blob: 80ab346954ac5b3a6578ba3d9a08bfae5aa0de0c [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800219 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
222 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
225 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700228 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
229 "src/f32-spmm/gen/1x1-minmax-scalar.c",
230 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar.c",
232 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar.c",
234 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar.c",
236 "src/f32-spmm/gen/8x2-minmax-scalar.c",
237 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700238 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
239 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700241 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
242 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
243 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700244 "src/f32-vbinary/gen/vadd-scalar-x1.c",
245 "src/f32-vbinary/gen/vadd-scalar-x2.c",
246 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700247 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
248 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
249 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700250 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
251 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
252 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700253 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
254 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
255 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700256 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
257 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
258 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700259 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
260 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
261 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700262 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
263 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
264 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700265 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
266 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
267 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700268 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
269 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
270 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
272 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
273 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800274 "src/f32-vbinary/gen/vmax-scalar-x1.c",
275 "src/f32-vbinary/gen/vmax-scalar-x2.c",
276 "src/f32-vbinary/gen/vmax-scalar-x4.c",
277 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
278 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
279 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
280 "src/f32-vbinary/gen/vmin-scalar-x1.c",
281 "src/f32-vbinary/gen/vmin-scalar-x2.c",
282 "src/f32-vbinary/gen/vmin-scalar-x4.c",
283 "src/f32-vbinary/gen/vminc-scalar-x1.c",
284 "src/f32-vbinary/gen/vminc-scalar-x2.c",
285 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700286 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
287 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
288 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700289 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
290 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
291 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700292 "src/f32-vbinary/gen/vmul-scalar-x1.c",
293 "src/f32-vbinary/gen/vmul-scalar-x2.c",
294 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700295 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
296 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
297 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700298 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
299 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
300 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700301 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
302 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
303 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700304 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
305 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
306 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700310 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
311 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
312 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700313 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
314 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
315 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700316 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
317 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
318 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700319 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
320 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
321 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700322 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
323 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
324 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
325 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
326 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
327 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700328 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
329 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
330 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700334 "src/f32-vbinary/gen/vsub-scalar-x1.c",
335 "src/f32-vbinary/gen/vsub-scalar-x2.c",
336 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700337 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
338 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
339 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700340 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
341 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
342 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700343 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
344 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
345 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700346 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
347 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
348 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700349 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
350 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
351 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700352 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
353 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
354 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
355 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
356 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
357 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
358 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
359 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
360 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700361 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
362 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
363 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700364 "src/f32-vunary/gen/vabs-scalar-x1.c",
365 "src/f32-vunary/gen/vabs-scalar-x2.c",
366 "src/f32-vunary/gen/vabs-scalar-x4.c",
367 "src/f32-vunary/gen/vneg-scalar-x1.c",
368 "src/f32-vunary/gen/vneg-scalar-x2.c",
369 "src/f32-vunary/gen/vneg-scalar-x4.c",
370 "src/f32-vunary/gen/vsqr-scalar-x1.c",
371 "src/f32-vunary/gen/vsqr-scalar-x2.c",
372 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800373 "src/math/expminus-scalar-lut2048-p1.c",
374 "src/math/expminus-scalar-lut64-p2.c",
375 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700376 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700377 "src/math/roundne-scalar-nearbyint.c",
378 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700379 "src/math/roundd-scalar-addsub.c",
380 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700381 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700382 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700383 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700384 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700385 "src/math/roundz-scalar-addsub.c",
386 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700387 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800388 "src/math/sigmoid-scalar-lut2048-p1-div.c",
389 "src/math/sigmoid-scalar-lut64-p2-div.c",
390 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700391 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
392 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700393 "src/qs8-requantization/fp32-scalar-lrintf.c",
394 "src/qs8-requantization/fp32-scalar-magic.c",
395 "src/qs8-requantization/precise-scalar-signed64.c",
396 "src/qs8-requantization/precise-scalar-unsigned32.c",
397 "src/qs8-requantization/precise-scalar-unsigned64.c",
398 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700399 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
400 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
401 "src/qu8-dwconv/up1x9-minmax-scalar.c",
402 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
403 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700404 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
405 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700406 "src/qu8-gemm/2x2-minmax-scalar.c",
407 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700408 "src/qu8-requantization/fp32-scalar-lrintf.c",
409 "src/qu8-requantization/fp32-scalar-magic.c",
410 "src/qu8-requantization/precise-scalar-signed64.c",
411 "src/qu8-requantization/precise-scalar-unsigned32.c",
412 "src/qu8-requantization/precise-scalar-unsigned64.c",
413 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700414 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700415 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700416 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700417 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700418 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700419 "src/x32-fill/scalar-float.c",
420 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700421 "src/x32-packx/x2-scalar.c",
422 "src/x32-packx/x3-scalar.c",
423 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700424 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700425 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700426 "src/x32-unpool/scalar.c",
427 "src/x32-zip/x2-scalar.c",
428 "src/x32-zip/x3-scalar.c",
429 "src/x32-zip/x4-scalar.c",
430 "src/x32-zip/xm-scalar.c",
431 "src/x8-lut/scalar.c",
432 "src/x8-zip/x2-scalar.c",
433 "src/x8-zip/x3-scalar.c",
434 "src/x8-zip/x4-scalar.c",
435 "src/x8-zip/xm-scalar.c",
436]
437
Marat Dukhan436ebe62019-12-04 15:10:12 -0800438WASM_UKERNELS = [
Frank Barchardd91624e2020-08-12 13:45:56 -0700439 "src/f32-relu/wasm.S",
Marat Dukhan99936602020-04-11 16:47:01 -0700440 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
441 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700442 "src/f32-clamp/gen/wasm-x1.c",
443 "src/f32-clamp/gen/wasm-x2.c",
444 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700445 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
446 "src/f32-dwconv/gen/up1x4-wasm.c",
447 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
448 "src/f32-dwconv/gen/up1x9-wasm.c",
449 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
450 "src/f32-dwconv/gen/up1x25-wasm.c",
451 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
452 "src/f32-dwconv/gen/up2x4-wasm.c",
453 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
454 "src/f32-dwconv/gen/up2x9-wasm.c",
455 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
456 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700457 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
458 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
459 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
460 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700461 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
462 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700463 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
464 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
465 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
466 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700467 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
468 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700469 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
470 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700471 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
472 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
473 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700474 "src/f32-gemm/gen/1x4-wasm.c",
475 "src/f32-gemm/gen/2x4-wasm.c",
476 "src/f32-gemm/gen/4x2-wasm.c",
477 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700478 "src/f32-gemm/gen/1x4-relu-wasm.c",
479 "src/f32-gemm/gen/2x4-relu-wasm.c",
480 "src/f32-gemm/gen/4x2-relu-wasm.c",
481 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700482 "src/f32-gemm/gen/1x4-minmax-wasm.c",
483 "src/f32-gemm/gen/2x4-minmax-wasm.c",
484 "src/f32-gemm/gen/4x2-minmax-wasm.c",
485 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700486 "src/f32-hswish/gen/hswish-wasm-x1.c",
487 "src/f32-hswish/gen/hswish-wasm-x2.c",
488 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700489 "src/f32-igemm/gen/1x4-wasm.c",
490 "src/f32-igemm/gen/2x4-wasm.c",
491 "src/f32-igemm/gen/4x2-wasm.c",
492 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700493 "src/f32-igemm/gen/1x4-relu-wasm.c",
494 "src/f32-igemm/gen/2x4-relu-wasm.c",
495 "src/f32-igemm/gen/4x2-relu-wasm.c",
496 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700497 "src/f32-igemm/gen/1x4-minmax-wasm.c",
498 "src/f32-igemm/gen/2x4-minmax-wasm.c",
499 "src/f32-igemm/gen/4x2-minmax-wasm.c",
500 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700501 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
502 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
503 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700504 "src/f32-prelu/gen/wasm-2x1.c",
505 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700506 "src/f32-relu/gen/wasm-x1.c",
507 "src/f32-relu/gen/wasm-x2.c",
508 "src/f32-relu/gen/wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700509 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
510 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
511 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
512 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
513 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
514 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
515 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
516 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
517 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
518 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
519 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
520 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700521 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
522 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
523 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
524 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
525 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
526 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
527 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
528 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
529 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
530 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
531 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
532 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800533 "src/f32-vbinary/gen/vmax-wasm-x1.c",
534 "src/f32-vbinary/gen/vmax-wasm-x2.c",
535 "src/f32-vbinary/gen/vmax-wasm-x4.c",
536 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
537 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
538 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
539 "src/f32-vbinary/gen/vmin-wasm-x1.c",
540 "src/f32-vbinary/gen/vmin-wasm-x2.c",
541 "src/f32-vbinary/gen/vmin-wasm-x4.c",
542 "src/f32-vbinary/gen/vminc-wasm-x1.c",
543 "src/f32-vbinary/gen/vminc-wasm-x2.c",
544 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700545 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
546 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
547 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
548 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
549 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
550 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
551 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
552 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
553 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
554 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
557 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
558 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
559 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
560 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
561 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
562 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700563 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
564 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
565 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
566 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
569 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
570 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
571 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
572 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
573 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
574 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
575 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
576 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
577 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
578 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
579 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
580 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700581 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
582 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
583 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700584 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
585 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
586 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800587]
588
Marat Dukhan290055c2020-06-09 12:24:29 -0700589WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700590 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
591 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
592 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700593 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
594 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
595 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
596 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700597 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
598 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
599 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
600 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700601 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
602 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
603 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
604 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
605 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
606 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700607 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
608 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
609 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
610 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
611 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
612 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
613 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
614 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
615 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
616 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
617 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
618 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
619 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
620 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
621 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
622 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
623 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
624 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
625 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
626 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
627 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
628 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
629 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
630 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700631 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
632 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
633 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700634 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700635 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
636 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
637 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
638 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700639 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
640 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
641 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
642 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
643 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
644 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
645 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
646 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
647 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
648 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
649 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
650 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
651 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
652 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
653 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
654 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
655 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
656 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
657 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
658 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
659 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
660 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
661 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
662 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
663 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
664 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
665 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
666 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
667 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
668 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700669 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
670 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700671 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
672 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
673 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
674 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
675 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
676 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
677 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
678 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
679 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
680 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
681 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
682 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
683 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
684 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
685 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
686 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
687 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
688 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
689 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
690 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
691 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
692 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
693 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
694 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
695 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
696 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
697 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
698 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
699 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
700 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700701 "src/f32-ibilinear/gen/wasmsimd-c4.c",
702 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700703 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
704 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
705 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700706 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700707 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
708 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
709 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
710 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700711 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
712 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
713 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
714 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
715 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
716 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
717 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
718 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
719 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
720 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
721 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
722 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
723 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
724 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
725 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
726 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
727 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
728 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
729 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
730 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
731 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
732 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
733 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
734 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
735 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
736 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
737 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
738 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
739 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
740 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700741 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
742 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700743 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
744 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
745 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
746 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700747 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
748 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700749 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
750 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
751 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700752 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
753 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700754 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
755 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
756 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
757 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700758 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
759 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700760 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
761 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
762 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700763 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
764 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700765 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
766 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
767 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
768 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
769 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
770 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
771 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700772 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
773 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700774 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
775 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
776 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
777 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700778 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
779 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
780 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
781 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
782 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
783 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
784 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
785 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
786 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
787 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
788 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
789 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700790 "src/f32-relu/gen/wasmsimd-x4.c",
791 "src/f32-relu/gen/wasmsimd-x8.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700792 "src/f32-rmax/wasmsimd-arm.c",
793 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700794 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
795 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
796 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
