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1985 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1986 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1987 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1988 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1989 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1990 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1991 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1992 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1993 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1994 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1995 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1996 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1997 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1998 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1999 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2000 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2001 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2002 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2003 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002004 "src/f32-hswish/gen/hswish-avx-x8.c",
2005 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002006 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2007 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2008 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2009 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2010 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2011 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2012 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2013 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2014 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002015 "src/f32-prelu/gen/avx-2x8.c",
2016 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002017 "src/f32-relu/gen/avx-x16.c",
2018 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002019 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002020 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2021 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2022 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2023 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2024 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2025 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2026 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2027 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002028 "src/f32-vbinary/gen/vmax-avx-x8.c",
2029 "src/f32-vbinary/gen/vmax-avx-x16.c",
2030 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2031 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2032 "src/f32-vbinary/gen/vmin-avx-x8.c",
2033 "src/f32-vbinary/gen/vmin-avx-x16.c",
2034 "src/f32-vbinary/gen/vminc-avx-x8.c",
2035 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002036 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2037 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2038 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2039 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2040 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2041 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2042 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2043 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002044 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2045 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2046 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2047 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002048 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2049 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2050 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2051 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002052 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2053 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002054 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002055 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2056 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2057 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2058 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2059 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2060 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2061 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2062 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002063 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2064 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002065 "src/f32-vunary/gen/vabs-avx-x8.c",
2066 "src/f32-vunary/gen/vabs-avx-x16.c",
2067 "src/f32-vunary/gen/vneg-avx-x8.c",
2068 "src/f32-vunary/gen/vneg-avx-x16.c",
2069 "src/f32-vunary/gen/vsqr-avx-x8.c",
2070 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002071]
2072
Marat Dukhan1566fee2020-08-02 21:55:41 -07002073XOP_UKERNELS = [
2074 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2075 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2076 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2077 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002078 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2079 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002080 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2081 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002082 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002083 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2084 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002085 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002086 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2087 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2088 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002089 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2090 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2091 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2092 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2093 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2094 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002095 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002096 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2097 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002098 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002099]
2100
Marat Dukhanfda12b82019-11-21 12:27:59 -08002101FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002102 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2103 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2104 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2105 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2106 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2107 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2108 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2109 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2110 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2111 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2112 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2113 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2114 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2115 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2116 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2117 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2118 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2119 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2120 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2121 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2122 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2123 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2124 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2125 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2126 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2127 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2128 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2129 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2130 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2131 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2132 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2133 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2134 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2135 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2136 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2137 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2138 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2139 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2140 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2141 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002142 "src/f32-hswish/gen/hswish-fma3-x8.c",
2143 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002144 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2145 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2146 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2147 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2148 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2149 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2150 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2151 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2152 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2153 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2154 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2155 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2156 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2157 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002158 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2159 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2160 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2161 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2162 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2163 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2164 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2165 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002166 "src/math/sqrt-fma3-nr1fma.c",
2167 "src/math/sqrt-fma3-nr2fma.c",
2168 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002169]
2170
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002171AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002172 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2173 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2174 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2175 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2176 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2177 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2178 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2179 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2180 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2181 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2182 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2183 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2184 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2185 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2186 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2187 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2188 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2189 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2190 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2191 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2192 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2193 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2194 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2195 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2196 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2197 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2201 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2202 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2203 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2204 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2205 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2206 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2207 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002208 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2209 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2210 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2211 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2212 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2213 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2214 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2215 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2216 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2217 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2218 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2219 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2220 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2221 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2222 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2223 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2224 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2225 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2226 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2227 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2228 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2229 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2230 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2231 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2232 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2233 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2234 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2235 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2236 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2237 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002238 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2239 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2247 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2248 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2249 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2250 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2251 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2252 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2253 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2254 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2255 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2256 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2257 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2258 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2259 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2260 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2261 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002262 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2263 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002264 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2265 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2266 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2267 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002268 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2269 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2270 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002271 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2272 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2273 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002274 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2275 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2276 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002277 "src/math/exp-avx2-p5.c",
2278 "src/math/exp-avx2-perm-p3.c",
2279 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002280 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002281 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002282 "src/math/sigmoid-avx2-rr2-p5-div.c",
2283 "src/math/sigmoid-avx2-rr1-p5-div.c",
2284 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2285 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2286 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2287 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002288]
2289
Marat Dukhan08c4a432019-10-03 09:29:21 -07002290AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002291 "src/f32-clamp/gen/avx512f-x16.c",
2292 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002293 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2294 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2295 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2296 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2297 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2298 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2299 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2300 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2301 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2302 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2303 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2304 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2305 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2306 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2307 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2308 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2309 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2310 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2311 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2312 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2313 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2314 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2315 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2316 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002317 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2318 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002319 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2320 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2321 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2322 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2323 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2324 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002325 "src/f32-prelu/gen/avx512f-2x16.c",
