blob: 4a6a0c0fcfe7b29075120c7f063095cdaf78f62d [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 Barchard8e229db2020-07-06 23:31:35 -0700241 "src/f32-vbinary/gen/vadd-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700244 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
245 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
246 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
248 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
249 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700250 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
251 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
252 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700253 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
254 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
255 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
260 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
261 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800262 "src/f32-vbinary/gen/vmax-scalar-x1.c",
263 "src/f32-vbinary/gen/vmax-scalar-x2.c",
264 "src/f32-vbinary/gen/vmax-scalar-x4.c",
265 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
266 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
267 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
268 "src/f32-vbinary/gen/vmin-scalar-x1.c",
269 "src/f32-vbinary/gen/vmin-scalar-x2.c",
270 "src/f32-vbinary/gen/vmin-scalar-x4.c",
271 "src/f32-vbinary/gen/vminc-scalar-x1.c",
272 "src/f32-vbinary/gen/vminc-scalar-x2.c",
273 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700274 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700277 "src/f32-vbinary/gen/vmul-scalar-x1.c",
278 "src/f32-vbinary/gen/vmul-scalar-x2.c",
279 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700280 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
281 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
282 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
284 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
285 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700289 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
290 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
291 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700292 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
293 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
294 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700295 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
296 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
297 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700298 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
299 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
300 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
301 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
302 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
303 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700307 "src/f32-vbinary/gen/vsub-scalar-x1.c",
308 "src/f32-vbinary/gen/vsub-scalar-x2.c",
309 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700310 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
311 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
312 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700313 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
314 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
315 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700316 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
317 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
318 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700319 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
320 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
321 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700322 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
323 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
324 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
325 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
326 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
327 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
328 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
329 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
330 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700331 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
332 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
333 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700334 "src/f32-vunary/gen/vabs-scalar-x1.c",
335 "src/f32-vunary/gen/vabs-scalar-x2.c",
336 "src/f32-vunary/gen/vabs-scalar-x4.c",
337 "src/f32-vunary/gen/vneg-scalar-x1.c",
338 "src/f32-vunary/gen/vneg-scalar-x2.c",
339 "src/f32-vunary/gen/vneg-scalar-x4.c",
340 "src/f32-vunary/gen/vsqr-scalar-x1.c",
341 "src/f32-vunary/gen/vsqr-scalar-x2.c",
342 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800343 "src/math/expminus-scalar-lut2048-p1.c",
344 "src/math/expminus-scalar-lut64-p2.c",
345 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700346 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700347 "src/math/roundne-scalar-nearbyint.c",
348 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700349 "src/math/roundd-scalar-addsub.c",
350 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700351 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700352 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700353 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700354 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700355 "src/math/roundz-scalar-addsub.c",
356 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700357 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800358 "src/math/sigmoid-scalar-lut2048-p1-div.c",
359 "src/math/sigmoid-scalar-lut64-p2-div.c",
360 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700361 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
362 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700363 "src/qs8-requantization/fp32-scalar-lrintf.c",
364 "src/qs8-requantization/fp32-scalar-magic.c",
365 "src/qs8-requantization/precise-scalar-signed64.c",
366 "src/qs8-requantization/precise-scalar-unsigned32.c",
367 "src/qs8-requantization/precise-scalar-unsigned64.c",
368 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700369 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
370 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
371 "src/qu8-dwconv/up1x9-minmax-scalar.c",
372 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
373 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700374 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
375 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700376 "src/qu8-gemm/2x2-minmax-scalar.c",
377 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700378 "src/qu8-requantization/fp32-scalar-lrintf.c",
379 "src/qu8-requantization/fp32-scalar-magic.c",
380 "src/qu8-requantization/precise-scalar-signed64.c",
381 "src/qu8-requantization/precise-scalar-unsigned32.c",
382 "src/qu8-requantization/precise-scalar-unsigned64.c",
383 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700384 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700385 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700386 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700387 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700388 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700389 "src/x32-fill/scalar-float.c",
390 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700391 "src/x32-packx/x2-scalar.c",
392 "src/x32-packx/x3-scalar.c",
393 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700394 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700395 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700396 "src/x32-unpool/scalar.c",
397 "src/x32-zip/x2-scalar.c",
398 "src/x32-zip/x3-scalar.c",
399 "src/x32-zip/x4-scalar.c",
400 "src/x32-zip/xm-scalar.c",
401 "src/x8-lut/scalar.c",
402 "src/x8-zip/x2-scalar.c",
403 "src/x8-zip/x3-scalar.c",
404 "src/x8-zip/x4-scalar.c",
405 "src/x8-zip/xm-scalar.c",
406]
407
Marat Dukhan436ebe62019-12-04 15:10:12 -0800408WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700409 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
410 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700411 "src/f32-clamp/gen/wasm-x1.c",
412 "src/f32-clamp/gen/wasm-x2.c",
413 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700414 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
415 "src/f32-dwconv/gen/up1x4-wasm.c",
416 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
417 "src/f32-dwconv/gen/up1x9-wasm.c",
418 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
419 "src/f32-dwconv/gen/up1x25-wasm.c",
420 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
421 "src/f32-dwconv/gen/up2x4-wasm.c",
422 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
423 "src/f32-dwconv/gen/up2x9-wasm.c",
424 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
425 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700426 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
427 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
428 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
429 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700430 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
431 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700432 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
433 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
434 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
435 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700436 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
437 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700438 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
439 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700440 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
441 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
442 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700443 "src/f32-gemm/gen/1x4-wasm.c",
444 "src/f32-gemm/gen/2x4-wasm.c",
445 "src/f32-gemm/gen/4x2-wasm.c",
446 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700447 "src/f32-gemm/gen/1x4-relu-wasm.c",
448 "src/f32-gemm/gen/2x4-relu-wasm.c",
449 "src/f32-gemm/gen/4x2-relu-wasm.c",
450 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700451 "src/f32-gemm/gen/1x4-minmax-wasm.c",
452 "src/f32-gemm/gen/2x4-minmax-wasm.c",
453 "src/f32-gemm/gen/4x2-minmax-wasm.c",
454 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700455 "src/f32-hswish/gen/hswish-wasm-x1.c",
456 "src/f32-hswish/gen/hswish-wasm-x2.c",
457 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700458 "src/f32-igemm/gen/1x4-wasm.c",
459 "src/f32-igemm/gen/2x4-wasm.c",
460 "src/f32-igemm/gen/4x2-wasm.c",
461 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700462 "src/f32-igemm/gen/1x4-relu-wasm.c",
463 "src/f32-igemm/gen/2x4-relu-wasm.c",
464 "src/f32-igemm/gen/4x2-relu-wasm.c",
465 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700466 "src/f32-igemm/gen/1x4-minmax-wasm.c",
467 "src/f32-igemm/gen/2x4-minmax-wasm.c",
468 "src/f32-igemm/gen/4x2-minmax-wasm.c",
469 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700470 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
471 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
472 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700473 "src/f32-prelu/gen/wasm-2x1.c",
474 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700475 "src/f32-relu/gen/wasm-x1.c",
476 "src/f32-relu/gen/wasm-x2.c",
477 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700478 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
479 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
480 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
481 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
482 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
483 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
484 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
485 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
486 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
487 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
488 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
489 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800490 "src/f32-vbinary/gen/vmax-wasm-x1.c",
491 "src/f32-vbinary/gen/vmax-wasm-x2.c",
492 "src/f32-vbinary/gen/vmax-wasm-x4.c",
493 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
494 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
495 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
496 "src/f32-vbinary/gen/vmin-wasm-x1.c",
497 "src/f32-vbinary/gen/vmin-wasm-x2.c",
498 "src/f32-vbinary/gen/vmin-wasm-x4.c",
499 "src/f32-vbinary/gen/vminc-wasm-x1.c",
500 "src/f32-vbinary/gen/vminc-wasm-x2.c",
501 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700502 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
503 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
504 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
505 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
506 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
507 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
508 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
509 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
510 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
511 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
512 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
513 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
514 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
515 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
516 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
517 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
518 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
519 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700520 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
521 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
522 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700523 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
524 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
525 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800526]
527
Marat Dukhan290055c2020-06-09 12:24:29 -0700528WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700529 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
530 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
531 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700532 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
533 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
534 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
535 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700536 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
537 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
538 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
539 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700540 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
541 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
542 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
543 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
544 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
545 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700546 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
547 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
548 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
549 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
550 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
551 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
552 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
553 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
554 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
555 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
556 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
557 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
558 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
559 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
560 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
561 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
562 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
563 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
564 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
565 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
566 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
567 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
568 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
569 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700570 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
571 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
572 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700573 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700574 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
575 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
576 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
577 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700578 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
579 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
580 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
581 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
582 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
583 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
584 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
585 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
586 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
587 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
588 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
589 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
590 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
591 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
592 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
593 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
594 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
595 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
596 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
597 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
598 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
599 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
600 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
601 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
602 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
603 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
604 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
605 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
606 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
607 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700608 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
609 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700610 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
611 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
612 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
613 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
614 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
615 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
616 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
617 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
618 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
619 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
620 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
621 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
622 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
623 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
624 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
625 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
626 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
627 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
628 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
629 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
630 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
631 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
632 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
633 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
634 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
635 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
636 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
637 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
638 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
639 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700640 "src/f32-ibilinear/gen/wasmsimd-c4.c",
641 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700642 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
643 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
644 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700645 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700646 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
647 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
648 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
649 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700650 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
651 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
652 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
653 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
654 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
655 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
656 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
657 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
658 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
659 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
660 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
661 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
662 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
663 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
664 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
665 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
666 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
667 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
668 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
669 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
670 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
671 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
672 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
673 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
674 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
675 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
676 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
677 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
678 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
679 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700680 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
681 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700682 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
683 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
684 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
685 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700686 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
687 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700688 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
689 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
690 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700691 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
692 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700693 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
694 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
695 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
696 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700697 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
698 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700699 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
700 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
701 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700702 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
703 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700704 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
705 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
706 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
707 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
708 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
709 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
710 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700711 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
712 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700713 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
714 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
715 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
716 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700717 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
718 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
719 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
720 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
721 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
722 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
723 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
724 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
725 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
726 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
727 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
728 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700729 "src/f32-relu/gen/wasmsimd-x4.c",
730 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700731 "src/f32-rmax/wasmsimd-arm.c",
732 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700733 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
734 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
735 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
736 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
737 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
738 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700739 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
740 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
741 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
742 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
743 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
744 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700745 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
746 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
747 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
748 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700749 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
750 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700751 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
752 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
753 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
754 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700755 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
756 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700757 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
758 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
759 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
760 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700761 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
762 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700763 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
764 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
765 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
766 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700767 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
768 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700769 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
770 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
771 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
772 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700773 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
774 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
775 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
776 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700777 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
778 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
779 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
780 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700781 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
782 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
783 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
784 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700785 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
786 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
787 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
788 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700789 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