797 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
798 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
799 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700800 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
801 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
802 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
803 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
804 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
805 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700806 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
807 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
808 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
809 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700810 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
811 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700812 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
813 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700814 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
815 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
816 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
817 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700818 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
819 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700820 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
821 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700822 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
823 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
824 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
825 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700826 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
827 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700828 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
829 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700830 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
831 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
832 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
833 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700834 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
835 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700836 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
837 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700838 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
839 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
840 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
841 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700842 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
843 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
844 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
845 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700846 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
847 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
848 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
849 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700850 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
851 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
852 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
853 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700854 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
855 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
856 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
857 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700858 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
859 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700860 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
861 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700862 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
863 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
864 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
865 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700866 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
867 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700868 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
869 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700870 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
871 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
872 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
873 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700876 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
877 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700878 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
879 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
880 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
881 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700882 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
883 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700884 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
885 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700886 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700888 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
889 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700890 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
891 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
892 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
893 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700894 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
895 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700896 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
897 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700898 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
899 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
900 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
901 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700902 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
903 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700904 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
905 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700906 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
907 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
908 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
909 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700910 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
911 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
912 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
913 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700914 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
915 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
916 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
917 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
918 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
919 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
920 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
921 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
922 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
923 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
924 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
925 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
926 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
927 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700928 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
929 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700930 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
931 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
932 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
933 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
934 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
935 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700936 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
937 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
938 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700939 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
940 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
941 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700942 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
943 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
944 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700945 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
946 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
947 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
948 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
949 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
950 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
951 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
952 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
953 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
954 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
955 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
956 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
957 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
958 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
959 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700960 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700961 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700962 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700963 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700964 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700965 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700966 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700967 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700968 "src/x32-zip/x2-wasmsimd.c",
969 "src/x32-zip/x3-wasmsimd.c",
970 "src/x32-zip/x4-wasmsimd.c",
971 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700972 "src/math/roundne-wasmsimd-addsub.c",
973 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700974 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700975 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700976 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700977 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700978 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700979 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
980 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700981]
982
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800983PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700984 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700985 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700986 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700987 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700988 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700989 "src/f32-spmm/gen/4x1-minmax-psimd.c",
990 "src/f32-spmm/gen/8x1-minmax-psimd.c",
991 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700992]
993
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800994PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700995 "src/qs8-requantization/precise-psimd.c",
996 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700997 "src/qu8-requantization/precise-psimd.c",
998 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800999]
1000
Marat Dukhan08c4a432019-10-03 09:29:21 -07001001# ISA-specific micro-kernels
1002NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001003 "src/f32-argmaxpool/4x-neon-c4.c",
1004 "src/f32-argmaxpool/9p8x-neon-c4.c",
1005 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001006 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1007 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001008 "src/f32-clamp/gen/neon-x4.c",
1009 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001010 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1011 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001012 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1013 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001014 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1015 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001016 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1017 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001018 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1019 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1020 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1021 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001022 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1023 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1024 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1025 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001026 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1027 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1028 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1029 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001030 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001031 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1032 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001033 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1034 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1035 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1036 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1038 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1039 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1040 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1041 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1042 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1044 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1045 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1046 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1047 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1048 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1049 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1050 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1051 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1052 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1053 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1054 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1055 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1056 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1057 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1058 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1059 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1060 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1061 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1062 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1063 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001064 "src/f32-hswish/gen/hswish-neon-x4.c",
1065 "src/f32-hswish/gen/hswish-neon-x8.c",
1066 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001067 "src/f32-ibilinear/gen/neon-c4.c",
1068 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001069 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1070 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1071 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1072 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1073 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1074 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1075 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1076 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1077 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1078 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1079 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1080 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1081 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1082 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1083 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1084 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001085 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1086 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1087 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001088 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1089 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001090 "src/f32-prelu/gen/neon-1x4.c",
1091 "src/f32-prelu/gen/neon-1x8.c",
1092 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001093 "src/f32-prelu/gen/neon-2x4.c",
1094 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001095 "src/f32-prelu/gen/neon-2x16.c",
1096 "src/f32-prelu/gen/neon-4x4.c",
1097 "src/f32-prelu/gen/neon-4x8.c",
1098 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001099 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1100 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1101 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1103 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1104 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1105 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1106 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1107 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1108 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1109 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1110 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1111 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1112 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1113 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1114 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1115 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1116 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1117 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1118 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1119 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1120 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1121 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1122 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001123 "src/f32-relu/gen/neon-x4.c",
1124 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001125 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001126 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001127 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1128 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1129 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1130 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1131 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1132 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1133 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1134 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1135 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1136 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1137 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1138 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1139 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1140 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1141 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1142 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1143 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1144 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001145 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1146 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1147 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1148 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001149 "src/f32-vbinary/gen/vmax-neon-x4.c",
1150 "src/f32-vbinary/gen/vmax-neon-x8.c",
1151 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1152 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1153 "src/f32-vbinary/gen/vmin-neon-x4.c",
1154 "src/f32-vbinary/gen/vmin-neon-x8.c",
1155 "src/f32-vbinary/gen/vminc-neon-x4.c",
1156 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001157 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1158 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1159 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1160 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1161 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1162 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001163 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1164 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1165 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1166 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001167 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1168 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1169 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1170 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001171 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1172 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001173 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1174 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001175 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1176 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1177 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1178 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1179 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1180 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1181 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1182 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001183 "src/f32-vunary/gen/vabs-neon-x4.c",
1184 "src/f32-vunary/gen/vabs-neon-x8.c",
1185 "src/f32-vunary/gen/vneg-neon-x4.c",
1186 "src/f32-vunary/gen/vneg-neon-x8.c",
1187 "src/f32-vunary/gen/vsqr-neon-x4.c",
1188 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001189 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1190 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1191 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1192 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001193 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1194 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1195 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1196 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1197 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1198 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1199 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1200 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001201 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1202 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1203 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1204 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1205 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1206 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1207 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1208 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001209 "src/qs8-requantization/precise-neon.c",
1210 "src/qs8-requantization/fp32-neon.c",
1211 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001212 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1213 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1214 "src/qu8-dwconv/up8x9-minmax-neon.c",
1215 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1216 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1217 "src/qu8-gemm/4x8-minmax-neon.c",
1218 "src/qu8-gemm/8x8-minmax-neon.c",
1219 "src/qu8-igemm/4x8-minmax-neon.c",
1220 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001221 "src/qu8-requantization/precise-neon.c",
1222 "src/qu8-requantization/fp32-neon.c",
1223 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001224 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001225 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001226 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001227 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001228 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001229 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001230 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001231 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001232 "src/x32-zip/x2-neon.c",
1233 "src/x32-zip/x3-neon.c",
1234 "src/x32-zip/x4-neon.c",
1235 "src/x32-zip/xm-neon.c",
1236 "src/x8-zip/x2-neon.c",
1237 "src/x8-zip/x3-neon.c",
1238 "src/x8-zip/x4-neon.c",
1239 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001240 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001241 "src/math/roundd-neon-addsub.c",
1242 "src/math/roundd-neon-cvt.c",
1243 "src/math/roundu-neon-addsub.c",
1244 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001245 "src/math/roundz-neon-addsub.c",
1246 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001247 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001248 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1249 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1250 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1251 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1252 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1253 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001254 "src/math/sqrt-neon-nr1rsqrts.c",
1255 "src/math/sqrt-neon-nr2rsqrts.c",
1256 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001257]
1258
1259NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001260 "src/f32-ibilinear/gen/neonfma-c4.c",
1261 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001262 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1263 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1264 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1265 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1266 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1267 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1268 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1269 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1270 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001271 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1272 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1273 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1274 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001275 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1276 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1277 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1278 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001279 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1280 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1281 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1282 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001283 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1284 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1285 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1286 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1287 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1288 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1289 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1290 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1291 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1292 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1293 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1294 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1295 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1296 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1297 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1298 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1299 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1300 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001301 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1302 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001303 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1304 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1305 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1306 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1307 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1308 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1309 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1310 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1311 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1312 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1313 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1314 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1315 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1316 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1317 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1318 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1319 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1320 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1321 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1322 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1323 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1324 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1325 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1326 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001327 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1328 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1329 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001381 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1382 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001383 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1384 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1385 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1386 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1387 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1388 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1389 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1390 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1391 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1392 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1393 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1394 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1395 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1396 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1397 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1398 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1399 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1400 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1401 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1402 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001403 "src/math/exp-neonfma-lut64-p2.c",
1404 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001405 "src/math/expminus-neonfma-lut2048-p1.c",