2326 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002327 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2328 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2329 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2330 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2331 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2332 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2333 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2334 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2335 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2336 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2337 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2338 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2339 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2340 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2341 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2342 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2343 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2344 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2345 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2346 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2347 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2348 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2349 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2350 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2351 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2352 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2353 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2354 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2355 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2356 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2357 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2358 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2359 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2360 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2361 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2362 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002363 "src/f32-relu/gen/avx512f-x16.c",
2364 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002365 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002366 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2367 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2368 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2369 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2370 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2371 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2372 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2373 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002374 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2375 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2376 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2377 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2378 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2379 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2380 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2381 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002382 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2383 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2384 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2385 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2386 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2387 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2388 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2389 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002390 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2391 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2392 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2393 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002394 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2395 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2396 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2397 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002398 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2399 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002400 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002401 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2402 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2403 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2404 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2405 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2406 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2407 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2408 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2409 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2410 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2411 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2412 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2413 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2414 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2415 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2416 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2417 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2418 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2419 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2420 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2421 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2422 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2423 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2424 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002425 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2426 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2427 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2428 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2429 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2430 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2431 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2432 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002433 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2434 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2435 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2436 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2437 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2438 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2439 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2440 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002441 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2442 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2443 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2444 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2445 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2446 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002447 "src/math/exp-avx512f-p5-scalef.c",
2448 "src/math/exp-avx512f-p5.c",
2449 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002450 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002451 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002452 "src/math/sqrt-avx512f-nr1fma.c",
2453 "src/math/sqrt-avx512f-nr2fma.c",
2454 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002455]
2456
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002457AVX512SKX_UKERNELS = [
2458 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2459 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2460 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2461 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2462 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2463 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2464 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2465 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2466]
2467
Marat Dukhan08c4a432019-10-03 09:29:21 -07002468AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002469 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002470 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002471 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2472 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002473 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002474 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002475 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002476 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002477 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2478 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002479 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2480 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2481 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2482 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002483]
2484
2485AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002486 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002487 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2488 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002489 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002490 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2491 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002492 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2493 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2494 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002495 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002496 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2497 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2498 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002499 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002500 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2501 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002502 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002503 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2504 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2505 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2506 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2507 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2508 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2509 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2510 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2511 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2512 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2513 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2514 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2515 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2516 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2517 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2518 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2519 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2520 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002521 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2522 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002523 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002524 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2525 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2526 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2527 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2528 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2529 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2530 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2531 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2532 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2533 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2534 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2535 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2536 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2537 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2538 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2539 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2540 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2541 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002542 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2543 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2544 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2545 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2546 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2547 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2548 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2549 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2550 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2551 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2552 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2553 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2554 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2555 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2556 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2557 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2558 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2559 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560]
2561
Marat Dukhan1b354632020-03-23 12:50:22 -07002562INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002563 "src/xnnpack/argmaxpool.h",
2564 "src/xnnpack/avgpool.h",
2565 "src/xnnpack/clamp.h",
2566 "src/xnnpack/common.h",
2567 "src/xnnpack/conv.h",
2568 "src/xnnpack/dwconv.h",
2569 "src/xnnpack/gavgpool.h",
2570 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002571 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002572 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002573 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002574 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002575 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002576 "src/xnnpack/lut.h",
2577 "src/xnnpack/math.h",
2578 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002579 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 "src/xnnpack/packx.h",
2581 "src/xnnpack/pad.h",
2582 "src/xnnpack/params.h",
2583 "src/xnnpack/pavgpool.h",
2584 "src/xnnpack/ppmm.h",
2585 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002586 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002587 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002588 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002589 "src/xnnpack/rmax.h",
2590 "src/xnnpack/scalar-utils.h",
2591 "src/xnnpack/spmm.h",
2592 "src/xnnpack/unpool.h",
2593 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002594 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002595 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002596 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002597 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002598 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002599 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002600 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002601]
2602
2603INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002604 "include/xnnpack.h",
2605 "src/xnnpack/allocator.h",
2606 "src/xnnpack/compute.h",
2607 "src/xnnpack/im2col.h",
2608 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002609 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002610 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611 "src/xnnpack/operator.h",
2612 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002613 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002614 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002615 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002616 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002617]
2618
Marat Dukhan1b354632020-03-23 12:50:22 -07002619ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002620 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002621]
2622
Marat Dukhan1b354632020-03-23 12:50:22 -07002623MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002624 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002625 "include/xnnpack.h",
2626]
2627
Marat Dukhan1b354632020-03-23 12:50:22 -07002628MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002629 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002630 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002631 "src/xnnpack/requantization.h",
2632 "include/xnnpack.h",
2633]
2634
2635OPERATOR_TEST_PARAMS_HDRS = [
2636 "src/xnnpack/params.h",
2637 "src/xnnpack/common.h",
2638]
2639
2640WEIGHTS_PACK_HDRS = [
2641 "src/xnnpack/pack.h",
2642 "src/xnnpack/operator.h",
2643 "src/xnnpack/compute.h",
2644]
2645
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002646LOGGING_COPTS = select({
2647 # No logging in optimized mode
2648 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2649 # Full logging in debug mode
2650 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2651 # Error-only logging in default (fastbuild) mode
2652 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2653})
2654
Marat Dukhan3b59de22020-06-03 20:15:19 -07002655LOGGING_SRCS = select({
2656 # No logging in optimized mode
2657 ":optimized_build": [],
2658 "//conditions:default": [
2659 "src/operator-strings.c",
2660 "src/subgraph-strings.c",
2661 ],
2662})
2663
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002664LOGGING_HDRS = [
2665 "src/xnnpack/log.h",
2666]
2667
Marat Dukhan08c4a432019-10-03 09:29:21 -07002668xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002669 name = "tables",
2670 srcs = TABLE_SRCS,
2671 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002672 gcc_copts = xnnpack_gcc_std_copts(),
2673 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002674)
2675
2676xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002677 name = "scalar_ukernels",
2678 srcs = SCALAR_UKERNELS,
2679 hdrs = INTERNAL_HDRS,
2680 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002681 gcc_copts = xnnpack_gcc_std_copts(),
2682 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002683 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002684 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 "@FP16",
2686 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002687 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002688 ],
2689)
2690
2691xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002692 name = "scalar_ukernels_test_mode",
2693 srcs = SCALAR_UKERNELS,
2694 hdrs = INTERNAL_HDRS,
2695 aarch32_copts = ["-marm"],
2696 copts = [
2697 "-UNDEBUG",
2698 "-DXNN_TEST_MODE=1",
2699 ],
2700 gcc_copts = xnnpack_gcc_std_copts(),
2701 msvc_copts = xnnpack_msvc_std_copts(),
2702 deps = [
2703 ":tables",
2704 "@FP16",
2705 "@FXdiv",
2706 "@pthreadpool",
2707 ],
2708)
2709
2710xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002711 name = "wasm_ukernels",
2712 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002713 gcc_copts = xnnpack_gcc_std_copts(),
2714 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002715 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002716 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002717 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002718 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002719 "@FP16",
2720 "@FXdiv",
2721 "@pthreadpool",
2722 ],
2723)
2724
2725xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002726 name = "wasm_ukernels_test_mode",
2727 hdrs = INTERNAL_HDRS,
2728 copts = [
2729 "-UNDEBUG",
2730 "-DXNN_TEST_MODE=1",
2731 ],
2732 gcc_copts = xnnpack_gcc_std_copts(),
2733 msvc_copts = xnnpack_msvc_std_copts(),
2734 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002735 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002736 deps = [
2737 ":tables",
2738 "@FP16",
2739 "@FXdiv",