790 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700791 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
792 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
793 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
794 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700795 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
796 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700797 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
798 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
799 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
800 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700801 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
802 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700803 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
804 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
805 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
806 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700807 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
808 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700809 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
810 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700811 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
812 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700813 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
814 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
815 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
816 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700817 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
818 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700819 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
820 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
821 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
822 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700823 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
824 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700825 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
826 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
827 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
828 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700829 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
830 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
831 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
832 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700833 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
834 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
835 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
836 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
837 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
838 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
839 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
840 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
841 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
842 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
843 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
844 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
845 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
846 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700847 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
848 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700849 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
850 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
851 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
852 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
853 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
854 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700855 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
856 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
857 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700858 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
859 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
860 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700861 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
862 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
863 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700864 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
865 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
866 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
867 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
868 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
869 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
870 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
871 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
872 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
873 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
874 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
875 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
876 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
877 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
878 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700879 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700880 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700881 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700882 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700883 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700884 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700885 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700886 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700887 "src/x32-zip/x2-wasmsimd.c",
888 "src/x32-zip/x3-wasmsimd.c",
889 "src/x32-zip/x4-wasmsimd.c",
890 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700891 "src/math/roundne-wasmsimd-addsub.c",
892 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700893 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700894 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700895 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700896 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700897 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700898 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
899 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700900]
901
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800902PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700903 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700904 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700905 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700906 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700907 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700908 "src/f32-spmm/gen/4x1-minmax-psimd.c",
909 "src/f32-spmm/gen/8x1-minmax-psimd.c",
910 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700911]
912
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800913PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700914 "src/qs8-requantization/precise-psimd.c",
915 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700916 "src/qu8-requantization/precise-psimd.c",
917 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800918]
919
Marat Dukhan08c4a432019-10-03 09:29:21 -0700920# ISA-specific micro-kernels
921NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -0700922 "src/f32-argmaxpool/4x-neon-c4.c",
923 "src/f32-argmaxpool/9p8x-neon-c4.c",
924 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700925 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
926 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700927 "src/f32-clamp/gen/neon-x4.c",
928 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700929 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
930 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700931 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
932 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -0700933 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
934 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -0700935 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
936 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700937 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
938 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
939 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
940 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700941 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
942 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
943 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
944 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -0700945 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
946 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
947 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
948 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700949 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700950 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
951 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700952 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
953 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
954 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
955 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
956 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
957 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
958 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
959 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
960 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
961 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
962 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
963 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
964 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
965 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
966 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
967 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
968 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
969 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
970 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
971 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
972 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
973 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
974 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
975 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
976 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
977 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
978 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
979 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
980 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
981 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
982 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700983 "src/f32-hswish/gen/hswish-neon-x4.c",
984 "src/f32-hswish/gen/hswish-neon-x8.c",
985 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700986 "src/f32-ibilinear/gen/neon-c4.c",
987 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700988 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
989 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
990 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
991 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
992 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
993 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
994 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
995 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
996 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
997 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
998 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
999 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1000 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1001 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1002 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1003 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001004 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1005 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1006 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001007 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1008 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001009 "src/f32-prelu/gen/neon-1x4.c",
1010 "src/f32-prelu/gen/neon-1x8.c",
1011 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001012 "src/f32-prelu/gen/neon-2x4.c",
1013 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001014 "src/f32-prelu/gen/neon-2x16.c",
1015 "src/f32-prelu/gen/neon-4x4.c",
1016 "src/f32-prelu/gen/neon-4x8.c",
1017 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001018 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1019 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1020 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1021 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1022 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1023 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1024 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1025 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1026 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1027 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1028 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1029 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1030 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1031 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1032 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1033 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1034 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1035 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1036 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1037 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1038 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1039 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1040 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1041 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001042 "src/f32-relu/gen/neon-x4.c",
1043 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001044 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001045 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001046 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1047 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1048 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1049 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1050 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1051 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1052 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1053 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1054 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1055 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1056 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1057 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1058 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1059 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1060 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1061 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1062 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1063 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001064 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1065 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1066 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1067 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001068 "src/f32-vbinary/gen/vmax-neon-x4.c",
1069 "src/f32-vbinary/gen/vmax-neon-x8.c",
1070 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1071 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1072 "src/f32-vbinary/gen/vmin-neon-x4.c",
1073 "src/f32-vbinary/gen/vmin-neon-x8.c",
1074 "src/f32-vbinary/gen/vminc-neon-x4.c",
1075 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001076 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1077 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1078 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1079 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1080 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1081 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001082 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1083 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1084 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1085 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001086 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1087 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1088 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1089 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001090 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1091 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001092 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1093 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001094 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1095 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1096 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1097 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1098 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1099 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1100 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1101 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001102 "src/f32-vunary/gen/vabs-neon-x4.c",
1103 "src/f32-vunary/gen/vabs-neon-x8.c",
1104 "src/f32-vunary/gen/vneg-neon-x4.c",
1105 "src/f32-vunary/gen/vneg-neon-x8.c",
1106 "src/f32-vunary/gen/vsqr-neon-x4.c",
1107 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001108 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1109 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1110 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1111 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1112 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1113 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1114 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1115 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001116 "src/qs8-requantization/precise-neon.c",
1117 "src/qs8-requantization/fp32-neon.c",
1118 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001119 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1120 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1121 "src/qu8-dwconv/up8x9-minmax-neon.c",
1122 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1123 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1124 "src/qu8-gemm/4x8-minmax-neon.c",
1125 "src/qu8-gemm/8x8-minmax-neon.c",
1126 "src/qu8-igemm/4x8-minmax-neon.c",
1127 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001128 "src/qu8-requantization/precise-neon.c",
1129 "src/qu8-requantization/fp32-neon.c",
1130 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001131 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001132 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001133 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001134 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001135 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001136 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001137 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001138 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001139 "src/x32-zip/x2-neon.c",
1140 "src/x32-zip/x3-neon.c",
1141 "src/x32-zip/x4-neon.c",
1142 "src/x32-zip/xm-neon.c",
1143 "src/x8-zip/x2-neon.c",
1144 "src/x8-zip/x3-neon.c",
1145 "src/x8-zip/x4-neon.c",
1146 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001147 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001148 "src/math/roundd-neon-addsub.c",
1149 "src/math/roundd-neon-cvt.c",
1150 "src/math/roundu-neon-addsub.c",
1151 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001152 "src/math/roundz-neon-addsub.c",
1153 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001154 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001155 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1156 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1157 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1158 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1159 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1160 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001161 "src/math/sqrt-neon-nr1rsqrts.c",
1162 "src/math/sqrt-neon-nr2rsqrts.c",
1163 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001164]
1165
1166NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001167 "src/f32-ibilinear/gen/neonfma-c4.c",
1168 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001169 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1170 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1171 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1172 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1173 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1174 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1175 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1176 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1177 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001178 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1179 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1180 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1181 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001182 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1183 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1184 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1185 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001186 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1187 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1188 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1189 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001190 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1191 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1192 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1193 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1194 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1195 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1196 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1197 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1198 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1199 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1200 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1201 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1202 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1203 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1204 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1205 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1206 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1207 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001210 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1211 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1212 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1213 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1214 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1215 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1216 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1217 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1218 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1219 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1220 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1221 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1222 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1223 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1224 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1225 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1226 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1228 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1229 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1231 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1232 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1233 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001234 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1245 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1251 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1252 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1253 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1254 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1255 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1256 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1257 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1258 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1259 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1260 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1261 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1262 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1263 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1264 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1265 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1266 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1267 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1268 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1269 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1270 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1271 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1272 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1273 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1274 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1275 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1276 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1277 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1278 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1279 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1280 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1281 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1282 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1283 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1284 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1285 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1286 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1287 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001288 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1289 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001290 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1291 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1292 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1293 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1294 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1295 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1296 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1297 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1298 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1299 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1300 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1301 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1302 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1303 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1304 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1305 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1306 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1307 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1308 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1309 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001310 "src/math/exp-neonfma-lut64-p2.c",
1311 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001312 "src/math/expminus-neonfma-lut2048-p1.c",
1313 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001314 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001315 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1316 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1317 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1318 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1319 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1320 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1321 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1322 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1323 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1324 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1325 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1326 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1327 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1328 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1329 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1330 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1331 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1332 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001333 "src/math/sqrt-neonfma-nr1fma.c",
1334 "src/math/sqrt-neonfma-nr2fma.c",
1335 "src/math/sqrt-neonfma-nr3fma.c",
1336 "src/math/sqrt-neonfma-nr2fma1adj.c",
1337 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001338]
1339