1406 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001407 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001408 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1409 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1410 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1411 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1412 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1413 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1414 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1415 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1416 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1417 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1418 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1419 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1420 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1421 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1422 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1423 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1424 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1425 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001426 "src/math/sqrt-neonfma-nr1fma.c",
1427 "src/math/sqrt-neonfma-nr2fma.c",
1428 "src/math/sqrt-neonfma-nr3fma.c",
1429 "src/math/sqrt-neonfma-nr2fma1adj.c",
1430 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001431]
1432
1433AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001434 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1435 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1436 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1437 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1438 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1439 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001440 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1441 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1442 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1443 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1444 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1445 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1446 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1447 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1448 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1449 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1450 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1451 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1452 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1453 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1454 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1455 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1456 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1457 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1458 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1459 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001460 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1461 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001462 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1463 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001464 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1465 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001466 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1467 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001468 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1469 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1470 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1471 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1472 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001473 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001491 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1492 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1493 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1494 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1495 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1496 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1497 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1498 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1499 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1500 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1501 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1502 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1503 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1504 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1505 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1506 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1507 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1508 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001509 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1510 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001511 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1512 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1513 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1514 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1515 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1516 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001517]
1518
Marat Dukhan8853b822020-05-07 12:19:01 -07001519NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001520 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1521 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1522 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1523 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1524 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1525 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1526 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1527 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001528 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001529 "src/math/roundd-neonv8.c",
1530 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001531 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001532]
1533
Marat Dukhan08c4a432019-10-03 09:29:21 -07001534AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001535 "src/f16-clamp/gen/neonfp16arith-x16.c",
1536 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001537 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1538 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1539 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1540 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1541 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1542 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1543 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1544 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1545 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1546 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1547 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1548 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001549 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1550 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001551 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1552 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001553 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001554 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001555 "src/f16-relu/gen/neonfp16arith-x16.c",
1556 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001557 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1558 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1559 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1560 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1561 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1562 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1563 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1564 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1565 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1566 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1567 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1568 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1569 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1570 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1571 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1572 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1573 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1574 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1575 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1576 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1577 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1578 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1579 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1580 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1581 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1582 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1583 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1584 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001585 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1586 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001587 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1588 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1589 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1590 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1591 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1592 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1593 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1594 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001595 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1596 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1597 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1598 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001599 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1600 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1601 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1602 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1603 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1604 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1605 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1606 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1607 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1608 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1609 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1610 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001611 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1612 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1613 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1614 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1615 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1616 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1617 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1618 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001619]
1620
Benoit Jacoba9644732020-08-13 12:48:55 -07001621NEONDOT_UKERNELS = [
1622 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1623 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1624 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
1625]
1626
Marat Dukhan08c4a432019-10-03 09:29:21 -07001627SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001628 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1629 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001630 "src/f32-clamp/gen/sse-x4.c",
1631 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001632 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1633 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001634 "src/f32-dwconv-chw/3x3p1-sse.c",
1635 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001636 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1637 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1638 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1639 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1640 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1641 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1642 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1643 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1644 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1645 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1646 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1647 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001648 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001649 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1650 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001651 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1652 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1653 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1654 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1655 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1656 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1657 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1658 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1659 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1660 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1661 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1662 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1663 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001664 "src/f32-hswish/gen/hswish-sse-x4.c",
1665 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001666 "src/f32-ibilinear/gen/sse-c4.c",
1667 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001668 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1669 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1670 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1671 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1672 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1673 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1674 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001675 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1676 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1677 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001678 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001679 "src/f32-prelu/gen/sse-2x4.c",
1680 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001681 "src/f32-relu/gen/sse-x4.c",
1682 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001683 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001684 "src/f32-spmm/gen/4x1-minmax-sse.c",
1685 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001686 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001687 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1688 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1689 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1690 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1691 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1692 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1693 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1694 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001695 "src/f32-vbinary/gen/vmax-sse-x4.c",
1696 "src/f32-vbinary/gen/vmax-sse-x8.c",
1697 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1698 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1699 "src/f32-vbinary/gen/vmin-sse-x4.c",
1700 "src/f32-vbinary/gen/vmin-sse-x8.c",
1701 "src/f32-vbinary/gen/vminc-sse-x4.c",
1702 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001703 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1704 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1705 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1706 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1707 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1708 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1709 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1710 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001711 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1712 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1713 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1714 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001715 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1716 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1717 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1718 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001719 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1720 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001721 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1722 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001723 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1724 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001725 "src/f32-vunary/gen/vabs-sse-x4.c",
1726 "src/f32-vunary/gen/vabs-sse-x8.c",
1727 "src/f32-vunary/gen/vneg-sse-x4.c",
1728 "src/f32-vunary/gen/vneg-sse-x8.c",
1729 "src/f32-vunary/gen/vsqr-sse-x4.c",
1730 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001731 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001732 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001733 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001734 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001735 "src/math/roundd-sse-addsub.c",
1736 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001737 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001738 "src/math/sqrt-sse-hh1mac.c",
1739 "src/math/sqrt-sse-nr1mac.c",
1740 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001741]
1742
1743SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001744 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1745 "src/f32-argmaxpool/4x-sse2-c4.c",
1746 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001747 "src/f32-prelu/gen/sse2-2x4.c",
1748 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001749 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1750 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1751 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1752 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1753 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1754 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1755 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1756 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1757 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1758 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1759 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1760 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001761 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001762 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001763 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001764 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001765 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1766 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001767 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1768 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001769 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1770 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1771 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1772 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1773 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1774 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1775 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1776 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001777 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1778 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1779 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001780 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1781 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1782 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001783 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1784 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1785 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001786 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1787 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001788 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1789 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001790 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1791 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001792 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1793 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001794 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001795 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1796 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001797 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001798 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1799 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1800 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001801 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1802 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001803 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1804 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001805 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1806 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001807 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001808 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1809 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001810 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001811 "src/qs8-requantization/precise-sse2.c",
1812 "src/qs8-requantization/fp32-sse2.c",
1813 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001814 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1815 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1816 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1817 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001818 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1819 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1820 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1821 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001822 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1823 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1824 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1825 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1826 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1827 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1828 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1829 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001830 "src/qu8-requantization/precise-sse2.c",
1831 "src/qu8-requantization/fp32-sse2.c",
1832 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001833 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001834 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001835 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001836 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001837 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001838 "src/x32-zip/x2-sse2.c",
1839 "src/x32-zip/x3-sse2.c",
1840 "src/x32-zip/x4-sse2.c",
1841 "src/x32-zip/xm-sse2.c",
1842 "src/x8-zip/x2-sse2.c",
1843 "src/x8-zip/x3-sse2.c",
1844 "src/x8-zip/x4-sse2.c",
1845 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001846 "src/math/exp-sse2-p5.c",
1847 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001848 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001849 "src/math/roundd-sse2-cvt.c",
1850 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001851 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001852 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001853]
1854
1855SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001856 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1857 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1858 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001859 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1860 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1861 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001862 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1863 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1864 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001865 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1866 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001867 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1868 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001869 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1870 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001871 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1872 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001873 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001874 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1875 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001876 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001877 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1878 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1879 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001880 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1881 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001882 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1883 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001884 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1885 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001886 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001887 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1888 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001889 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001890 "src/qs8-requantization/precise-ssse3.c",
1891 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001892 "src/qu8-requantization/precise-ssse3.c",
1893 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001894]
1895
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001896SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001897 "src/f32-prelu/gen/sse41-2x4.c",
1898 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001899 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001900 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001901 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001902 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001903 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1904 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001905 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1906 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001907 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1908 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1909 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1910 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1911 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1912 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1913 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1914 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001915 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1916 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1917 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001918 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1919 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1920 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001921 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1922 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1923 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001924 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1925 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001926 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1927 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001928 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1929 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001930 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1931 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001932 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001933 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1934 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001935 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001936 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1937 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1938 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001939 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1940 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001941 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1942 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001943 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1944 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001945 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001946 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1947 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001948 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001949 "src/qs8-requantization/fp32-sse4.c",
1950 "src/qs8-requantization/precise-sse4.c",
1951 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001952 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1953 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1954 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1955 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001956 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
1957 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
1958 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
1959 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001960 "src/qu8-requantization/precise-sse4.c",
1961 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001962 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001963 "src/math/roundd-sse41.c",
1964 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001965 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001966]
1967
Marat Dukhan08c4a432019-10-03 09:29:21 -07001968AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001969 "src/f32-clamp/gen/avx-x8.c",