2740 "@pthreadpool",
2741 ],
2742)
2743
2744xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002745 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002746 hdrs = INTERNAL_HDRS,
2747 aarch32_copts = [
2748 "-marm",
2749 "-mfpu=neon",
2750 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002751 gcc_copts = xnnpack_gcc_std_copts(),
2752 gcc_x86_copts = ["-msse2"],
2753 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002754 optimized_copts = [
2755 "-O3",
2756 "-ffast-math",
2757 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002758 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002760 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002761 "@FP16",
2762 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002763 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002764 ],
2765)
2766
2767xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002768 name = "psimd_fastmath_ukernels_test_mode",
2769 hdrs = INTERNAL_HDRS,
2770 aarch32_copts = [
2771 "-marm",
2772 "-mfpu=neon",
2773 ],
2774 copts = [
2775 "-UNDEBUG",
2776 "-DXNN_TEST_MODE=1",
2777 ],
2778 gcc_copts = xnnpack_gcc_std_copts(),
2779 gcc_x86_copts = ["-msse2"],
2780 msvc_copts = xnnpack_msvc_std_copts(),
2781 optimized_copts = [
2782 "-O3",
2783 "-ffast-math",
2784 ],
2785 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2786 deps = [
2787 ":tables",
2788 "@FP16",
2789 "@psimd",
2790 "@pthreadpool",
2791 ],
2792)
2793
2794xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002795 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002796 hdrs = INTERNAL_HDRS,
2797 aarch32_copts = [
2798 "-marm",
2799 "-mfpu=neon",
2800 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002801 gcc_copts = xnnpack_gcc_std_copts(),
2802 gcc_x86_copts = ["-msse2"],
2803 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002804 optimized_copts = [
2805 "-O3",
2806 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002807 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002808 deps = [
2809 ":tables",
2810 "@FP16",
2811 "@psimd",
2812 "@pthreadpool",
2813 ],
2814)
2815
2816xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002817 name = "psimd_accmath_ukernels_test_mode",
2818 hdrs = INTERNAL_HDRS,
2819 aarch32_copts = [
2820 "-marm",
2821 "-mfpu=neon",
2822 ],
2823 copts = [
2824 "-UNDEBUG",
2825 "-DXNN_TEST_MODE=1",
2826 ],
2827 gcc_copts = xnnpack_gcc_std_copts(),
2828 gcc_x86_copts = ["-msse2"],
2829 msvc_copts = xnnpack_msvc_std_copts(),
2830 optimized_copts = [
2831 "-O3",
2832 ],
2833 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2834 deps = [
2835 ":tables",
2836 "@FP16",
2837 "@psimd",
2838 "@pthreadpool",
2839 ],
2840)
2841
2842xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002843 name = "neon_ukernels",
2844 hdrs = INTERNAL_HDRS,
2845 aarch32_copts = [
2846 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002847 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002848 "-mfpu=neon",
2849 ],
2850 aarch32_srcs = NEON_UKERNELS,
2851 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002852 gcc_copts = xnnpack_gcc_std_copts(),
2853 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002854 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002855 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002856 "@FP16",
2857 "@pthreadpool",
2858 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002859)
2860
2861xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002862 name = "neon_ukernels_test_mode",
2863 hdrs = INTERNAL_HDRS,
2864 aarch32_copts = [
2865 "-marm",
2866 "-march=armv7-a",
2867 "-mfpu=neon",
2868 ],
2869 aarch32_srcs = NEON_UKERNELS,
2870 aarch64_srcs = NEON_UKERNELS,
2871 copts = [
2872 "-UNDEBUG",
2873 "-DXNN_TEST_MODE=1",
2874 ],
2875 gcc_copts = xnnpack_gcc_std_copts(),
2876 msvc_copts = xnnpack_msvc_std_copts(),
2877 deps = [
2878 ":tables",
2879 "@FP16",
2880 "@pthreadpool",
2881 ],
2882)
2883
2884xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002885 name = "neonfma_ukernels",
2886 hdrs = INTERNAL_HDRS,
2887 aarch32_copts = [
2888 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002889 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002890 "-mfpu=neon-vfpv4",
2891 ],
2892 aarch32_srcs = NEONFMA_UKERNELS,
2893 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002894 apple_aarch32_copts = [
2895 "-mcpu=swift",
2896 "-mtune=generic",
2897 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002898 gcc_copts = xnnpack_gcc_std_copts(),
2899 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002900 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002901 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002902 "@FP16",
2903 "@pthreadpool",
2904 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002905)
2906
2907xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002908 name = "neonfma_ukernels_test_mode",
2909 hdrs = INTERNAL_HDRS,
2910 aarch32_copts = [
2911 "-marm",
2912 "-march=armv7-a",
2913 "-mfpu=neon-vfpv4",
2914 ],
2915 aarch32_srcs = NEONFMA_UKERNELS,
2916 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002917 apple_aarch32_copts = [
2918 "-mcpu=swift",
2919 "-mtune=generic",
2920 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002921 copts = [
2922 "-UNDEBUG",
2923 "-DXNN_TEST_MODE=1",
2924 ],
2925 gcc_copts = xnnpack_gcc_std_copts(),
2926 msvc_copts = xnnpack_msvc_std_copts(),
2927 deps = [
2928 ":tables",
2929 "@FP16",
2930 "@pthreadpool",
2931 ],
2932)
2933
2934xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002935 name = "neonv8_ukernels",
2936 hdrs = INTERNAL_HDRS,
2937 aarch32_copts = [
2938 "-marm",
2939 "-march=armv8-a",
2940 "-mfpu=neon-fp-armv8",
2941 ],
2942 aarch32_srcs = NEONV8_UKERNELS,
2943 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002944 apple_aarch32_copts = [
2945 "-mcpu=cyclone",
2946 "-mtune=generic",
2947 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002948 gcc_copts = xnnpack_gcc_std_copts(),
2949 msvc_copts = xnnpack_msvc_std_copts(),
2950 deps = [
2951 ":tables",
2952 "@FP16",
2953 "@pthreadpool",
2954 ],
2955)
2956
2957xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002958 name = "neonv8_ukernels_test_mode",
2959 hdrs = INTERNAL_HDRS,
2960 aarch32_copts = [
2961 "-marm",
2962 "-march=armv8-a",
2963 "-mfpu=neon-fp-armv8",
2964 ],
2965 aarch32_srcs = NEONV8_UKERNELS,
2966 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002967 apple_aarch32_copts = [
2968 "-mcpu=cyclone",
2969 "-mtune=generic",
2970 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002971 copts = [
2972 "-UNDEBUG",
2973 "-DXNN_TEST_MODE=1",
2974 ],
2975 gcc_copts = xnnpack_gcc_std_copts(),
2976 msvc_copts = xnnpack_msvc_std_copts(),
2977 deps = [
2978 ":tables",
2979 "@FP16",
2980 "@pthreadpool",
2981 ],
2982)
2983
2984xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002985 name = "neonfp16arith_ukernels",
2986 hdrs = INTERNAL_HDRS,
2987 aarch64_copts = ["-march=armv8.2-a+fp16"],
2988 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002989 gcc_copts = xnnpack_gcc_std_copts(),
2990 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002991 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002992 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002993 "@FP16",
2994 "@pthreadpool",
2995 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002996)
2997
2998xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002999 name = "neonfp16arith_ukernels_test_mode",
3000 hdrs = INTERNAL_HDRS,
3001 aarch64_copts = ["-march=armv8.2-a+fp16"],
3002 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3003 copts = [
3004 "-UNDEBUG",
3005 "-DXNN_TEST_MODE=1",
3006 ],
3007 gcc_copts = xnnpack_gcc_std_copts(),
3008 msvc_copts = xnnpack_msvc_std_copts(),
3009 deps = [
3010 ":tables",
3011 "@FP16",
3012 "@pthreadpool",
3013 ],
3014)
3015
3016xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003017 name = "neondot_ukernels",
3018 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003019 aarch32_copts = [
3020 "-marm",
3021 "-march=armv8.2-a+dotprod",
3022 "-mfpu=neon-fp-armv8",
3023 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003024 aarch32_srcs = NEONDOT_UKERNELS,
3025 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3026 aarch64_srcs = NEONDOT_UKERNELS,
3027 gcc_copts = xnnpack_gcc_std_copts(),
3028 msvc_copts = xnnpack_msvc_std_copts(),
3029 deps = [
3030 ":tables",
3031 "@FP16",
3032 "@pthreadpool",
3033 ],
3034)
3035
3036xnnpack_cc_library(
3037 name = "neondot_ukernels_test_mode",
3038 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003039 aarch32_copts = [
3040 "-marm",
3041 "-march=armv8.2-a+dotprod",
3042 "-mfpu=neon-fp-armv8",
3043 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003044 aarch32_srcs = NEONDOT_UKERNELS,
3045 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3046 aarch64_srcs = NEONDOT_UKERNELS,
3047 copts = [
3048 "-UNDEBUG",
3049 "-DXNN_TEST_MODE=1",
3050 ],
3051 gcc_copts = xnnpack_gcc_std_copts(),
3052 msvc_copts = xnnpack_msvc_std_copts(),
3053 deps = [
3054 ":tables",
3055 "@FP16",
3056 "@pthreadpool",
3057 ],
3058)
3059
3060xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003061 name = "sse2_ukernels",
3062 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003063 gcc_copts = xnnpack_gcc_std_copts(),
3064 gcc_x86_copts = ["-msse2"],
3065 msvc_copts = xnnpack_msvc_std_copts(),
3066 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003068 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003069 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003070 "@FP16",
3071 "@pthreadpool",
3072 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003073)
3074
3075xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003076 name = "sse2_ukernels_test_mode",
3077 hdrs = INTERNAL_HDRS,
3078 copts = [
3079 "-UNDEBUG",
3080 "-DXNN_TEST_MODE=1",
3081 ],
3082 gcc_copts = xnnpack_gcc_std_copts(),
3083 gcc_x86_copts = ["-msse2"],
3084 msvc_copts = xnnpack_msvc_std_copts(),
3085 msvc_x86_32_copts = ["/arch:SSE2"],
3086 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3087 deps = [
3088 ":tables",
3089 "@FP16",
3090 "@pthreadpool",
3091 ],
3092)
3093
3094xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003095 name = "ssse3_ukernels",
3096 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003097 gcc_copts = xnnpack_gcc_std_copts(),
3098 gcc_x86_copts = ["-mssse3"],
3099 msvc_copts = xnnpack_msvc_std_copts(),
3100 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003101 x86_srcs = SSSE3_UKERNELS,
3102 deps = [
3103 ":tables",
3104 "@FP16",
3105 "@pthreadpool",
3106 ],
3107)
3108
3109xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003110 name = "ssse3_ukernels_test_mode",
3111 hdrs = INTERNAL_HDRS,
3112 copts = [
3113 "-UNDEBUG",
3114 "-DXNN_TEST_MODE=1",
3115 ],
3116 gcc_copts = xnnpack_gcc_std_copts(),
3117 gcc_x86_copts = ["-mssse3"],
3118 msvc_copts = xnnpack_msvc_std_copts(),
3119 msvc_x86_32_copts = ["/arch:SSE2"],
3120 x86_srcs = SSSE3_UKERNELS,
3121 deps = [
3122 ":tables",
3123 "@FP16",
3124 "@pthreadpool",
3125 ],
3126)
3127
3128xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003129 name = "sse41_ukernels",
3130 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003131 gcc_copts = xnnpack_gcc_std_copts(),
3132 gcc_x86_copts = ["-msse4.1"],
3133 msvc_copts = xnnpack_msvc_std_copts(),
3134 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003135 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003136 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003137 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003138 "@FP16",
3139 "@pthreadpool",
3140 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003141)
3142
3143xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003144 name = "sse41_ukernels_test_mode",
3145 hdrs = INTERNAL_HDRS,
3146 copts = [
3147 "-UNDEBUG",
3148 "-DXNN_TEST_MODE=1",
3149 ],
3150 gcc_copts = xnnpack_gcc_std_copts(),
3151 gcc_x86_copts = ["-msse4.1"],
3152 msvc_copts = xnnpack_msvc_std_copts(),
3153 msvc_x86_32_copts = ["/arch:SSE2"],
3154 x86_srcs = SSE41_UKERNELS,
3155 deps = [
3156 ":tables",
3157 "@FP16",
3158 "@pthreadpool",
3159 ],
3160)
3161
3162xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003163 name = "avx_ukernels",
3164 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003165 gcc_copts = xnnpack_gcc_std_copts(),
3166 gcc_x86_copts = ["-mavx"],
3167 msvc_copts = xnnpack_msvc_std_copts(),
3168 msvc_x86_32_copts = ["/arch:AVX"],
3169 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003170 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003171 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003172 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003173 "@FP16",
3174 "@pthreadpool",
3175 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003176)
3177
3178xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003179 name = "avx_ukernels_test_mode",
3180 hdrs = INTERNAL_HDRS,
3181 copts = [
3182 "-UNDEBUG",
3183 "-DXNN_TEST_MODE=1",
3184 ],
3185 gcc_copts = xnnpack_gcc_std_copts(),
3186 gcc_x86_copts = ["-mavx"],
3187 msvc_copts = xnnpack_msvc_std_copts(),
3188 msvc_x86_32_copts = ["/arch:AVX"],
3189 msvc_x86_64_copts = ["/arch:AVX"],
3190 x86_srcs = AVX_UKERNELS,
3191 deps = [
3192 ":tables",
3193 "@FP16",
3194 "@pthreadpool",
3195 ],
3196)
3197
3198xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003199 name = "xop_ukernels",
3200 hdrs = INTERNAL_HDRS,
3201 gcc_copts = xnnpack_gcc_std_copts(),
3202 gcc_x86_copts = ["-mxop"],
3203 msvc_copts = xnnpack_msvc_std_copts(),
3204 msvc_x86_32_copts = ["/arch:AVX"],
3205 msvc_x86_64_copts = ["/arch:AVX"],
3206 x86_srcs = XOP_UKERNELS,
3207 deps = [
3208 ":tables",
3209 "@FP16",
3210 "@pthreadpool",
3211 ],
3212)
3213
3214xnnpack_cc_library(
3215 name = "xop_ukernels_test_mode",
3216 hdrs = INTERNAL_HDRS,
3217 copts = [
3218 "-UNDEBUG",
3219 "-DXNN_TEST_MODE=1",
3220 ],
3221 gcc_copts = xnnpack_gcc_std_copts(),
3222 gcc_x86_copts = ["-mxop"],
3223 msvc_copts = xnnpack_msvc_std_copts(),
3224 msvc_x86_32_copts = ["/arch:AVX"],
3225 msvc_x86_64_copts = ["/arch:AVX"],
3226 x86_srcs = XOP_UKERNELS,
3227 deps = [
3228 ":tables",
3229 "@FP16",
3230 "@pthreadpool",
3231 ],
3232)
3233
3234xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003235 name = "fma3_ukernels",
3236 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003237 gcc_copts = xnnpack_gcc_std_copts(),
3238 gcc_x86_copts = ["-mfma"],
3239 msvc_copts = xnnpack_msvc_std_copts(),
3240 msvc_x86_32_copts = ["/arch:AVX"],
3241 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003242 x86_srcs = FMA3_UKERNELS,
3243 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003244 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003245 "@FP16",
3246 "@pthreadpool",
3247 ],
3248)
3249
3250xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003251 name = "fma3_ukernels_test_mode",
3252 hdrs = INTERNAL_HDRS,
3253 copts = [
3254 "-UNDEBUG",
3255 "-DXNN_TEST_MODE=1",
3256 ],
3257 gcc_copts = xnnpack_gcc_std_copts(),
3258 gcc_x86_copts = ["-mfma"],
3259 msvc_copts = xnnpack_msvc_std_copts(),
3260 msvc_x86_32_copts = ["/arch:AVX"],
3261 msvc_x86_64_copts = ["/arch:AVX"],
3262 x86_srcs = FMA3_UKERNELS,
3263 deps = [
3264 ":tables",
3265 "@FP16",
3266 "@pthreadpool",
3267 ],
3268)
3269
3270xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003271 name = "avx2_ukernels",
3272 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003273 gcc_copts = xnnpack_gcc_std_copts(),
3274 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003275 "-mfma",
3276 "-mavx2",
3277 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003278 msvc_copts = xnnpack_msvc_std_copts(),
3279 msvc_x86_32_copts = ["/arch:AVX2"],
3280 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003281 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003282 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003283 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003284 "@FP16",
3285 "@pthreadpool",
3286 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003287)
3288
3289xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003290 name = "avx2_ukernels_test_mode",
3291 hdrs = INTERNAL_HDRS,
3292 copts = [
3293 "-UNDEBUG",
3294 "-DXNN_TEST_MODE=1",
3295 ],
3296 gcc_copts = xnnpack_gcc_std_copts(),
3297 gcc_x86_copts = [
3298 "-mfma",
3299 "-mavx2",
3300 ],
3301 msvc_copts = xnnpack_msvc_std_copts(),
3302 msvc_x86_32_copts = ["/arch:AVX2"],
3303 msvc_x86_64_copts = ["/arch:AVX2"],
3304 x86_srcs = AVX2_UKERNELS,
3305 deps = [
3306 ":tables",
3307 "@FP16",
3308 "@pthreadpool",
3309 ],
3310)
3311
3312xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003313 name = "avx512f_ukernels",
3314 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003315 gcc_copts = xnnpack_gcc_std_copts(),
3316 gcc_x86_copts = ["-mavx512f"],
3317 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3318 msvc_copts = xnnpack_msvc_std_copts(),
3319 msvc_x86_32_copts = ["/arch:AVX512"],
3320 msvc_x86_64_copts = ["/arch:AVX512"],
3321 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003322 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003323 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003324 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003325 "@FP16",
3326 "@pthreadpool",
3327 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003328)
3329
3330xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003331 name = "avx512f_ukernels_test_mode",
3332 hdrs = INTERNAL_HDRS,
3333 copts = [
3334 "-UNDEBUG",
3335 "-DXNN_TEST_MODE=1",
3336 ],
3337 gcc_copts = xnnpack_gcc_std_copts(),
3338 gcc_x86_copts = ["-mavx512f"],
3339 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3340 msvc_copts = xnnpack_msvc_std_copts(),
3341 msvc_x86_32_copts = ["/arch:AVX512"],
3342 msvc_x86_64_copts = ["/arch:AVX512"],
3343 msys_copts = ["-fno-asynchronous-unwind-tables"],
3344 x86_srcs = AVX512F_UKERNELS,
3345 deps = [
3346 ":tables",
3347 "@FP16",
3348 "@pthreadpool",
3349 ],
3350)
3351
3352xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003353 name = "avx512skx_ukernels",
3354 hdrs = INTERNAL_HDRS,
3355 gcc_copts = xnnpack_gcc_std_copts(),
3356 gcc_x86_copts = [
3357 "-mavx512f",
3358 "-mavx512cd",
3359 "-mavx512bw",
3360 "-mavx512dq",
3361 "-mavx512vl",
3362 ],
3363 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3364 msvc_copts = xnnpack_msvc_std_copts(),
3365 msvc_x86_32_copts = ["/arch:AVX512"],
3366 msvc_x86_64_copts = ["/arch:AVX512"],
3367 msys_copts = ["-fno-asynchronous-unwind-tables"],
3368 x86_srcs = AVX512SKX_UKERNELS,
3369 deps = [
3370 ":tables",
3371 "@FP16",
3372 "@pthreadpool",
3373 ],
3374)
3375
3376xnnpack_cc_library(
3377 name = "avx512skx_ukernels_test_mode",
3378 hdrs = INTERNAL_HDRS,
3379 copts = [
3380 "-UNDEBUG",
3381 "-DXNN_TEST_MODE=1",
3382 ],
3383 gcc_copts = xnnpack_gcc_std_copts(),
3384 gcc_x86_copts = [
3385 "-mavx512f",
3386 "-mavx512cd",
3387 "-mavx512bw",
3388 "-mavx512dq",
3389 "-mavx512vl",
3390 ],
3391 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3392 msvc_copts = xnnpack_msvc_std_copts(),
3393 msvc_x86_32_copts = ["/arch:AVX512"],
3394 msvc_x86_64_copts = ["/arch:AVX512"],
3395 msys_copts = ["-fno-asynchronous-unwind-tables"],
3396 x86_srcs = AVX512SKX_UKERNELS,
3397 deps = [
3398 ":tables",
3399 "@FP16",
3400 "@pthreadpool",
3401 ],
3402)
3403