1340AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001341 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1342 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1343 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1344 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1345 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1346 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001347 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1348 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1349 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1350 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1351 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1352 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1353 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1354 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1355 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1356 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1357 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1358 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1359 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1360 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1361 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1362 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1363 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1364 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1365 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1366 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001367 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1368 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001369 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1370 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001371 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1372 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001373 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1374 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001375 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1376 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1377 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1378 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1379 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001380 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1383 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1384 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1385 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1386 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1387 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1388 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1389 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1390 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1391 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1392 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1393 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1394 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1395 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1396 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1397 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001398 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1399 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1400 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1401 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1402 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1403 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1404 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1405 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1406 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1407 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1408 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1409 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1410 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1411 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1412 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1413 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1414 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1415 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001416 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1417 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001418 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1419 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1420 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1421 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1422 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1423 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001424]
1425
Marat Dukhan8853b822020-05-07 12:19:01 -07001426NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001427 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1428 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1429 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1430 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1431 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1432 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1433 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1434 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001435 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001436 "src/math/roundd-neonv8.c",
1437 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001438 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001439]
1440
Marat Dukhan08c4a432019-10-03 09:29:21 -07001441AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001442 "src/f16-clamp/gen/neonfp16arith-x16.c",
1443 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001444 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1445 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1446 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1447 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1448 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1449 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1450 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1451 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1452 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1453 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1454 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1455 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001456 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1457 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001458 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1459 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001460 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001461 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001462 "src/f16-relu/gen/neonfp16arith-x16.c",
1463 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001464 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1465 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1466 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1467 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1468 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1469 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1470 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1471 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1472 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1473 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1474 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1475 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1476 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1477 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1478 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1479 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1480 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1481 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1482 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1483 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1484 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1485 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1486 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1487 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1488 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1489 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1490 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1491 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001492 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1493 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001494 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1495 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1496 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1497 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1498 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1499 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1500 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1501 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001502 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1503 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1504 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1505 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001506 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1507 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1508 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1509 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1510 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1511 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1512 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1513 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1514 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1515 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1516 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1517 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001518 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1519 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1520 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1521 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1522 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1523 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1524 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1525 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001526]
1527
1528SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001529 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1530 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001531 "src/f32-clamp/gen/sse-x4.c",
1532 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001533 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1534 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001535 "src/f32-dwconv-chw/3x3p1-sse.c",
1536 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001537 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1538 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1539 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1540 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1541 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1542 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1543 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1544 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1545 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1546 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1547 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1548 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001549 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001550 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1551 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001552 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1553 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1554 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1555 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1556 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1557 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1558 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1559 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1560 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1561 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1562 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1563 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1564 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001565 "src/f32-hswish/gen/hswish-sse-x4.c",
1566 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001567 "src/f32-ibilinear/gen/sse-c4.c",
1568 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001569 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1570 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1571 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1572 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1573 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1574 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1575 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001576 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1577 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1578 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001579 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001580 "src/f32-prelu/gen/sse-2x4.c",
1581 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001582 "src/f32-relu/gen/sse-x4.c",
1583 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001584 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001585 "src/f32-spmm/gen/4x1-minmax-sse.c",
1586 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001587 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001588 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1589 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1590 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1591 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1592 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1593 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1594 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1595 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001596 "src/f32-vbinary/gen/vmax-sse-x4.c",
1597 "src/f32-vbinary/gen/vmax-sse-x8.c",
1598 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1599 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1600 "src/f32-vbinary/gen/vmin-sse-x4.c",
1601 "src/f32-vbinary/gen/vmin-sse-x8.c",
1602 "src/f32-vbinary/gen/vminc-sse-x4.c",
1603 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001604 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1605 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1606 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1607 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1608 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1609 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1610 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1611 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001612 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1613 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1614 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1615 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001616 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1617 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1618 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1619 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001620 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1621 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001622 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1623 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001624 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1625 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001626 "src/f32-vunary/gen/vabs-sse-x4.c",
1627 "src/f32-vunary/gen/vabs-sse-x8.c",
1628 "src/f32-vunary/gen/vneg-sse-x4.c",
1629 "src/f32-vunary/gen/vneg-sse-x8.c",
1630 "src/f32-vunary/gen/vsqr-sse-x4.c",
1631 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001632 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001633 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001634 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001635 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001636 "src/math/roundd-sse-addsub.c",
1637 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001638 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001639 "src/math/sqrt-sse-hh1mac.c",
1640 "src/math/sqrt-sse-nr1mac.c",
1641 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001642]
1643
1644SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001645 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1646 "src/f32-argmaxpool/4x-sse2-c4.c",
1647 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001648 "src/f32-prelu/gen/sse2-2x4.c",
1649 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001650 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1651 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1652 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1653 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1654 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1655 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1656 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1657 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1658 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1659 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1660 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1661 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001662 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001663 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001664 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001665 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001666 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1667 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001668 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1669 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001670 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1671 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1672 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1673 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1674 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1675 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1676 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1677 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001678 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1679 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1680 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001681 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1682 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1683 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001684 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1685 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1686 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001687 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1688 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001689 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1690 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001691 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1692 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001693 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1694 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001695 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001696 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1697 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001698 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001699 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1700 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1701 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001702 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1703 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001704 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1705 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001706 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1707 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001708 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001709 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1710 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001711 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001712 "src/qs8-requantization/precise-sse2.c",
1713 "src/qs8-requantization/fp32-sse2.c",
1714 "src/qs8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001715 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1716 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1717 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1718 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1719 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1720 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1721 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1722 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001723 "src/qu8-requantization/precise-sse2.c",
1724 "src/qu8-requantization/fp32-sse2.c",
1725 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001726 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001727 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001728 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001729 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001730 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001731 "src/x32-zip/x2-sse2.c",
1732 "src/x32-zip/x3-sse2.c",
1733 "src/x32-zip/x4-sse2.c",
1734 "src/x32-zip/xm-sse2.c",
1735 "src/x8-zip/x2-sse2.c",
1736 "src/x8-zip/x3-sse2.c",
1737 "src/x8-zip/x4-sse2.c",
1738 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001739 "src/math/exp-sse2-p5.c",
1740 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001741 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001742 "src/math/roundd-sse2-cvt.c",
1743 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001744 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001745 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001746]
1747
1748SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001749 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1750 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1751 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001752 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1753 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1754 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001755 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1756 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1757 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001758 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1759 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001760 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1761 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001762 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1763 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001764 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1765 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001766 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001767 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1768 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001769 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001770 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1771 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1772 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001773 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1774 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001775 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1776 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001777 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1778 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001779 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001780 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1781 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001782 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001783 "src/qs8-requantization/precise-ssse3.c",
1784 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001785 "src/qu8-requantization/precise-ssse3.c",
1786 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001787]
1788
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001789SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001790 "src/f32-prelu/gen/sse41-2x4.c",
1791 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001792 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001793 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001794 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001795 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001796 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1797 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001798 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1799 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001800 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1801 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1802 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1803 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1804 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1805 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1806 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1807 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001808 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1809 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1810 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001811 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1812 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1813 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001814 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1815 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1816 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001817 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1818 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001819 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1820 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001821 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1822 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001823 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1824 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001825 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001826 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1827 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001828 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001829 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1830 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1831 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001832 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1833 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001834 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1835 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001836 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1837 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001838 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001839 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1840 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001841 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001842 "src/qs8-requantization/fp32-sse4.c",
1843 "src/qs8-requantization/precise-sse4.c",
1844 "src/qs8-requantization/q31-sse4.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001845 "src/qu8-requantization/precise-sse4.c",
1846 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001847 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001848 "src/math/roundd-sse41.c",
1849 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001850 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001851]
1852
Marat Dukhan08c4a432019-10-03 09:29:21 -07001853AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001854 "src/f32-clamp/gen/avx-x8.c",
1855 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001856 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1857 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1858 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1859 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1860 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1861 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1862 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1863 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1864 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1865 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1866 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1867 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1868 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1869 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1870 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1871 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1872 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1873 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1874 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1875 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1876 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1877 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1878 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1879 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1880 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1881 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1882 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1883 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1884 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