1970 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001971 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1972 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1973 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1974 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1975 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1976 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1977 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1978 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1979 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1980 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1981 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1982 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1983 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1984 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1985 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1986 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1987 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1988 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1989 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1990 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1991 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1992 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1993 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1994 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
1995 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1996 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1997 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1998 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
1999 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2000 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002001 "src/f32-hswish/gen/hswish-avx-x8.c",
2002 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002003 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2004 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2005 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2006 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2007 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2008 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2009 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2010 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2011 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002012 "src/f32-prelu/gen/avx-2x8.c",
2013 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002014 "src/f32-relu/gen/avx-x16.c",
2015 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002016 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002017 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2018 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2019 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2020 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2021 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2022 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2023 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2024 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002025 "src/f32-vbinary/gen/vmax-avx-x8.c",
2026 "src/f32-vbinary/gen/vmax-avx-x16.c",
2027 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2028 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2029 "src/f32-vbinary/gen/vmin-avx-x8.c",
2030 "src/f32-vbinary/gen/vmin-avx-x16.c",
2031 "src/f32-vbinary/gen/vminc-avx-x8.c",
2032 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002033 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2034 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2035 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2036 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2037 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2038 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2039 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2040 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002041 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2042 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2043 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2044 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002045 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2046 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2047 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2048 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002049 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2050 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002051 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002052 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2053 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2054 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2055 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2056 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2057 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2058 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2059 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002060 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2061 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002062 "src/f32-vunary/gen/vabs-avx-x8.c",
2063 "src/f32-vunary/gen/vabs-avx-x16.c",
2064 "src/f32-vunary/gen/vneg-avx-x8.c",
2065 "src/f32-vunary/gen/vneg-avx-x16.c",
2066 "src/f32-vunary/gen/vsqr-avx-x8.c",
2067 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002068]
2069
Marat Dukhan1566fee2020-08-02 21:55:41 -07002070XOP_UKERNELS = [
2071 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2072 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2073 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2074 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002075 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2076 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002077 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2078 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002079 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002080 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2081 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002082 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002083 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2084 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2085 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002086 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2087 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2088 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2089 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2090 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2091 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002092 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002093 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2094 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002095 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002096]
2097
Marat Dukhanfda12b82019-11-21 12:27:59 -08002098FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002099 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2100 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2101 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2102 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2103 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2104 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2105 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2106 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2107 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2108 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2109 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2110 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2111 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2112 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2113 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2114 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2115 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2116 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2117 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2118 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2119 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2120 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2121 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2122 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2123 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2124 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2125 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2126 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2127 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2128 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2129 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2130 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2131 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2132 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2133 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2134 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2135 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2136 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2137 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2138 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002139 "src/f32-hswish/gen/hswish-fma3-x8.c",
2140 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002141 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2142 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2143 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2144 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2145 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2146 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2147 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2148 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2149 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2150 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2151 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2152 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2153 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2154 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002155 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2156 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2157 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2158 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2159 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2160 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2161 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2162 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002163 "src/math/sqrt-fma3-nr1fma.c",
2164 "src/math/sqrt-fma3-nr2fma.c",
2165 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002166]
2167
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002168AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002169 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2170 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2171 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2172 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2173 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2174 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2175 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2176 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2177 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2178 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2179 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2180 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2181 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2182 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2183 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2184 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2185 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2186 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2187 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2188 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2189 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2190 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2191 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2192 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2193 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2194 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2195 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2196 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2197 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2201 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2202 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2203 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2204 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002205 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2206 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2207 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2208 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2209 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2210 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2211 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2212 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2213 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2214 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2215 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2216 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2217 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2218 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2219 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2220 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2221 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2222 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2223 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2224 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2225 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2226 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2227 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2228 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2229 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2230 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2231 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2232 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2233 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2234 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002235 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2236 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2237 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2238 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2239 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2247 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2248 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2249 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2250 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2251 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2252 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2253 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2254 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2255 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2256 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2257 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2258 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002259 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2260 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002261 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2262 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2263 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2264 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002265 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2266 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2267 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002268 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2269 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2270 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002271 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2272 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2273 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002274 "src/math/exp-avx2-p5.c",
2275 "src/math/exp-avx2-perm-p3.c",
2276 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002277 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002278 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002279 "src/math/sigmoid-avx2-rr2-p5-div.c",
2280 "src/math/sigmoid-avx2-rr1-p5-div.c",
2281 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2282 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2283 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2284 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002285]
2286
Marat Dukhan08c4a432019-10-03 09:29:21 -07002287AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002288 "src/f32-clamp/gen/avx512f-x16.c",
2289 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002290 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2291 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2292 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2293 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2294 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2295 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2296 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2297 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2298 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2299 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2300 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2301 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2302 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2303 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2304 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2305 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2306 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2307 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2308 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2309 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2310 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2311 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2312 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2313 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002314 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2315 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002316 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2317 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2318 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2319 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2320 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2321 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002322 "src/f32-prelu/gen/avx512f-2x16.c",
2323 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002324 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2325 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2326 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2327 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2328 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2329 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2330 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2331 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2332 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2333 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2334 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2335 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2336 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2337 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2338 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2339 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2340 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2341 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2342 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2343 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2344 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2345 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2346 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2347 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2348 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2349 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2350 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2351 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2352 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2353 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2354 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2355 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2356 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2357 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2358 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2359 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002360 "src/f32-relu/gen/avx512f-x16.c",
2361 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002362 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002363 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2364 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2365 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2366 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2367 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2368 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2369 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2370 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002371 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2372 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2373 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2374 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2375 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2376 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2377 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2378 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002379 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2380 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2381 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2382 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2383 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2384 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2385 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2386 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002387 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2388 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2389 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2390 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002391 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2392 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2393 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2394 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002395 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2396 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002397 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002398 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2399 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2400 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2401 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2402 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2403 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2404 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2405 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2406 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2407 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2408 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2409 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2410 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2411 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2412 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2413 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2414 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2415 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2416 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2417 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2418 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2419 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2420 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2421 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002422 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2423 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2424 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2425 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2426 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2427 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2428 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2429 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002430 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2431 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2432 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2433 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2434 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2435 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2436 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2437 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002438 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2439 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2440 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2441 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2442 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2443 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002444 "src/math/exp-avx512f-p5-scalef.c",
2445 "src/math/exp-avx512f-p5.c",
2446 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002447 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002448 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002449 "src/math/sqrt-avx512f-nr1fma.c",
2450 "src/math/sqrt-avx512f-nr2fma.c",
2451 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002452]
2453
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002454AVX512SKX_UKERNELS = [
2455 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2456 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2457 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2458 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2459 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2460 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2461 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2462 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2463]
2464
Marat Dukhan08c4a432019-10-03 09:29:21 -07002465AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002466 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002467 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002468 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2469 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002470 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002471 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002472 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002473 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002474 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2475 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002476 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2477 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2478 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2479 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002480]
2481
2482AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002483 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002484 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2485 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002486 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002487 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2488 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002489 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2490 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2491 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002492 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002493 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2494 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2495 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002496 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002497 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2498 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002499 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002500 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2501 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2502 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2503 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2504 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2505 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2506 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2507 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2508 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2509 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2510 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2511 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2512 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2513 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2514 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2515 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2516 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2517 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002518 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2519 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002520 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002521 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2522 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2523 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2524 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2525 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2526 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2527 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2528 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2529 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2530 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2531 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2532 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2533 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2534 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2535 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2536 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2537 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2538 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002539 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2540 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2541 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2542 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2543 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2544 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2545 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2546 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2547 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2548 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2549 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2550 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2551 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2552 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2553 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2554 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2555 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2556 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002557]