3404xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 name = "asm_ukernels",
3406 hdrs = ["src/xnnpack/assembly.h"],
3407 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003408 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003409 aarch64_srcs = AARCH64_ASM_UKERNELS,
3410)
3411
Marat Dukhan3b59de22020-06-03 20:15:19 -07003412xnnpack_cc_library(
3413 name = "logging_utils",
3414 srcs = LOGGING_SRCS,
3415 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3416 copts = LOGGING_COPTS + [
3417 "-Isrc",
3418 "-Iinclude",
3419 ] + select({
3420 ":debug_build": [],
3421 "//conditions:default": xnnpack_min_size_copts(),
3422 }),
3423 gcc_copts = xnnpack_gcc_std_copts(),
3424 msvc_copts = xnnpack_msvc_std_copts(),
3425 visibility = xnnpack_visibility(),
3426 deps = [
3427 "@FP16",
3428 "@clog",
3429 "@pthreadpool",
3430 ],
3431)
3432
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433xnnpack_aggregate_library(
3434 name = "ukernels",
3435 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003436 ":neon_ukernels",
3437 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003438 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003439 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440 ":asm_ukernels",
3441 ],
3442 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443 ":neon_ukernels",
3444 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003445 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003447 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003448 ":asm_ukernels",
3449 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003450 generic_deps = [
3451 ":scalar_ukernels",
3452 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003453 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003454 ":psimd_fastmath_ukernels",
3455 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003456 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003457 wasm_deps = [
3458 ":wasm_ukernels",
3459 ],
3460 wasmsimd_deps = [
3461 ":wasm_ukernels",
3462 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003463 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003465 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003466 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003467 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003468 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003469 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003470 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003471 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003472 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003473 ],
3474)
3475
Marat Dukhan33fcf782020-05-24 14:27:15 -07003476xnnpack_aggregate_library(
3477 name = "ukernels_test_mode",
3478 aarch32_deps = [
3479 ":neon_ukernels_test_mode",
3480 ":neonfma_ukernels_test_mode",
3481 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003482 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003483 ":asm_ukernels",
3484 ],
3485 aarch64_deps = [
3486 ":neon_ukernels_test_mode",
3487 ":neonfma_ukernels_test_mode",
3488 ":neonv8_ukernels_test_mode",
3489 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003490 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003491 ":asm_ukernels",
3492 ],
3493 generic_deps = [
3494 ":scalar_ukernels_test_mode",
3495 ],
3496 psimd_deps = [
3497 ":psimd_fastmath_ukernels_test_mode",
3498 ":psimd_accmath_ukernels_test_mode",
3499 ],
3500 wasm_deps = [
3501 ":wasm_ukernels_test_mode",
3502 ],
3503 wasmsimd_deps = [
3504 ":wasm_ukernels_test_mode",
3505 ],
3506 x86_deps = [
3507 ":sse2_ukernels_test_mode",
3508 ":ssse3_ukernels_test_mode",
3509 ":sse41_ukernels_test_mode",
3510 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003511 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003512 ":fma3_ukernels_test_mode",
3513 ":avx2_ukernels_test_mode",
3514 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003515 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003516 ],
3517)
3518
Marat Dukhan86df3842020-06-17 14:03:31 -07003519# This target must be compatible with Android NDK r17 and compile with Clang 6.
3520# It is similar to :ukernels target, but excludes microkernels which require
3521# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003522xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003523 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003524 aarch32_deps = [
3525 ":neon_ukernels",
3526 ":neonfma_ukernels",
3527 ":neonv8_ukernels",
3528 ":asm_ukernels",
3529 ],
3530 aarch64_deps = [
3531 ":neon_ukernels",
3532 ":neonfma_ukernels",
3533 ":neonv8_ukernels",
3534 ":asm_ukernels",
3535 ],
3536 generic_deps = [
3537 ":scalar_ukernels",
3538 ],
3539 psimd_deps = [
3540 ":psimd_fastmath_ukernels",
3541 ":psimd_accmath_ukernels",
3542 ],
3543 wasm_deps = [
3544 ":wasm_ukernels",
3545 ],
3546 wasmsimd_deps = [
3547 ":wasm_ukernels",
3548 ],
3549 x86_deps = [
3550 ":sse2_ukernels",
3551 ":ssse3_ukernels",
3552 ":sse41_ukernels",
3553 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003554 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003555 ":fma3_ukernels",
3556 ":avx2_ukernels",
3557 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003558 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003559 ],
3560)
3561
Marat Dukhan08c4a432019-10-03 09:29:21 -07003562xnnpack_cc_library(
3563 name = "im2col",
3564 srcs = ["src/im2col.c"],
3565 hdrs = [
3566 "src/xnnpack/common.h",
3567 "src/xnnpack/im2col.h",
3568 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003569 gcc_copts = xnnpack_gcc_std_copts(),
3570 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003571)
3572
3573xnnpack_cc_library(
3574 name = "indirection",
3575 srcs = ["src/indirection.c"],
3576 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003577 gcc_copts = xnnpack_gcc_std_copts(),
3578 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 deps = [
3580 "@FP16",
3581 "@FXdiv",
3582 "@pthreadpool",
3583 ],
3584)
3585
3586xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003587 name = "indirection_test_mode",
3588 srcs = ["src/indirection.c"],
3589 hdrs = INTERNAL_HDRS,
3590 copts = [
3591 "-UNDEBUG",
3592 "-DXNN_TEST_MODE=1",
3593 ],
3594 gcc_copts = xnnpack_gcc_std_copts(),
3595 msvc_copts = xnnpack_msvc_std_copts(),
3596 deps = [
3597 "@FP16",
3598 "@FXdiv",
3599 "@pthreadpool",
3600 ],
3601)
3602
3603xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003604 name = "packing",
3605 srcs = ["src/packing.c"],
3606 hdrs = INTERNAL_HDRS,
3607 gcc_copts = xnnpack_gcc_std_copts(),
3608 msvc_copts = xnnpack_msvc_std_copts(),
3609 deps = [
3610 "@FP16",
3611 "@FXdiv",
3612 "@pthreadpool",
3613 ],
3614)
3615
3616xnnpack_cc_library(
3617 name = "packing_test_mode",
3618 srcs = ["src/packing.c"],
3619 hdrs = INTERNAL_HDRS,
3620 copts = [
3621 "-UNDEBUG",
3622 "-DXNN_TEST_MODE=1",
3623 ],
3624 gcc_copts = xnnpack_gcc_std_copts(),
3625 msvc_copts = xnnpack_msvc_std_copts(),
3626 deps = [
3627 "@FP16",
3628 "@FXdiv",
3629 "@pthreadpool",
3630 ],
3631)
3632
3633xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003634 name = "operator_run",
3635 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003636 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003637 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003638 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3639 "//conditions:default": [],
3640 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003641 gcc_copts = xnnpack_gcc_std_copts(),
3642 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003643 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003644 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 "@FP16",
3646 "@FXdiv",
3647 "@clog",
3648 "@pthreadpool",
3649 ],
3650)
3651
Chao Mei6ddfc602020-05-13 22:29:36 -07003652xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003653 name = "operator_run_test_mode",
3654 srcs = ["src/operator-run.c"],
3655 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3656 copts = LOGGING_COPTS + [
3657 "-UNDEBUG",
3658 "-DXNN_TEST_MODE=1",
3659 ] + select({
3660 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3661 "//conditions:default": [],
3662 }),
3663 gcc_copts = xnnpack_gcc_std_copts(),
3664 msvc_copts = xnnpack_msvc_std_copts(),
3665 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003666 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003667 "@FP16",
3668 "@FXdiv",
3669 "@clog",
3670 "@pthreadpool",
3671 ],
3672)
3673
3674xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003675 name = "memory_planner",
3676 srcs = ["src/memory-planner.c"],
3677 hdrs = INTERNAL_HDRS,
3678 defines = select({
3679 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3680 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3681 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3682 }),
3683 gcc_copts = xnnpack_gcc_std_copts(),
3684 msvc_copts = xnnpack_msvc_std_copts(),
3685 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003686 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003687 "@pthreadpool",
3688 ],
3689)
3690
Marat Dukhan33fcf782020-05-24 14:27:15 -07003691xnnpack_cc_library(
3692 name = "memory_planner_test_mode",
3693 srcs = ["src/memory-planner.c"],
3694 hdrs = INTERNAL_HDRS,
3695 copts = [
3696 "-UNDEBUG",
3697 "-DXNN_TEST_MODE=1",
3698 ],
3699 defines = select({
3700 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3701 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3702 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3703 }),
3704 gcc_copts = xnnpack_gcc_std_copts(),
3705 msvc_copts = xnnpack_msvc_std_copts(),
3706 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003707 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003708 "@pthreadpool",
3709 ],
3710)
3711
Marat Dukhan08c4a432019-10-03 09:29:21 -07003712cc_library(
3713 name = "enable_assembly",
3714 defines = select({
3715 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3716 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003717 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003718 }),
3719)
3720
Marat Dukhan9de90e02020-06-18 16:04:12 -07003721cc_library(
3722 name = "enable_sparse",
3723 defines = select({
3724 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3725 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003726 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003727 }),
3728)
3729
Marat Dukhancf056b22019-10-07 10:26:29 -07003730xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003731 name = "operators",
3732 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003733 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003734 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003735 ],
3736 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003737 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003738 "-Isrc",
3739 "-Iinclude",
3740 ] + select({
3741 ":debug_build": [],
3742 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003743 }) + select({
3744 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3745 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003747 gcc_copts = xnnpack_gcc_std_copts(),
3748 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003749 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003751 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003752 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 "@FP16",
3754 "@FXdiv",
3755 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003756 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003757 ],
3758)
3759
Marat Dukhan10a38082020-04-17 03:58:35 -07003760xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003761 name = "operators_test_mode",
3762 srcs = OPERATOR_SRCS + [
3763 "src/memory.c",
3764 "src/operator-delete.c",
3765 ],
3766 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3767 copts = LOGGING_COPTS + [
3768 "-Isrc",
3769 "-Iinclude",
3770 "-UNDEBUG",
3771 "-DXNN_TEST_MODE=1",
3772 ] + select({
3773 ":debug_build": [],
3774 "//conditions:default": xnnpack_min_size_copts(),
3775 }) + select({
3776 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3777 "//conditions:default": [],
3778 }),
3779 gcc_copts = xnnpack_gcc_std_copts(),
3780 msvc_copts = xnnpack_msvc_std_copts(),
3781 deps = [
3782 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003783 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003784 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003785 "@FP16",
3786 "@FXdiv",
3787 "@clog",
3788 "@pthreadpool",
3789 ],
3790)
3791
3792xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003793 name = "XNNPACK",
3794 srcs = [
3795 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003796 "src/runtime.c",
3797 "src/subgraph.c",
3798 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003799 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003800 hdrs = ["include/xnnpack.h"],
3801 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003802 "-Isrc",
3803 "-Iinclude",
3804 ] + select({
3805 ":debug_build": [],
3806 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003807 }) + select({
3808 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3809 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003810 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003811 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003812 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003813 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003814 visibility = xnnpack_visibility(),
3815 deps = [
3816 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003817 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003818 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003819 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003820 ":operator_run",
3821 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003822 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003823 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003824 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003825 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003826 ] + select({
3827 ":emscripten": [],
3828 "//conditions:default": ["@cpuinfo"],
3829 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003830)
3831
Marat Dukhan10a38082020-04-17 03:58:35 -07003832xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003833 name = "XNNPACK_test_mode",
3834 srcs = [
3835 "src/init.c",
3836 "src/runtime.c",
3837 "src/subgraph.c",
3838 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003839 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003840 hdrs = ["include/xnnpack.h"],
3841 copts = LOGGING_COPTS + [
3842 "-Isrc",
3843 "-Iinclude",
3844 "-UNDEBUG",
3845 "-DXNN_TEST_MODE=1",
3846 ] + select({
3847 ":debug_build": [],
3848 "//conditions:default": xnnpack_min_size_copts(),
3849 }) + select({
3850 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3851 "//conditions:default": [],
3852 }),
3853 gcc_copts = xnnpack_gcc_std_copts(),
3854 includes = ["include"],
3855 msvc_copts = xnnpack_msvc_std_copts(),
3856 visibility = xnnpack_visibility(),
3857 deps = [
3858 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003859 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003860 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003861 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003862 ":operator_run_test_mode",
3863 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003864 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003865 "@clog",
3866 "@FP16",
3867 "@pthreadpool",
3868 ] + select({
3869 ":emscripten": [],
3870 "//conditions:default": ["@cpuinfo"],
3871 }),
3872)
3873
3874xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003875 name = "xnnpack_f32",
3876 srcs = [
3877 "src/init.c",
3878 "src/runtime.c",
3879 "src/subgraph.c",
3880 "src/tensor.c",
3881 ] + SUBGRAPH_SRCS,
3882 hdrs = ["include/xnnpack.h"],
3883 copts = LOGGING_COPTS + [
3884 "-Isrc",
3885 "-Iinclude",
3886 ] + select({
3887 ":debug_build": [],
3888 "//conditions:default": xnnpack_min_size_copts(),
3889 }) + select({
3890 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3891 "//conditions:default": [],
3892 }),
3893 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003894 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003895 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003896 "XNN_NO_U8_OPERATORS",
3897 "XNN_NO_X8_OPERATORS",
3898 "XNN_NO_F16_OPERATORS",
3899 "XNN_NO_X16_OPERATORS",
3900 ],
3901 gcc_copts = xnnpack_gcc_std_copts(),
3902 includes = ["include"],
3903 msvc_copts = xnnpack_msvc_std_copts(),
3904 visibility = xnnpack_visibility(),
3905 deps = [
3906 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003907 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003908 ":logging_utils",
3909 ":memory_planner",
3910 ":operator_run",
3911 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003912 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003913 "@clog",
3914 "@FP16",
3915 "@pthreadpool",
3916 ] + select({
3917 ":emscripten": [],
3918 "//conditions:default": ["@cpuinfo"],
3919 }),
3920)
3921
3922xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003923 name = "xnnpack_operators_nhwc_f32",
3924 srcs = [
3925 "src/init.c",
3926 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003927 hdrs = ["include/xnnpack.h"],
3928 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003929 "-Isrc",
3930 "-Iinclude",
3931 ] + select({
3932 ":debug_build": [],
3933 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003934 }) + select({
3935 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3936 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003937 }),
3938 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003939 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003940 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003941 "XNN_NO_U8_OPERATORS",
3942 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003943 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003944 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003945 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003946 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003947 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003948 visibility = xnnpack_visibility(),
3949 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003950 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003951 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003952 ":operator_run",
3953 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003954 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003955 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003956 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003957 ] + select({
3958 ":emscripten": [],
3959 "//conditions:default": ["@cpuinfo"],
3960 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961)
3962
Marat Dukhancf056b22019-10-07 10:26:29 -07003963xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003964 name = "bench_utils",
3965 srcs = ["bench/utils.cc"],
3966 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003967 deps = [
3968 "@com_google_benchmark//:benchmark",
3969 "@cpuinfo",
3970 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971)
3972
Frank Barchard7e955972019-10-11 10:34:25 -07003973######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003974