1885 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001886 "src/f32-hswish/gen/hswish-avx-x8.c",
1887 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001888 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
1889 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
1890 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
1891 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1892 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
1893 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1894 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
1895 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
1896 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07001897 "src/f32-prelu/gen/avx-2x8.c",
1898 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001899 "src/f32-relu/gen/avx-x16.c",
1900 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001901 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001902 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1903 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1904 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1905 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1906 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1907 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1908 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1909 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08001910 "src/f32-vbinary/gen/vmax-avx-x8.c",
1911 "src/f32-vbinary/gen/vmax-avx-x16.c",
1912 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1913 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1914 "src/f32-vbinary/gen/vmin-avx-x8.c",
1915 "src/f32-vbinary/gen/vmin-avx-x16.c",
1916 "src/f32-vbinary/gen/vminc-avx-x8.c",
1917 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001918 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1919 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1920 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1921 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1922 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1923 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1924 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1925 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001926 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1927 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1928 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1929 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001930 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1931 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1932 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1933 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001934 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1935 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07001936 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001937 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1938 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1939 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1940 "src/f32-vrnd/gen/vrndz-avx-x16.c",
1941 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1942 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1943 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1944 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001945 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
1946 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001947 "src/f32-vunary/gen/vabs-avx-x8.c",
1948 "src/f32-vunary/gen/vabs-avx-x16.c",
1949 "src/f32-vunary/gen/vneg-avx-x8.c",
1950 "src/f32-vunary/gen/vneg-avx-x16.c",
1951 "src/f32-vunary/gen/vsqr-avx-x8.c",
1952 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001953]
1954
Marat Dukhan1566fee2020-08-02 21:55:41 -07001955XOP_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001956 "src/qs8-dwconv/gen/up8x9-minmax-xop-mul16.c",
1957 "src/qs8-dwconv/gen/up16x9-minmax-xop-mul16.c",
1958 "src/qs8-dwconv/gen/up24x9-minmax-xop-mul16.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001959 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
1960 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
1961 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
1962 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001963 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
1964 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001965 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
1966 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001967 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001968 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
1969 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001970 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001971 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
1972 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
1973 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001974 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
1975 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
1976 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
1977 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
1978 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
1979 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001980 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001981 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
1982 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001983 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07001984]
1985
Marat Dukhanfda12b82019-11-21 12:27:59 -08001986FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07001987 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
1988 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
1989 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
1990 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
1991 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
1992 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
1993 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
1994 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
1995 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
1996 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
1997 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
1998 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
1999 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2000 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2001 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2002 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2003 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2004 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2005 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2006 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2007 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2008 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2009 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2010 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2011 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2012 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2013 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2014 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2015 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2016 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2017 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2018 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2019 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2020 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2021 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2022 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2023 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2024 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2025 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2026 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002027 "src/f32-hswish/gen/hswish-fma3-x8.c",
2028 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002029 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2030 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2031 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2032 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2033 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2034 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2035 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2036 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2037 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2038 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2039 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2040 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2041 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2042 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002043 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2044 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2045 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2046 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2047 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2048 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2049 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2050 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002051 "src/math/sqrt-fma3-nr1fma.c",
2052 "src/math/sqrt-fma3-nr2fma.c",
2053 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002054]
2055
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002056AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002057 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2058 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2059 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2060 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2061 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2062 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2063 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2064 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2065 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2066 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2067 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2068 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2069 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2070 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2071 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2072 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2073 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2074 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2075 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2076 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2077 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2078 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2079 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2080 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2081 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2082 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2083 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2084 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2085 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2086 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2087 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2088 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2089 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2090 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2091 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2092 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002093 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2094 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2095 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2096 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2097 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2098 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2099 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2100 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2101 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2102 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2103 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2104 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2105 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2106 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2107 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2108 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2109 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2110 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2111 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2112 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2113 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2114 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2115 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2116 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2117 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2118 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2119 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2120 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2121 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2122 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002123 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2124 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2125 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2126 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2127 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2128 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2129 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2130 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2131 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2132 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2133 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2134 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2135 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2136 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2137 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2138 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2139 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2140 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2141 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2142 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2143 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2144 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2145 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2146 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002147 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2148 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002149 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2150 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2151 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2152 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002153 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2154 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2155 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002156 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2157 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2158 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002159 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2160 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2161 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002162 "src/math/exp-avx2-p5.c",
2163 "src/math/exp-avx2-perm-p3.c",
2164 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002165 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002166 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002167 "src/math/sigmoid-avx2-rr2-p5-div.c",
2168 "src/math/sigmoid-avx2-rr1-p5-div.c",
2169 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2170 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2171 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2172 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002173]
2174
Marat Dukhan08c4a432019-10-03 09:29:21 -07002175AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002176 "src/f32-clamp/gen/avx512f-x16.c",
2177 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002178 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2179 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2180 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2181 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2182 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2183 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2184 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2185 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2186 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2187 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2188 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2189 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2190 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2191 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2192 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2193 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2194 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2195 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2196 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2197 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2198 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2199 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2200 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2201 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002202 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2203 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002204 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2205 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2206 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2207 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2208 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2209 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002210 "src/f32-prelu/gen/avx512f-2x16.c",
2211 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002212 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2213 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2214 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2215 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2216 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2217 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2218 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2219 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2220 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2221 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2222 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2223 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2224 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2225 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2226 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2227 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2228 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2229 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2230 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2231 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2232 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2233 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2234 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2235 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2236 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2237 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2238 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2239 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2240 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2241 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2242 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2243 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2244 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2245 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2246 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2247 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002248 "src/f32-relu/gen/avx512f-x16.c",
2249 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002250 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002251 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2252 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2253 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2254 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2255 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2256 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2257 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2258 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002259 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2260 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2261 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2262 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2263 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2264 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2265 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2266 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002267 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2268 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2269 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2270 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2271 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2272 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2273 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2274 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002275 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2276 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2277 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2278 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002279 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2280 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2281 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2282 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002283 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2284 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002285 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002286 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2287 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2288 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2289 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2290 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2291 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2292 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2293 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2294 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2295 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2296 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2297 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2298 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2299 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2300 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2301 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2302 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2303 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2304 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2305 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2306 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2307 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2308 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2309 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002310 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2311 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2312 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2313 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2314 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2315 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2316 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2317 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002318 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2319 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2320 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2321 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2322 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2323 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2324 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2325 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002326 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2327 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2328 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2329 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2330 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2331 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002332 "src/math/exp-avx512f-p5-scalef.c",
2333 "src/math/exp-avx512f-p5.c",
2334 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002335 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002336 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002337 "src/math/sqrt-avx512f-nr1fma.c",
2338 "src/math/sqrt-avx512f-nr2fma.c",
2339 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002340]
2341
2342AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002343 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002344 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002345 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2346 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002347 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002348 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002349 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002350 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002351 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2352 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002353 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2354 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2355 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2356 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002357]
2358
2359AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002360 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002361 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2362 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002363 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002364 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2365 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002366 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2367 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2368 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002369 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002370 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2371 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2372 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002373 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002374 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2375 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002376 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002377 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2378 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2379 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2380 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2381 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2382 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2383 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2384 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2385 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2386 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2387 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2388 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2389 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2390 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2391 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2392 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2393 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2394 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002395 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2396 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002397 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002398 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2399 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2400 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2401 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2402 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2403 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2404 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2405 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2406 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2407 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2408 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2409 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2410 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2411 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2412 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2413 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2414 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2415 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002416 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2417 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2418 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2419 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2420 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2421 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2422 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2423 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2424 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2425 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2426 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2427 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2428 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2429 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2430 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2431 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2432 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2433 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002434]