2558
Marat Dukhan1b354632020-03-23 12:50:22 -07002559INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002560 "src/xnnpack/argmaxpool.h",
2561 "src/xnnpack/avgpool.h",
2562 "src/xnnpack/clamp.h",
2563 "src/xnnpack/common.h",
2564 "src/xnnpack/conv.h",
2565 "src/xnnpack/dwconv.h",
2566 "src/xnnpack/gavgpool.h",
2567 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002568 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002569 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002570 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002571 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002572 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002573 "src/xnnpack/lut.h",
2574 "src/xnnpack/math.h",
2575 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002576 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002577 "src/xnnpack/packx.h",
2578 "src/xnnpack/pad.h",
2579 "src/xnnpack/params.h",
2580 "src/xnnpack/pavgpool.h",
2581 "src/xnnpack/ppmm.h",
2582 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002583 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002584 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002585 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002586 "src/xnnpack/rmax.h",
2587 "src/xnnpack/scalar-utils.h",
2588 "src/xnnpack/spmm.h",
2589 "src/xnnpack/unpool.h",
2590 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002591 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002592 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002593 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002594 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002595 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002596 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002597 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002598]
2599
2600INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002601 "include/xnnpack.h",
2602 "src/xnnpack/allocator.h",
2603 "src/xnnpack/compute.h",
2604 "src/xnnpack/im2col.h",
2605 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002606 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002607 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608 "src/xnnpack/operator.h",
2609 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002610 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002611 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002612 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002613 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002614]
2615
Marat Dukhan1b354632020-03-23 12:50:22 -07002616ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002617 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002618]
2619
Marat Dukhan1b354632020-03-23 12:50:22 -07002620MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002621 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002622 "include/xnnpack.h",
2623]
2624
Marat Dukhan1b354632020-03-23 12:50:22 -07002625MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002626 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002627 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002628 "src/xnnpack/requantization.h",
2629 "include/xnnpack.h",
2630]
2631
2632OPERATOR_TEST_PARAMS_HDRS = [
2633 "src/xnnpack/params.h",
2634 "src/xnnpack/common.h",
2635]
2636
2637WEIGHTS_PACK_HDRS = [
2638 "src/xnnpack/pack.h",
2639 "src/xnnpack/operator.h",
2640 "src/xnnpack/compute.h",
2641]
2642
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002643LOGGING_COPTS = select({
2644 # No logging in optimized mode
2645 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2646 # Full logging in debug mode
2647 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2648 # Error-only logging in default (fastbuild) mode
2649 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2650})
2651
Marat Dukhan3b59de22020-06-03 20:15:19 -07002652LOGGING_SRCS = select({
2653 # No logging in optimized mode
2654 ":optimized_build": [],
2655 "//conditions:default": [
2656 "src/operator-strings.c",
2657 "src/subgraph-strings.c",
2658 ],
2659})
2660
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002661LOGGING_HDRS = [
2662 "src/xnnpack/log.h",
2663]
2664
Marat Dukhan08c4a432019-10-03 09:29:21 -07002665xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002666 name = "tables",
2667 srcs = TABLE_SRCS,
2668 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002669 gcc_copts = xnnpack_gcc_std_copts(),
2670 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002671)
2672
2673xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002674 name = "scalar_ukernels",
2675 srcs = SCALAR_UKERNELS,
2676 hdrs = INTERNAL_HDRS,
2677 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002678 gcc_copts = xnnpack_gcc_std_copts(),
2679 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002680 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002681 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002682 "@FP16",
2683 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002684 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 ],
2686)
2687
2688xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002689 name = "scalar_ukernels_test_mode",
2690 srcs = SCALAR_UKERNELS,
2691 hdrs = INTERNAL_HDRS,
2692 aarch32_copts = ["-marm"],
2693 copts = [
2694 "-UNDEBUG",
2695 "-DXNN_TEST_MODE=1",
2696 ],
2697 gcc_copts = xnnpack_gcc_std_copts(),
2698 msvc_copts = xnnpack_msvc_std_copts(),
2699 deps = [
2700 ":tables",
2701 "@FP16",
2702 "@FXdiv",
2703 "@pthreadpool",
2704 ],
2705)
2706
2707xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002708 name = "wasm_ukernels",
2709 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002710 gcc_copts = xnnpack_gcc_std_copts(),
2711 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002712 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002713 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002714 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002715 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002716 "@FP16",
2717 "@FXdiv",
2718 "@pthreadpool",
2719 ],
2720)
2721
2722xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002723 name = "wasm_ukernels_test_mode",
2724 hdrs = INTERNAL_HDRS,
2725 copts = [
2726 "-UNDEBUG",
2727 "-DXNN_TEST_MODE=1",
2728 ],
2729 gcc_copts = xnnpack_gcc_std_copts(),
2730 msvc_copts = xnnpack_msvc_std_copts(),
2731 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002732 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002733 deps = [
2734 ":tables",
2735 "@FP16",
2736 "@FXdiv",
2737 "@pthreadpool",
2738 ],
2739)
2740
2741xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002742 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002743 hdrs = INTERNAL_HDRS,
2744 aarch32_copts = [
2745 "-marm",
2746 "-mfpu=neon",
2747 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002748 gcc_copts = xnnpack_gcc_std_copts(),
2749 gcc_x86_copts = ["-msse2"],
2750 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 optimized_copts = [
2752 "-O3",
2753 "-ffast-math",
2754 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002755 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002756 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002757 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002758 "@FP16",
2759 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002760 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002761 ],
2762)
2763
2764xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002765 name = "psimd_fastmath_ukernels_test_mode",
2766 hdrs = INTERNAL_HDRS,
2767 aarch32_copts = [
2768 "-marm",
2769 "-mfpu=neon",
2770 ],
2771 copts = [
2772 "-UNDEBUG",
2773 "-DXNN_TEST_MODE=1",
2774 ],
2775 gcc_copts = xnnpack_gcc_std_copts(),
2776 gcc_x86_copts = ["-msse2"],
2777 msvc_copts = xnnpack_msvc_std_copts(),
2778 optimized_copts = [
2779 "-O3",
2780 "-ffast-math",
2781 ],
2782 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2783 deps = [
2784 ":tables",
2785 "@FP16",
2786 "@psimd",
2787 "@pthreadpool",
2788 ],
2789)
2790
2791xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002792 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002793 hdrs = INTERNAL_HDRS,
2794 aarch32_copts = [
2795 "-marm",
2796 "-mfpu=neon",
2797 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002798 gcc_copts = xnnpack_gcc_std_copts(),
2799 gcc_x86_copts = ["-msse2"],
2800 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002801 optimized_copts = [
2802 "-O3",
2803 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002804 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002805 deps = [
2806 ":tables",
2807 "@FP16",
2808 "@psimd",
2809 "@pthreadpool",
2810 ],
2811)
2812
2813xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002814 name = "psimd_accmath_ukernels_test_mode",
2815 hdrs = INTERNAL_HDRS,
2816 aarch32_copts = [
2817 "-marm",
2818 "-mfpu=neon",
2819 ],
2820 copts = [
2821 "-UNDEBUG",
2822 "-DXNN_TEST_MODE=1",
2823 ],
2824 gcc_copts = xnnpack_gcc_std_copts(),
2825 gcc_x86_copts = ["-msse2"],
2826 msvc_copts = xnnpack_msvc_std_copts(),
2827 optimized_copts = [
2828 "-O3",
2829 ],
2830 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2831 deps = [
2832 ":tables",
2833 "@FP16",
2834 "@psimd",
2835 "@pthreadpool",
2836 ],
2837)
2838
2839xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002840 name = "neon_ukernels",
2841 hdrs = INTERNAL_HDRS,
2842 aarch32_copts = [
2843 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002844 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002845 "-mfpu=neon",
2846 ],
2847 aarch32_srcs = NEON_UKERNELS,
2848 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002849 gcc_copts = xnnpack_gcc_std_copts(),
2850 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002851 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002852 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002853 "@FP16",
2854 "@pthreadpool",
2855 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002856)
2857
2858xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002859 name = "neon_ukernels_test_mode",
2860 hdrs = INTERNAL_HDRS,
2861 aarch32_copts = [
2862 "-marm",
2863 "-march=armv7-a",
2864 "-mfpu=neon",
2865 ],
2866 aarch32_srcs = NEON_UKERNELS,
2867 aarch64_srcs = NEON_UKERNELS,
2868 copts = [
2869 "-UNDEBUG",
2870 "-DXNN_TEST_MODE=1",
2871 ],
2872 gcc_copts = xnnpack_gcc_std_copts(),
2873 msvc_copts = xnnpack_msvc_std_copts(),
2874 deps = [
2875 ":tables",
2876 "@FP16",
2877 "@pthreadpool",
2878 ],
2879)
2880
2881xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002882 name = "neonfma_ukernels",
2883 hdrs = INTERNAL_HDRS,
2884 aarch32_copts = [
2885 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002886 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002887 "-mfpu=neon-vfpv4",
2888 ],
2889 aarch32_srcs = NEONFMA_UKERNELS,
2890 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002891 apple_aarch32_copts = [
2892 "-mcpu=swift",
2893 "-mtune=generic",
2894 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002895 gcc_copts = xnnpack_gcc_std_copts(),
2896 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002897 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002898 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002899 "@FP16",
2900 "@pthreadpool",
2901 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002902)
2903
2904xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002905 name = "neonfma_ukernels_test_mode",
2906 hdrs = INTERNAL_HDRS,
2907 aarch32_copts = [
2908 "-marm",
2909 "-march=armv7-a",
2910 "-mfpu=neon-vfpv4",
2911 ],
2912 aarch32_srcs = NEONFMA_UKERNELS,
2913 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002914 apple_aarch32_copts = [
2915 "-mcpu=swift",
2916 "-mtune=generic",
2917 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002918 copts = [
2919 "-UNDEBUG",
2920 "-DXNN_TEST_MODE=1",
2921 ],
2922 gcc_copts = xnnpack_gcc_std_copts(),
2923 msvc_copts = xnnpack_msvc_std_copts(),
2924 deps = [
2925 ":tables",
2926 "@FP16",
2927 "@pthreadpool",
2928 ],
2929)
2930
2931xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002932 name = "neonv8_ukernels",
2933 hdrs = INTERNAL_HDRS,
2934 aarch32_copts = [
2935 "-marm",
2936 "-march=armv8-a",
2937 "-mfpu=neon-fp-armv8",
2938 ],
2939 aarch32_srcs = NEONV8_UKERNELS,
2940 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002941 apple_aarch32_copts = [
2942 "-mcpu=cyclone",
2943 "-mtune=generic",
2944 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002945 gcc_copts = xnnpack_gcc_std_copts(),
2946 msvc_copts = xnnpack_msvc_std_copts(),
2947 deps = [
2948 ":tables",
2949 "@FP16",
2950 "@pthreadpool",
2951 ],
2952)
2953
2954xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002955 name = "neonv8_ukernels_test_mode",
2956 hdrs = INTERNAL_HDRS,
2957 aarch32_copts = [
2958 "-marm",
2959 "-march=armv8-a",
2960 "-mfpu=neon-fp-armv8",
2961 ],
2962 aarch32_srcs = NEONV8_UKERNELS,
2963 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002964 apple_aarch32_copts = [
2965 "-mcpu=cyclone",
2966 "-mtune=generic",
2967 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002968 copts = [
2969 "-UNDEBUG",
2970 "-DXNN_TEST_MODE=1",
2971 ],
2972 gcc_copts = xnnpack_gcc_std_copts(),
2973 msvc_copts = xnnpack_msvc_std_copts(),
2974 deps = [
2975 ":tables",
2976 "@FP16",
2977 "@pthreadpool",
2978 ],
2979)
2980
2981xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 name = "neonfp16arith_ukernels",
2983 hdrs = INTERNAL_HDRS,
2984 aarch64_copts = ["-march=armv8.2-a+fp16"],
2985 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002986 gcc_copts = xnnpack_gcc_std_copts(),
2987 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002988 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002989 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002990 "@FP16",
2991 "@pthreadpool",
2992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993)
2994
2995xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002996 name = "neonfp16arith_ukernels_test_mode",
2997 hdrs = INTERNAL_HDRS,
2998 aarch64_copts = ["-march=armv8.2-a+fp16"],
2999 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3000 copts = [
3001 "-UNDEBUG",
3002 "-DXNN_TEST_MODE=1",
3003 ],
3004 gcc_copts = xnnpack_gcc_std_copts(),
3005 msvc_copts = xnnpack_msvc_std_copts(),
3006 deps = [
3007 ":tables",
3008 "@FP16",
3009 "@pthreadpool",
3010 ],
3011)
3012
3013xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003014 name = "neondot_ukernels",
3015 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003016 aarch32_copts = [
3017 "-marm",
3018 "-march=armv8.2-a+dotprod",
3019 "-mfpu=neon-fp-armv8",
3020 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003021 aarch32_srcs = NEONDOT_UKERNELS,
3022 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3023 aarch64_srcs = NEONDOT_UKERNELS,
3024 gcc_copts = xnnpack_gcc_std_copts(),
3025 msvc_copts = xnnpack_msvc_std_copts(),
3026 deps = [
3027 ":tables",
3028 "@FP16",
3029 "@pthreadpool",
3030 ],
3031)
3032
3033xnnpack_cc_library(
3034 name = "neondot_ukernels_test_mode",
3035 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003036 aarch32_copts = [
3037 "-marm",
3038 "-march=armv8.2-a+dotprod",
3039 "-mfpu=neon-fp-armv8",
3040 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003041 aarch32_srcs = NEONDOT_UKERNELS,
3042 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3043 aarch64_srcs = NEONDOT_UKERNELS,
3044 copts = [
3045 "-UNDEBUG",
3046 "-DXNN_TEST_MODE=1",
3047 ],
3048 gcc_copts = xnnpack_gcc_std_copts(),
3049 msvc_copts = xnnpack_msvc_std_copts(),
3050 deps = [
3051 ":tables",
3052 "@FP16",
3053 "@pthreadpool",
3054 ],
3055)
3056
3057xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003058 name = "sse2_ukernels",
3059 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003060 gcc_copts = xnnpack_gcc_std_copts(),
3061 gcc_x86_copts = ["-msse2"],
3062 msvc_copts = xnnpack_msvc_std_copts(),
3063 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003064 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003065 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003066 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003067 "@FP16",
3068 "@pthreadpool",
3069 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003070)
3071
3072xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003073 name = "sse2_ukernels_test_mode",
3074 hdrs = INTERNAL_HDRS,
3075 copts = [
3076 "-UNDEBUG",
3077 "-DXNN_TEST_MODE=1",
3078 ],
3079 gcc_copts = xnnpack_gcc_std_copts(),
3080 gcc_x86_copts = ["-msse2"],
3081 msvc_copts = xnnpack_msvc_std_copts(),
3082 msvc_x86_32_copts = ["/arch:SSE2"],
3083 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3084 deps = [
3085 ":tables",
3086 "@FP16",
3087 "@pthreadpool",
3088 ],
3089)
3090
3091xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003092 name = "ssse3_ukernels",
3093 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003094 gcc_copts = xnnpack_gcc_std_copts(),
3095 gcc_x86_copts = ["-mssse3"],
3096 msvc_copts = xnnpack_msvc_std_copts(),
3097 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003098 x86_srcs = SSSE3_UKERNELS,
3099 deps = [
3100 ":tables",
3101 "@FP16",
3102 "@pthreadpool",
3103 ],
3104)
3105
3106xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003107 name = "ssse3_ukernels_test_mode",
3108 hdrs = INTERNAL_HDRS,
3109 copts = [
3110 "-UNDEBUG",
3111 "-DXNN_TEST_MODE=1",
3112 ],
3113 gcc_copts = xnnpack_gcc_std_copts(),
3114 gcc_x86_copts = ["-mssse3"],
3115 msvc_copts = xnnpack_msvc_std_copts(),
3116 msvc_x86_32_copts = ["/arch:SSE2"],
3117 x86_srcs = SSSE3_UKERNELS,
3118 deps = [
3119 ":tables",
3120 "@FP16",
3121 "@pthreadpool",
3122 ],
3123)
3124
3125xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003126 name = "sse41_ukernels",
3127 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003128 gcc_copts = xnnpack_gcc_std_copts(),
3129 gcc_x86_copts = ["-msse4.1"],
3130 msvc_copts = xnnpack_msvc_std_copts(),
3131 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003132 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003133 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003134 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003135 "@FP16",
3136 "@pthreadpool",
3137 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003138)
3139
3140xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003141 name = "sse41_ukernels_test_mode",
3142 hdrs = INTERNAL_HDRS,
3143 copts = [
3144 "-UNDEBUG",
3145 "-DXNN_TEST_MODE=1",
3146 ],
3147 gcc_copts = xnnpack_gcc_std_copts(),
3148 gcc_x86_copts = ["-msse4.1"],
3149 msvc_copts = xnnpack_msvc_std_copts(),
3150 msvc_x86_32_copts = ["/arch:SSE2"],
3151 x86_srcs = SSE41_UKERNELS,
3152 deps = [
3153 ":tables",
3154 "@FP16",
3155 "@pthreadpool",
3156 ],
3157)
3158
3159xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003160 name = "avx_ukernels",
3161 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003162 gcc_copts = xnnpack_gcc_std_copts(),
3163 gcc_x86_copts = ["-mavx"],
3164 msvc_copts = xnnpack_msvc_std_copts(),
3165 msvc_x86_32_copts = ["/arch:AVX"],
3166 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003167 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003168 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003169 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003170 "@FP16",
3171 "@pthreadpool",
3172 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173)
3174
3175xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003176 name = "avx_ukernels_test_mode",
3177 hdrs = INTERNAL_HDRS,
3178 copts = [
3179 "-UNDEBUG",
3180 "-DXNN_TEST_MODE=1",
3181 ],
3182 gcc_copts = xnnpack_gcc_std_copts(),
3183 gcc_x86_copts = ["-mavx"],
3184 msvc_copts = xnnpack_msvc_std_copts(),
3185 msvc_x86_32_copts = ["/arch:AVX"],
3186 msvc_x86_64_copts = ["/arch:AVX"],
3187 x86_srcs = AVX_UKERNELS,
3188 deps = [
3189 ":tables",
3190 "@FP16",
3191 "@pthreadpool",
3192 ],
3193)
3194
3195xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003196 name = "xop_ukernels",
3197 hdrs = INTERNAL_HDRS,
3198 gcc_copts = xnnpack_gcc_std_copts(),
3199 gcc_x86_copts = ["-mxop"],
3200 msvc_copts = xnnpack_msvc_std_copts(),
3201 msvc_x86_32_copts = ["/arch:AVX"],
3202 msvc_x86_64_copts = ["/arch:AVX"],
3203 x86_srcs = XOP_UKERNELS,
3204 deps = [
3205 ":tables",
3206 "@FP16",
3207 "@pthreadpool",
3208 ],
3209)
3210
3211xnnpack_cc_library(
3212 name = "xop_ukernels_test_mode",
3213 hdrs = INTERNAL_HDRS,
3214 copts = [
3215 "-UNDEBUG",
3216 "-DXNN_TEST_MODE=1",
3217 ],
3218 gcc_copts = xnnpack_gcc_std_copts(),
3219 gcc_x86_copts = ["-mxop"],
3220 msvc_copts = xnnpack_msvc_std_copts(),
3221 msvc_x86_32_copts = ["/arch:AVX"],
3222 msvc_x86_64_copts = ["/arch:AVX"],
3223 x86_srcs = XOP_UKERNELS,
3224 deps = [
3225 ":tables",
3226 "@FP16",
3227 "@pthreadpool",
3228 ],
3229)
3230
3231xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003232 name = "fma3_ukernels",
3233 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003234 gcc_copts = xnnpack_gcc_std_copts(),
3235 gcc_x86_copts = ["-mfma"],
3236 msvc_copts = xnnpack_msvc_std_copts(),
3237 msvc_x86_32_copts = ["/arch:AVX"],
3238 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003239 x86_srcs = FMA3_UKERNELS,
3240 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003241 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003242 "@FP16",
3243 "@pthreadpool",
3244 ],
3245)
3246
3247xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003248 name = "fma3_ukernels_test_mode",
3249 hdrs = INTERNAL_HDRS,
3250 copts = [
3251 "-UNDEBUG",
3252 "-DXNN_TEST_MODE=1",
3253 ],
3254 gcc_copts = xnnpack_gcc_std_copts(),
3255 gcc_x86_copts = ["-mfma"],
3256 msvc_copts = xnnpack_msvc_std_copts(),
3257 msvc_x86_32_copts = ["/arch:AVX"],
3258 msvc_x86_64_copts = ["/arch:AVX"],
3259 x86_srcs = FMA3_UKERNELS,
3260 deps = [
3261 ":tables",
3262 "@FP16",
3263 "@pthreadpool",
3264 ],
3265)
3266
3267xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003268 name = "avx2_ukernels",
3269 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003270 gcc_copts = xnnpack_gcc_std_copts(),
3271 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003272 "-mfma",
3273 "-mavx2",
3274 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003275 msvc_copts = xnnpack_msvc_std_copts(),
3276 msvc_x86_32_copts = ["/arch:AVX2"],
3277 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003278 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003279 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003280 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003281 "@FP16",
3282 "@pthreadpool",
3283 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003284)
3285
3286xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003287 name = "avx2_ukernels_test_mode",
3288 hdrs = INTERNAL_HDRS,
3289 copts = [
3290 "-UNDEBUG",
3291 "-DXNN_TEST_MODE=1",
3292 ],
3293 gcc_copts = xnnpack_gcc_std_copts(),
3294 gcc_x86_copts = [
3295 "-mfma",
3296 "-mavx2",
3297 ],
3298 msvc_copts = xnnpack_msvc_std_copts(),
3299 msvc_x86_32_copts = ["/arch:AVX2"],
3300 msvc_x86_64_copts = ["/arch:AVX2"],
3301 x86_srcs = AVX2_UKERNELS,
3302 deps = [
3303 ":tables",
3304 "@FP16",
3305 "@pthreadpool",
3306 ],
3307)
3308
3309xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003310 name = "avx512f_ukernels",
3311 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003312 gcc_copts = xnnpack_gcc_std_copts(),
3313 gcc_x86_copts = ["-mavx512f"],
3314 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3315 msvc_copts = xnnpack_msvc_std_copts(),
3316 msvc_x86_32_copts = ["/arch:AVX512"],
3317 msvc_x86_64_copts = ["/arch:AVX512"],
3318 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003319 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003320 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003321 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003322 "@FP16",
3323 "@pthreadpool",
3324 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003325)
3326
3327xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003328 name = "avx512f_ukernels_test_mode",
3329 hdrs = INTERNAL_HDRS,
3330 copts = [
3331 "-UNDEBUG",
3332 "-DXNN_TEST_MODE=1",
3333 ],
3334 gcc_copts = xnnpack_gcc_std_copts(),
3335 gcc_x86_copts = ["-mavx512f"],
3336 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3337 msvc_copts = xnnpack_msvc_std_copts(),
3338 msvc_x86_32_copts = ["/arch:AVX512"],
3339 msvc_x86_64_copts = ["/arch:AVX512"],
3340 msys_copts = ["-fno-asynchronous-unwind-tables"],
3341 x86_srcs = AVX512F_UKERNELS,
3342 deps = [
3343 ":tables",
3344 "@FP16",
3345 "@pthreadpool",
3346 ],
3347)
3348
3349xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003350 name = "avx512skx_ukernels",
3351 hdrs = INTERNAL_HDRS,
3352 gcc_copts = xnnpack_gcc_std_copts(),
3353 gcc_x86_copts = [
3354 "-mavx512f",
3355 "-mavx512cd",
3356 "-mavx512bw",
3357 "-mavx512dq",
3358 "-mavx512vl",
3359 ],
3360 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3361 msvc_copts = xnnpack_msvc_std_copts(),
3362 msvc_x86_32_copts = ["/arch:AVX512"],
3363 msvc_x86_64_copts = ["/arch:AVX512"],
3364 msys_copts = ["-fno-asynchronous-unwind-tables"],
3365 x86_srcs = AVX512SKX_UKERNELS,
3366 deps = [
3367 ":tables",
3368 "@FP16",
3369 "@pthreadpool",
3370 ],
3371)
3372
3373xnnpack_cc_library(
3374 name = "avx512skx_ukernels_test_mode",
3375 hdrs = INTERNAL_HDRS,
3376 copts = [
3377 "-UNDEBUG",
3378 "-DXNN_TEST_MODE=1",
3379 ],
3380 gcc_copts = xnnpack_gcc_std_copts(),
3381 gcc_x86_copts = [
3382 "-mavx512f",
3383 "-mavx512cd",
3384 "-mavx512bw",
3385 "-mavx512dq",
3386 "-mavx512vl",
3387 ],
3388 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3389 msvc_copts = xnnpack_msvc_std_copts(),
3390 msvc_x86_32_copts = ["/arch:AVX512"],
3391 msvc_x86_64_copts = ["/arch:AVX512"],
3392 msys_copts = ["-fno-asynchronous-unwind-tables"],
3393 x86_srcs = AVX512SKX_UKERNELS,
3394 deps = [
3395 ":tables",
3396 "@FP16",
3397 "@pthreadpool",
3398 ],
3399)
3400
3401xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003402 name = "asm_ukernels",
3403 hdrs = ["src/xnnpack/assembly.h"],
3404 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003405 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003406 aarch64_srcs = AARCH64_ASM_UKERNELS,
3407)
3408
Marat Dukhan3b59de22020-06-03 20:15:19 -07003409xnnpack_cc_library(
3410 name = "logging_utils",
3411 srcs = LOGGING_SRCS,
3412 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3413 copts = LOGGING_COPTS + [
3414 "-Isrc",
3415 "-Iinclude",
3416 ] + select({
3417 ":debug_build": [],
3418 "//conditions:default": xnnpack_min_size_copts(),
3419 }),
3420 gcc_copts = xnnpack_gcc_std_copts(),
3421 msvc_copts = xnnpack_msvc_std_copts(),
3422 visibility = xnnpack_visibility(),
3423 deps = [
3424 "@FP16",
3425 "@clog",
3426 "@pthreadpool",
3427 ],
3428)
3429
Marat Dukhan08c4a432019-10-03 09:29:21 -07003430xnnpack_aggregate_library(
3431 name = "ukernels",
3432 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003433 ":neon_ukernels",
3434 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003435 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003436 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003437 ":asm_ukernels",
3438 ],
3439 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003440 ":neon_ukernels",
3441 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003442 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003444 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003445 ":asm_ukernels",
3446 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003447 generic_deps = [
3448 ":scalar_ukernels",
3449 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003450 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003451 ":psimd_fastmath_ukernels",
3452 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003454 wasm_deps = [
3455 ":wasm_ukernels",
3456 ],
3457 wasmsimd_deps = [
3458 ":wasm_ukernels",
3459 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003460 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003461 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003462 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003463 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003464 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003465 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003466 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003467 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003468 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003469 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003470 ],
3471)
3472
Marat Dukhan33fcf782020-05-24 14:27:15 -07003473xnnpack_aggregate_library(
3474 name = "ukernels_test_mode",
3475 aarch32_deps = [
3476 ":neon_ukernels_test_mode",
3477 ":neonfma_ukernels_test_mode",
3478 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003479 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003480 ":asm_ukernels",
3481 ],
3482 aarch64_deps = [
3483 ":neon_ukernels_test_mode",
3484 ":neonfma_ukernels_test_mode",
3485 ":neonv8_ukernels_test_mode",
3486 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003487 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003488 ":asm_ukernels",
3489 ],
3490 generic_deps = [
3491 ":scalar_ukernels_test_mode",
3492 ],
3493 psimd_deps = [
3494 ":psimd_fastmath_ukernels_test_mode",
3495 ":psimd_accmath_ukernels_test_mode",
3496 ],
3497 wasm_deps = [
3498 ":wasm_ukernels_test_mode",
3499 ],
3500 wasmsimd_deps = [
3501 ":wasm_ukernels_test_mode",
3502 ],
3503 x86_deps = [
3504 ":sse2_ukernels_test_mode",
3505 ":ssse3_ukernels_test_mode",
3506 ":sse41_ukernels_test_mode",
3507 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003508 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003509 ":fma3_ukernels_test_mode",
3510 ":avx2_ukernels_test_mode",
3511 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003512 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003513 ],
3514)
3515
Marat Dukhan86df3842020-06-17 14:03:31 -07003516# This target must be compatible with Android NDK r17 and compile with Clang 6.