3975xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003976 name = "qs8_gemm_bench",
3977 srcs = [
3978 "bench/gemm.h",
3979 "bench/qs8-gemm.cc",
3980 "src/xnnpack/AlignedAllocator.h",
3981 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3982 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3983)
3984
3985xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003986 name = "qs8_requantization_bench",
3987 srcs = [
3988 "bench/qs8-requantization.cc",
3989 "src/xnnpack/requantization-stubs.h",
3990 "src/xnnpack/AlignedAllocator.h",
3991 ] + MICROKERNEL_BENCHMARK_HDRS,
3992 deps = MICROKERNEL_BENCHMARK_DEPS,
3993)
3994
3995xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003996 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997 srcs = [
3998 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003999 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004000 "src/xnnpack/AlignedAllocator.h",
4001 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004002 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004003 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004004)
4005
4006xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004007 name = "qu8_requantization_bench",
4008 srcs = [
4009 "bench/qu8-requantization.cc",
4010 "src/xnnpack/requantization-stubs.h",
4011 "src/xnnpack/AlignedAllocator.h",
4012 ] + MICROKERNEL_BENCHMARK_HDRS,
4013 deps = MICROKERNEL_BENCHMARK_DEPS,
4014)
4015
4016xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004017 name = "f16_igemm_bench",
4018 srcs = [
4019 "bench/f16-igemm.cc",
4020 "bench/conv.h",
4021 "bench/google/conv.h",
4022 "src/xnnpack/AlignedAllocator.h",
4023 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004024 deps = MICROKERNEL_BENCHMARK_DEPS + [
4025 ":indirection",
4026 ":packing",
4027 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004028)
4029
4030xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004031 name = "f16_gemm_bench",
4032 srcs = [
4033 "bench/f16-gemm.cc",
4034 "bench/gemm.h",
4035 "src/xnnpack/AlignedAllocator.h",
4036 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004037 deps = MICROKERNEL_BENCHMARK_DEPS + [
4038 ":packing",
4039 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040)
4041
4042xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004043 name = "f16_spmm_bench",
4044 srcs = [
4045 "bench/f16-spmm.cc",
4046 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004047 "src/xnnpack/AlignedAllocator.h",
4048 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004049 deps = MICROKERNEL_BENCHMARK_DEPS,
4050)
4051
4052xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004053 name = "f32_igemm_bench",
4054 srcs = [
4055 "bench/f32-igemm.cc",
4056 "bench/conv.h",
4057 "src/xnnpack/AlignedAllocator.h",
4058 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004059 deps = MICROKERNEL_BENCHMARK_DEPS + [
4060 ":indirection",
4061 ":packing",
4062 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004063)
4064
4065xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004066 name = "f16_relu_bench",
4067 srcs = [
4068 "bench/f16-relu.cc",
4069 "src/xnnpack/AlignedAllocator.h",
4070 ] + MICROKERNEL_BENCHMARK_HDRS,
4071 deps = MICROKERNEL_BENCHMARK_DEPS,
4072)
4073
4074xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004075 name = "f32_conv_hwc_bench",
4076 srcs = [
4077 "bench/f32-conv-hwc.cc",
4078 "bench/dconv.h",
4079 "src/xnnpack/AlignedAllocator.h",
4080 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004081 deps = MICROKERNEL_BENCHMARK_DEPS + [
4082 ":packing",
4083 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004084)
4085
4086xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004087 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004088 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004089 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004090 "bench/dconv.h",
4091 "src/xnnpack/AlignedAllocator.h",
4092 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004093 deps = MICROKERNEL_BENCHMARK_DEPS + [
4094 ":packing",
4095 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004096)
4097
4098xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004099 name = "f16_dwconv_bench",
4100 srcs = [
4101 "bench/f16-dwconv.cc",
4102 "bench/dwconv.h",
4103 "bench/google/dwconv.h",
4104 "src/xnnpack/AlignedAllocator.h",
4105 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004106 deps = MICROKERNEL_BENCHMARK_DEPS + [
4107 ":indirection",
4108 ":packing",
4109 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004110)
4111
4112xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004113 name = "f32_dwconv_bench",
4114 srcs = [
4115 "bench/f32-dwconv.cc",
4116 "bench/dwconv.h",
4117 "src/xnnpack/AlignedAllocator.h",
4118 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004119 deps = MICROKERNEL_BENCHMARK_DEPS + [
4120 ":indirection",
4121 ":packing",
4122 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123)
4124
4125xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004126 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004128 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 "bench/dwconv.h",
4130 "src/xnnpack/AlignedAllocator.h",
4131 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004132 deps = MICROKERNEL_BENCHMARK_DEPS + [
4133 ":indirection",
4134 ":packing",
4135 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004136)
4137
4138xnnpack_benchmark(
4139 name = "f32_gemm_bench",
4140 srcs = [
4141 "bench/f32-gemm.cc",
4142 "bench/gemm.h",
4143 "src/xnnpack/AlignedAllocator.h",
4144 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004145 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004146 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147)
4148
4149xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004150 name = "f32_hswish_bench",
4151 srcs = [
4152 "bench/f32-hswish.cc",
4153 "src/xnnpack/AlignedAllocator.h",
4154 ] + MICROKERNEL_BENCHMARK_HDRS,
4155 deps = MICROKERNEL_BENCHMARK_DEPS,
4156)
4157
4158xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004159 name = "f32_raddexpminusmax_bench",
4160 srcs = [
4161 "bench/f32-raddexpminusmax.cc",
4162 "src/xnnpack/AlignedAllocator.h",
4163 ] + MICROKERNEL_BENCHMARK_HDRS,
4164 deps = MICROKERNEL_BENCHMARK_DEPS,
4165)
4166
4167xnnpack_benchmark(
4168 name = "f32_raddextexp_bench",
4169 srcs = [
4170 "bench/f32-raddextexp.cc",
4171 "src/xnnpack/AlignedAllocator.h",
4172 ] + MICROKERNEL_BENCHMARK_HDRS,
4173 deps = MICROKERNEL_BENCHMARK_DEPS,
4174)
4175
4176xnnpack_benchmark(
4177 name = "f32_raddstoreexpminusmax_bench",
4178 srcs = [
4179 "bench/f32-raddstoreexpminusmax.cc",
4180 "src/xnnpack/AlignedAllocator.h",
4181 ] + MICROKERNEL_BENCHMARK_HDRS,
4182 deps = MICROKERNEL_BENCHMARK_DEPS,
4183)
4184
4185xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004186 name = "f32_relu_bench",
4187 srcs = [
4188 "bench/f32-relu.cc",
4189 "src/xnnpack/AlignedAllocator.h",
4190 ] + MICROKERNEL_BENCHMARK_HDRS,
4191 deps = MICROKERNEL_BENCHMARK_DEPS,
4192)
4193
4194xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004195 name = "f32_rmax_bench",
4196 srcs = [
4197 "bench/f32-rmax.cc",
4198 "src/xnnpack/AlignedAllocator.h",
4199 ] + MICROKERNEL_BENCHMARK_HDRS,
4200 deps = MICROKERNEL_BENCHMARK_DEPS,
4201)
4202
4203xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004204 name = "f32_sigmoid_bench",
4205 srcs = [
4206 "bench/f32-sigmoid.cc",
4207 "src/xnnpack/AlignedAllocator.h",
4208 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004209 deps = MICROKERNEL_BENCHMARK_DEPS,
4210)
4211
4212xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004213 name = "f32_spmm_bench",
4214 srcs = [
4215 "bench/f32-spmm.cc",
4216 "bench/gemm.h",
4217 "src/xnnpack/AlignedAllocator.h",
4218 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004219 deps = MICROKERNEL_BENCHMARK_DEPS,
4220)
4221
4222xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004223 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004224 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004225 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004226 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004227 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004228 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004229)
4230
4231xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004232 name = "f32_vscaleexpminusmax_bench",
4233 srcs = [
4234 "bench/f32-vscaleexpminusmax.cc",
4235 "src/xnnpack/AlignedAllocator.h",
4236 ] + MICROKERNEL_BENCHMARK_HDRS,
4237 deps = MICROKERNEL_BENCHMARK_DEPS,
4238)
4239
4240xnnpack_benchmark(
4241 name = "f32_vscaleextexp_bench",
4242 srcs = [
4243 "bench/f32-vscaleextexp.cc",
4244 "src/xnnpack/AlignedAllocator.h",
4245 ] + MICROKERNEL_BENCHMARK_HDRS,
4246 deps = MICROKERNEL_BENCHMARK_DEPS,
4247)
4248
4249xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004250 name = "f32_vsqrt_bench",
4251 srcs = [
4252 "bench/f32-vsqrt.cc",
4253 "src/xnnpack/AlignedAllocator.h",
4254 ] + MICROKERNEL_BENCHMARK_HDRS,
4255 deps = MICROKERNEL_BENCHMARK_DEPS,
4256)
4257
4258xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004259 name = "f32_im2col_gemm_bench",
4260 srcs = [
4261 "bench/f32-im2col-gemm.cc",
4262 "bench/conv.h",
4263 "src/xnnpack/AlignedAllocator.h",
4264 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004265 deps = MICROKERNEL_BENCHMARK_DEPS + [
4266 ":im2col",
4267 ":packing",
4268 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269)
4270
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004271xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004272 name = "rounding_bench",
4273 srcs = [
4274 "bench/rounding.cc",
4275 "src/xnnpack/math-stubs.h",
4276 "src/xnnpack/AlignedAllocator.h",
4277 ] + MICROKERNEL_BENCHMARK_HDRS,
4278 deps = MICROKERNEL_BENCHMARK_DEPS,
4279)
4280
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281########################### Benchmarks for operators ###########################
4282
4283xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004284 name = "average_pooling_bench",
4285 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004286 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004287 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004288 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004289)
4290
4291xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004292 name = "bankers_rounding_bench",
4293 srcs = ["bench/bankers-rounding.cc"],
4294 copts = xnnpack_optional_tflite_copts(),
4295 tags = ["nowin32"],
4296 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4297)
4298
4299xnnpack_benchmark(
4300 name = "ceiling_bench",
4301 srcs = ["bench/ceiling.cc"],
4302 copts = xnnpack_optional_tflite_copts(),
4303 tags = ["nowin32"],
4304 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4305)
4306
4307xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308 name = "channel_shuffle_bench",
4309 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004310 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004311)
4312
4313xnnpack_benchmark(
4314 name = "convolution_bench",
4315 srcs = ["bench/convolution.cc"],
4316 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004317 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004318 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319)
4320
4321xnnpack_benchmark(
4322 name = "deconvolution_bench",
4323 srcs = ["bench/deconvolution.cc"],
4324 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004325 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004326 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004327)
4328
4329xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004330 name = "floor_bench",
4331 srcs = ["bench/floor.cc"],
4332 copts = xnnpack_optional_tflite_copts(),
4333 tags = ["nowin32"],
4334 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4335)
4336
4337xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338 name = "global_average_pooling_bench",
4339 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004340 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341)
4342
4343xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004344 name = "hardswish_bench",
4345 srcs = ["bench/hardswish.cc"],
4346 copts = xnnpack_optional_tflite_copts(),
4347 tags = ["nowin32"],
4348 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4349)
4350
4351xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352 name = "max_pooling_bench",
4353 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004354 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355)
4356
4357xnnpack_benchmark(
4358 name = "sigmoid_bench",
4359 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004360 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004361 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004362 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004363)
4364
4365xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004366 name = "prelu_bench",
4367 srcs = ["bench/prelu.cc"],
4368 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004369 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004370 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004371)
4372
4373xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004374 name = "softmax_bench",
4375 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004376 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004377 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004378 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004379)
4380
Marat Dukhan87727142020-06-24 15:24:10 -07004381xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004382 name = "square_root_bench",
4383 srcs = ["bench/square-root.cc"],
4384 copts = xnnpack_optional_tflite_copts(),
4385 tags = ["nowin32"],
4386 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4387)
4388
4389xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004390 name = "truncation_bench",
4391 srcs = ["bench/truncation.cc"],
4392 deps = OPERATOR_BENCHMARK_DEPS,
4393)
4394
Marat Dukhanc068bb62019-10-04 13:24:39 -07004395############################# End-to-end benchmarks ############################
4396
4397cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004398 name = "fp32_mobilenet_v1",
4399 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004400 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004401 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004402 linkstatic = True,
4403 deps = [
4404 ":XNNPACK",
4405 "@pthreadpool",
4406 ],
4407)
4408
4409cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004410 name = "fp16_mobilenet_v1",
4411 srcs = ["models/fp16-mobilenet-v1.cc"],
4412 hdrs = ["models/models.h"],
4413 copts = xnnpack_std_cxxopts(),
4414 linkstatic = True,
4415 deps = [
4416 ":XNNPACK",
4417 "@FP16",
4418 "@pthreadpool",
4419 ],
4420)
4421
4422cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004423 name = "qs8_mobilenet_v1",
4424 srcs = ["models/qs8-mobilenet-v1.cc"],
4425 hdrs = ["models/models.h"],
4426 copts = xnnpack_std_cxxopts(),
4427 linkstatic = True,
4428 deps = [
4429 ":XNNPACK",
4430 "@pthreadpool",
4431 ],
4432)
4433
4434cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004435 name = "fp32_mobilenet_v2",
4436 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004437 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004438 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004439 linkstatic = True,
4440 deps = [
4441 ":XNNPACK",
4442 "@pthreadpool",
4443 ],
4444)
4445
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004446cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004447 name = "fp16_mobilenet_v2",
4448 srcs = ["models/fp16-mobilenet-v2.cc"],
4449 hdrs = ["models/models.h"],
4450 copts = xnnpack_std_cxxopts(),
4451 linkstatic = True,
4452 deps = [
4453 ":XNNPACK",
4454 "@FP16",
4455 "@pthreadpool",
4456 ],
4457)
4458
4459cc_library(
4460 name = "fp32_mobilenet_v3_large",
4461 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004462 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004463 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004464 linkstatic = True,
4465 deps = [
4466 ":XNNPACK",
4467 "@pthreadpool",
4468 ],
4469)
4470
4471cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004472 name = "fp32_mobilenet_v3_small",
4473 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004474 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004475 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004476 linkstatic = True,
4477 deps = [
4478 ":XNNPACK",
4479 "@pthreadpool",
4480 ],
4481)
4482
Marat Dukhanc068bb62019-10-04 13:24:39 -07004483xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004484 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004485 srcs = [
4486 "bench/f32-dwconv-e2e.cc",
4487 "bench/end2end.h",
4488 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004489 deps = MICROKERNEL_BENCHMARK_DEPS + [
4490 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004491 ":fp32_mobilenet_v1",
4492 ":fp32_mobilenet_v2",
4493 ":fp32_mobilenet_v3_large",
4494 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004495 ],
4496)
4497
4498xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004499 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004500 srcs = [
4501 "bench/f32-gemm-e2e.cc",
4502 "bench/end2end.h",
4503 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004504 deps = MICROKERNEL_BENCHMARK_DEPS + [
4505 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004506 ":fp32_mobilenet_v1",
4507 ":fp32_mobilenet_v2",
4508 ":fp32_mobilenet_v3_large",
4509 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004510 ],
4511)
4512
4513xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004514 name = "end2end_bench",
4515 srcs = ["bench/end2end.cc"],
4516 deps = [
4517 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004518 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004519 ":fp16_mobilenet_v1",
4520 ":fp16_mobilenet_v2",
4521 ":fp32_mobilenet_v1",
4522 ":fp32_mobilenet_v2",
4523 ":fp32_mobilenet_v3_large",
4524 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004525 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004526 "@pthreadpool",
4527 ],
4528)
4529
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004530#################### Accuracy evaluation for math functions ####################
4531
4532xnnpack_benchmark(
4533 name = "f32_exp_eval",
4534 srcs = [
4535 "eval/f32-exp.cc",
4536 "src/xnnpack/AlignedAllocator.h",
4537 ] + ACCURACY_EVAL_HDRS,
4538 deps = ACCURACY_EVAL_DEPS,
4539)
4540
Marat Dukhan515c9772019-10-17 18:07:57 -07004541xnnpack_benchmark(
4542 name = "f32_expminus_eval",
4543 srcs = [
4544 "eval/f32-expminus.cc",
4545 "src/xnnpack/AlignedAllocator.h",
4546 ] + ACCURACY_EVAL_HDRS,
4547 deps = ACCURACY_EVAL_DEPS,
4548)
4549
Marat Dukhan98ba4412019-10-23 02:14:28 -07004550xnnpack_benchmark(
4551 name = "f32_extexp_eval",
4552 srcs = [
4553 "eval/f32-extexp.cc",
4554 "src/xnnpack/AlignedAllocator.h",
4555 ] + ACCURACY_EVAL_HDRS,
4556 deps = ACCURACY_EVAL_DEPS,
4557)
4558
Marat Dukhan8853b822020-05-07 12:19:01 -07004559xnnpack_unit_test(
4560 name = "f32_roundne_eval",
4561 srcs = [