2435
Marat Dukhan1b354632020-03-23 12:50:22 -07002436INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002437 "src/xnnpack/argmaxpool.h",
2438 "src/xnnpack/avgpool.h",
2439 "src/xnnpack/clamp.h",
2440 "src/xnnpack/common.h",
2441 "src/xnnpack/conv.h",
2442 "src/xnnpack/dwconv.h",
2443 "src/xnnpack/gavgpool.h",
2444 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002445 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002446 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002447 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002448 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002449 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002450 "src/xnnpack/lut.h",
2451 "src/xnnpack/math.h",
2452 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002453 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002454 "src/xnnpack/packx.h",
2455 "src/xnnpack/pad.h",
2456 "src/xnnpack/params.h",
2457 "src/xnnpack/pavgpool.h",
2458 "src/xnnpack/ppmm.h",
2459 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002460 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002461 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002462 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002463 "src/xnnpack/rmax.h",
2464 "src/xnnpack/scalar-utils.h",
2465 "src/xnnpack/spmm.h",
2466 "src/xnnpack/unpool.h",
2467 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002468 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002469 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002470 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002471 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002472 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002473 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002474 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002475]
2476
2477INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002478 "include/xnnpack.h",
2479 "src/xnnpack/allocator.h",
2480 "src/xnnpack/compute.h",
2481 "src/xnnpack/im2col.h",
2482 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002483 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002484 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002485 "src/xnnpack/operator.h",
2486 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002487 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002488 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002489 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002490 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002491]
2492
Marat Dukhan1b354632020-03-23 12:50:22 -07002493ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002494 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002495]
2496
Marat Dukhan1b354632020-03-23 12:50:22 -07002497MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002498 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002499 "include/xnnpack.h",
2500]
2501
Marat Dukhan1b354632020-03-23 12:50:22 -07002502MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002503 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002504 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002505 "src/xnnpack/requantization.h",
2506 "include/xnnpack.h",
2507]
2508
2509OPERATOR_TEST_PARAMS_HDRS = [
2510 "src/xnnpack/params.h",
2511 "src/xnnpack/common.h",
2512]
2513
2514WEIGHTS_PACK_HDRS = [
2515 "src/xnnpack/pack.h",
2516 "src/xnnpack/operator.h",
2517 "src/xnnpack/compute.h",
2518]
2519
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002520LOGGING_COPTS = select({
2521 # No logging in optimized mode
2522 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2523 # Full logging in debug mode
2524 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2525 # Error-only logging in default (fastbuild) mode
2526 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2527})
2528
Marat Dukhan3b59de22020-06-03 20:15:19 -07002529LOGGING_SRCS = select({
2530 # No logging in optimized mode
2531 ":optimized_build": [],
2532 "//conditions:default": [
2533 "src/operator-strings.c",
2534 "src/subgraph-strings.c",
2535 ],
2536})
2537
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002538LOGGING_HDRS = [
2539 "src/xnnpack/log.h",
2540]
2541
Marat Dukhan08c4a432019-10-03 09:29:21 -07002542xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002543 name = "tables",
2544 srcs = TABLE_SRCS,
2545 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002546 gcc_copts = xnnpack_gcc_std_copts(),
2547 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002548)
2549
2550xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002551 name = "scalar_ukernels",
2552 srcs = SCALAR_UKERNELS,
2553 hdrs = INTERNAL_HDRS,
2554 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002555 gcc_copts = xnnpack_gcc_std_copts(),
2556 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002558 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002559 "@FP16",
2560 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002561 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002562 ],
2563)
2564
2565xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002566 name = "scalar_ukernels_test_mode",
2567 srcs = SCALAR_UKERNELS,
2568 hdrs = INTERNAL_HDRS,
2569 aarch32_copts = ["-marm"],
2570 copts = [
2571 "-UNDEBUG",
2572 "-DXNN_TEST_MODE=1",
2573 ],
2574 gcc_copts = xnnpack_gcc_std_copts(),
2575 msvc_copts = xnnpack_msvc_std_copts(),
2576 deps = [
2577 ":tables",
2578 "@FP16",
2579 "@FXdiv",
2580 "@pthreadpool",
2581 ],
2582)
2583
2584xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002585 name = "wasm_ukernels",
2586 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002587 gcc_copts = xnnpack_gcc_std_copts(),
2588 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002589 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002590 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002591 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002592 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002593 "@FP16",
2594 "@FXdiv",
2595 "@pthreadpool",
2596 ],
2597)
2598
2599xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002600 name = "wasm_ukernels_test_mode",
2601 hdrs = INTERNAL_HDRS,
2602 copts = [
2603 "-UNDEBUG",
2604 "-DXNN_TEST_MODE=1",
2605 ],
2606 gcc_copts = xnnpack_gcc_std_copts(),
2607 msvc_copts = xnnpack_msvc_std_copts(),
2608 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002609 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002610 deps = [
2611 ":tables",
2612 "@FP16",
2613 "@FXdiv",
2614 "@pthreadpool",
2615 ],
2616)
2617
2618xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002619 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002620 hdrs = INTERNAL_HDRS,
2621 aarch32_copts = [
2622 "-marm",
2623 "-mfpu=neon",
2624 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002625 gcc_copts = xnnpack_gcc_std_copts(),
2626 gcc_x86_copts = ["-msse2"],
2627 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 optimized_copts = [
2629 "-O3",
2630 "-ffast-math",
2631 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002632 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002634 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002635 "@FP16",
2636 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002637 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002638 ],
2639)
2640
2641xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002642 name = "psimd_fastmath_ukernels_test_mode",
2643 hdrs = INTERNAL_HDRS,
2644 aarch32_copts = [
2645 "-marm",
2646 "-mfpu=neon",
2647 ],
2648 copts = [
2649 "-UNDEBUG",
2650 "-DXNN_TEST_MODE=1",
2651 ],
2652 gcc_copts = xnnpack_gcc_std_copts(),
2653 gcc_x86_copts = ["-msse2"],
2654 msvc_copts = xnnpack_msvc_std_copts(),
2655 optimized_copts = [
2656 "-O3",
2657 "-ffast-math",
2658 ],
2659 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2660 deps = [
2661 ":tables",
2662 "@FP16",
2663 "@psimd",
2664 "@pthreadpool",
2665 ],
2666)
2667
2668xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002669 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002670 hdrs = INTERNAL_HDRS,
2671 aarch32_copts = [
2672 "-marm",
2673 "-mfpu=neon",
2674 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002675 gcc_copts = xnnpack_gcc_std_copts(),
2676 gcc_x86_copts = ["-msse2"],
2677 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002678 optimized_copts = [
2679 "-O3",
2680 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002681 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002682 deps = [
2683 ":tables",
2684 "@FP16",
2685 "@psimd",
2686 "@pthreadpool",
2687 ],
2688)
2689
2690xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002691 name = "psimd_accmath_ukernels_test_mode",
2692 hdrs = INTERNAL_HDRS,
2693 aarch32_copts = [
2694 "-marm",
2695 "-mfpu=neon",
2696 ],
2697 copts = [
2698 "-UNDEBUG",
2699 "-DXNN_TEST_MODE=1",
2700 ],
2701 gcc_copts = xnnpack_gcc_std_copts(),
2702 gcc_x86_copts = ["-msse2"],
2703 msvc_copts = xnnpack_msvc_std_copts(),
2704 optimized_copts = [
2705 "-O3",
2706 ],
2707 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2708 deps = [
2709 ":tables",
2710 "@FP16",
2711 "@psimd",
2712 "@pthreadpool",
2713 ],
2714)
2715
2716xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002717 name = "neon_ukernels",
2718 hdrs = INTERNAL_HDRS,
2719 aarch32_copts = [
2720 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002721 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002722 "-mfpu=neon",
2723 ],
2724 aarch32_srcs = NEON_UKERNELS,
2725 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002726 gcc_copts = xnnpack_gcc_std_copts(),
2727 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002728 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002729 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002730 "@FP16",
2731 "@pthreadpool",
2732 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002733)
2734
2735xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002736 name = "neon_ukernels_test_mode",
2737 hdrs = INTERNAL_HDRS,
2738 aarch32_copts = [
2739 "-marm",
2740 "-march=armv7-a",
2741 "-mfpu=neon",
2742 ],
2743 aarch32_srcs = NEON_UKERNELS,
2744 aarch64_srcs = NEON_UKERNELS,
2745 copts = [
2746 "-UNDEBUG",
2747 "-DXNN_TEST_MODE=1",
2748 ],
2749 gcc_copts = xnnpack_gcc_std_copts(),
2750 msvc_copts = xnnpack_msvc_std_copts(),
2751 deps = [
2752 ":tables",
2753 "@FP16",
2754 "@pthreadpool",
2755 ],
2756)
2757
2758xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 name = "neonfma_ukernels",
2760 hdrs = INTERNAL_HDRS,
2761 aarch32_copts = [
2762 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002763 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002764 "-mfpu=neon-vfpv4",
2765 ],
2766 aarch32_srcs = NEONFMA_UKERNELS,
2767 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002768 apple_aarch32_copts = [
2769 "-mcpu=swift",
2770 "-mtune=generic",
2771 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002772 gcc_copts = xnnpack_gcc_std_copts(),
2773 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002774 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002775 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002776 "@FP16",
2777 "@pthreadpool",
2778 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002779)
2780
2781xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002782 name = "neonfma_ukernels_test_mode",
2783 hdrs = INTERNAL_HDRS,
2784 aarch32_copts = [
2785 "-marm",
2786 "-march=armv7-a",
2787 "-mfpu=neon-vfpv4",
2788 ],
2789 aarch32_srcs = NEONFMA_UKERNELS,
2790 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002791 apple_aarch32_copts = [
2792 "-mcpu=swift",
2793 "-mtune=generic",
2794 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002795 copts = [
2796 "-UNDEBUG",
2797 "-DXNN_TEST_MODE=1",
2798 ],
2799 gcc_copts = xnnpack_gcc_std_copts(),
2800 msvc_copts = xnnpack_msvc_std_copts(),
2801 deps = [
2802 ":tables",
2803 "@FP16",
2804 "@pthreadpool",
2805 ],
2806)
2807
2808xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002809 name = "neonv8_ukernels",
2810 hdrs = INTERNAL_HDRS,
2811 aarch32_copts = [
2812 "-marm",
2813 "-march=armv8-a",
2814 "-mfpu=neon-fp-armv8",
2815 ],
2816 aarch32_srcs = NEONV8_UKERNELS,
2817 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002818 apple_aarch32_copts = [
2819 "-mcpu=cyclone",
2820 "-mtune=generic",
2821 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002822 gcc_copts = xnnpack_gcc_std_copts(),
2823 msvc_copts = xnnpack_msvc_std_copts(),
2824 deps = [
2825 ":tables",
2826 "@FP16",
2827 "@pthreadpool",
2828 ],
2829)
2830
2831xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002832 name = "neonv8_ukernels_test_mode",
2833 hdrs = INTERNAL_HDRS,
2834 aarch32_copts = [
2835 "-marm",
2836 "-march=armv8-a",
2837 "-mfpu=neon-fp-armv8",
2838 ],
2839 aarch32_srcs = NEONV8_UKERNELS,
2840 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002841 apple_aarch32_copts = [
2842 "-mcpu=cyclone",
2843 "-mtune=generic",
2844 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002845 copts = [
2846 "-UNDEBUG",
2847 "-DXNN_TEST_MODE=1",
2848 ],
2849 gcc_copts = xnnpack_gcc_std_copts(),
2850 msvc_copts = xnnpack_msvc_std_copts(),
2851 deps = [
2852 ":tables",
2853 "@FP16",
2854 "@pthreadpool",
2855 ],
2856)
2857
2858xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002859 name = "neonfp16arith_ukernels",
2860 hdrs = INTERNAL_HDRS,
2861 aarch64_copts = ["-march=armv8.2-a+fp16"],
2862 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002863 gcc_copts = xnnpack_gcc_std_copts(),
2864 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002865 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002866 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002867 "@FP16",
2868 "@pthreadpool",
2869 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002870)
2871
2872xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002873 name = "neonfp16arith_ukernels_test_mode",
2874 hdrs = INTERNAL_HDRS,
2875 aarch64_copts = ["-march=armv8.2-a+fp16"],
2876 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
2877 copts = [
2878 "-UNDEBUG",
2879 "-DXNN_TEST_MODE=1",
2880 ],
2881 gcc_copts = xnnpack_gcc_std_copts(),
2882 msvc_copts = xnnpack_msvc_std_copts(),
2883 deps = [
2884 ":tables",
2885 "@FP16",
2886 "@pthreadpool",
2887 ],
2888)
2889
2890xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002891 name = "sse2_ukernels",
2892 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002893 gcc_copts = xnnpack_gcc_std_copts(),
2894 gcc_x86_copts = ["-msse2"],
2895 msvc_copts = xnnpack_msvc_std_copts(),
2896 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002897 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002898 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002899 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002900 "@FP16",
2901 "@pthreadpool",
2902 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002903)
2904
2905xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002906 name = "sse2_ukernels_test_mode",
2907 hdrs = INTERNAL_HDRS,
2908 copts = [
2909 "-UNDEBUG",
2910 "-DXNN_TEST_MODE=1",
2911 ],
2912 gcc_copts = xnnpack_gcc_std_copts(),
2913 gcc_x86_copts = ["-msse2"],
2914 msvc_copts = xnnpack_msvc_std_copts(),
2915 msvc_x86_32_copts = ["/arch:SSE2"],
2916 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
2917 deps = [
2918 ":tables",
2919 "@FP16",
2920 "@pthreadpool",
2921 ],
2922)
2923
2924xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002925 name = "ssse3_ukernels",
2926 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002927 gcc_copts = xnnpack_gcc_std_copts(),
2928 gcc_x86_copts = ["-mssse3"],
2929 msvc_copts = xnnpack_msvc_std_copts(),
2930 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07002931 x86_srcs = SSSE3_UKERNELS,
2932 deps = [
2933 ":tables",
2934 "@FP16",
2935 "@pthreadpool",
2936 ],
2937)
2938
2939xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002940 name = "ssse3_ukernels_test_mode",
2941 hdrs = INTERNAL_HDRS,
2942 copts = [
2943 "-UNDEBUG",
2944 "-DXNN_TEST_MODE=1",
2945 ],
2946 gcc_copts = xnnpack_gcc_std_copts(),
2947 gcc_x86_copts = ["-mssse3"],
2948 msvc_copts = xnnpack_msvc_std_copts(),
2949 msvc_x86_32_copts = ["/arch:SSE2"],
2950 x86_srcs = SSSE3_UKERNELS,
2951 deps = [
2952 ":tables",
2953 "@FP16",
2954 "@pthreadpool",
2955 ],
2956)
2957
2958xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002959 name = "sse41_ukernels",
2960 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002961 gcc_copts = xnnpack_gcc_std_copts(),
2962 gcc_x86_copts = ["-msse4.1"],
2963 msvc_copts = xnnpack_msvc_std_copts(),
2964 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002965 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08002966 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002967 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002968 "@FP16",
2969 "@pthreadpool",
2970 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002971)
2972
2973xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002974 name = "sse41_ukernels_test_mode",
2975 hdrs = INTERNAL_HDRS,
2976 copts = [
2977 "-UNDEBUG",
2978 "-DXNN_TEST_MODE=1",
2979 ],
2980 gcc_copts = xnnpack_gcc_std_copts(),
2981 gcc_x86_copts = ["-msse4.1"],
2982 msvc_copts = xnnpack_msvc_std_copts(),
2983 msvc_x86_32_copts = ["/arch:SSE2"],
2984 x86_srcs = SSE41_UKERNELS,
2985 deps = [
2986 ":tables",
2987 "@FP16",
2988 "@pthreadpool",
2989 ],
2990)
2991
2992xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 name = "avx_ukernels",
2994 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002995 gcc_copts = xnnpack_gcc_std_copts(),
2996 gcc_x86_copts = ["-mavx"],
2997 msvc_copts = xnnpack_msvc_std_copts(),
2998 msvc_x86_32_copts = ["/arch:AVX"],
2999 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003000 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003001 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003002 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003003 "@FP16",
3004 "@pthreadpool",
3005 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003006)
3007
3008xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003009 name = "avx_ukernels_test_mode",
3010 hdrs = INTERNAL_HDRS,
3011 copts = [
3012 "-UNDEBUG",
3013 "-DXNN_TEST_MODE=1",
3014 ],
3015 gcc_copts = xnnpack_gcc_std_copts(),
3016 gcc_x86_copts = ["-mavx"],
3017 msvc_copts = xnnpack_msvc_std_copts(),
3018 msvc_x86_32_copts = ["/arch:AVX"],
3019 msvc_x86_64_copts = ["/arch:AVX"],
3020 x86_srcs = AVX_UKERNELS,
3021 deps = [
3022 ":tables",
3023 "@FP16",
3024 "@pthreadpool",
3025 ],
3026)
3027
3028xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003029 name = "xop_ukernels",
3030 hdrs = INTERNAL_HDRS,
3031 gcc_copts = xnnpack_gcc_std_copts(),
3032 gcc_x86_copts = ["-mxop"],
3033 msvc_copts = xnnpack_msvc_std_copts(),
3034 msvc_x86_32_copts = ["/arch:AVX"],
3035 msvc_x86_64_copts = ["/arch:AVX"],
3036 x86_srcs = XOP_UKERNELS,
3037 deps = [
3038 ":tables",
3039 "@FP16",
3040 "@pthreadpool",
3041 ],
3042)
3043
3044xnnpack_cc_library(
3045 name = "xop_ukernels_test_mode",
3046 hdrs = INTERNAL_HDRS,
3047 copts = [
3048 "-UNDEBUG",
3049 "-DXNN_TEST_MODE=1",
3050 ],
3051 gcc_copts = xnnpack_gcc_std_copts(),
3052 gcc_x86_copts = ["-mxop"],
3053 msvc_copts = xnnpack_msvc_std_copts(),
3054 msvc_x86_32_copts = ["/arch:AVX"],
3055 msvc_x86_64_copts = ["/arch:AVX"],
3056 x86_srcs = XOP_UKERNELS,
3057 deps = [
3058 ":tables",
3059 "@FP16",
3060 "@pthreadpool",
3061 ],
3062)
3063
3064xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003065 name = "fma3_ukernels",
3066 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003067 gcc_copts = xnnpack_gcc_std_copts(),
3068 gcc_x86_copts = ["-mfma"],
3069 msvc_copts = xnnpack_msvc_std_copts(),
3070 msvc_x86_32_copts = ["/arch:AVX"],
3071 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003072 x86_srcs = FMA3_UKERNELS,
3073 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003074 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003075 "@FP16",
3076 "@pthreadpool",
3077 ],
3078)
3079
3080xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003081 name = "fma3_ukernels_test_mode",
3082 hdrs = INTERNAL_HDRS,
3083 copts = [
3084 "-UNDEBUG",
3085 "-DXNN_TEST_MODE=1",
3086 ],
3087 gcc_copts = xnnpack_gcc_std_copts(),
3088 gcc_x86_copts = ["-mfma"],
3089 msvc_copts = xnnpack_msvc_std_copts(),
3090 msvc_x86_32_copts = ["/arch:AVX"],
3091 msvc_x86_64_copts = ["/arch:AVX"],
3092 x86_srcs = FMA3_UKERNELS,
3093 deps = [
3094 ":tables",
3095 "@FP16",
3096 "@pthreadpool",
3097 ],
3098)
3099
3100xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003101 name = "avx2_ukernels",
3102 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003103 gcc_copts = xnnpack_gcc_std_copts(),
3104 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003105 "-mfma",
3106 "-mavx2",
3107 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003108 msvc_copts = xnnpack_msvc_std_copts(),
3109 msvc_x86_32_copts = ["/arch:AVX2"],
3110 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003111 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003112 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003113 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003114 "@FP16",
3115 "@pthreadpool",
3116 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003117)
3118
3119xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003120 name = "avx2_ukernels_test_mode",
3121 hdrs = INTERNAL_HDRS,
3122 copts = [
3123 "-UNDEBUG",
3124 "-DXNN_TEST_MODE=1",
3125 ],
3126 gcc_copts = xnnpack_gcc_std_copts(),
3127 gcc_x86_copts = [
3128 "-mfma",
3129 "-mavx2",
3130 ],
3131 msvc_copts = xnnpack_msvc_std_copts(),
3132 msvc_x86_32_copts = ["/arch:AVX2"],
3133 msvc_x86_64_copts = ["/arch:AVX2"],
3134 x86_srcs = AVX2_UKERNELS,
3135 deps = [
3136 ":tables",
3137 "@FP16",
3138 "@pthreadpool",
3139 ],
3140)
3141
3142xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003143 name = "avx512f_ukernels",
3144 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003145 gcc_copts = xnnpack_gcc_std_copts(),
3146 gcc_x86_copts = ["-mavx512f"],
3147 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3148 msvc_copts = xnnpack_msvc_std_copts(),
3149 msvc_x86_32_copts = ["/arch:AVX512"],
3150 msvc_x86_64_copts = ["/arch:AVX512"],
3151 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003152 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003153 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003154 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003155 "@FP16",
3156 "@pthreadpool",
3157 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003158)
3159
3160xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003161 name = "avx512f_ukernels_test_mode",
3162 hdrs = INTERNAL_HDRS,
3163 copts = [
3164 "-UNDEBUG",
3165 "-DXNN_TEST_MODE=1",
3166 ],
3167 gcc_copts = xnnpack_gcc_std_copts(),
3168 gcc_x86_copts = ["-mavx512f"],
3169 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3170 msvc_copts = xnnpack_msvc_std_copts(),
3171 msvc_x86_32_copts = ["/arch:AVX512"],
3172 msvc_x86_64_copts = ["/arch:AVX512"],
3173 msys_copts = ["-fno-asynchronous-unwind-tables"],
3174 x86_srcs = AVX512F_UKERNELS,
3175 deps = [
3176 ":tables",
3177 "@FP16",
3178 "@pthreadpool",
3179 ],
3180)
3181
3182xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003183 name = "asm_ukernels",
3184 hdrs = ["src/xnnpack/assembly.h"],
3185 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003186 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003187 aarch64_srcs = AARCH64_ASM_UKERNELS,
3188)
3189
Marat Dukhan3b59de22020-06-03 20:15:19 -07003190xnnpack_cc_library(
3191 name = "logging_utils",
3192 srcs = LOGGING_SRCS,
3193 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3194 copts = LOGGING_COPTS + [
3195 "-Isrc",
3196 "-Iinclude",
3197 ] + select({
3198 ":debug_build": [],
3199 "//conditions:default": xnnpack_min_size_copts(),
3200 }),
3201 gcc_copts = xnnpack_gcc_std_copts(),
3202 msvc_copts = xnnpack_msvc_std_copts(),
3203 visibility = xnnpack_visibility(),
3204 deps = [
3205 "@FP16",
3206 "@clog",
3207 "@pthreadpool",
3208 ],
3209)
3210
Marat Dukhan08c4a432019-10-03 09:29:21 -07003211xnnpack_aggregate_library(
3212 name = "ukernels",
3213 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003214 ":neon_ukernels",
3215 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003216 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003217 ":asm_ukernels",
3218 ],
3219 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003220 ":neon_ukernels",
3221 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003222 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003223 ":neonfp16arith_ukernels",
3224 ":asm_ukernels",
3225 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003226 generic_deps = [
3227 ":scalar_ukernels",
3228 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003229 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003230 ":psimd_fastmath_ukernels",
3231 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003232 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003233 wasm_deps = [
3234 ":wasm_ukernels",
3235 ],
3236 wasmsimd_deps = [
3237 ":wasm_ukernels",
3238 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003239 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003240 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003241 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003242 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003243 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003244 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003245 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003246 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003247 ":avx512f_ukernels",
3248 ],
3249)
3250
Marat Dukhan33fcf782020-05-24 14:27:15 -07003251xnnpack_aggregate_library(
3252 name = "ukernels_test_mode",
3253 aarch32_deps = [
3254 ":neon_ukernels_test_mode",
3255 ":neonfma_ukernels_test_mode",
3256 ":neonv8_ukernels_test_mode",
3257 ":asm_ukernels",
3258 ],
3259 aarch64_deps = [
3260 ":neon_ukernels_test_mode",
3261 ":neonfma_ukernels_test_mode",
3262 ":neonv8_ukernels_test_mode",
3263 ":neonfp16arith_ukernels_test_mode",
3264 ":asm_ukernels",
3265 ],
3266 generic_deps = [
3267 ":scalar_ukernels_test_mode",
3268 ],
3269 psimd_deps = [
3270 ":psimd_fastmath_ukernels_test_mode",
3271 ":psimd_accmath_ukernels_test_mode",
3272 ],
3273 wasm_deps = [
3274 ":wasm_ukernels_test_mode",
3275 ],
3276 wasmsimd_deps = [
3277 ":wasm_ukernels_test_mode",
3278 ],
3279 x86_deps = [
3280 ":sse2_ukernels_test_mode",
3281 ":ssse3_ukernels_test_mode",
3282 ":sse41_ukernels_test_mode",
3283 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003284 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003285 ":fma3_ukernels_test_mode",
3286 ":avx2_ukernels_test_mode",
3287 ":avx512f_ukernels_test_mode",
3288 ],
3289)
3290
Marat Dukhan86df3842020-06-17 14:03:31 -07003291# This target must be compatible with Android NDK r17 and compile with Clang 6.
3292# It is similar to :ukernels target, but excludes microkernels which require
3293# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003294xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003295 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003296 aarch32_deps = [
3297 ":neon_ukernels",
3298 ":neonfma_ukernels",
3299 ":neonv8_ukernels",
3300 ":asm_ukernels",
3301 ],
3302 aarch64_deps = [
3303 ":neon_ukernels",
3304 ":neonfma_ukernels",
3305 ":neonv8_ukernels",
3306 ":asm_ukernels",
3307 ],
3308 generic_deps = [
3309 ":scalar_ukernels",
3310 ],
3311 psimd_deps = [
3312 ":psimd_fastmath_ukernels",