3517# It is similar to :ukernels target, but excludes microkernels which require
3518# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003519xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003520 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003521 aarch32_deps = [
3522 ":neon_ukernels",
3523 ":neonfma_ukernels",
3524 ":neonv8_ukernels",
3525 ":asm_ukernels",
3526 ],
3527 aarch64_deps = [
3528 ":neon_ukernels",
3529 ":neonfma_ukernels",
3530 ":neonv8_ukernels",
3531 ":asm_ukernels",
3532 ],
3533 generic_deps = [
3534 ":scalar_ukernels",
3535 ],
3536 psimd_deps = [
3537 ":psimd_fastmath_ukernels",
3538 ":psimd_accmath_ukernels",
3539 ],
3540 wasm_deps = [
3541 ":wasm_ukernels",
3542 ],
3543 wasmsimd_deps = [
3544 ":wasm_ukernels",
3545 ],
3546 x86_deps = [
3547 ":sse2_ukernels",
3548 ":ssse3_ukernels",
3549 ":sse41_ukernels",
3550 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003551 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003552 ":fma3_ukernels",
3553 ":avx2_ukernels",
3554 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003555 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003556 ],
3557)
3558
Marat Dukhan08c4a432019-10-03 09:29:21 -07003559xnnpack_cc_library(
3560 name = "im2col",
3561 srcs = ["src/im2col.c"],
3562 hdrs = [
3563 "src/xnnpack/common.h",
3564 "src/xnnpack/im2col.h",
3565 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003566 gcc_copts = xnnpack_gcc_std_copts(),
3567 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003568)
3569
3570xnnpack_cc_library(
3571 name = "indirection",
3572 srcs = ["src/indirection.c"],
3573 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003574 gcc_copts = xnnpack_gcc_std_copts(),
3575 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 deps = [
3577 "@FP16",
3578 "@FXdiv",
3579 "@pthreadpool",
3580 ],
3581)
3582
3583xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003584 name = "indirection_test_mode",
3585 srcs = ["src/indirection.c"],
3586 hdrs = INTERNAL_HDRS,
3587 copts = [
3588 "-UNDEBUG",
3589 "-DXNN_TEST_MODE=1",
3590 ],
3591 gcc_copts = xnnpack_gcc_std_copts(),
3592 msvc_copts = xnnpack_msvc_std_copts(),
3593 deps = [
3594 "@FP16",
3595 "@FXdiv",
3596 "@pthreadpool",
3597 ],
3598)
3599
3600xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003601 name = "packing",
3602 srcs = ["src/packing.c"],
3603 hdrs = INTERNAL_HDRS,
3604 gcc_copts = xnnpack_gcc_std_copts(),
3605 msvc_copts = xnnpack_msvc_std_copts(),
3606 deps = [
3607 "@FP16",
3608 "@FXdiv",
3609 "@pthreadpool",
3610 ],
3611)
3612
3613xnnpack_cc_library(
3614 name = "packing_test_mode",
3615 srcs = ["src/packing.c"],
3616 hdrs = INTERNAL_HDRS,
3617 copts = [
3618 "-UNDEBUG",
3619 "-DXNN_TEST_MODE=1",
3620 ],
3621 gcc_copts = xnnpack_gcc_std_copts(),
3622 msvc_copts = xnnpack_msvc_std_copts(),
3623 deps = [
3624 "@FP16",
3625 "@FXdiv",
3626 "@pthreadpool",
3627 ],
3628)
3629
3630xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003631 name = "operator_run",
3632 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003633 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003634 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003635 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3636 "//conditions:default": [],
3637 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003638 gcc_copts = xnnpack_gcc_std_copts(),
3639 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003640 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003641 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003642 "@FP16",
3643 "@FXdiv",
3644 "@clog",
3645 "@pthreadpool",
3646 ],
3647)
3648
Chao Mei6ddfc602020-05-13 22:29:36 -07003649xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003650 name = "operator_run_test_mode",
3651 srcs = ["src/operator-run.c"],
3652 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3653 copts = LOGGING_COPTS + [
3654 "-UNDEBUG",
3655 "-DXNN_TEST_MODE=1",
3656 ] + select({
3657 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3658 "//conditions:default": [],
3659 }),
3660 gcc_copts = xnnpack_gcc_std_copts(),
3661 msvc_copts = xnnpack_msvc_std_copts(),
3662 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003663 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003664 "@FP16",
3665 "@FXdiv",
3666 "@clog",
3667 "@pthreadpool",
3668 ],
3669)
3670
3671xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003672 name = "memory_planner",
3673 srcs = ["src/memory-planner.c"],
3674 hdrs = INTERNAL_HDRS,
3675 defines = select({
3676 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3677 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3678 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3679 }),
3680 gcc_copts = xnnpack_gcc_std_copts(),
3681 msvc_copts = xnnpack_msvc_std_copts(),
3682 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003683 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003684 "@pthreadpool",
3685 ],
3686)
3687
Marat Dukhan33fcf782020-05-24 14:27:15 -07003688xnnpack_cc_library(
3689 name = "memory_planner_test_mode",
3690 srcs = ["src/memory-planner.c"],
3691 hdrs = INTERNAL_HDRS,
3692 copts = [
3693 "-UNDEBUG",
3694 "-DXNN_TEST_MODE=1",
3695 ],
3696 defines = select({
3697 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3698 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3699 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3700 }),
3701 gcc_copts = xnnpack_gcc_std_copts(),
3702 msvc_copts = xnnpack_msvc_std_copts(),
3703 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003704 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003705 "@pthreadpool",
3706 ],
3707)
3708
Marat Dukhan08c4a432019-10-03 09:29:21 -07003709cc_library(
3710 name = "enable_assembly",
3711 defines = select({
3712 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3713 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003714 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003715 }),
3716)
3717
Marat Dukhan9de90e02020-06-18 16:04:12 -07003718cc_library(
3719 name = "enable_sparse",
3720 defines = select({
3721 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3722 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003723 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003724 }),
3725)
3726
Marat Dukhancf056b22019-10-07 10:26:29 -07003727xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003728 name = "operators",
3729 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003730 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003731 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003732 ],
3733 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003734 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003735 "-Isrc",
3736 "-Iinclude",
3737 ] + select({
3738 ":debug_build": [],
3739 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003740 }) + select({
3741 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3742 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003743 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003744 gcc_copts = xnnpack_gcc_std_copts(),
3745 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003746 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003747 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003748 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003749 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003750 "@FP16",
3751 "@FXdiv",
3752 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003753 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003754 ],
3755)
3756
Marat Dukhan10a38082020-04-17 03:58:35 -07003757xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003758 name = "operators_test_mode",
3759 srcs = OPERATOR_SRCS + [
3760 "src/memory.c",
3761 "src/operator-delete.c",
3762 ],
3763 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3764 copts = LOGGING_COPTS + [
3765 "-Isrc",
3766 "-Iinclude",
3767 "-UNDEBUG",
3768 "-DXNN_TEST_MODE=1",
3769 ] + select({
3770 ":debug_build": [],
3771 "//conditions:default": xnnpack_min_size_copts(),
3772 }) + select({
3773 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3774 "//conditions:default": [],
3775 }),
3776 gcc_copts = xnnpack_gcc_std_copts(),
3777 msvc_copts = xnnpack_msvc_std_copts(),
3778 deps = [
3779 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003780 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003781 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003782 "@FP16",
3783 "@FXdiv",
3784 "@clog",
3785 "@pthreadpool",
3786 ],
3787)
3788
3789xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003790 name = "XNNPACK",
3791 srcs = [
3792 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003793 "src/runtime.c",
3794 "src/subgraph.c",
3795 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003796 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003797 hdrs = ["include/xnnpack.h"],
3798 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003799 "-Isrc",
3800 "-Iinclude",
3801 ] + select({
3802 ":debug_build": [],
3803 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003804 }) + select({
3805 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3806 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003807 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003808 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003809 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003810 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003811 visibility = xnnpack_visibility(),
3812 deps = [
3813 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003814 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003815 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003816 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003817 ":operator_run",
3818 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003819 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003820 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003821 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003822 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003823 ] + select({
3824 ":emscripten": [],
3825 "//conditions:default": ["@cpuinfo"],
3826 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003827)
3828
Marat Dukhan10a38082020-04-17 03:58:35 -07003829xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003830 name = "XNNPACK_test_mode",
3831 srcs = [
3832 "src/init.c",
3833 "src/runtime.c",
3834 "src/subgraph.c",
3835 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003836 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003837 hdrs = ["include/xnnpack.h"],
3838 copts = LOGGING_COPTS + [
3839 "-Isrc",
3840 "-Iinclude",
3841 "-UNDEBUG",
3842 "-DXNN_TEST_MODE=1",
3843 ] + select({
3844 ":debug_build": [],
3845 "//conditions:default": xnnpack_min_size_copts(),
3846 }) + select({
3847 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3848 "//conditions:default": [],
3849 }),
3850 gcc_copts = xnnpack_gcc_std_copts(),
3851 includes = ["include"],
3852 msvc_copts = xnnpack_msvc_std_copts(),
3853 visibility = xnnpack_visibility(),
3854 deps = [
3855 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003856 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003857 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003858 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003859 ":operator_run_test_mode",
3860 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003861 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003862 "@clog",
3863 "@FP16",
3864 "@pthreadpool",
3865 ] + select({
3866 ":emscripten": [],
3867 "//conditions:default": ["@cpuinfo"],
3868 }),
3869)
3870
3871xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003872 name = "xnnpack_f32",
3873 srcs = [
3874 "src/init.c",
3875 "src/runtime.c",
3876 "src/subgraph.c",
3877 "src/tensor.c",
3878 ] + SUBGRAPH_SRCS,
3879 hdrs = ["include/xnnpack.h"],
3880 copts = LOGGING_COPTS + [
3881 "-Isrc",
3882 "-Iinclude",
3883 ] + select({
3884 ":debug_build": [],
3885 "//conditions:default": xnnpack_min_size_copts(),
3886 }) + select({
3887 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3888 "//conditions:default": [],
3889 }),
3890 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003891 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003892 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003893 "XNN_NO_U8_OPERATORS",
3894 "XNN_NO_X8_OPERATORS",
3895 "XNN_NO_F16_OPERATORS",
3896 "XNN_NO_X16_OPERATORS",
3897 ],
3898 gcc_copts = xnnpack_gcc_std_copts(),
3899 includes = ["include"],
3900 msvc_copts = xnnpack_msvc_std_copts(),
3901 visibility = xnnpack_visibility(),
3902 deps = [
3903 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003904 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003905 ":logging_utils",
3906 ":memory_planner",
3907 ":operator_run",
3908 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003909 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003910 "@clog",
3911 "@FP16",
3912 "@pthreadpool",
3913 ] + select({
3914 ":emscripten": [],
3915 "//conditions:default": ["@cpuinfo"],
3916 }),
3917)
3918
3919xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003920 name = "xnnpack_operators_nhwc_f32",
3921 srcs = [
3922 "src/init.c",
3923 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003924 hdrs = ["include/xnnpack.h"],
3925 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003926 "-Isrc",
3927 "-Iinclude",
3928 ] + select({
3929 ":debug_build": [],
3930 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003931 }) + select({
3932 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3933 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003934 }),
3935 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003936 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003937 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003938 "XNN_NO_U8_OPERATORS",