4562 "eval/f32-roundne.cc",
4563 "src/xnnpack/AlignedAllocator.h",
4564 "src/xnnpack/math-stubs.h",
4565 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004566 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004567 deps = MICROKERNEL_TEST_DEPS,
4568)
4569
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004570xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004571 name = "f32_roundd_eval",
4572 srcs = [
4573 "eval/f32-roundd.cc",
4574 "src/xnnpack/AlignedAllocator.h",
4575 "src/xnnpack/math-stubs.h",
4576 ] + MICROKERNEL_TEST_HDRS,
4577 automatic = False,
4578 deps = MICROKERNEL_TEST_DEPS,
4579)
4580
4581xnnpack_unit_test(
4582 name = "f32_roundu_eval",
4583 srcs = [
4584 "eval/f32-roundu.cc",
4585 "src/xnnpack/AlignedAllocator.h",
4586 "src/xnnpack/math-stubs.h",
4587 ] + MICROKERNEL_TEST_HDRS,
4588 automatic = False,
4589 deps = MICROKERNEL_TEST_DEPS,
4590)
4591
4592xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004593 name = "f32_roundz_eval",
4594 srcs = [
4595 "eval/f32-roundz.cc",
4596 "src/xnnpack/AlignedAllocator.h",
4597 "src/xnnpack/math-stubs.h",
4598 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004599 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004600 deps = MICROKERNEL_TEST_DEPS,
4601)
4602
Marat Dukhan346a9e52019-11-15 09:06:30 -08004603xnnpack_benchmark(
4604 name = "f32_sigmoid_eval",
4605 srcs = [
4606 "eval/f32-sigmoid.cc",
4607 "src/xnnpack/AlignedAllocator.h",
4608 ] + ACCURACY_EVAL_HDRS,
4609 deps = ACCURACY_EVAL_DEPS,
4610)
4611
Marat Dukhan84000762020-06-29 18:38:43 -07004612xnnpack_benchmark(
4613 name = "f32_sqrt_eval",
4614 srcs = [
4615 "eval/f32-sqrt.cc",
4616 "src/xnnpack/AlignedAllocator.h",
4617 ] + ACCURACY_EVAL_HDRS,
4618 deps = ACCURACY_EVAL_DEPS + [
4619 ":bench_utils",
4620 ],
4621)
4622
Marat Dukhan08c4a432019-10-03 09:29:21 -07004623######################### Unit tests for micro-kernels #########################
4624
4625xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004626 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004627 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004628 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004629 "test/gemm-microkernel-tester.h",
4630 "src/xnnpack/AlignedAllocator.h",
4631 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004632 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004633)
4634
4635xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004636 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004637 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004638 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004639 "test/spmm-microkernel-tester.h",
4640 "src/xnnpack/AlignedAllocator.h",
4641 ] + MICROKERNEL_TEST_HDRS,
4642 deps = MICROKERNEL_TEST_DEPS,
4643)
4644
4645xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004646 name = "f32_argmaxpool_test",
4647 srcs = [
4648 "test/f32-argmaxpool.cc",
4649 "test/argmaxpool-microkernel-tester.h",
4650 "src/xnnpack/AlignedAllocator.h",
4651 ] + MICROKERNEL_TEST_HDRS,
4652 deps = MICROKERNEL_TEST_DEPS,
4653)
4654
4655xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004656 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004657 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004658 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004659 "test/avgpool-microkernel-tester.h",
4660 "src/xnnpack/AlignedAllocator.h",
4661 ] + MICROKERNEL_TEST_HDRS,
4662 deps = MICROKERNEL_TEST_DEPS,
4663)
4664
4665xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004666 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004667 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004668 "test/f32-ibilinear.cc",
4669 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004670 "src/xnnpack/AlignedAllocator.h",
4671 ] + MICROKERNEL_TEST_HDRS,
4672 deps = MICROKERNEL_TEST_DEPS,
4673)
4674
4675xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004676 name = "f16_clamp_test",
4677 srcs = [
4678 "test/f16-clamp.cc",
4679 "test/clamp-microkernel-tester.h",
4680 ] + MICROKERNEL_TEST_HDRS,
4681 deps = MICROKERNEL_TEST_DEPS,
4682)
4683
4684xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004685 name = "f32_clamp_test",
4686 srcs = [
4687 "test/f32-clamp.cc",
4688 "test/clamp-microkernel-tester.h",
4689 ] + MICROKERNEL_TEST_HDRS,
4690 deps = MICROKERNEL_TEST_DEPS,
4691)
4692
4693xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004694 name = "f32_igemm_test",
4695 srcs = [
4696 "test/f32-igemm.cc",
4697 "test/gemm-microkernel-tester.h",
4698 "src/xnnpack/AlignedAllocator.h",
4699 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004700 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004701)
4702
4703xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004704 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004705 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004706 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004707 "test/gemm-microkernel-tester.h",
4708 "src/xnnpack/AlignedAllocator.h",
4709 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004710 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004711)
4712
4713xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004714 name = "f32_igemm_minmax_test",
4715 srcs = [
4716 "test/f32-igemm-minmax.cc",
4717 "test/gemm-microkernel-tester.h",
4718 "src/xnnpack/AlignedAllocator.h",
4719 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004720 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004721)
4722
4723xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004724 name = "f16_igemm_minmax_test",
4725 srcs = [
4726 "test/f16-igemm-minmax.cc",
4727 "test/gemm-microkernel-tester.h",
4728 "src/xnnpack/AlignedAllocator.h",
4729 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004730 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004731)
4732
4733xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004734 name = "f32_conv_hwc_test",
4735 srcs = [
4736 "test/f32-conv-hwc.cc",
4737 "test/conv-hwc-microkernel-tester.h",
4738 "src/xnnpack/AlignedAllocator.h",
4739 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004740 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004741)
4742
4743xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004744 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004745 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004746 "test/f32-conv-hwc2chw.cc",
4747 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004748 "src/xnnpack/AlignedAllocator.h",
4749 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004750 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004751)
4752
4753xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004754 name = "f32_dwconv_test",
4755 srcs = [
4756 "test/f32-dwconv.cc",
4757 "test/dwconv-microkernel-tester.h",
4758 "src/xnnpack/AlignedAllocator.h",
4759 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004760 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004761)
4762
4763xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004764 name = "f16_dwconv_minmax_test",
4765 srcs = [
4766 "test/f16-dwconv-minmax.cc",
4767 "test/dwconv-microkernel-tester.h",
4768 "src/xnnpack/AlignedAllocator.h",
4769 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004770 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004771)
4772
4773xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004774 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004775 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004776 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004777 "test/dwconv-microkernel-tester.h",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004780 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004781)
4782
4783xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004784 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004786 "test/f32-dwconv-chw.cc",
4787 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004788 "src/xnnpack/AlignedAllocator.h",
4789 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004790 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004791)
4792
4793xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004794 name = "f16_gavgpool_minmax_test",
4795 srcs = [
4796 "test/f16-gavgpool-minmax.cc",
4797 "test/gavgpool-microkernel-tester.h",
4798 "src/xnnpack/AlignedAllocator.h",
4799 ] + MICROKERNEL_TEST_HDRS,
4800 deps = MICROKERNEL_TEST_DEPS,
4801)
4802
4803xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004804 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004805 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004806 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004807 "test/gavgpool-microkernel-tester.h",
4808 "src/xnnpack/AlignedAllocator.h",
4809 ] + MICROKERNEL_TEST_HDRS,
4810 deps = MICROKERNEL_TEST_DEPS,
4811)
4812
4813xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004814 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004815 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004816 "test/f32-gavgpool-cw.cc",
4817 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004818 "src/xnnpack/AlignedAllocator.h",
4819 ] + MICROKERNEL_TEST_HDRS,
4820 deps = MICROKERNEL_TEST_DEPS,
4821)
4822
4823xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004824 name = "f32_gemm_test",
4825 srcs = [
4826 "test/f32-gemm.cc",
4827 "test/gemm-microkernel-tester.h",
4828 "src/xnnpack/AlignedAllocator.h",
4829 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004830 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004831)
4832
4833xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004834 name = "f32_gemm_relu_test",
4835 srcs = [
4836 "test/f32-gemm-relu.cc",
4837 "test/gemm-microkernel-tester.h",
4838 "src/xnnpack/AlignedAllocator.h",
4839 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004840 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004841)
4842
4843xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004844 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004845 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004846 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004847 "test/gemm-microkernel-tester.h",
4848 "src/xnnpack/AlignedAllocator.h",
4849 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004850 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851)
4852
4853xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004854 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004855 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004856 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004857 "test/gemm-microkernel-tester.h",
4858 "src/xnnpack/AlignedAllocator.h",
4859 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004860 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861)
4862
4863xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004864 name = "f16_hswish_test",
4865 srcs = [
4866 "test/f16-hswish.cc",
4867 "test/hswish-microkernel-tester.h",
4868 ] + MICROKERNEL_TEST_HDRS,
4869 deps = MICROKERNEL_TEST_DEPS,
4870)
4871
4872xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004873 name = "f32_hswish_test",
4874 srcs = [
4875 "test/f32-hswish.cc",
4876 "test/hswish-microkernel-tester.h",
4877 ] + MICROKERNEL_TEST_HDRS,
4878 deps = MICROKERNEL_TEST_DEPS,
4879)
4880
4881xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004882 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004884 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004885 "test/maxpool-microkernel-tester.h",
4886 ] + MICROKERNEL_TEST_HDRS,
4887 deps = MICROKERNEL_TEST_DEPS,
4888)
4889
4890xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004891 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004892 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004893 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004894 "test/avgpool-microkernel-tester.h",
4895 "src/xnnpack/AlignedAllocator.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004901 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004902 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004903 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004904 "test/gemm-microkernel-tester.h",
4905 "src/xnnpack/AlignedAllocator.h",
4906 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004907 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004908)
4909
4910xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004911 name = "f16_prelu_test",
4912 srcs = [
4913 "test/f16-prelu.cc",
4914 "test/prelu-microkernel-tester.h",
4915 "src/xnnpack/AlignedAllocator.h",
4916 ] + MICROKERNEL_TEST_HDRS,
4917 deps = MICROKERNEL_TEST_DEPS,
4918)
4919
4920xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004921 name = "f32_prelu_test",
4922 srcs = [
4923 "test/f32-prelu.cc",
4924 "test/prelu-microkernel-tester.h",
4925 "src/xnnpack/AlignedAllocator.h",
4926 ] + MICROKERNEL_TEST_HDRS,
4927 deps = MICROKERNEL_TEST_DEPS,
4928)
4929
4930xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004931 name = "f32_raddexpminusmax_test",
4932 srcs = [
4933 "test/f32-raddexpminusmax.cc",
4934 "test/raddexpminusmax-microkernel-tester.h",
4935 ] + MICROKERNEL_TEST_HDRS,
4936 deps = MICROKERNEL_TEST_DEPS,
4937)
4938
4939xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004940 name = "f32_raddextexp_test",
4941 srcs = [
4942 "test/f32-raddextexp.cc",
4943 "test/raddextexp-microkernel-tester.h",
4944 ] + MICROKERNEL_TEST_HDRS,
4945 deps = MICROKERNEL_TEST_DEPS,
4946)
4947
4948xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004949 name = "f32_raddstoreexpminusmax_test",
4950 srcs = [
4951 "test/f32-raddstoreexpminusmax.cc",
4952 "test/raddstoreexpminusmax-microkernel-tester.h",
4953 ] + MICROKERNEL_TEST_HDRS,
4954 deps = MICROKERNEL_TEST_DEPS,
4955)
4956
4957xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004958 name = "f32_relu_test",
4959 srcs = [
4960 "test/f32-relu.cc",
4961 "test/vunary-microkernel-tester.h",
4962 ] + MICROKERNEL_TEST_HDRS,
4963 deps = MICROKERNEL_TEST_DEPS,
4964)
4965
4966xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004967 name = "f32_rmax_test",
4968 srcs = [
4969 "test/f32-rmax.cc",
4970 "test/rmax-microkernel-tester.h",
4971 ] + MICROKERNEL_TEST_HDRS,
4972 deps = MICROKERNEL_TEST_DEPS,
4973)
4974
4975xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004976 name = "f32_sigmoid_test",
4977 srcs = [
4978 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004979 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004980 ] + MICROKERNEL_TEST_HDRS,
4981 deps = MICROKERNEL_TEST_DEPS,
4982)
4983
4984xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004985 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004986 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004987 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004988 "test/spmm-microkernel-tester.h",
4989 "src/xnnpack/AlignedAllocator.h",
4990 ] + MICROKERNEL_TEST_HDRS,
4991 deps = MICROKERNEL_TEST_DEPS,
4992)
4993
4994xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004995 name = "f16_vadd_minmax_test",
4996 srcs = [
4997 "test/f16-vadd-minmax.cc",
4998 "test/vbinary-microkernel-tester.h",
4999 ] + MICROKERNEL_TEST_HDRS,
5000 deps = MICROKERNEL_TEST_DEPS,
5001)
5002
5003xnnpack_unit_test(
5004 name = "f16_vaddc_minmax_test",
5005 srcs = [
5006 "test/f16-vaddc-minmax.cc",
5007 "test/vbinaryc-microkernel-tester.h",
5008 ] + MICROKERNEL_TEST_HDRS,
5009 deps = MICROKERNEL_TEST_DEPS,
5010)
5011
5012xnnpack_unit_test(
5013 name = "f16_vdiv_minmax_test",
5014 srcs = [
5015 "test/f16-vdiv-minmax.cc",
5016 "test/vbinary-microkernel-tester.h",
5017 ] + MICROKERNEL_TEST_HDRS,
5018 deps = MICROKERNEL_TEST_DEPS,
5019)
5020
5021xnnpack_unit_test(
5022 name = "f16_vdivc_minmax_test",
5023 srcs = [
5024 "test/f16-vdivc-minmax.cc",
5025 "test/vbinaryc-microkernel-tester.h",
5026 ] + MICROKERNEL_TEST_HDRS,
5027 deps = MICROKERNEL_TEST_DEPS,
5028)
5029
5030xnnpack_unit_test(
5031 name = "f16_vrdivc_minmax_test",
5032 srcs = [
5033 "test/f16-vrdivc-minmax.cc",
5034 "test/vbinaryc-microkernel-tester.h",
5035 ] + MICROKERNEL_TEST_HDRS,
5036 deps = MICROKERNEL_TEST_DEPS,
5037)
5038
5039xnnpack_unit_test(
5040 name = "f16_vmax_test",
5041 srcs = [
5042 "test/f16-vmax.cc",
5043 "test/vbinary-microkernel-tester.h",
5044 ] + MICROKERNEL_TEST_HDRS,
5045 deps = MICROKERNEL_TEST_DEPS,
5046)
5047
5048xnnpack_unit_test(
5049 name = "f16_vmaxc_test",
5050 srcs = [
5051 "test/f16-vmaxc.cc",
5052 "test/vbinaryc-microkernel-tester.h",
5053 ] + MICROKERNEL_TEST_HDRS,
5054 deps = MICROKERNEL_TEST_DEPS,
5055)
5056
5057xnnpack_unit_test(
5058 name = "f16_vmin_test",
5059 srcs = [
5060 "test/f16-vmin.cc",
5061 "test/vbinary-microkernel-tester.h",
5062 ] + MICROKERNEL_TEST_HDRS,
5063 deps = MICROKERNEL_TEST_DEPS,
5064)
5065
5066xnnpack_unit_test(
5067 name = "f16_vminc_test",
5068 srcs = [
5069 "test/f16-vminc.cc",
5070 "test/vbinaryc-microkernel-tester.h",
5071 ] + MICROKERNEL_TEST_HDRS,
5072 deps = MICROKERNEL_TEST_DEPS,
5073)
5074
5075xnnpack_unit_test(
5076 name = "f16_vmul_minmax_test",
5077 srcs = [
5078 "test/f16-vmul-minmax.cc",
5079 "test/vbinary-microkernel-tester.h",
5080 ] + MICROKERNEL_TEST_HDRS,
5081 deps = MICROKERNEL_TEST_DEPS,
5082)
5083
5084xnnpack_unit_test(
5085 name = "f16_vmulc_minmax_test",
5086 srcs = [
5087 "test/f16-vmulc-minmax.cc",
5088 "test/vbinaryc-microkernel-tester.h",
5089 ] + MICROKERNEL_TEST_HDRS,
5090 deps = MICROKERNEL_TEST_DEPS,
5091)
5092
5093xnnpack_unit_test(
5094 name = "f16_vsub_minmax_test",
5095 srcs = [
5096 "test/f16-vsub-minmax.cc",
5097 "test/vbinary-microkernel-tester.h",
5098 ] + MICROKERNEL_TEST_HDRS,
5099 deps = MICROKERNEL_TEST_DEPS,
5100)
5101
5102xnnpack_unit_test(
5103 name = "f16_vsubc_minmax_test",
5104 srcs = [
5105 "test/f16-vsubc-minmax.cc",
5106 "test/vbinaryc-microkernel-tester.h",
5107 ] + MICROKERNEL_TEST_HDRS,
5108 deps = MICROKERNEL_TEST_DEPS,
5109)
5110
5111xnnpack_unit_test(
5112 name = "f16_vrsubc_minmax_test",
5113 srcs = [
5114 "test/f16-vrsubc-minmax.cc",
5115 "test/vbinaryc-microkernel-tester.h",
5116 ] + MICROKERNEL_TEST_HDRS,
5117 deps = MICROKERNEL_TEST_DEPS,
5118)
5119
5120xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005121 name = "f32_vabs_test",
5122 srcs = [
5123 "test/f32-vabs.cc",
5124 "test/vunary-microkernel-tester.h",
5125 ] + MICROKERNEL_TEST_HDRS,
5126 deps = MICROKERNEL_TEST_DEPS,
5127)
5128
5129xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005130 name = "f32_vadd_test",
5131 srcs = [
5132 "test/f32-vadd.cc",
5133 "test/vbinary-microkernel-tester.h",
5134 ] + MICROKERNEL_TEST_HDRS,
5135 deps = MICROKERNEL_TEST_DEPS,
5136)
5137
5138xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005139 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005140 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005141 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005142 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005143 ] + MICROKERNEL_TEST_HDRS,
5144 deps = MICROKERNEL_TEST_DEPS,
5145)
5146
5147xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005148 name = "f32_vadd_relu_test",
5149 srcs = [
5150 "test/f32-vadd-relu.cc",
5151 "test/vbinary-microkernel-tester.h",
5152 ] + MICROKERNEL_TEST_HDRS,
5153 deps = MICROKERNEL_TEST_DEPS,
5154)
5155