3313 ":psimd_accmath_ukernels",
3314 ],
3315 wasm_deps = [
3316 ":wasm_ukernels",
3317 ],
3318 wasmsimd_deps = [
3319 ":wasm_ukernels",
3320 ],
3321 x86_deps = [
3322 ":sse2_ukernels",
3323 ":ssse3_ukernels",
3324 ":sse41_ukernels",
3325 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003326 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003327 ":fma3_ukernels",
3328 ":avx2_ukernels",
3329 ":avx512f_ukernels",
3330 ],
3331)
3332
Marat Dukhan08c4a432019-10-03 09:29:21 -07003333xnnpack_cc_library(
3334 name = "im2col",
3335 srcs = ["src/im2col.c"],
3336 hdrs = [
3337 "src/xnnpack/common.h",
3338 "src/xnnpack/im2col.h",
3339 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003340 gcc_copts = xnnpack_gcc_std_copts(),
3341 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003342)
3343
3344xnnpack_cc_library(
3345 name = "indirection",
3346 srcs = ["src/indirection.c"],
3347 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003348 gcc_copts = xnnpack_gcc_std_copts(),
3349 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003350 deps = [
3351 "@FP16",
3352 "@FXdiv",
3353 "@pthreadpool",
3354 ],
3355)
3356
3357xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003358 name = "indirection_test_mode",
3359 srcs = ["src/indirection.c"],
3360 hdrs = INTERNAL_HDRS,
3361 copts = [
3362 "-UNDEBUG",
3363 "-DXNN_TEST_MODE=1",
3364 ],
3365 gcc_copts = xnnpack_gcc_std_copts(),
3366 msvc_copts = xnnpack_msvc_std_copts(),
3367 deps = [
3368 "@FP16",
3369 "@FXdiv",
3370 "@pthreadpool",
3371 ],
3372)
3373
3374xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003375 name = "packing",
3376 srcs = ["src/packing.c"],
3377 hdrs = INTERNAL_HDRS,
3378 gcc_copts = xnnpack_gcc_std_copts(),
3379 msvc_copts = xnnpack_msvc_std_copts(),
3380 deps = [
3381 "@FP16",
3382 "@FXdiv",
3383 "@pthreadpool",
3384 ],
3385)
3386
3387xnnpack_cc_library(
3388 name = "packing_test_mode",
3389 srcs = ["src/packing.c"],
3390 hdrs = INTERNAL_HDRS,
3391 copts = [
3392 "-UNDEBUG",
3393 "-DXNN_TEST_MODE=1",
3394 ],
3395 gcc_copts = xnnpack_gcc_std_copts(),
3396 msvc_copts = xnnpack_msvc_std_copts(),
3397 deps = [
3398 "@FP16",
3399 "@FXdiv",
3400 "@pthreadpool",
3401 ],
3402)
3403
3404xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003405 name = "operator_run",
3406 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003407 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003408 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003409 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3410 "//conditions:default": [],
3411 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003412 gcc_copts = xnnpack_gcc_std_copts(),
3413 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003414 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003415 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003416 "@FP16",
3417 "@FXdiv",
3418 "@clog",
3419 "@pthreadpool",
3420 ],
3421)
3422
Chao Mei6ddfc602020-05-13 22:29:36 -07003423xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003424 name = "operator_run_test_mode",
3425 srcs = ["src/operator-run.c"],
3426 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3427 copts = LOGGING_COPTS + [
3428 "-UNDEBUG",
3429 "-DXNN_TEST_MODE=1",
3430 ] + select({
3431 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3432 "//conditions:default": [],
3433 }),
3434 gcc_copts = xnnpack_gcc_std_copts(),
3435 msvc_copts = xnnpack_msvc_std_copts(),
3436 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003437 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003438 "@FP16",
3439 "@FXdiv",
3440 "@clog",
3441 "@pthreadpool",
3442 ],
3443)
3444
3445xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003446 name = "memory_planner",
3447 srcs = ["src/memory-planner.c"],
3448 hdrs = INTERNAL_HDRS,
3449 defines = select({
3450 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3451 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3452 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3453 }),
3454 gcc_copts = xnnpack_gcc_std_copts(),
3455 msvc_copts = xnnpack_msvc_std_copts(),
3456 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003457 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003458 "@pthreadpool",
3459 ],
3460)
3461
Marat Dukhan33fcf782020-05-24 14:27:15 -07003462xnnpack_cc_library(
3463 name = "memory_planner_test_mode",
3464 srcs = ["src/memory-planner.c"],
3465 hdrs = INTERNAL_HDRS,
3466 copts = [
3467 "-UNDEBUG",
3468 "-DXNN_TEST_MODE=1",
3469 ],
3470 defines = select({
3471 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3472 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3473 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3474 }),
3475 gcc_copts = xnnpack_gcc_std_copts(),
3476 msvc_copts = xnnpack_msvc_std_copts(),
3477 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003478 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003479 "@pthreadpool",
3480 ],
3481)
3482
Marat Dukhan08c4a432019-10-03 09:29:21 -07003483cc_library(
3484 name = "enable_assembly",
3485 defines = select({
3486 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3487 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003488 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003489 }),
3490)
3491
Marat Dukhan9de90e02020-06-18 16:04:12 -07003492cc_library(
3493 name = "enable_sparse",
3494 defines = select({
3495 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3496 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003497 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003498 }),
3499)
3500
Marat Dukhancf056b22019-10-07 10:26:29 -07003501xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003502 name = "operators",
3503 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003504 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003506 ],
3507 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003508 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003509 "-Isrc",
3510 "-Iinclude",
3511 ] + select({
3512 ":debug_build": [],
3513 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003514 }) + select({
3515 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3516 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003517 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003518 gcc_copts = xnnpack_gcc_std_copts(),
3519 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003520 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003521 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003522 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003523 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003524 "@FP16",
3525 "@FXdiv",
3526 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003527 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003528 ],
3529)
3530
Marat Dukhan10a38082020-04-17 03:58:35 -07003531xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003532 name = "operators_test_mode",
3533 srcs = OPERATOR_SRCS + [
3534 "src/memory.c",
3535 "src/operator-delete.c",
3536 ],
3537 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3538 copts = LOGGING_COPTS + [
3539 "-Isrc",
3540 "-Iinclude",
3541 "-UNDEBUG",
3542 "-DXNN_TEST_MODE=1",
3543 ] + select({
3544 ":debug_build": [],
3545 "//conditions:default": xnnpack_min_size_copts(),
3546 }) + select({
3547 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3548 "//conditions:default": [],
3549 }),
3550 gcc_copts = xnnpack_gcc_std_copts(),
3551 msvc_copts = xnnpack_msvc_std_copts(),
3552 deps = [
3553 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003554 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003555 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003556 "@FP16",
3557 "@FXdiv",
3558 "@clog",
3559 "@pthreadpool",
3560 ],
3561)
3562
3563xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003564 name = "XNNPACK",
3565 srcs = [
3566 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003567 "src/runtime.c",
3568 "src/subgraph.c",
3569 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003570 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003571 hdrs = ["include/xnnpack.h"],
3572 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003573 "-Isrc",
3574 "-Iinclude",
3575 ] + select({
3576 ":debug_build": [],
3577 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003578 }) + select({
3579 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3580 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003581 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003582 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003583 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003584 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003585 visibility = xnnpack_visibility(),
3586 deps = [
3587 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003588 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003589 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003590 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003591 ":operator_run",
3592 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003593 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003594 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003595 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003596 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003597 ] + select({
3598 ":emscripten": [],
3599 "//conditions:default": ["@cpuinfo"],
3600 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601)
3602
Marat Dukhan10a38082020-04-17 03:58:35 -07003603xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003604 name = "XNNPACK_test_mode",
3605 srcs = [
3606 "src/init.c",
3607 "src/runtime.c",
3608 "src/subgraph.c",
3609 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003610 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003611 hdrs = ["include/xnnpack.h"],
3612 copts = LOGGING_COPTS + [
3613 "-Isrc",
3614 "-Iinclude",
3615 "-UNDEBUG",
3616 "-DXNN_TEST_MODE=1",
3617 ] + select({
3618 ":debug_build": [],
3619 "//conditions:default": xnnpack_min_size_copts(),
3620 }) + select({
3621 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3622 "//conditions:default": [],
3623 }),
3624 gcc_copts = xnnpack_gcc_std_copts(),
3625 includes = ["include"],
3626 msvc_copts = xnnpack_msvc_std_copts(),
3627 visibility = xnnpack_visibility(),
3628 deps = [
3629 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003630 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003631 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003632 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003633 ":operator_run_test_mode",
3634 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003635 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003636 "@clog",
3637 "@FP16",
3638 "@pthreadpool",
3639 ] + select({
3640 ":emscripten": [],
3641 "//conditions:default": ["@cpuinfo"],
3642 }),
3643)
3644
3645xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003646 name = "xnnpack_f32",
3647 srcs = [
3648 "src/init.c",
3649 "src/runtime.c",
3650 "src/subgraph.c",
3651 "src/tensor.c",
3652 ] + SUBGRAPH_SRCS,
3653 hdrs = ["include/xnnpack.h"],
3654 copts = LOGGING_COPTS + [
3655 "-Isrc",
3656 "-Iinclude",
3657 ] + select({
3658 ":debug_build": [],
3659 "//conditions:default": xnnpack_min_size_copts(),
3660 }) + select({
3661 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3662 "//conditions:default": [],
3663 }),
3664 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003665 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003666 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003667 "XNN_NO_U8_OPERATORS",
3668 "XNN_NO_X8_OPERATORS",
3669 "XNN_NO_F16_OPERATORS",
3670 "XNN_NO_X16_OPERATORS",
3671 ],
3672 gcc_copts = xnnpack_gcc_std_copts(),
3673 includes = ["include"],
3674 msvc_copts = xnnpack_msvc_std_copts(),
3675 visibility = xnnpack_visibility(),
3676 deps = [
3677 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003678 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003679 ":logging_utils",
3680 ":memory_planner",
3681 ":operator_run",
3682 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003683 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003684 "@clog",
3685 "@FP16",
3686 "@pthreadpool",
3687 ] + select({
3688 ":emscripten": [],
3689 "//conditions:default": ["@cpuinfo"],
3690 }),
3691)
3692
3693xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003694 name = "xnnpack_operators_nhwc_f32",
3695 srcs = [
3696 "src/init.c",
3697 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003698 hdrs = ["include/xnnpack.h"],
3699 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003700 "-Isrc",
3701 "-Iinclude",
3702 ] + select({
3703 ":debug_build": [],
3704 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003705 }) + select({
3706 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3707 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003708 }),
3709 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003710 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003711 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003712 "XNN_NO_U8_OPERATORS",
3713 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003714 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003715 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003716 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003717 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003718 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 visibility = xnnpack_visibility(),
3720 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003721 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003722 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003723 ":operator_run",
3724 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003725 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003726 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003727 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003728 ] + select({
3729 ":emscripten": [],
3730 "//conditions:default": ["@cpuinfo"],
3731 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003732)
3733
Marat Dukhancf056b22019-10-07 10:26:29 -07003734xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 name = "bench_utils",
3736 srcs = ["bench/utils.cc"],
3737 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003738 deps = [
3739 "@com_google_benchmark//:benchmark",
3740 "@cpuinfo",
3741 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003742)
3743
Frank Barchard7e955972019-10-11 10:34:25 -07003744######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745
3746xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003747 name = "qs8_gemm_bench",
3748 srcs = [
3749 "bench/gemm.h",
3750 "bench/qs8-gemm.cc",
3751 "src/xnnpack/AlignedAllocator.h",
3752 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3753 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3754)
3755
3756xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003757 name = "qs8_requantization_bench",
3758 srcs = [
3759 "bench/qs8-requantization.cc",
3760 "src/xnnpack/requantization-stubs.h",
3761 "src/xnnpack/AlignedAllocator.h",
3762 ] + MICROKERNEL_BENCHMARK_HDRS,
3763 deps = MICROKERNEL_BENCHMARK_DEPS,
3764)
3765
3766xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003767 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003768 srcs = [
3769 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003770 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003771 "src/xnnpack/AlignedAllocator.h",
3772 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003773 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003774 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003775)
3776
3777xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07003778 name = "qu8_requantization_bench",
3779 srcs = [
3780 "bench/qu8-requantization.cc",
3781 "src/xnnpack/requantization-stubs.h",
3782 "src/xnnpack/AlignedAllocator.h",
3783 ] + MICROKERNEL_BENCHMARK_HDRS,
3784 deps = MICROKERNEL_BENCHMARK_DEPS,
3785)
3786
3787xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07003788 name = "f16_igemm_bench",
3789 srcs = [
3790 "bench/f16-igemm.cc",
3791 "bench/conv.h",
3792 "bench/google/conv.h",
3793 "src/xnnpack/AlignedAllocator.h",
3794 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003795 deps = MICROKERNEL_BENCHMARK_DEPS + [
3796 ":indirection",
3797 ":packing",
3798 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07003799)
3800
3801xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003802 name = "f16_gemm_bench",
3803 srcs = [
3804 "bench/f16-gemm.cc",
3805 "bench/gemm.h",
3806 "src/xnnpack/AlignedAllocator.h",
3807 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003808 deps = MICROKERNEL_BENCHMARK_DEPS + [
3809 ":packing",
3810 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003811)
3812
3813xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003814 name = "f16_spmm_bench",
3815 srcs = [
3816 "bench/f16-spmm.cc",
3817 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003818 "src/xnnpack/AlignedAllocator.h",
3819 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08003820 deps = MICROKERNEL_BENCHMARK_DEPS,
3821)
3822
3823xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003824 name = "f32_igemm_bench",
3825 srcs = [
3826 "bench/f32-igemm.cc",
3827 "bench/conv.h",
3828 "src/xnnpack/AlignedAllocator.h",
3829 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003830 deps = MICROKERNEL_BENCHMARK_DEPS + [
3831 ":indirection",
3832 ":packing",
3833 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003834)
3835
3836xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07003837 name = "f16_relu_bench",
3838 srcs = [
3839 "bench/f16-relu.cc",
3840 "src/xnnpack/AlignedAllocator.h",
3841 ] + MICROKERNEL_BENCHMARK_HDRS,
3842 deps = MICROKERNEL_BENCHMARK_DEPS,
3843)
3844
3845xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003846 name = "f32_conv_hwc_bench",
3847 srcs = [
3848 "bench/f32-conv-hwc.cc",
3849 "bench/dconv.h",
3850 "src/xnnpack/AlignedAllocator.h",
3851 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003852 deps = MICROKERNEL_BENCHMARK_DEPS + [
3853 ":packing",
3854 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003855)
3856
3857xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003858 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07003859 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003860 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07003861 "bench/dconv.h",
3862 "src/xnnpack/AlignedAllocator.h",
3863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003864 deps = MICROKERNEL_BENCHMARK_DEPS + [
3865 ":packing",
3866 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07003867)
3868
3869xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07003870 name = "f16_dwconv_bench",
3871 srcs = [
3872 "bench/f16-dwconv.cc",
3873 "bench/dwconv.h",
3874 "bench/google/dwconv.h",
3875 "src/xnnpack/AlignedAllocator.h",
3876 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003877 deps = MICROKERNEL_BENCHMARK_DEPS + [
3878 ":indirection",
3879 ":packing",
3880 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07003881)
3882
3883xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003884 name = "f32_dwconv_bench",
3885 srcs = [
3886 "bench/f32-dwconv.cc",
3887 "bench/dwconv.h",
3888 "src/xnnpack/AlignedAllocator.h",
3889 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003890 deps = MICROKERNEL_BENCHMARK_DEPS + [
3891 ":indirection",
3892 ":packing",
3893 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003894)
3895
3896xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07003897 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003898 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07003899 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003900 "bench/dwconv.h",
3901 "src/xnnpack/AlignedAllocator.h",
3902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07003903 deps = MICROKERNEL_BENCHMARK_DEPS + [
3904 ":indirection",
3905 ":packing",
3906 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003907)
3908
3909xnnpack_benchmark(
3910 name = "f32_gemm_bench",
3911 srcs = [
3912 "bench/f32-gemm.cc",
3913 "bench/gemm.h",
3914 "src/xnnpack/AlignedAllocator.h",
3915 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003916 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07003917 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003918)
3919
3920xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07003921 name = "f32_hswish_bench",
3922 srcs = [
3923 "bench/f32-hswish.cc",
3924 "src/xnnpack/AlignedAllocator.h",
3925 ] + MICROKERNEL_BENCHMARK_HDRS,
3926 deps = MICROKERNEL_BENCHMARK_DEPS,
3927)
3928
3929xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08003930 name = "f32_raddexpminusmax_bench",
3931 srcs = [
3932 "bench/f32-raddexpminusmax.cc",
3933 "src/xnnpack/AlignedAllocator.h",
3934 ] + MICROKERNEL_BENCHMARK_HDRS,
3935 deps = MICROKERNEL_BENCHMARK_DEPS,
3936)
3937
3938xnnpack_benchmark(
3939 name = "f32_raddextexp_bench",
3940 srcs = [
3941 "bench/f32-raddextexp.cc",
3942 "src/xnnpack/AlignedAllocator.h",
3943 ] + MICROKERNEL_BENCHMARK_HDRS,
3944 deps = MICROKERNEL_BENCHMARK_DEPS,
3945)
3946
3947xnnpack_benchmark(
3948 name = "f32_raddstoreexpminusmax_bench",
3949 srcs = [
3950 "bench/f32-raddstoreexpminusmax.cc",
3951 "src/xnnpack/AlignedAllocator.h",
3952 ] + MICROKERNEL_BENCHMARK_HDRS,
3953 deps = MICROKERNEL_BENCHMARK_DEPS,
3954)
3955
3956xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07003957 name = "f32_relu_bench",
3958 srcs = [
3959 "bench/f32-relu.cc",
3960 "src/xnnpack/AlignedAllocator.h",
3961 ] + MICROKERNEL_BENCHMARK_HDRS,
3962 deps = MICROKERNEL_BENCHMARK_DEPS,
3963)
3964
3965xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 name = "f32_rmax_bench",
3967 srcs = [
3968 "bench/f32-rmax.cc",
3969 "src/xnnpack/AlignedAllocator.h",
3970 ] + MICROKERNEL_BENCHMARK_HDRS,
3971 deps = MICROKERNEL_BENCHMARK_DEPS,
3972)
3973
3974xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08003975 name = "f32_sigmoid_bench",
3976 srcs = [
3977 "bench/f32-sigmoid.cc",
3978 "src/xnnpack/AlignedAllocator.h",
3979 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08003980 deps = MICROKERNEL_BENCHMARK_DEPS,
3981)
3982
3983xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003984 name = "f32_spmm_bench",
3985 srcs = [
3986 "bench/f32-spmm.cc",
3987 "bench/gemm.h",
3988 "src/xnnpack/AlignedAllocator.h",
3989 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003990 deps = MICROKERNEL_BENCHMARK_DEPS,
3991)
3992
3993xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003994 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003995 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08003996 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07003997 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003998 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08003999 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004000)
4001
4002xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004003 name = "f32_vscaleexpminusmax_bench",
4004 srcs = [
4005 "bench/f32-vscaleexpminusmax.cc",
4006 "src/xnnpack/AlignedAllocator.h",
4007 ] + MICROKERNEL_BENCHMARK_HDRS,
4008 deps = MICROKERNEL_BENCHMARK_DEPS,
4009)
4010
4011xnnpack_benchmark(
4012 name = "f32_vscaleextexp_bench",
4013 srcs = [
4014 "bench/f32-vscaleextexp.cc",
4015 "src/xnnpack/AlignedAllocator.h",
4016 ] + MICROKERNEL_BENCHMARK_HDRS,
4017 deps = MICROKERNEL_BENCHMARK_DEPS,
4018)
4019
4020xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004021 name = "f32_vsqrt_bench",
4022 srcs = [
4023 "bench/f32-vsqrt.cc",
4024 "src/xnnpack/AlignedAllocator.h",
4025 ] + MICROKERNEL_BENCHMARK_HDRS,
4026 deps = MICROKERNEL_BENCHMARK_DEPS,
4027)
4028
4029xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004030 name = "f32_im2col_gemm_bench",
4031 srcs = [
4032 "bench/f32-im2col-gemm.cc",
4033 "bench/conv.h",
4034 "src/xnnpack/AlignedAllocator.h",
4035 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004036 deps = MICROKERNEL_BENCHMARK_DEPS + [
4037 ":im2col",
4038 ":packing",
4039 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004040)
4041
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004042xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004043 name = "rounding_bench",
4044 srcs = [
4045 "bench/rounding.cc",
4046 "src/xnnpack/math-stubs.h",
4047 "src/xnnpack/AlignedAllocator.h",
4048 ] + MICROKERNEL_BENCHMARK_HDRS,
4049 deps = MICROKERNEL_BENCHMARK_DEPS,
4050)
4051
Marat Dukhan08c4a432019-10-03 09:29:21 -07004052########################### Benchmarks for operators ###########################
4053
4054xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004055 name = "average_pooling_bench",
4056 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004057 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004058 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004059 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060)
4061
4062xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004063 name = "bankers_rounding_bench",
4064 srcs = ["bench/bankers-rounding.cc"],
4065 copts = xnnpack_optional_tflite_copts(),
4066 tags = ["nowin32"],
4067 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4068)
4069
4070xnnpack_benchmark(
4071 name = "ceiling_bench",
4072 srcs = ["bench/ceiling.cc"],
4073 copts = xnnpack_optional_tflite_copts(),
4074 tags = ["nowin32"],
4075 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4076)
4077
4078xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004079 name = "channel_shuffle_bench",
4080 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004081 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004082)
4083
4084xnnpack_benchmark(
4085 name = "convolution_bench",
4086 srcs = ["bench/convolution.cc"],
4087 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004088 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004089 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004090)
4091
4092xnnpack_benchmark(
4093 name = "deconvolution_bench",
4094 srcs = ["bench/deconvolution.cc"],
4095 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004096 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004097 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098)
4099
4100xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004101 name = "floor_bench",
4102 srcs = ["bench/floor.cc"],
4103 copts = xnnpack_optional_tflite_copts(),
4104 tags = ["nowin32"],
4105 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4106)
4107
4108xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004109 name = "global_average_pooling_bench",
4110 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004111 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004112)
4113
4114xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004115 name = "hardswish_bench",
4116 srcs = ["bench/hardswish.cc"],
4117 copts = xnnpack_optional_tflite_copts(),
4118 tags = ["nowin32"],
4119 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4120)
4121
4122xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004123 name = "max_pooling_bench",
4124 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004125 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004126)
4127
4128xnnpack_benchmark(
4129 name = "sigmoid_bench",
4130 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004131 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004132 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004133 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004134)
4135
4136xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004137 name = "prelu_bench",
4138 srcs = ["bench/prelu.cc"],
4139 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004140 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004141 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004142)
4143
4144xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004145 name = "softmax_bench",
4146 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004147 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004148 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004149 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004150)
4151
Marat Dukhan87727142020-06-24 15:24:10 -07004152xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004153 name = "square_root_bench",
4154 srcs = ["bench/square-root.cc"],
4155 copts = xnnpack_optional_tflite_copts(),
4156 tags = ["nowin32"],
4157 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4158)
4159
4160xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004161 name = "truncation_bench",
4162 srcs = ["bench/truncation.cc"],
4163 deps = OPERATOR_BENCHMARK_DEPS,
4164)
4165
Marat Dukhanc068bb62019-10-04 13:24:39 -07004166############################# End-to-end benchmarks ############################
4167
4168cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004169 name = "fp32_mobilenet_v1",
4170 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004171 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004172 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004173 linkstatic = True,
4174 deps = [
4175 ":XNNPACK",
4176 "@pthreadpool",
4177 ],
4178)
4179
4180cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004181 name = "fp16_mobilenet_v1",
4182 srcs = ["models/fp16-mobilenet-v1.cc"],
4183 hdrs = ["models/models.h"],
4184 copts = xnnpack_std_cxxopts(),
4185 linkstatic = True,
4186 deps = [
4187 ":XNNPACK",
4188 "@FP16",
4189 "@pthreadpool",
4190 ],
4191)
4192
4193cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004194 name = "qs8_mobilenet_v1",
4195 srcs = ["models/qs8-mobilenet-v1.cc"],