3939 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003940 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003941 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003942 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003943 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003944 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003945 visibility = xnnpack_visibility(),
3946 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003947 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003948 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003949 ":operator_run",
3950 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003951 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003952 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003953 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003954 ] + select({
3955 ":emscripten": [],
3956 "//conditions:default": ["@cpuinfo"],
3957 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003958)
3959
Marat Dukhancf056b22019-10-07 10:26:29 -07003960xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003961 name = "bench_utils",
3962 srcs = ["bench/utils.cc"],
3963 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003964 deps = [
3965 "@com_google_benchmark//:benchmark",
3966 "@cpuinfo",
3967 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003968)
3969
Frank Barchard7e955972019-10-11 10:34:25 -07003970######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003971
3972xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003973 name = "qs8_gemm_bench",
3974 srcs = [
3975 "bench/gemm.h",
3976 "bench/qs8-gemm.cc",
3977 "src/xnnpack/AlignedAllocator.h",
3978 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3979 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3980)
3981
3982xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07003983 name = "qs8_requantization_bench",
3984 srcs = [
3985 "bench/qs8-requantization.cc",
3986 "src/xnnpack/requantization-stubs.h",
3987 "src/xnnpack/AlignedAllocator.h",
3988 ] + MICROKERNEL_BENCHMARK_HDRS,
3989 deps = MICROKERNEL_BENCHMARK_DEPS,
3990)
3991
3992xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07003993 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003994 srcs = [
3995 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003996 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003997 "src/xnnpack/AlignedAllocator.h",
3998 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003999 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004000 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004001)
4002
4003xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004004 name = "qu8_requantization_bench",
4005 srcs = [
4006 "bench/qu8-requantization.cc",
4007 "src/xnnpack/requantization-stubs.h",
4008 "src/xnnpack/AlignedAllocator.h",
4009 ] + MICROKERNEL_BENCHMARK_HDRS,
4010 deps = MICROKERNEL_BENCHMARK_DEPS,
4011)
4012
4013xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004014 name = "f16_igemm_bench",
4015 srcs = [
4016 "bench/f16-igemm.cc",
4017 "bench/conv.h",
4018 "bench/google/conv.h",
4019 "src/xnnpack/AlignedAllocator.h",
4020 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004021 deps = MICROKERNEL_BENCHMARK_DEPS + [
4022 ":indirection",
4023 ":packing",
4024 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004025)
4026
4027xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004028 name = "f16_gemm_bench",
4029 srcs = [
4030 "bench/f16-gemm.cc",
4031 "bench/gemm.h",
4032 "src/xnnpack/AlignedAllocator.h",
4033 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004034 deps = MICROKERNEL_BENCHMARK_DEPS + [
4035 ":packing",
4036 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004037)
4038
4039xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004040 name = "f16_spmm_bench",
4041 srcs = [
4042 "bench/f16-spmm.cc",
4043 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004044 "src/xnnpack/AlignedAllocator.h",
4045 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004046 deps = MICROKERNEL_BENCHMARK_DEPS,
4047)
4048
4049xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004050 name = "f32_igemm_bench",
4051 srcs = [
4052 "bench/f32-igemm.cc",
4053 "bench/conv.h",
4054 "src/xnnpack/AlignedAllocator.h",
4055 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004056 deps = MICROKERNEL_BENCHMARK_DEPS + [
4057 ":indirection",
4058 ":packing",
4059 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004060)
4061
4062xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004063 name = "f16_relu_bench",
4064 srcs = [
4065 "bench/f16-relu.cc",
4066 "src/xnnpack/AlignedAllocator.h",
4067 ] + MICROKERNEL_BENCHMARK_HDRS,
4068 deps = MICROKERNEL_BENCHMARK_DEPS,
4069)
4070
4071xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004072 name = "f32_conv_hwc_bench",
4073 srcs = [
4074 "bench/f32-conv-hwc.cc",
4075 "bench/dconv.h",
4076 "src/xnnpack/AlignedAllocator.h",
4077 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004078 deps = MICROKERNEL_BENCHMARK_DEPS + [
4079 ":packing",
4080 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004081)
4082
4083xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004084 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004085 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004086 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004087 "bench/dconv.h",
4088 "src/xnnpack/AlignedAllocator.h",
4089 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004090 deps = MICROKERNEL_BENCHMARK_DEPS + [
4091 ":packing",
4092 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004093)
4094
4095xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004096 name = "f16_dwconv_bench",
4097 srcs = [
4098 "bench/f16-dwconv.cc",
4099 "bench/dwconv.h",
4100 "bench/google/dwconv.h",
4101 "src/xnnpack/AlignedAllocator.h",
4102 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004103 deps = MICROKERNEL_BENCHMARK_DEPS + [
4104 ":indirection",
4105 ":packing",
4106 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004107)
4108
4109xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004110 name = "f32_dwconv_bench",
4111 srcs = [
4112 "bench/f32-dwconv.cc",
4113 "bench/dwconv.h",
4114 "src/xnnpack/AlignedAllocator.h",
4115 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004116 deps = MICROKERNEL_BENCHMARK_DEPS + [
4117 ":indirection",
4118 ":packing",
4119 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004120)
4121
4122xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004123 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004124 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004125 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004126 "bench/dwconv.h",
4127 "src/xnnpack/AlignedAllocator.h",
4128 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004129 deps = MICROKERNEL_BENCHMARK_DEPS + [
4130 ":indirection",
4131 ":packing",
4132 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004133)
4134
4135xnnpack_benchmark(
4136 name = "f32_gemm_bench",
4137 srcs = [
4138 "bench/f32-gemm.cc",
4139 "bench/gemm.h",
4140 "src/xnnpack/AlignedAllocator.h",
4141 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004142 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004143 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144)
4145
4146xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004147 name = "f32_hswish_bench",
4148 srcs = [
4149 "bench/f32-hswish.cc",
4150 "src/xnnpack/AlignedAllocator.h",
4151 ] + MICROKERNEL_BENCHMARK_HDRS,
4152 deps = MICROKERNEL_BENCHMARK_DEPS,
4153)
4154
4155xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004156 name = "f32_raddexpminusmax_bench",
4157 srcs = [
4158 "bench/f32-raddexpminusmax.cc",
4159 "src/xnnpack/AlignedAllocator.h",
4160 ] + MICROKERNEL_BENCHMARK_HDRS,
4161 deps = MICROKERNEL_BENCHMARK_DEPS,
4162)
4163
4164xnnpack_benchmark(
4165 name = "f32_raddextexp_bench",
4166 srcs = [
4167 "bench/f32-raddextexp.cc",
4168 "src/xnnpack/AlignedAllocator.h",
4169 ] + MICROKERNEL_BENCHMARK_HDRS,
4170 deps = MICROKERNEL_BENCHMARK_DEPS,
4171)
4172
4173xnnpack_benchmark(
4174 name = "f32_raddstoreexpminusmax_bench",
4175 srcs = [
4176 "bench/f32-raddstoreexpminusmax.cc",
4177 "src/xnnpack/AlignedAllocator.h",
4178 ] + MICROKERNEL_BENCHMARK_HDRS,
4179 deps = MICROKERNEL_BENCHMARK_DEPS,
4180)
4181
4182xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004183 name = "f32_relu_bench",
4184 srcs = [
4185 "bench/f32-relu.cc",
4186 "src/xnnpack/AlignedAllocator.h",
4187 ] + MICROKERNEL_BENCHMARK_HDRS,
4188 deps = MICROKERNEL_BENCHMARK_DEPS,
4189)
4190
4191xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004192 name = "f32_rmax_bench",
4193 srcs = [
4194 "bench/f32-rmax.cc",
4195 "src/xnnpack/AlignedAllocator.h",
4196 ] + MICROKERNEL_BENCHMARK_HDRS,
4197 deps = MICROKERNEL_BENCHMARK_DEPS,
4198)
4199
4200xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004201 name = "f32_sigmoid_bench",
4202 srcs = [
4203 "bench/f32-sigmoid.cc",
4204 "src/xnnpack/AlignedAllocator.h",
4205 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004206 deps = MICROKERNEL_BENCHMARK_DEPS,
4207)
4208
4209xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004210 name = "f32_spmm_bench",
4211 srcs = [
4212 "bench/f32-spmm.cc",
4213 "bench/gemm.h",
4214 "src/xnnpack/AlignedAllocator.h",
4215 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004216 deps = MICROKERNEL_BENCHMARK_DEPS,
4217)
4218
4219xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004220 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004221 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004222 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004223 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004224 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004225 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004226)
4227
4228xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004229 name = "f32_vscaleexpminusmax_bench",
4230 srcs = [
4231 "bench/f32-vscaleexpminusmax.cc",
4232 "src/xnnpack/AlignedAllocator.h",
4233 ] + MICROKERNEL_BENCHMARK_HDRS,
4234 deps = MICROKERNEL_BENCHMARK_DEPS,
4235)
4236
4237xnnpack_benchmark(
4238 name = "f32_vscaleextexp_bench",
4239 srcs = [
4240 "bench/f32-vscaleextexp.cc",
4241 "src/xnnpack/AlignedAllocator.h",
4242 ] + MICROKERNEL_BENCHMARK_HDRS,
4243 deps = MICROKERNEL_BENCHMARK_DEPS,
4244)
4245
4246xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004247 name = "f32_vsqrt_bench",
4248 srcs = [
4249 "bench/f32-vsqrt.cc",
4250 "src/xnnpack/AlignedAllocator.h",
4251 ] + MICROKERNEL_BENCHMARK_HDRS,
4252 deps = MICROKERNEL_BENCHMARK_DEPS,
4253)
4254
4255xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004256 name = "f32_im2col_gemm_bench",
4257 srcs = [
4258 "bench/f32-im2col-gemm.cc",
4259 "bench/conv.h",
4260 "src/xnnpack/AlignedAllocator.h",
4261 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004262 deps = MICROKERNEL_BENCHMARK_DEPS + [
4263 ":im2col",
4264 ":packing",
4265 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004266)
4267
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004268xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004269 name = "rounding_bench",
4270 srcs = [
4271 "bench/rounding.cc",
4272 "src/xnnpack/math-stubs.h",
4273 "src/xnnpack/AlignedAllocator.h",
4274 ] + MICROKERNEL_BENCHMARK_HDRS,
4275 deps = MICROKERNEL_BENCHMARK_DEPS,
4276)
4277
Marat Dukhan08c4a432019-10-03 09:29:21 -07004278########################### Benchmarks for operators ###########################
4279
4280xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004281 name = "average_pooling_bench",
4282 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004283 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004284 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004285 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004286)
4287
4288xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004289 name = "bankers_rounding_bench",
4290 srcs = ["bench/bankers-rounding.cc"],
4291 copts = xnnpack_optional_tflite_copts(),
4292 tags = ["nowin32"],
4293 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4294)
4295
4296xnnpack_benchmark(
4297 name = "ceiling_bench",
4298 srcs = ["bench/ceiling.cc"],
4299 copts = xnnpack_optional_tflite_copts(),
4300 tags = ["nowin32"],
4301 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4302)
4303
4304xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004305 name = "channel_shuffle_bench",
4306 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004307 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004308)
4309
4310xnnpack_benchmark(
4311 name = "convolution_bench",
4312 srcs = ["bench/convolution.cc"],
4313 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004314 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004315 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004316)
4317
4318xnnpack_benchmark(
4319 name = "deconvolution_bench",
4320 srcs = ["bench/deconvolution.cc"],
4321 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004322 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004323 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004324)
4325
4326xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004327 name = "floor_bench",
4328 srcs = ["bench/floor.cc"],
4329 copts = xnnpack_optional_tflite_copts(),
4330 tags = ["nowin32"],
4331 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4332)
4333
4334xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 name = "global_average_pooling_bench",
4336 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004337 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004338)
4339