5156xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005157 name = "f32_vaddc_test",
5158 srcs = [
5159 "test/f32-vaddc.cc",
5160 "test/vbinaryc-microkernel-tester.h",
5161 ] + MICROKERNEL_TEST_HDRS,
5162 deps = MICROKERNEL_TEST_DEPS,
5163)
5164
5165xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005166 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005167 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005168 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005169 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005170 ] + MICROKERNEL_TEST_HDRS,
5171 deps = MICROKERNEL_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005175 name = "f32_vaddc_relu_test",
5176 srcs = [
5177 "test/f32-vaddc-relu.cc",
5178 "test/vbinaryc-microkernel-tester.h",
5179 ] + MICROKERNEL_TEST_HDRS,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005184 name = "f32_vdiv_test",
5185 srcs = [
5186 "test/f32-vdiv.cc",
5187 "test/vbinary-microkernel-tester.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005193 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005194 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005195 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005196 "test/vbinary-microkernel-tester.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005202 name = "f32_vdiv_relu_test",
5203 srcs = [
5204 "test/f32-vdiv-relu.cc",
5205 "test/vbinary-microkernel-tester.h",
5206 ] + MICROKERNEL_TEST_HDRS,
5207 deps = MICROKERNEL_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005211 name = "f32_vdivc_test",
5212 srcs = [
5213 "test/f32-vdivc.cc",
5214 "test/vbinaryc-microkernel-tester.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005220 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005221 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005222 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005223 "test/vbinaryc-microkernel-tester.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005229 name = "f32_vdivc_relu_test",
5230 srcs = [
5231 "test/f32-vdivc-relu.cc",
5232 "test/vbinaryc-microkernel-tester.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005238 name = "f32_vrdivc_test",
5239 srcs = [
5240 "test/f32-vrdivc.cc",
5241 "test/vbinaryc-microkernel-tester.h",
5242 ] + MICROKERNEL_TEST_HDRS,
5243 deps = MICROKERNEL_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005247 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005248 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005249 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005250 "test/vbinaryc-microkernel-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005256 name = "f32_vrdivc_relu_test",
5257 srcs = [
5258 "test/f32-vrdivc-relu.cc",
5259 "test/vbinaryc-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005265 name = "f32_vmax_test",
5266 srcs = [
5267 "test/f32-vmax.cc",
5268 "test/vbinary-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
5274 name = "f32_vmaxc_test",
5275 srcs = [
5276 "test/f32-vmaxc.cc",
5277 "test/vbinaryc-microkernel-tester.h",
5278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
5283 name = "f32_vmin_test",
5284 srcs = [
5285 "test/f32-vmin.cc",
5286 "test/vbinary-microkernel-tester.h",
5287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
5292 name = "f32_vminc_test",
5293 srcs = [
5294 "test/f32-vminc.cc",
5295 "test/vbinaryc-microkernel-tester.h",
5296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005301 name = "f32_vmul_test",
5302 srcs = [
5303 "test/f32-vmul.cc",
5304 "test/vbinary-microkernel-tester.h",
5305 ] + MICROKERNEL_TEST_HDRS,
5306 deps = MICROKERNEL_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005310 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005311 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005312 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005313 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005314 ] + MICROKERNEL_TEST_HDRS,
5315 deps = MICROKERNEL_TEST_DEPS,
5316)
5317
5318xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005319 name = "f32_vmul_relu_test",
5320 srcs = [
5321 "test/f32-vmul-relu.cc",
5322 "test/vbinary-microkernel-tester.h",
5323 ] + MICROKERNEL_TEST_HDRS,
5324 deps = MICROKERNEL_TEST_DEPS,
5325)
5326
5327xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005328 name = "f32_vmulc_test",
5329 srcs = [
5330 "test/f32-vmulc.cc",
5331 "test/vbinaryc-microkernel-tester.h",
5332 ] + MICROKERNEL_TEST_HDRS,
5333 deps = MICROKERNEL_TEST_DEPS,
5334)
5335
5336xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005337 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005338 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005339 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005340 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005341 ] + MICROKERNEL_TEST_HDRS,
5342 deps = MICROKERNEL_TEST_DEPS,
5343)
5344
5345xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005346 name = "f32_vmulc_relu_test",
5347 srcs = [
5348 "test/f32-vmulc-relu.cc",
5349 "test/vbinaryc-microkernel-tester.h",
5350 ] + MICROKERNEL_TEST_HDRS,
5351 deps = MICROKERNEL_TEST_DEPS,
5352)
5353
5354xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005355 name = "f16_vmulcaddc_minmax_test",
5356 srcs = [
5357 "test/f16-vmulcaddc-minmax.cc",
5358 "test/vmulcaddc-microkernel-tester.h",
5359 "src/xnnpack/AlignedAllocator.h",
5360 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005361 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005362)
5363
5364xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005365 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005366 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005367 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005368 "test/vmulcaddc-microkernel-tester.h",
5369 "src/xnnpack/AlignedAllocator.h",
5370 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005371 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005372)
5373
5374xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005375 name = "f32_vlrelu_test",
5376 srcs = [
5377 "test/f32-vlrelu.cc",
5378 "test/vunary-microkernel-tester.h",
5379 ] + MICROKERNEL_TEST_HDRS,
5380 deps = MICROKERNEL_TEST_DEPS,
5381)
5382
5383xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005384 name = "f32_vneg_test",
5385 srcs = [
5386 "test/f32-vneg.cc",
5387 "test/vunary-microkernel-tester.h",
5388 ] + MICROKERNEL_TEST_HDRS,
5389 deps = MICROKERNEL_TEST_DEPS,
5390)
5391
5392xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005393 name = "f32_vrndne_test",
5394 srcs = [
5395 "test/f32-vrndne.cc",
5396 "test/vunary-microkernel-tester.h",
5397 ] + MICROKERNEL_TEST_HDRS,
5398 deps = MICROKERNEL_TEST_DEPS,
5399)
5400
5401xnnpack_unit_test(
5402 name = "f32_vrndz_test",
5403 srcs = [
5404 "test/f32-vrndz.cc",
5405 "test/vunary-microkernel-tester.h",
5406 ] + MICROKERNEL_TEST_HDRS,
5407 deps = MICROKERNEL_TEST_DEPS,
5408)
5409
5410xnnpack_unit_test(
5411 name = "f32_vrndu_test",
5412 srcs = [
5413 "test/f32-vrndu.cc",
5414 "test/vunary-microkernel-tester.h",
5415 ] + MICROKERNEL_TEST_HDRS,
5416 deps = MICROKERNEL_TEST_DEPS,
5417)
5418
5419xnnpack_unit_test(
5420 name = "f32_vrndd_test",
5421 srcs = [
5422 "test/f32-vrndd.cc",
5423 "test/vunary-microkernel-tester.h",
5424 ] + MICROKERNEL_TEST_HDRS,
5425 deps = MICROKERNEL_TEST_DEPS,
5426)
5427
5428xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005429 name = "f32_vscale_test",
5430 srcs = [
5431 "test/f32-vscale.cc",
5432 "test/vscale-microkernel-tester.h",
5433 ] + MICROKERNEL_TEST_HDRS,
5434 deps = MICROKERNEL_TEST_DEPS,
5435)
5436
5437xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005438 name = "f32_vscaleexpminusmax_test",
5439 srcs = [
5440 "test/f32-vscaleexpminusmax.cc",
5441 "test/vscaleexpminusmax-microkernel-tester.h",
5442 ] + MICROKERNEL_TEST_HDRS,
5443 deps = MICROKERNEL_TEST_DEPS,
5444)
5445
5446xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005447 name = "f32_vscaleextexp_test",
5448 srcs = [
5449 "test/f32-vscaleextexp.cc",
5450 "test/vscaleextexp-microkernel-tester.h",
5451 ] + MICROKERNEL_TEST_HDRS,
5452 deps = MICROKERNEL_TEST_DEPS,
5453)
5454
5455xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005456 name = "f32_vsqr_test",
5457 srcs = [
5458 "test/f32-vsqr.cc",
5459 "test/vunary-microkernel-tester.h",
5460 ] + MICROKERNEL_TEST_HDRS,
5461 deps = MICROKERNEL_TEST_DEPS,
5462)
5463
5464xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005465 name = "f32_vsqrdiff_test",
5466 srcs = [
5467 "test/f32-vsqrdiff.cc",
5468 "test/vbinary-microkernel-tester.h",
5469 ] + MICROKERNEL_TEST_HDRS,
5470 deps = MICROKERNEL_TEST_DEPS,
5471)
5472
5473xnnpack_unit_test(
5474 name = "f32_vsqrdiffc_test",
5475 srcs = [
5476 "test/f32-vsqrdiffc.cc",
5477 "test/vbinaryc-microkernel-tester.h",
5478 ] + MICROKERNEL_TEST_HDRS,
5479 deps = MICROKERNEL_TEST_DEPS,
5480)
5481
5482xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005483 name = "f32_vsqrt_test",
5484 srcs = [
5485 "test/f32-vsqrt.cc",
5486 "test/vunary-microkernel-tester.h",
5487 ] + MICROKERNEL_TEST_HDRS,
5488 deps = MICROKERNEL_TEST_DEPS,
5489)
5490
5491xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005492 name = "f32_vsub_test",
5493 srcs = [
5494 "test/f32-vsub.cc",
5495 "test/vbinary-microkernel-tester.h",
5496 ] + MICROKERNEL_TEST_HDRS,
5497 deps = MICROKERNEL_TEST_DEPS,
5498)
5499
5500xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005501 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005502 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005503 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005504 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005505 ] + MICROKERNEL_TEST_HDRS,
5506 deps = MICROKERNEL_TEST_DEPS,
5507)
5508
5509xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005510 name = "f32_vsub_relu_test",
5511 srcs = [
5512 "test/f32-vsub-relu.cc",
5513 "test/vbinary-microkernel-tester.h",
5514 ] + MICROKERNEL_TEST_HDRS,
5515 deps = MICROKERNEL_TEST_DEPS,
5516)
5517
5518xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005519 name = "f32_vsubc_test",
5520 srcs = [
5521 "test/f32-vsubc.cc",
5522 "test/vbinaryc-microkernel-tester.h",
5523 ] + MICROKERNEL_TEST_HDRS,
5524 deps = MICROKERNEL_TEST_DEPS,
5525)
5526
5527xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005528 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005529 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005530 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005531 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005537 name = "f32_vsubc_relu_test",
5538 srcs = [
5539 "test/f32-vsubc-relu.cc",
5540 "test/vbinaryc-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005546 name = "f32_vrsubc_test",
5547 srcs = [
5548 "test/f32-vrsubc.cc",
5549 "test/vbinaryc-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005555 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005556 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005557 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005558 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005564 name = "f32_vrsubc_relu_test",
5565 srcs = [
5566 "test/f32-vrsubc-relu.cc",
5567 "test/vbinaryc-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005573 name = "qs8_dwconv_minmax_test",
5574 srcs = [
5575 "test/qs8-dwconv-minmax.cc",
5576 "test/dwconv-microkernel-tester.h",
5577 "src/xnnpack/AlignedAllocator.h",
5578 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5579 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5580)
5581
5582xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005583 name = "qs8_gavgpool_minmax_test",
5584 srcs = [
5585 "test/qs8-gavgpool-minmax.cc",
5586 "test/gavgpool-microkernel-tester.h",
5587 "src/xnnpack/AlignedAllocator.h",
5588 ] + MICROKERNEL_TEST_HDRS,
5589 deps = MICROKERNEL_TEST_DEPS,
5590)
5591
5592xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005593 name = "qs8_gemm_minmax_test",
5594 srcs = [
5595 "test/qs8-gemm-minmax.cc",
5596 "test/gemm-microkernel-tester.h",
5597 "src/xnnpack/AlignedAllocator.h",
5598 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5599 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5600)
5601
5602xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005603 name = "qs8_igemm_minmax_test",
5604 srcs = [
5605 "test/qs8-igemm-minmax.cc",
5606 "test/gemm-microkernel-tester.h",
5607 "src/xnnpack/AlignedAllocator.h",
5608 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5609 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5610)
5611
5612xnnpack_unit_test(
5613 name = "qs8_requantization_test",
5614 srcs = [
5615 "src/xnnpack/requantization-stubs.h",
5616 "test/qs8-requantization.cc",
5617 "test/requantization-tester.h",
5618 ] + MICROKERNEL_TEST_HDRS,
5619 deps = MICROKERNEL_TEST_DEPS,
5620)
5621
5622xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005623 name = "qs8_vadd_minmax_test",
5624 srcs = [
5625 "test/qs8-vadd-minmax.cc",
5626 "test/vadd-microkernel-tester.h",
5627 ] + MICROKERNEL_TEST_HDRS,
5628 deps = MICROKERNEL_TEST_DEPS,
5629)
5630
5631xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005632 name = "qs8_vaddc_minmax_test",
5633 srcs = [
5634 "test/qs8-vaddc-minmax.cc",
5635 "test/vaddc-microkernel-tester.h",
5636 ] + MICROKERNEL_TEST_HDRS,
5637 deps = MICROKERNEL_TEST_DEPS,
5638)
5639
5640xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005641 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005642 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005643 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005644 "test/avgpool-microkernel-tester.h",
5645 "src/xnnpack/AlignedAllocator.h",
5646 ] + MICROKERNEL_TEST_HDRS,
5647 deps = MICROKERNEL_TEST_DEPS,
5648)
5649
5650xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005651 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005653 "test/qu8-dwconv-minmax.cc",
5654 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655 "src/xnnpack/AlignedAllocator.h",
5656 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005657 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658)
5659
5660xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005661 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005662 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005663 "test/qu8-igemm-minmax.cc",
5664 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005665 "src/xnnpack/AlignedAllocator.h",
5666 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005667 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005668)
5669
5670xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005671 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005673 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005674 "test/gavgpool-microkernel-tester.h",
5675 "src/xnnpack/AlignedAllocator.h",
5676 ] + MICROKERNEL_TEST_HDRS,
5677 deps = MICROKERNEL_TEST_DEPS,
5678)
5679
5680xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005681 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005682 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005683 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005684 "test/gemm-microkernel-tester.h",
5685 "src/xnnpack/AlignedAllocator.h",
5686 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005687 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005688)
5689
5690xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005691 name = "qu8_requantization_test",
5692 srcs = [
5693 "src/xnnpack/requantization-stubs.h",
5694 "test/qu8-requantization.cc",
5695 "test/requantization-tester.h",
5696 ] + MICROKERNEL_TEST_HDRS,
5697 deps = MICROKERNEL_TEST_DEPS,
5698)
5699
5700xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005701 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005703 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005704 "test/vadd-microkernel-tester.h",
5705 ] + MICROKERNEL_TEST_HDRS,
5706 deps = MICROKERNEL_TEST_DEPS,
5707)
5708
5709xnnpack_unit_test(
5710 name = "u8_clamp_test",
5711 srcs = [
5712 "test/u8-clamp.cc",
5713 "test/clamp-microkernel-tester.h",
5714 ] + MICROKERNEL_TEST_HDRS,
5715 deps = MICROKERNEL_TEST_DEPS,
5716)
5717
5718xnnpack_unit_test(
5719 name = "u8_lut32norm_test",
5720 srcs = [
5721 "test/u8-lut32norm.cc",
5722 "test/lut-norm-microkernel-tester.h",
5723 ] + MICROKERNEL_TEST_HDRS,
5724 deps = MICROKERNEL_TEST_DEPS,
5725)
5726
5727xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005728 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005729 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005730 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005731 "test/maxpool-microkernel-tester.h",
5732 ] + MICROKERNEL_TEST_HDRS,
5733 deps = MICROKERNEL_TEST_DEPS,
5734)
5735
5736xnnpack_unit_test(
5737 name = "u8_rmax_test",
5738 srcs = [
5739 "test/u8-rmax.cc",
5740 "test/rmax-microkernel-tester.h",
5741 ] + MICROKERNEL_TEST_HDRS,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005746 name = "x32_fill_test",
5747 srcs = [
5748 "test/x32-fill.cc",
5749 "test/fill-microkernel-tester.h",
5750 ] + MICROKERNEL_TEST_HDRS,
5751 deps = MICROKERNEL_TEST_DEPS,
5752)
5753
5754xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005755 name = "x32_packx_test",
5756 srcs = [
5757 "test/x32-packx.cc",
5758 "test/pack-microkernel-tester.h",
5759 "src/xnnpack/AlignedAllocator.h",
5760 ] + MICROKERNEL_TEST_HDRS,
5761 deps = MICROKERNEL_TEST_DEPS,
5762)
5763
5764xnnpack_unit_test(
5765 name = "x32_pad_test",
5766 srcs = [
5767 "test/x32-pad.cc",
5768 "test/pad-microkernel-tester.h",
5769 ] + MICROKERNEL_TEST_HDRS,
5770 deps = MICROKERNEL_TEST_DEPS,
5771)
5772
5773xnnpack_unit_test(
5774 name = "x32_unpool_test",
5775 srcs = [
5776 "test/x32-unpool.cc",
5777 "test/unpool-microkernel-tester.h",
5778 ] + MICROKERNEL_TEST_HDRS,
5779 deps = MICROKERNEL_TEST_DEPS,
5780)
5781
5782xnnpack_unit_test(
5783 name = "x32_zip_test",
5784 srcs = [
5785 "test/x32-zip.cc",
5786 "test/zip-microkernel-tester.h",
5787 ] + MICROKERNEL_TEST_HDRS,
5788 deps = MICROKERNEL_TEST_DEPS,
5789)
5790
5791xnnpack_unit_test(
5792 name = "x8_lut_test",
5793 srcs = [
5794 "test/x8-lut.cc",
5795 "test/lut-microkernel-tester.h",
5796 ] + MICROKERNEL_TEST_HDRS,
5797 deps = MICROKERNEL_TEST_DEPS,
5798)
5799
5800xnnpack_unit_test(
5801 name = "x8_zip_test",
5802 srcs = [
5803 "test/x8-zip.cc",
5804 "test/zip-microkernel-tester.h",
5805 ] + MICROKERNEL_TEST_HDRS,
5806 deps = MICROKERNEL_TEST_DEPS,
5807)
5808
Marat Dukhan20c3b922020-03-10 03:45:06 -07005809########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005810
5811xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005812 name = "operator_size_test",
5813 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005814 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815)
5816
Marat Dukhan20c3b922020-03-10 03:45:06 -07005817xnnpack_binary(
5818 name = "subgraph_size_test",
5819 srcs = ["test/subgraph-size.c"],
5820 deps = [":XNNPACK"],
5821)
5822
5823########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005824
5825xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005826 name = "abs_nc_test",
5827 srcs = [
5828 "test/abs-nc.cc",
5829 "test/abs-operator-tester.h",
5830 ],
5831 deps = OPERATOR_TEST_DEPS,
5832)
5833
5834xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005835 name = "add_nd_test",