4196 hdrs = ["models/models.h"],
4197 copts = xnnpack_std_cxxopts(),
4198 linkstatic = True,
4199 deps = [
4200 ":XNNPACK",
4201 "@pthreadpool",
4202 ],
4203)
4204
4205cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004206 name = "fp32_mobilenet_v2",
4207 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004208 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004209 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004210 linkstatic = True,
4211 deps = [
4212 ":XNNPACK",
4213 "@pthreadpool",
4214 ],
4215)
4216
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004217cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004218 name = "fp16_mobilenet_v2",
4219 srcs = ["models/fp16-mobilenet-v2.cc"],
4220 hdrs = ["models/models.h"],
4221 copts = xnnpack_std_cxxopts(),
4222 linkstatic = True,
4223 deps = [
4224 ":XNNPACK",
4225 "@FP16",
4226 "@pthreadpool",
4227 ],
4228)
4229
4230cc_library(
4231 name = "fp32_mobilenet_v3_large",
4232 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004233 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004234 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004235 linkstatic = True,
4236 deps = [
4237 ":XNNPACK",
4238 "@pthreadpool",
4239 ],
4240)
4241
4242cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004243 name = "fp32_mobilenet_v3_small",
4244 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004245 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004246 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004247 linkstatic = True,
4248 deps = [
4249 ":XNNPACK",
4250 "@pthreadpool",
4251 ],
4252)
4253
Marat Dukhanc068bb62019-10-04 13:24:39 -07004254xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004255 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004256 srcs = [
4257 "bench/f32-dwconv-e2e.cc",
4258 "bench/end2end.h",
4259 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004260 deps = MICROKERNEL_BENCHMARK_DEPS + [
4261 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004262 ":fp32_mobilenet_v1",
4263 ":fp32_mobilenet_v2",
4264 ":fp32_mobilenet_v3_large",
4265 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004266 ],
4267)
4268
4269xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004270 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004271 srcs = [
4272 "bench/f32-gemm-e2e.cc",
4273 "bench/end2end.h",
4274 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004275 deps = MICROKERNEL_BENCHMARK_DEPS + [
4276 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004277 ":fp32_mobilenet_v1",
4278 ":fp32_mobilenet_v2",
4279 ":fp32_mobilenet_v3_large",
4280 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004281 ],
4282)
4283
4284xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004285 name = "end2end_bench",
4286 srcs = ["bench/end2end.cc"],
4287 deps = [
4288 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004289 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004290 ":fp16_mobilenet_v1",
4291 ":fp16_mobilenet_v2",
4292 ":fp32_mobilenet_v1",
4293 ":fp32_mobilenet_v2",
4294 ":fp32_mobilenet_v3_large",
4295 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004296 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004297 "@pthreadpool",
4298 ],
4299)
4300
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004301#################### Accuracy evaluation for math functions ####################
4302
4303xnnpack_benchmark(
4304 name = "f32_exp_eval",
4305 srcs = [
4306 "eval/f32-exp.cc",
4307 "src/xnnpack/AlignedAllocator.h",
4308 ] + ACCURACY_EVAL_HDRS,
4309 deps = ACCURACY_EVAL_DEPS,
4310)
4311
Marat Dukhan515c9772019-10-17 18:07:57 -07004312xnnpack_benchmark(
4313 name = "f32_expminus_eval",
4314 srcs = [
4315 "eval/f32-expminus.cc",
4316 "src/xnnpack/AlignedAllocator.h",
4317 ] + ACCURACY_EVAL_HDRS,
4318 deps = ACCURACY_EVAL_DEPS,
4319)
4320
Marat Dukhan98ba4412019-10-23 02:14:28 -07004321xnnpack_benchmark(
4322 name = "f32_extexp_eval",
4323 srcs = [
4324 "eval/f32-extexp.cc",
4325 "src/xnnpack/AlignedAllocator.h",
4326 ] + ACCURACY_EVAL_HDRS,
4327 deps = ACCURACY_EVAL_DEPS,
4328)
4329
Marat Dukhan8853b822020-05-07 12:19:01 -07004330xnnpack_unit_test(
4331 name = "f32_roundne_eval",
4332 srcs = [
4333 "eval/f32-roundne.cc",
4334 "src/xnnpack/AlignedAllocator.h",
4335 "src/xnnpack/math-stubs.h",
4336 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004337 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004338 deps = MICROKERNEL_TEST_DEPS,
4339)
4340
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004341xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004342 name = "f32_roundd_eval",
4343 srcs = [
4344 "eval/f32-roundd.cc",
4345 "src/xnnpack/AlignedAllocator.h",
4346 "src/xnnpack/math-stubs.h",
4347 ] + MICROKERNEL_TEST_HDRS,
4348 automatic = False,
4349 deps = MICROKERNEL_TEST_DEPS,
4350)
4351
4352xnnpack_unit_test(
4353 name = "f32_roundu_eval",
4354 srcs = [
4355 "eval/f32-roundu.cc",
4356 "src/xnnpack/AlignedAllocator.h",
4357 "src/xnnpack/math-stubs.h",
4358 ] + MICROKERNEL_TEST_HDRS,
4359 automatic = False,
4360 deps = MICROKERNEL_TEST_DEPS,
4361)
4362
4363xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004364 name = "f32_roundz_eval",
4365 srcs = [
4366 "eval/f32-roundz.cc",
4367 "src/xnnpack/AlignedAllocator.h",
4368 "src/xnnpack/math-stubs.h",
4369 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004370 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004371 deps = MICROKERNEL_TEST_DEPS,
4372)
4373
Marat Dukhan346a9e52019-11-15 09:06:30 -08004374xnnpack_benchmark(
4375 name = "f32_sigmoid_eval",
4376 srcs = [
4377 "eval/f32-sigmoid.cc",
4378 "src/xnnpack/AlignedAllocator.h",
4379 ] + ACCURACY_EVAL_HDRS,
4380 deps = ACCURACY_EVAL_DEPS,
4381)
4382
Marat Dukhan84000762020-06-29 18:38:43 -07004383xnnpack_benchmark(
4384 name = "f32_sqrt_eval",
4385 srcs = [
4386 "eval/f32-sqrt.cc",
4387 "src/xnnpack/AlignedAllocator.h",
4388 ] + ACCURACY_EVAL_HDRS,
4389 deps = ACCURACY_EVAL_DEPS + [
4390 ":bench_utils",
4391 ],
4392)
4393
Marat Dukhan08c4a432019-10-03 09:29:21 -07004394######################### Unit tests for micro-kernels #########################
4395
4396xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004397 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004398 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004399 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004400 "test/gemm-microkernel-tester.h",
4401 "src/xnnpack/AlignedAllocator.h",
4402 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004403 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004404)
4405
4406xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004407 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004408 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004409 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004410 "test/spmm-microkernel-tester.h",
4411 "src/xnnpack/AlignedAllocator.h",
4412 ] + MICROKERNEL_TEST_HDRS,
4413 deps = MICROKERNEL_TEST_DEPS,
4414)
4415
4416xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004417 name = "f32_argmaxpool_test",
4418 srcs = [
4419 "test/f32-argmaxpool.cc",
4420 "test/argmaxpool-microkernel-tester.h",
4421 "src/xnnpack/AlignedAllocator.h",
4422 ] + MICROKERNEL_TEST_HDRS,
4423 deps = MICROKERNEL_TEST_DEPS,
4424)
4425
4426xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004427 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004428 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004429 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004430 "test/avgpool-microkernel-tester.h",
4431 "src/xnnpack/AlignedAllocator.h",
4432 ] + MICROKERNEL_TEST_HDRS,
4433 deps = MICROKERNEL_TEST_DEPS,
4434)
4435
4436xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004437 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004438 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004439 "test/f32-ibilinear.cc",
4440 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004441 "src/xnnpack/AlignedAllocator.h",
4442 ] + MICROKERNEL_TEST_HDRS,
4443 deps = MICROKERNEL_TEST_DEPS,
4444)
4445
4446xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004447 name = "f16_clamp_test",
4448 srcs = [
4449 "test/f16-clamp.cc",
4450 "test/clamp-microkernel-tester.h",
4451 ] + MICROKERNEL_TEST_HDRS,
4452 deps = MICROKERNEL_TEST_DEPS,
4453)
4454
4455xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004456 name = "f32_clamp_test",
4457 srcs = [
4458 "test/f32-clamp.cc",
4459 "test/clamp-microkernel-tester.h",
4460 ] + MICROKERNEL_TEST_HDRS,
4461 deps = MICROKERNEL_TEST_DEPS,
4462)
4463
4464xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004465 name = "f32_igemm_test",
4466 srcs = [
4467 "test/f32-igemm.cc",
4468 "test/gemm-microkernel-tester.h",
4469 "src/xnnpack/AlignedAllocator.h",
4470 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004471 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004472)
4473
4474xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004475 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004476 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004477 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 "test/gemm-microkernel-tester.h",
4479 "src/xnnpack/AlignedAllocator.h",
4480 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004481 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004482)
4483
4484xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004485 name = "f32_igemm_minmax_test",
4486 srcs = [
4487 "test/f32-igemm-minmax.cc",
4488 "test/gemm-microkernel-tester.h",
4489 "src/xnnpack/AlignedAllocator.h",
4490 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004491 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004492)
4493
4494xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004495 name = "f16_igemm_minmax_test",
4496 srcs = [
4497 "test/f16-igemm-minmax.cc",
4498 "test/gemm-microkernel-tester.h",
4499 "src/xnnpack/AlignedAllocator.h",
4500 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004501 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004502)
4503
4504xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004505 name = "f32_conv_hwc_test",
4506 srcs = [
4507 "test/f32-conv-hwc.cc",
4508 "test/conv-hwc-microkernel-tester.h",
4509 "src/xnnpack/AlignedAllocator.h",
4510 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004511 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004512)
4513
4514xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004515 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004516 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004517 "test/f32-conv-hwc2chw.cc",
4518 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004519 "src/xnnpack/AlignedAllocator.h",
4520 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004521 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004522)
4523
4524xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004525 name = "f32_dwconv_test",
4526 srcs = [
4527 "test/f32-dwconv.cc",
4528 "test/dwconv-microkernel-tester.h",
4529 "src/xnnpack/AlignedAllocator.h",
4530 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004531 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004532)
4533
4534xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004535 name = "f16_dwconv_minmax_test",
4536 srcs = [
4537 "test/f16-dwconv-minmax.cc",
4538 "test/dwconv-microkernel-tester.h",
4539 "src/xnnpack/AlignedAllocator.h",
4540 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004541 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004542)
4543
4544xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004545 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004547 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004548 "test/dwconv-microkernel-tester.h",
4549 "src/xnnpack/AlignedAllocator.h",
4550 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004551 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004552)
4553
4554xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004555 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004556 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004557 "test/f32-dwconv-chw.cc",
4558 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004559 "src/xnnpack/AlignedAllocator.h",
4560 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004561 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004562)
4563
4564xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004565 name = "f16_gavgpool_minmax_test",
4566 srcs = [
4567 "test/f16-gavgpool-minmax.cc",
4568 "test/gavgpool-microkernel-tester.h",
4569 "src/xnnpack/AlignedAllocator.h",
4570 ] + MICROKERNEL_TEST_HDRS,
4571 deps = MICROKERNEL_TEST_DEPS,
4572)
4573
4574xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004575 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004576 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004577 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004578 "test/gavgpool-microkernel-tester.h",
4579 "src/xnnpack/AlignedAllocator.h",
4580 ] + MICROKERNEL_TEST_HDRS,
4581 deps = MICROKERNEL_TEST_DEPS,
4582)
4583
4584xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004585 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004586 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004587 "test/f32-gavgpool-cw.cc",
4588 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004589 "src/xnnpack/AlignedAllocator.h",
4590 ] + MICROKERNEL_TEST_HDRS,
4591 deps = MICROKERNEL_TEST_DEPS,
4592)
4593
4594xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004595 name = "f32_gemm_test",
4596 srcs = [
4597 "test/f32-gemm.cc",
4598 "test/gemm-microkernel-tester.h",
4599 "src/xnnpack/AlignedAllocator.h",
4600 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004601 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004602)
4603
4604xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004605 name = "f32_gemm_relu_test",
4606 srcs = [
4607 "test/f32-gemm-relu.cc",
4608 "test/gemm-microkernel-tester.h",
4609 "src/xnnpack/AlignedAllocator.h",
4610 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004611 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004612)
4613
4614xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004615 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004616 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004617 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004618 "test/gemm-microkernel-tester.h",
4619 "src/xnnpack/AlignedAllocator.h",
4620 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004621 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004622)
4623
4624xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004625 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004627 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004628 "test/gemm-microkernel-tester.h",
4629 "src/xnnpack/AlignedAllocator.h",
4630 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004631 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004632)
4633
4634xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004635 name = "f16_hswish_test",
4636 srcs = [
4637 "test/f16-hswish.cc",
4638 "test/hswish-microkernel-tester.h",
4639 ] + MICROKERNEL_TEST_HDRS,
4640 deps = MICROKERNEL_TEST_DEPS,
4641)
4642
4643xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004644 name = "f32_hswish_test",
4645 srcs = [
4646 "test/f32-hswish.cc",
4647 "test/hswish-microkernel-tester.h",
4648 ] + MICROKERNEL_TEST_HDRS,
4649 deps = MICROKERNEL_TEST_DEPS,
4650)
4651
4652xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004653 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004654 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004655 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656 "test/maxpool-microkernel-tester.h",
4657 ] + MICROKERNEL_TEST_HDRS,
4658 deps = MICROKERNEL_TEST_DEPS,
4659)
4660
4661xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004662 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004663 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004664 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004665 "test/avgpool-microkernel-tester.h",
4666 "src/xnnpack/AlignedAllocator.h",
4667 ] + MICROKERNEL_TEST_HDRS,
4668 deps = MICROKERNEL_TEST_DEPS,
4669)
4670
4671xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004672 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004674 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004675 "test/gemm-microkernel-tester.h",
4676 "src/xnnpack/AlignedAllocator.h",
4677 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004678 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004679)
4680
4681xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004682 name = "f16_prelu_test",
4683 srcs = [
4684 "test/f16-prelu.cc",
4685 "test/prelu-microkernel-tester.h",
4686 "src/xnnpack/AlignedAllocator.h",
4687 ] + MICROKERNEL_TEST_HDRS,
4688 deps = MICROKERNEL_TEST_DEPS,
4689)
4690
4691xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004692 name = "f32_prelu_test",
4693 srcs = [
4694 "test/f32-prelu.cc",
4695 "test/prelu-microkernel-tester.h",
4696 "src/xnnpack/AlignedAllocator.h",
4697 ] + MICROKERNEL_TEST_HDRS,
4698 deps = MICROKERNEL_TEST_DEPS,
4699)
4700
4701xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004702 name = "f32_raddexpminusmax_test",
4703 srcs = [
4704 "test/f32-raddexpminusmax.cc",
4705 "test/raddexpminusmax-microkernel-tester.h",
4706 ] + MICROKERNEL_TEST_HDRS,
4707 deps = MICROKERNEL_TEST_DEPS,
4708)
4709
4710xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004711 name = "f32_raddextexp_test",
4712 srcs = [
4713 "test/f32-raddextexp.cc",
4714 "test/raddextexp-microkernel-tester.h",
4715 ] + MICROKERNEL_TEST_HDRS,
4716 deps = MICROKERNEL_TEST_DEPS,
4717)
4718
4719xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004720 name = "f32_raddstoreexpminusmax_test",
4721 srcs = [
4722 "test/f32-raddstoreexpminusmax.cc",
4723 "test/raddstoreexpminusmax-microkernel-tester.h",
4724 ] + MICROKERNEL_TEST_HDRS,
4725 deps = MICROKERNEL_TEST_DEPS,
4726)
4727
4728xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004729 name = "f32_relu_test",
4730 srcs = [
4731 "test/f32-relu.cc",
4732 "test/vunary-microkernel-tester.h",
4733 ] + MICROKERNEL_TEST_HDRS,
4734 deps = MICROKERNEL_TEST_DEPS,
4735)
4736
4737xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738 name = "f32_rmax_test",
4739 srcs = [
4740 "test/f32-rmax.cc",
4741 "test/rmax-microkernel-tester.h",
4742 ] + MICROKERNEL_TEST_HDRS,
4743 deps = MICROKERNEL_TEST_DEPS,
4744)
4745
4746xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004747 name = "f32_sigmoid_test",
4748 srcs = [
4749 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004750 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004751 ] + MICROKERNEL_TEST_HDRS,
4752 deps = MICROKERNEL_TEST_DEPS,
4753)
4754
4755xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004756 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004757 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004758 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 "test/spmm-microkernel-tester.h",
4760 "src/xnnpack/AlignedAllocator.h",
4761 ] + MICROKERNEL_TEST_HDRS,
4762 deps = MICROKERNEL_TEST_DEPS,
4763)
4764
4765xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004766 name = "f16_vadd_minmax_test",
4767 srcs = [
4768 "test/f16-vadd-minmax.cc",
4769 "test/vbinary-microkernel-tester.h",
4770 ] + MICROKERNEL_TEST_HDRS,
4771 deps = MICROKERNEL_TEST_DEPS,
4772)
4773
4774xnnpack_unit_test(
4775 name = "f16_vaddc_minmax_test",
4776 srcs = [
4777 "test/f16-vaddc-minmax.cc",
4778 "test/vbinaryc-microkernel-tester.h",
4779 ] + MICROKERNEL_TEST_HDRS,
4780 deps = MICROKERNEL_TEST_DEPS,
4781)
4782
4783xnnpack_unit_test(
4784 name = "f16_vdiv_minmax_test",
4785 srcs = [
4786 "test/f16-vdiv-minmax.cc",
4787 "test/vbinary-microkernel-tester.h",
4788 ] + MICROKERNEL_TEST_HDRS,
4789 deps = MICROKERNEL_TEST_DEPS,
4790)
4791
4792xnnpack_unit_test(
4793 name = "f16_vdivc_minmax_test",
4794 srcs = [
4795 "test/f16-vdivc-minmax.cc",
4796 "test/vbinaryc-microkernel-tester.h",
4797 ] + MICROKERNEL_TEST_HDRS,
4798 deps = MICROKERNEL_TEST_DEPS,
4799)
4800
4801xnnpack_unit_test(
4802 name = "f16_vrdivc_minmax_test",
4803 srcs = [
4804 "test/f16-vrdivc-minmax.cc",
4805 "test/vbinaryc-microkernel-tester.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
4811 name = "f16_vmax_test",
4812 srcs = [
4813 "test/f16-vmax.cc",
4814 "test/vbinary-microkernel-tester.h",
4815 ] + MICROKERNEL_TEST_HDRS,
4816 deps = MICROKERNEL_TEST_DEPS,
4817)
4818
4819xnnpack_unit_test(
4820 name = "f16_vmaxc_test",
4821 srcs = [
4822 "test/f16-vmaxc.cc",
4823 "test/vbinaryc-microkernel-tester.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
4829 name = "f16_vmin_test",
4830 srcs = [
4831 "test/f16-vmin.cc",
4832 "test/vbinary-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
4838 name = "f16_vminc_test",
4839 srcs = [
4840 "test/f16-vminc.cc",
4841 "test/vbinaryc-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
4847 name = "f16_vmul_minmax_test",
4848 srcs = [
4849 "test/f16-vmul-minmax.cc",
4850 "test/vbinary-microkernel-tester.h",
4851 ] + MICROKERNEL_TEST_HDRS,
4852 deps = MICROKERNEL_TEST_DEPS,
4853)
4854
4855xnnpack_unit_test(
4856 name = "f16_vmulc_minmax_test",
4857 srcs = [
4858 "test/f16-vmulc-minmax.cc",
4859 "test/vbinaryc-microkernel-tester.h",
4860 ] + MICROKERNEL_TEST_HDRS,
4861 deps = MICROKERNEL_TEST_DEPS,
4862)
4863
4864xnnpack_unit_test(
4865 name = "f16_vsub_minmax_test",
4866 srcs = [
4867 "test/f16-vsub-minmax.cc",
4868 "test/vbinary-microkernel-tester.h",
4869 ] + MICROKERNEL_TEST_HDRS,
4870 deps = MICROKERNEL_TEST_DEPS,
4871)
4872
4873xnnpack_unit_test(
4874 name = "f16_vsubc_minmax_test",
4875 srcs = [
4876 "test/f16-vsubc-minmax.cc",
4877 "test/vbinaryc-microkernel-tester.h",
4878 ] + MICROKERNEL_TEST_HDRS,
4879 deps = MICROKERNEL_TEST_DEPS,
4880)
4881
4882xnnpack_unit_test(
4883 name = "f16_vrsubc_minmax_test",
4884 srcs = [
4885 "test/f16-vrsubc-minmax.cc",
4886 "test/vbinaryc-microkernel-tester.h",
4887 ] + MICROKERNEL_TEST_HDRS,
4888 deps = MICROKERNEL_TEST_DEPS,
4889)
4890
4891xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07004892 name = "f32_vabs_test",
4893 srcs = [
4894 "test/f32-vabs.cc",
4895 "test/vunary-microkernel-tester.h",
4896 ] + MICROKERNEL_TEST_HDRS,
4897 deps = MICROKERNEL_TEST_DEPS,
4898)
4899
4900xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004901 name = "f32_vadd_test",
4902 srcs = [
4903 "test/f32-vadd.cc",
4904 "test/vbinary-microkernel-tester.h",
4905 ] + MICROKERNEL_TEST_HDRS,
4906 deps = MICROKERNEL_TEST_DEPS,
4907)
4908
4909xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004910 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004911 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004912 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004913 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004914 ] + MICROKERNEL_TEST_HDRS,
4915 deps = MICROKERNEL_TEST_DEPS,
4916)
4917
4918xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004919 name = "f32_vaddc_test",
4920 srcs = [
4921 "test/f32-vaddc.cc",
4922 "test/vbinaryc-microkernel-tester.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004928 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08004929 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004930 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004931 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004937 name = "f32_vdiv_test",
4938 srcs = [
4939 "test/f32-vdiv.cc",
4940 "test/vbinary-microkernel-tester.h",
4941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004946 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004947 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004948 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004949 "test/vbinary-microkernel-tester.h",
4950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004955 name = "f32_vdivc_test",
4956 srcs = [
4957 "test/f32-vdivc.cc",
4958 "test/vbinaryc-microkernel-tester.h",
4959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004964 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004965 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004966 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004967 "test/vbinaryc-microkernel-tester.h",
4968 ] + MICROKERNEL_TEST_HDRS,
4969 deps = MICROKERNEL_TEST_DEPS,
4970)
4971
4972xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07004973 name = "f32_vrdivc_test",
4974 srcs = [
4975 "test/f32-vrdivc.cc",
4976 "test/vbinaryc-microkernel-tester.h",
4977 ] + MICROKERNEL_TEST_HDRS,
4978 deps = MICROKERNEL_TEST_DEPS,
4979)
4980
4981xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004982 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004983 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07004984 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08004985 "test/vbinaryc-microkernel-tester.h",
4986 ] + MICROKERNEL_TEST_HDRS,
4987 deps = MICROKERNEL_TEST_DEPS,
4988)
4989
4990xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08004991 name = "f32_vmax_test",
4992 srcs = [
4993 "test/f32-vmax.cc",
4994 "test/vbinary-microkernel-tester.h",
4995 ] + MICROKERNEL_TEST_HDRS,
4996 deps = MICROKERNEL_TEST_DEPS,
4997)
4998
4999xnnpack_unit_test(
5000 name = "f32_vmaxc_test",
5001 srcs = [
5002 "test/f32-vmaxc.cc",
5003 "test/vbinaryc-microkernel-tester.h",
5004 ] + MICROKERNEL_TEST_HDRS,
5005 deps = MICROKERNEL_TEST_DEPS,
5006)
5007
5008xnnpack_unit_test(
5009 name = "f32_vmin_test",
5010 srcs = [
5011 "test/f32-vmin.cc",
5012 "test/vbinary-microkernel-tester.h",
5013 ] + MICROKERNEL_TEST_HDRS,
5014 deps = MICROKERNEL_TEST_DEPS,
5015)
5016
5017xnnpack_unit_test(
5018 name = "f32_vminc_test",
5019 srcs = [
5020 "test/f32-vminc.cc",
5021 "test/vbinaryc-microkernel-tester.h",
5022 ] + MICROKERNEL_TEST_HDRS,
5023 deps = MICROKERNEL_TEST_DEPS,
5024)
5025
5026xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005027 name = "f32_vmul_test",
5028 srcs = [
5029 "test/f32-vmul.cc",
5030 "test/vbinary-microkernel-tester.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005036 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005037 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005038 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005039 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005045 name = "f32_vmulc_test",
5046 srcs = [
5047 "test/f32-vmulc.cc",
5048 "test/vbinaryc-microkernel-tester.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005054 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005055 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005056 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005057 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005063 name = "f16_vmulcaddc_minmax_test",
5064 srcs = [
5065 "test/f16-vmulcaddc-minmax.cc",
5066 "test/vmulcaddc-microkernel-tester.h",
5067 "src/xnnpack/AlignedAllocator.h",
5068 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005069 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005070)
5071
5072xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005073 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005075 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005076 "test/vmulcaddc-microkernel-tester.h",
5077 "src/xnnpack/AlignedAllocator.h",
5078 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005079 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005080)
5081
5082xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005083 name = "f32_vlrelu_test",
5084 srcs = [
5085 "test/f32-vlrelu.cc",
5086 "test/vunary-microkernel-tester.h",
5087 ] + MICROKERNEL_TEST_HDRS,
5088 deps = MICROKERNEL_TEST_DEPS,
5089)
5090
5091xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005092 name = "f32_vneg_test",
5093 srcs = [
5094 "test/f32-vneg.cc",
5095 "test/vunary-microkernel-tester.h",
5096 ] + MICROKERNEL_TEST_HDRS,
5097 deps = MICROKERNEL_TEST_DEPS,
5098)
5099
5100xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005101 name = "f32_vrndne_test",
5102 srcs = [
5103 "test/f32-vrndne.cc",
5104 "test/vunary-microkernel-tester.h",
5105 ] + MICROKERNEL_TEST_HDRS,
5106 deps = MICROKERNEL_TEST_DEPS,
5107)
5108
5109xnnpack_unit_test(
5110 name = "f32_vrndz_test",
5111 srcs = [
5112 "test/f32-vrndz.cc",
5113 "test/vunary-microkernel-tester.h",
5114 ] + MICROKERNEL_TEST_HDRS,
5115 deps = MICROKERNEL_TEST_DEPS,
5116)
5117
5118xnnpack_unit_test(
5119 name = "f32_vrndu_test",
5120 srcs = [
5121 "test/f32-vrndu.cc",
5122 "test/vunary-microkernel-tester.h",
5123 ] + MICROKERNEL_TEST_HDRS,
5124 deps = MICROKERNEL_TEST_DEPS,
5125)
5126
5127xnnpack_unit_test(
5128 name = "f32_vrndd_test",
5129 srcs = [
5130 "test/f32-vrndd.cc",
5131 "test/vunary-microkernel-tester.h",
5132 ] + MICROKERNEL_TEST_HDRS,
5133 deps = MICROKERNEL_TEST_DEPS,
5134)
5135
5136xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005137 name = "f32_vscale_test",
5138 srcs = [
5139 "test/f32-vscale.cc",
5140 "test/vscale-microkernel-tester.h",
5141 ] + MICROKERNEL_TEST_HDRS,
5142 deps = MICROKERNEL_TEST_DEPS,
5143)
5144
5145xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005146 name = "f32_vscaleexpminusmax_test",
5147 srcs = [
5148 "test/f32-vscaleexpminusmax.cc",
5149 "test/vscaleexpminusmax-microkernel-tester.h",
5150 ] + MICROKERNEL_TEST_HDRS,
5151 deps = MICROKERNEL_TEST_DEPS,
5152)
5153
5154xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005155 name = "f32_vscaleextexp_test",
5156 srcs = [
5157 "test/f32-vscaleextexp.cc",
5158 "test/vscaleextexp-microkernel-tester.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005164 name = "f32_vsqr_test",