5836 srcs = [
5837 "test/add-nd.cc",
5838 "test/binary-elementwise-operator-tester.h",
5839 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005840 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005841)
5842
5843xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005844 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005845 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005846 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005847 "test/argmax-pooling-operator-tester.h",
5848 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005849 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005850)
5851
5852xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005853 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005854 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005855 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856 "test/average-pooling-operator-tester.h",
5857 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005858 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005859)
5860
5861xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005862 name = "bankers_rounding_nc_test",
5863 srcs = [
5864 "test/bankers-rounding-nc.cc",
5865 "test/bankers-rounding-operator-tester.h",
5866 ],
5867 deps = OPERATOR_TEST_DEPS,
5868)
5869
5870xnnpack_unit_test(
5871 name = "ceiling_nc_test",
5872 srcs = [
5873 "test/ceiling-nc.cc",
5874 "test/ceiling-operator-tester.h",
5875 ],
5876 deps = OPERATOR_TEST_DEPS,
5877)
5878
5879xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005880 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005881 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005882 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005883 "test/channel-shuffle-operator-tester.h",
5884 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005885 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005886)
5887
5888xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005889 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005890 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005891 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005892 "test/clamp-operator-tester.h",
5893 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005894 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005895)
5896
5897xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005898 name = "constant_pad_nd_test",
5899 srcs = [
5900 "test/constant-pad-nd.cc",
5901 "test/constant-pad-operator-tester.h",
5902 ],
5903 deps = OPERATOR_TEST_DEPS,
5904)
5905
5906xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005907 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005909 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910 "test/convolution-operator-tester.h",
5911 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005912 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005913)
5914
5915xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005916 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005917 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005918 "test/convolution-nchw.cc",
5919 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005920 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005921 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922)
5923
5924xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005925 name = "copy_nc_test",
5926 srcs = [
5927 "test/copy-nc.cc",
5928 "test/copy-operator-tester.h",
5929 ],
5930 deps = OPERATOR_TEST_DEPS,
5931)
5932
5933xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005934 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005935 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005936 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005937 "test/deconvolution-operator-tester.h",
5938 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005939 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005940)
5941
5942xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005943 name = "divide_nd_test",
5944 srcs = [
5945 "test/binary-elementwise-operator-tester.h",
5946 "test/divide-nd.cc",
5947 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005948 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005949)
5950
5951xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005952 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005954 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005955 "test/fully-connected-operator-tester.h",
5956 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005957 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005958)
5959
5960xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005961 name = "floor_nc_test",
5962 srcs = [
5963 "test/floor-nc.cc",
5964 "test/floor-operator-tester.h",
5965 ],
5966 deps = OPERATOR_TEST_DEPS,
5967)
5968
5969xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005970 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005971 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005972 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005973 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005974 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005975 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005976)
5977
5978xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005979 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005981 "test/global-average-pooling-ncw.cc",
5982 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005983 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005984 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005985)
5986
5987xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005988 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005989 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005990 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005991 "test/hardswish-operator-tester.h",
5992 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005993 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994)
5995
5996xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005997 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005998 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005999 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 "test/leaky-relu-operator-tester.h",
6001 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006002 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006003)
6004
6005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006006 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006007 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006008 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009 "test/max-pooling-operator-tester.h",
6010 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006011 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006012)
6013
6014xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006015 name = "maximum_nd_test",
6016 srcs = [
6017 "test/binary-elementwise-operator-tester.h",
6018 "test/maximum-nd.cc",
6019 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006020 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006021)
6022
6023xnnpack_unit_test(
6024 name = "minimum_nd_test",
6025 srcs = [
6026 "test/binary-elementwise-operator-tester.h",
6027 "test/minimum-nd.cc",
6028 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006029 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006030)
6031
6032xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006033 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006034 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006035 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006036 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006037 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006038 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006039)
6040
6041xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006042 name = "negate_nc_test",
6043 srcs = [
6044 "test/negate-nc.cc",
6045 "test/negate-operator-tester.h",
6046 ],
6047 deps = OPERATOR_TEST_DEPS,
6048)
6049
6050xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006051 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006052 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006053 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054 "test/prelu-operator-tester.h",
6055 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006056 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006057)
6058
6059xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006060 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006061 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006062 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006063 "test/resize-bilinear-operator-tester.h",
6064 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006065 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006066)
6067
6068xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006069 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006070 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006071 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006072 "test/sigmoid-operator-tester.h",
6073 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006074 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006075)
6076
6077xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006078 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006080 "test/softmax-nc.cc",
6081 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006082 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006083 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006084)
6085
6086xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006087 name = "square_nc_test",
6088 srcs = [
6089 "test/square-nc.cc",
6090 "test/square-operator-tester.h",
6091 ],
6092 deps = OPERATOR_TEST_DEPS,
6093)
6094
6095xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006096 name = "square_root_nc_test",
6097 srcs = [
6098 "test/square-root-nc.cc",
6099 "test/square-root-operator-tester.h",
6100 ],
6101 deps = OPERATOR_TEST_DEPS,
6102)
6103
6104xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006105 name = "squared_difference_nd_test",
6106 srcs = [
6107 "test/binary-elementwise-operator-tester.h",
6108 "test/squared-difference-nd.cc",
6109 ],
6110 deps = OPERATOR_TEST_DEPS,
6111)
6112
6113xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006114 name = "subtract_nd_test",
6115 srcs = [
6116 "test/binary-elementwise-operator-tester.h",
6117 "test/subtract-nd.cc",
6118 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006119 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006120)
6121
6122xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006123 name = "truncation_nc_test",
6124 srcs = [
6125 "test/truncation-nc.cc",
6126 "test/truncation-operator-tester.h",
6127 ],
6128 deps = OPERATOR_TEST_DEPS,
6129)
6130
6131xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006132 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006133 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006134 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135 "test/unpooling-operator-tester.h",
6136 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006137 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006138)
6139
Chao Mei6ddfc602020-05-13 22:29:36 -07006140############################### Misc unit tests ###############################
6141
6142xnnpack_unit_test(
6143 name = "memory_planner_test",
6144 srcs = [
6145 "test/memory-planner-test.cc",
6146 ],
6147 deps = [
6148 ":XNNPACK",
6149 ":memory_planner",
6150 ],
6151)
6152
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153############################# Build configurations #############################
6154
Marat Dukhanb8642352019-10-30 15:43:02 -07006155# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006156config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006157 name = "xnn_enable_assembly_explicit_true",
6158 define_values = {"xnn_enable_assembly": "true"},
6159)
6160
6161# Disables usage of assembly kernels.
6162config_setting(
6163 name = "xnn_enable_assembly_explicit_false",
6164 define_values = {"xnn_enable_assembly": "false"},
6165)
6166
Marat Dukhan9de90e02020-06-18 16:04:12 -07006167# Enables usage of sparse inference.
6168config_setting(
6169 name = "xnn_enable_sparse_explicit_true",
6170 define_values = {"xnn_enable_sparse": "true"},
6171)
6172
6173# Disables usage of sparse inference.
6174config_setting(
6175 name = "xnn_enable_sparse_explicit_false",
6176 define_values = {"xnn_enable_sparse": "false"},
6177)
6178
Marat Dukhan05702cf2020-03-26 15:41:33 -07006179# Disables usage of HMP-aware optimizations.
6180config_setting(
6181 name = "xnn_enable_hmp_explicit_false",
6182 define_values = {"xnn_enable_hmp": "false"},
6183)
6184
Chao Mei6ddfc602020-05-13 22:29:36 -07006185# Enable usage of optimized memory allocation
6186config_setting(
6187 name = "xnn_enable_memopt_explicit_true",
6188 define_values = {"xnnpack_enable_memopt": "true"},
6189)
6190
6191# Disable usage of optimized memory allocation
6192config_setting(
6193 name = "xnn_enable_memopt_explicit_false",
6194 define_values = {"xnnpack_enable_memopt": "false"},
6195)
6196
Marat Dukhanb8642352019-10-30 15:43:02 -07006197# Builds with -c dbg
6198config_setting(
6199 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006201 "compilation_mode": "dbg",
6202 },
6203)
6204
6205# Builds with -c opt
6206config_setting(
6207 name = "optimized_build",
6208 values = {
6209 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006210 },
6211)
6212
6213config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006214 name = "linux_k8",
6215 values = {"cpu": "k8"},
6216)
6217
6218config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006219 name = "linux_arm",
6220 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006221)
6222
6223config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006224 name = "linux_armeabi",
6225 values = {"cpu": "armeabi"},
6226)
6227
6228config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006229 name = "linux_armhf",
6230 values = {"cpu": "armhf"},
6231)
6232
6233config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006234 name = "linux_armv7a",
6235 values = {"cpu": "armv7a"},
6236)
6237
6238config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006239 name = "linux_aarch64",
6240 values = {"cpu": "aarch64"},
6241)
6242
6243config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006244 name = "android",
6245 values = {"crosstool_top": "//external:android/crosstool"},
6246)
6247
6248config_setting(
6249 name = "android_armv7",
6250 values = {
6251 "crosstool_top": "//external:android/crosstool",
6252 "cpu": "armeabi-v7a",
6253 },
6254)
6255
6256config_setting(
6257 name = "android_arm64",
6258 values = {
6259 "crosstool_top": "//external:android/crosstool",
6260 "cpu": "arm64-v8a",
6261 },
6262)
6263
6264config_setting(
6265 name = "android_x86",
6266 values = {
6267 "crosstool_top": "//external:android/crosstool",
6268 "cpu": "x86",
6269 },
6270)
6271
6272config_setting(
6273 name = "android_x86_64",
6274 values = {
6275 "crosstool_top": "//external:android/crosstool",
6276 "cpu": "x86_64",
6277 },
6278)
6279
6280config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006281 name = "windows_x86_64",
6282 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006283)
6284
6285config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006286 name = "windows_x86_64_clang",
6287 values = {
6288 "compiler": "clang-cl",
6289 "cpu": "x64_windows",
6290 },
6291)
6292
6293config_setting(
6294 name = "windows_x86_64_mingw",
6295 values = {
6296 "compiler": "mingw-gcc",
6297 "cpu": "x64_windows",
6298 },
6299)
6300
6301config_setting(
6302 name = "windows_x86_64_msys",
6303 values = {
6304 "compiler": "msys-gcc",
6305 "cpu": "x64_windows",
6306 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006307)
6308
6309config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006310 name = "macos_x86_64",
6311 values = {
6312 "apple_platform_type": "macos",
6313 "cpu": "darwin",
6314 },
6315)
6316
6317config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006318 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006319 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006320)
6321
6322config_setting(
6323 name = "emscripten_wasm",
6324 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006325 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006326 "cpu": "wasm",
6327 },
6328)
6329
6330config_setting(
6331 name = "emscripten_wasmsimd",
6332 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006333 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006334 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006335 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006336 },
6337)
6338
6339config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006340 name = "ios_armv7",
6341 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006342 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006343 "cpu": "ios_armv7",
6344 },
6345)
6346
6347config_setting(
6348 name = "ios_arm64",
6349 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006350 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006351 "cpu": "ios_arm64",
6352 },
6353)
6354
6355config_setting(
6356 name = "ios_arm64e",
6357 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006358 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006359 "cpu": "ios_arm64e",
6360 },
6361)
6362
6363config_setting(
6364 name = "ios_x86",
6365 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006366 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006367 "cpu": "ios_i386",
6368 },
6369)
6370
6371config_setting(
6372 name = "ios_x86_64",
6373 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006374 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006375 "cpu": "ios_x86_64",
6376 },
6377)
6378
6379config_setting(
6380 name = "watchos_armv7k",
6381 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006382 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006383 "cpu": "watchos_armv7k",
6384 },
6385)
6386
6387config_setting(
6388 name = "watchos_arm64_32",
6389 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006390 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006391 "cpu": "watchos_arm64_32",
6392 },
6393)
6394
6395config_setting(
6396 name = "watchos_x86",
6397 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006398 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006399 "cpu": "watchos_i386",
6400 },
6401)
6402
6403config_setting(
6404 name = "watchos_x86_64",
6405 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006406 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006407 "cpu": "watchos_x86_64",
6408 },
6409)
6410
6411config_setting(
6412 name = "tvos_arm64",
6413 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006414 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006415 "cpu": "tvos_arm64",
6416 },
6417)
6418
6419config_setting(
6420 name = "tvos_x86_64",
6421 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006422 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006423 "cpu": "tvos_x86_64",
6424 },
6425)