5165 srcs = [
5166 "test/f32-vsqr.cc",
5167 "test/vunary-microkernel-tester.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005173 name = "f32_vsqrdiff_test",
5174 srcs = [
5175 "test/f32-vsqrdiff.cc",
5176 "test/vbinary-microkernel-tester.h",
5177 ] + MICROKERNEL_TEST_HDRS,
5178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
5182 name = "f32_vsqrdiffc_test",
5183 srcs = [
5184 "test/f32-vsqrdiffc.cc",
5185 "test/vbinaryc-microkernel-tester.h",
5186 ] + MICROKERNEL_TEST_HDRS,
5187 deps = MICROKERNEL_TEST_DEPS,
5188)
5189
5190xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005191 name = "f32_vsqrt_test",
5192 srcs = [
5193 "test/f32-vsqrt.cc",
5194 "test/vunary-microkernel-tester.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005200 name = "f32_vsub_test",
5201 srcs = [
5202 "test/f32-vsub.cc",
5203 "test/vbinary-microkernel-tester.h",
5204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005209 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005210 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005211 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005212 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005213 ] + MICROKERNEL_TEST_HDRS,
5214 deps = MICROKERNEL_TEST_DEPS,
5215)
5216
5217xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005218 name = "f32_vsubc_test",
5219 srcs = [
5220 "test/f32-vsubc.cc",
5221 "test/vbinaryc-microkernel-tester.h",
5222 ] + MICROKERNEL_TEST_HDRS,
5223 deps = MICROKERNEL_TEST_DEPS,
5224)
5225
5226xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005227 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005228 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005229 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005230 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005231 ] + MICROKERNEL_TEST_HDRS,
5232 deps = MICROKERNEL_TEST_DEPS,
5233)
5234
5235xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005236 name = "f32_vrsubc_test",
5237 srcs = [
5238 "test/f32-vrsubc.cc",
5239 "test/vbinaryc-microkernel-tester.h",
5240 ] + MICROKERNEL_TEST_HDRS,
5241 deps = MICROKERNEL_TEST_DEPS,
5242)
5243
5244xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005245 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005246 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005247 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005248 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005249 ] + MICROKERNEL_TEST_HDRS,
5250 deps = MICROKERNEL_TEST_DEPS,
5251)
5252
5253xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005254 name = "qs8_dwconv_minmax_test",
5255 srcs = [
5256 "test/qs8-dwconv-minmax.cc",
5257 "test/dwconv-microkernel-tester.h",
5258 "src/xnnpack/AlignedAllocator.h",
5259 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5260 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5261)
5262
5263xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005264 name = "qs8_gavgpool_minmax_test",
5265 srcs = [
5266 "test/qs8-gavgpool-minmax.cc",
5267 "test/gavgpool-microkernel-tester.h",
5268 "src/xnnpack/AlignedAllocator.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005274 name = "qs8_gemm_minmax_test",
5275 srcs = [
5276 "test/qs8-gemm-minmax.cc",
5277 "test/gemm-microkernel-tester.h",
5278 "src/xnnpack/AlignedAllocator.h",
5279 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5280 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5281)
5282
5283xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005284 name = "qs8_igemm_minmax_test",
5285 srcs = [
5286 "test/qs8-igemm-minmax.cc",
5287 "test/gemm-microkernel-tester.h",
5288 "src/xnnpack/AlignedAllocator.h",
5289 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5290 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5291)
5292
5293xnnpack_unit_test(
5294 name = "qs8_requantization_test",
5295 srcs = [
5296 "src/xnnpack/requantization-stubs.h",
5297 "test/qs8-requantization.cc",
5298 "test/requantization-tester.h",
5299 ] + MICROKERNEL_TEST_HDRS,
5300 deps = MICROKERNEL_TEST_DEPS,
5301)
5302
5303xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005304 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005305 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005306 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005307 "test/avgpool-microkernel-tester.h",
5308 "src/xnnpack/AlignedAllocator.h",
5309 ] + MICROKERNEL_TEST_HDRS,
5310 deps = MICROKERNEL_TEST_DEPS,
5311)
5312
5313xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005314 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005315 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005316 "test/qu8-dwconv-minmax.cc",
5317 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005318 "src/xnnpack/AlignedAllocator.h",
5319 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005320 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005321)
5322
5323xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005324 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005326 "test/qu8-igemm-minmax.cc",
5327 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005328 "src/xnnpack/AlignedAllocator.h",
5329 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005330 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005331)
5332
5333xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005334 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005335 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005336 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005337 "test/gavgpool-microkernel-tester.h",
5338 "src/xnnpack/AlignedAllocator.h",
5339 ] + MICROKERNEL_TEST_HDRS,
5340 deps = MICROKERNEL_TEST_DEPS,
5341)
5342
5343xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005344 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005345 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005346 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005347 "test/gemm-microkernel-tester.h",
5348 "src/xnnpack/AlignedAllocator.h",
5349 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005350 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005351)
5352
5353xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005354 name = "qu8_requantization_test",
5355 srcs = [
5356 "src/xnnpack/requantization-stubs.h",
5357 "test/qu8-requantization.cc",
5358 "test/requantization-tester.h",
5359 ] + MICROKERNEL_TEST_HDRS,
5360 deps = MICROKERNEL_TEST_DEPS,
5361)
5362
5363xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005364 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005366 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005367 "test/vadd-microkernel-tester.h",
5368 ] + MICROKERNEL_TEST_HDRS,
5369 deps = MICROKERNEL_TEST_DEPS,
5370)
5371
5372xnnpack_unit_test(
5373 name = "u8_clamp_test",
5374 srcs = [
5375 "test/u8-clamp.cc",
5376 "test/clamp-microkernel-tester.h",
5377 ] + MICROKERNEL_TEST_HDRS,
5378 deps = MICROKERNEL_TEST_DEPS,
5379)
5380
5381xnnpack_unit_test(
5382 name = "u8_lut32norm_test",
5383 srcs = [
5384 "test/u8-lut32norm.cc",
5385 "test/lut-norm-microkernel-tester.h",
5386 ] + MICROKERNEL_TEST_HDRS,
5387 deps = MICROKERNEL_TEST_DEPS,
5388)
5389
5390xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005391 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005392 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005393 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 "test/maxpool-microkernel-tester.h",
5395 ] + MICROKERNEL_TEST_HDRS,
5396 deps = MICROKERNEL_TEST_DEPS,
5397)
5398
5399xnnpack_unit_test(
5400 name = "u8_rmax_test",
5401 srcs = [
5402 "test/u8-rmax.cc",
5403 "test/rmax-microkernel-tester.h",
5404 ] + MICROKERNEL_TEST_HDRS,
5405 deps = MICROKERNEL_TEST_DEPS,
5406)
5407
5408xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005409 name = "x32_fill_test",
5410 srcs = [
5411 "test/x32-fill.cc",
5412 "test/fill-microkernel-tester.h",
5413 ] + MICROKERNEL_TEST_HDRS,
5414 deps = MICROKERNEL_TEST_DEPS,
5415)
5416
5417xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005418 name = "x32_packx_test",
5419 srcs = [
5420 "test/x32-packx.cc",
5421 "test/pack-microkernel-tester.h",
5422 "src/xnnpack/AlignedAllocator.h",
5423 ] + MICROKERNEL_TEST_HDRS,
5424 deps = MICROKERNEL_TEST_DEPS,
5425)
5426
5427xnnpack_unit_test(
5428 name = "x32_pad_test",
5429 srcs = [
5430 "test/x32-pad.cc",
5431 "test/pad-microkernel-tester.h",
5432 ] + MICROKERNEL_TEST_HDRS,
5433 deps = MICROKERNEL_TEST_DEPS,
5434)
5435
5436xnnpack_unit_test(
5437 name = "x32_unpool_test",
5438 srcs = [
5439 "test/x32-unpool.cc",
5440 "test/unpool-microkernel-tester.h",
5441 ] + MICROKERNEL_TEST_HDRS,
5442 deps = MICROKERNEL_TEST_DEPS,
5443)
5444
5445xnnpack_unit_test(
5446 name = "x32_zip_test",
5447 srcs = [
5448 "test/x32-zip.cc",
5449 "test/zip-microkernel-tester.h",
5450 ] + MICROKERNEL_TEST_HDRS,
5451 deps = MICROKERNEL_TEST_DEPS,
5452)
5453
5454xnnpack_unit_test(
5455 name = "x8_lut_test",
5456 srcs = [
5457 "test/x8-lut.cc",
5458 "test/lut-microkernel-tester.h",
5459 ] + MICROKERNEL_TEST_HDRS,
5460 deps = MICROKERNEL_TEST_DEPS,
5461)
5462
5463xnnpack_unit_test(
5464 name = "x8_zip_test",
5465 srcs = [
5466 "test/x8-zip.cc",
5467 "test/zip-microkernel-tester.h",
5468 ] + MICROKERNEL_TEST_HDRS,
5469 deps = MICROKERNEL_TEST_DEPS,
5470)
5471
Marat Dukhan20c3b922020-03-10 03:45:06 -07005472########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005473
5474xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005475 name = "operator_size_test",
5476 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005477 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005478)
5479
Marat Dukhan20c3b922020-03-10 03:45:06 -07005480xnnpack_binary(
5481 name = "subgraph_size_test",
5482 srcs = ["test/subgraph-size.c"],
5483 deps = [":XNNPACK"],
5484)
5485
5486########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005487
5488xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005489 name = "abs_nc_test",
5490 srcs = [
5491 "test/abs-nc.cc",
5492 "test/abs-operator-tester.h",
5493 ],
5494 deps = OPERATOR_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005498 name = "add_nd_test",
5499 srcs = [
5500 "test/add-nd.cc",
5501 "test/binary-elementwise-operator-tester.h",
5502 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005503 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005504)
5505
5506xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005507 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005508 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005509 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005510 "test/argmax-pooling-operator-tester.h",
5511 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005512 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005513)
5514
5515xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005516 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005517 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005518 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005519 "test/average-pooling-operator-tester.h",
5520 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005521 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005522)
5523
5524xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005525 name = "bankers_rounding_nc_test",
5526 srcs = [
5527 "test/bankers-rounding-nc.cc",
5528 "test/bankers-rounding-operator-tester.h",
5529 ],
5530 deps = OPERATOR_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
5534 name = "ceiling_nc_test",
5535 srcs = [
5536 "test/ceiling-nc.cc",
5537 "test/ceiling-operator-tester.h",
5538 ],
5539 deps = OPERATOR_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005543 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005544 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005545 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005546 "test/channel-shuffle-operator-tester.h",
5547 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005548 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005549)
5550
5551xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005552 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005553 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005554 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005555 "test/clamp-operator-tester.h",
5556 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005557 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005558)
5559
5560xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005561 name = "constant_pad_nd_test",
5562 srcs = [
5563 "test/constant-pad-nd.cc",
5564 "test/constant-pad-operator-tester.h",
5565 ],
5566 deps = OPERATOR_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005570 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005571 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005572 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005573 "test/convolution-operator-tester.h",
5574 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005575 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005576)
5577
5578xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005579 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005580 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005581 "test/convolution-nchw.cc",
5582 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005583 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005584 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005585)
5586
5587xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005588 name = "copy_nc_test",
5589 srcs = [
5590 "test/copy-nc.cc",
5591 "test/copy-operator-tester.h",
5592 ],
5593 deps = OPERATOR_TEST_DEPS,
5594)
5595
5596xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005597 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005598 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005599 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005600 "test/deconvolution-operator-tester.h",
5601 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005602 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005603)
5604
5605xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005606 name = "divide_nd_test",
5607 srcs = [
5608 "test/binary-elementwise-operator-tester.h",
5609 "test/divide-nd.cc",
5610 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005611 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005612)
5613
5614xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005615 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005616 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005617 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005618 "test/fully-connected-operator-tester.h",
5619 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005620 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005621)
5622
5623xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005624 name = "floor_nc_test",
5625 srcs = [
5626 "test/floor-nc.cc",
5627 "test/floor-operator-tester.h",
5628 ],
5629 deps = OPERATOR_TEST_DEPS,
5630)
5631
5632xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005633 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005634 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005635 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005636 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005637 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005638 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639)
5640
5641xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005642 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005643 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005644 "test/global-average-pooling-ncw.cc",
5645 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005646 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005647 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005648)
5649
5650xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005651 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005653 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005654 "test/hardswish-operator-tester.h",
5655 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005656 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005657)
5658
5659xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005660 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005661 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005662 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005663 "test/leaky-relu-operator-tester.h",
5664 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005665 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666)
5667
5668xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005669 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005670 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005671 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672 "test/max-pooling-operator-tester.h",
5673 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005674 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005675)
5676
5677xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08005678 name = "maximum_nd_test",
5679 srcs = [
5680 "test/binary-elementwise-operator-tester.h",
5681 "test/maximum-nd.cc",
5682 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005683 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005684)
5685
5686xnnpack_unit_test(
5687 name = "minimum_nd_test",
5688 srcs = [
5689 "test/binary-elementwise-operator-tester.h",
5690 "test/minimum-nd.cc",
5691 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005692 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08005693)
5694
5695xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005696 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005697 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005698 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08005699 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08005700 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005701 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08005702)
5703
5704xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005705 name = "negate_nc_test",
5706 srcs = [
5707 "test/negate-nc.cc",
5708 "test/negate-operator-tester.h",
5709 ],
5710 deps = OPERATOR_TEST_DEPS,
5711)
5712
5713xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005714 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005715 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005716 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005717 "test/prelu-operator-tester.h",
5718 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005719 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005720)
5721
5722xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005723 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08005724 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005725 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08005726 "test/resize-bilinear-operator-tester.h",
5727 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005728 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08005729)
5730
5731xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005732 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005733 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005734 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005735 "test/sigmoid-operator-tester.h",
5736 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005737 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005738)
5739
5740xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005741 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005742 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08005743 "test/softmax-nc.cc",
5744 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005745 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005746 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005747)
5748
5749xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005750 name = "square_nc_test",
5751 srcs = [
5752 "test/square-nc.cc",
5753 "test/square-operator-tester.h",
5754 ],
5755 deps = OPERATOR_TEST_DEPS,
5756)
5757
5758xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07005759 name = "square_root_nc_test",
5760 srcs = [
5761 "test/square-root-nc.cc",
5762 "test/square-root-operator-tester.h",
5763 ],
5764 deps = OPERATOR_TEST_DEPS,
5765)
5766
5767xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07005768 name = "squared_difference_nd_test",
5769 srcs = [
5770 "test/binary-elementwise-operator-tester.h",
5771 "test/squared-difference-nd.cc",
5772 ],
5773 deps = OPERATOR_TEST_DEPS,
5774)
5775
5776xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005777 name = "subtract_nd_test",
5778 srcs = [
5779 "test/binary-elementwise-operator-tester.h",
5780 "test/subtract-nd.cc",
5781 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005782 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08005783)
5784
5785xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005786 name = "truncation_nc_test",
5787 srcs = [
5788 "test/truncation-nc.cc",
5789 "test/truncation-operator-tester.h",
5790 ],
5791 deps = OPERATOR_TEST_DEPS,
5792)
5793
5794xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005795 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005796 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005797 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005798 "test/unpooling-operator-tester.h",
5799 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005800 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005801)
5802
Chao Mei6ddfc602020-05-13 22:29:36 -07005803############################### Misc unit tests ###############################
5804
5805xnnpack_unit_test(
5806 name = "memory_planner_test",
5807 srcs = [
5808 "test/memory-planner-test.cc",
5809 ],
5810 deps = [
5811 ":XNNPACK",
5812 ":memory_planner",
5813 ],
5814)
5815
Marat Dukhan08c4a432019-10-03 09:29:21 -07005816############################# Build configurations #############################
5817
Marat Dukhanb8642352019-10-30 15:43:02 -07005818# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07005819config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005820 name = "xnn_enable_assembly_explicit_true",
5821 define_values = {"xnn_enable_assembly": "true"},
5822)
5823
5824# Disables usage of assembly kernels.
5825config_setting(
5826 name = "xnn_enable_assembly_explicit_false",
5827 define_values = {"xnn_enable_assembly": "false"},
5828)
5829
Marat Dukhan9de90e02020-06-18 16:04:12 -07005830# Enables usage of sparse inference.
5831config_setting(
5832 name = "xnn_enable_sparse_explicit_true",
5833 define_values = {"xnn_enable_sparse": "true"},
5834)
5835
5836# Disables usage of sparse inference.
5837config_setting(
5838 name = "xnn_enable_sparse_explicit_false",
5839 define_values = {"xnn_enable_sparse": "false"},
5840)
5841
Marat Dukhan05702cf2020-03-26 15:41:33 -07005842# Disables usage of HMP-aware optimizations.
5843config_setting(
5844 name = "xnn_enable_hmp_explicit_false",
5845 define_values = {"xnn_enable_hmp": "false"},
5846)
5847
Chao Mei6ddfc602020-05-13 22:29:36 -07005848# Enable usage of optimized memory allocation
5849config_setting(
5850 name = "xnn_enable_memopt_explicit_true",
5851 define_values = {"xnnpack_enable_memopt": "true"},
5852)
5853
5854# Disable usage of optimized memory allocation
5855config_setting(
5856 name = "xnn_enable_memopt_explicit_false",
5857 define_values = {"xnnpack_enable_memopt": "false"},
5858)
5859
Marat Dukhanb8642352019-10-30 15:43:02 -07005860# Builds with -c dbg
5861config_setting(
5862 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005863 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07005864 "compilation_mode": "dbg",
5865 },
5866)
5867
5868# Builds with -c opt
5869config_setting(
5870 name = "optimized_build",
5871 values = {
5872 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005873 },
5874)
5875
5876config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07005877 name = "linux_k8",
5878 values = {"cpu": "k8"},
5879)
5880
5881config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005882 name = "linux_arm",
5883 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07005884)
5885
5886config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07005887 name = "linux_armeabi",
5888 values = {"cpu": "armeabi"},
5889)
5890
5891config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07005892 name = "linux_armhf",
5893 values = {"cpu": "armhf"},
5894)
5895
5896config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07005897 name = "linux_armv7a",
5898 values = {"cpu": "armv7a"},
5899)
5900
5901config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07005902 name = "linux_aarch64",
5903 values = {"cpu": "aarch64"},
5904)
5905
5906config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005907 name = "android",
5908 values = {"crosstool_top": "//external:android/crosstool"},
5909)
5910
5911config_setting(
5912 name = "android_armv7",
5913 values = {
5914 "crosstool_top": "//external:android/crosstool",
5915 "cpu": "armeabi-v7a",
5916 },
5917)
5918
5919config_setting(
5920 name = "android_arm64",
5921 values = {
5922 "crosstool_top": "//external:android/crosstool",
5923 "cpu": "arm64-v8a",
5924 },
5925)
5926
5927config_setting(
5928 name = "android_x86",
5929 values = {
5930 "crosstool_top": "//external:android/crosstool",
5931 "cpu": "x86",
5932 },
5933)
5934
5935config_setting(
5936 name = "android_x86_64",
5937 values = {
5938 "crosstool_top": "//external:android/crosstool",
5939 "cpu": "x86_64",
5940 },
5941)
5942
5943config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005944 name = "windows_x86_64",
5945 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005946)
5947
5948config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07005949 name = "windows_x86_64_clang",
5950 values = {
5951 "compiler": "clang-cl",
5952 "cpu": "x64_windows",
5953 },
5954)
5955
5956config_setting(
5957 name = "windows_x86_64_mingw",
5958 values = {
5959 "compiler": "mingw-gcc",
5960 "cpu": "x64_windows",
5961 },
5962)
5963
5964config_setting(
5965 name = "windows_x86_64_msys",
5966 values = {
5967 "compiler": "msys-gcc",
5968 "cpu": "x64_windows",
5969 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07005970)
5971
5972config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07005973 name = "macos_x86_64",
5974 values = {
5975 "apple_platform_type": "macos",
5976 "cpu": "darwin",
5977 },
5978)
5979
5980config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005981 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005982 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07005983)
5984
5985config_setting(
5986 name = "emscripten_wasm",
5987 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005988 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005989 "cpu": "wasm",
5990 },
5991)
5992
5993config_setting(
5994 name = "emscripten_wasmsimd",
5995 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07005996 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005997 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07005998 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005999 },
6000)
6001
6002config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006003 name = "ios_armv7",
6004 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006005 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006006 "cpu": "ios_armv7",
6007 },
6008)
6009
6010config_setting(
6011 name = "ios_arm64",
6012 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006013 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006014 "cpu": "ios_arm64",
6015 },
6016)
6017
6018config_setting(
6019 name = "ios_arm64e",
6020 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006021 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006022 "cpu": "ios_arm64e",
6023 },
6024)
6025
6026config_setting(
6027 name = "ios_x86",
6028 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006029 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006030 "cpu": "ios_i386",
6031 },
6032)
6033
6034config_setting(
6035 name = "ios_x86_64",
6036 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006037 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006038 "cpu": "ios_x86_64",
6039 },
6040)
6041
6042config_setting(
6043 name = "watchos_armv7k",
6044 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006045 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006046 "cpu": "watchos_armv7k",
6047 },
6048)
6049
6050config_setting(
6051 name = "watchos_arm64_32",
6052 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006053 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006054 "cpu": "watchos_arm64_32",
6055 },
6056)
6057
6058config_setting(
6059 name = "watchos_x86",
6060 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006061 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006062 "cpu": "watchos_i386",
6063 },
6064)
6065
6066config_setting(
6067 name = "watchos_x86_64",
6068 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006069 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006070 "cpu": "watchos_x86_64",
6071 },
6072)
6073
6074config_setting(
6075 name = "tvos_arm64",
6076 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006077 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006078 "cpu": "tvos_arm64",
6079 },
6080)
6081
6082config_setting(
6083 name = "tvos_x86_64",
6084 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006085 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006086 "cpu": "tvos_x86_64",
6087 },
6088)