4340xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004341 name = "hardswish_bench",
4342 srcs = ["bench/hardswish.cc"],
4343 copts = xnnpack_optional_tflite_copts(),
4344 tags = ["nowin32"],
4345 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4346)
4347
4348xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004349 name = "max_pooling_bench",
4350 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004351 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352)
4353
4354xnnpack_benchmark(
4355 name = "sigmoid_bench",
4356 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004357 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004358 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004359 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004360)
4361
4362xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004363 name = "prelu_bench",
4364 srcs = ["bench/prelu.cc"],
4365 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004366 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004367 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004368)
4369
4370xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004371 name = "softmax_bench",
4372 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004373 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004374 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004375 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004376)
4377
Marat Dukhan87727142020-06-24 15:24:10 -07004378xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004379 name = "square_root_bench",
4380 srcs = ["bench/square-root.cc"],
4381 copts = xnnpack_optional_tflite_copts(),
4382 tags = ["nowin32"],
4383 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4384)
4385
4386xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004387 name = "truncation_bench",
4388 srcs = ["bench/truncation.cc"],
4389 deps = OPERATOR_BENCHMARK_DEPS,
4390)
4391
Marat Dukhanc068bb62019-10-04 13:24:39 -07004392############################# End-to-end benchmarks ############################
4393
4394cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004395 name = "fp32_mobilenet_v1",
4396 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004397 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004398 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004399 linkstatic = True,
4400 deps = [
4401 ":XNNPACK",
4402 "@pthreadpool",
4403 ],
4404)
4405
4406cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004407 name = "fp16_mobilenet_v1",
4408 srcs = ["models/fp16-mobilenet-v1.cc"],
4409 hdrs = ["models/models.h"],
4410 copts = xnnpack_std_cxxopts(),
4411 linkstatic = True,
4412 deps = [
4413 ":XNNPACK",
4414 "@FP16",
4415 "@pthreadpool",
4416 ],
4417)
4418
4419cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004420 name = "qs8_mobilenet_v1",
4421 srcs = ["models/qs8-mobilenet-v1.cc"],
4422 hdrs = ["models/models.h"],
4423 copts = xnnpack_std_cxxopts(),
4424 linkstatic = True,
4425 deps = [
4426 ":XNNPACK",
4427 "@pthreadpool",
4428 ],
4429)
4430
4431cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004432 name = "fp32_mobilenet_v2",
4433 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004434 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004435 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004436 linkstatic = True,
4437 deps = [
4438 ":XNNPACK",
4439 "@pthreadpool",
4440 ],
4441)
4442
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004443cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004444 name = "fp16_mobilenet_v2",
4445 srcs = ["models/fp16-mobilenet-v2.cc"],
4446 hdrs = ["models/models.h"],
4447 copts = xnnpack_std_cxxopts(),
4448 linkstatic = True,
4449 deps = [
4450 ":XNNPACK",
4451 "@FP16",
4452 "@pthreadpool",
4453 ],
4454)
4455
4456cc_library(
4457 name = "fp32_mobilenet_v3_large",
4458 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004459 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004460 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004461 linkstatic = True,
4462 deps = [
4463 ":XNNPACK",
4464 "@pthreadpool",
4465 ],
4466)
4467
4468cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004469 name = "fp32_mobilenet_v3_small",
4470 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004471 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004472 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004473 linkstatic = True,
4474 deps = [
4475 ":XNNPACK",
4476 "@pthreadpool",
4477 ],
4478)
4479
Marat Dukhanc068bb62019-10-04 13:24:39 -07004480xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004481 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004482 srcs = [
4483 "bench/f32-dwconv-e2e.cc",
4484 "bench/end2end.h",
4485 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004486 deps = MICROKERNEL_BENCHMARK_DEPS + [
4487 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004488 ":fp32_mobilenet_v1",
4489 ":fp32_mobilenet_v2",
4490 ":fp32_mobilenet_v3_large",
4491 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004492 ],
4493)
4494
4495xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004496 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004497 srcs = [
4498 "bench/f32-gemm-e2e.cc",
4499 "bench/end2end.h",
4500 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004501 deps = MICROKERNEL_BENCHMARK_DEPS + [
4502 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004503 ":fp32_mobilenet_v1",
4504 ":fp32_mobilenet_v2",
4505 ":fp32_mobilenet_v3_large",
4506 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004507 ],
4508)
4509
4510xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004511 name = "end2end_bench",
4512 srcs = ["bench/end2end.cc"],
4513 deps = [
4514 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004515 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004516 ":fp16_mobilenet_v1",
4517 ":fp16_mobilenet_v2",
4518 ":fp32_mobilenet_v1",
4519 ":fp32_mobilenet_v2",
4520 ":fp32_mobilenet_v3_large",
4521 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004522 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004523 "@pthreadpool",
4524 ],
4525)
4526
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004527#################### Accuracy evaluation for math functions ####################
4528
4529xnnpack_benchmark(
4530 name = "f32_exp_eval",
4531 srcs = [
4532 "eval/f32-exp.cc",
4533 "src/xnnpack/AlignedAllocator.h",
4534 ] + ACCURACY_EVAL_HDRS,
4535 deps = ACCURACY_EVAL_DEPS,
4536)
4537
Marat Dukhan515c9772019-10-17 18:07:57 -07004538xnnpack_benchmark(
4539 name = "f32_expminus_eval",
4540 srcs = [
4541 "eval/f32-expminus.cc",
4542 "src/xnnpack/AlignedAllocator.h",
4543 ] + ACCURACY_EVAL_HDRS,
4544 deps = ACCURACY_EVAL_DEPS,
4545)
4546
Marat Dukhan98ba4412019-10-23 02:14:28 -07004547xnnpack_benchmark(
4548 name = "f32_extexp_eval",
4549 srcs = [
4550 "eval/f32-extexp.cc",
4551 "src/xnnpack/AlignedAllocator.h",
4552 ] + ACCURACY_EVAL_HDRS,
4553 deps = ACCURACY_EVAL_DEPS,
4554)
4555
Marat Dukhan8853b822020-05-07 12:19:01 -07004556xnnpack_unit_test(
4557 name = "f32_roundne_eval",
4558 srcs = [
4559 "eval/f32-roundne.cc",
4560 "src/xnnpack/AlignedAllocator.h",
4561 "src/xnnpack/math-stubs.h",
4562 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004563 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004564 deps = MICROKERNEL_TEST_DEPS,
4565)
4566
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004567xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004568 name = "f32_roundd_eval",
4569 srcs = [
4570 "eval/f32-roundd.cc",
4571 "src/xnnpack/AlignedAllocator.h",
4572 "src/xnnpack/math-stubs.h",
4573 ] + MICROKERNEL_TEST_HDRS,
4574 automatic = False,
4575 deps = MICROKERNEL_TEST_DEPS,
4576)
4577
4578xnnpack_unit_test(
4579 name = "f32_roundu_eval",
4580 srcs = [
4581 "eval/f32-roundu.cc",
4582 "src/xnnpack/AlignedAllocator.h",
4583 "src/xnnpack/math-stubs.h",
4584 ] + MICROKERNEL_TEST_HDRS,
4585 automatic = False,
4586 deps = MICROKERNEL_TEST_DEPS,
4587)
4588
4589xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004590 name = "f32_roundz_eval",
4591 srcs = [
4592 "eval/f32-roundz.cc",
4593 "src/xnnpack/AlignedAllocator.h",
4594 "src/xnnpack/math-stubs.h",
4595 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004596 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004597 deps = MICROKERNEL_TEST_DEPS,
4598)
4599
Marat Dukhan346a9e52019-11-15 09:06:30 -08004600xnnpack_benchmark(
4601 name = "f32_sigmoid_eval",
4602 srcs = [
4603 "eval/f32-sigmoid.cc",
4604 "src/xnnpack/AlignedAllocator.h",
4605 ] + ACCURACY_EVAL_HDRS,
4606 deps = ACCURACY_EVAL_DEPS,
4607)
4608
Marat Dukhan84000762020-06-29 18:38:43 -07004609xnnpack_benchmark(
4610 name = "f32_sqrt_eval",
4611 srcs = [
4612 "eval/f32-sqrt.cc",
4613 "src/xnnpack/AlignedAllocator.h",
4614 ] + ACCURACY_EVAL_HDRS,
4615 deps = ACCURACY_EVAL_DEPS + [
4616 ":bench_utils",
4617 ],
4618)
4619
Marat Dukhan08c4a432019-10-03 09:29:21 -07004620######################### Unit tests for micro-kernels #########################
4621
4622xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004623 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004624 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004625 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004626 "test/gemm-microkernel-tester.h",
4627 "src/xnnpack/AlignedAllocator.h",
4628 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004629 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004630)
4631
4632xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004633 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004634 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004635 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004636 "test/spmm-microkernel-tester.h",
4637 "src/xnnpack/AlignedAllocator.h",
4638 ] + MICROKERNEL_TEST_HDRS,
4639 deps = MICROKERNEL_TEST_DEPS,
4640)
4641
4642xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 name = "f32_argmaxpool_test",
4644 srcs = [
4645 "test/f32-argmaxpool.cc",
4646 "test/argmaxpool-microkernel-tester.h",
4647 "src/xnnpack/AlignedAllocator.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_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004654 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004655 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004656 "test/avgpool-microkernel-tester.h",
4657 "src/xnnpack/AlignedAllocator.h",
4658 ] + MICROKERNEL_TEST_HDRS,
4659 deps = MICROKERNEL_TEST_DEPS,
4660)
4661
4662xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004663 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004664 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004665 "test/f32-ibilinear.cc",
4666 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004667 "src/xnnpack/AlignedAllocator.h",
4668 ] + MICROKERNEL_TEST_HDRS,
4669 deps = MICROKERNEL_TEST_DEPS,
4670)
4671
4672xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004673 name = "f16_clamp_test",
4674 srcs = [
4675 "test/f16-clamp.cc",
4676 "test/clamp-microkernel-tester.h",
4677 ] + MICROKERNEL_TEST_HDRS,
4678 deps = MICROKERNEL_TEST_DEPS,
4679)
4680
4681xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004682 name = "f32_clamp_test",
4683 srcs = [
4684 "test/f32-clamp.cc",
4685 "test/clamp-microkernel-tester.h",
4686 ] + MICROKERNEL_TEST_HDRS,
4687 deps = MICROKERNEL_TEST_DEPS,
4688)
4689
4690xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004691 name = "f32_igemm_test",
4692 srcs = [
4693 "test/f32-igemm.cc",
4694 "test/gemm-microkernel-tester.h",
4695 "src/xnnpack/AlignedAllocator.h",
4696 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004697 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004698)
4699
4700xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004701 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004702 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004703 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004704 "test/gemm-microkernel-tester.h",
4705 "src/xnnpack/AlignedAllocator.h",
4706 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004707 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004708)
4709
4710xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004711 name = "f32_igemm_minmax_test",
4712 srcs = [
4713 "test/f32-igemm-minmax.cc",
4714 "test/gemm-microkernel-tester.h",
4715 "src/xnnpack/AlignedAllocator.h",
4716 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004717 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004718)
4719
4720xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004721 name = "f16_igemm_minmax_test",
4722 srcs = [
4723 "test/f16-igemm-minmax.cc",
4724 "test/gemm-microkernel-tester.h",
4725 "src/xnnpack/AlignedAllocator.h",
4726 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004727 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004728)
4729
4730xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004731 name = "f32_conv_hwc_test",
4732 srcs = [
4733 "test/f32-conv-hwc.cc",
4734 "test/conv-hwc-microkernel-tester.h",
4735 "src/xnnpack/AlignedAllocator.h",
4736 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004737 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004738)
4739
4740xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004741 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004742 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004743 "test/f32-conv-hwc2chw.cc",
4744 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004745 "src/xnnpack/AlignedAllocator.h",
4746 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004747 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004748)
4749
4750xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004751 name = "f32_dwconv_test",
4752 srcs = [
4753 "test/f32-dwconv.cc",
4754 "test/dwconv-microkernel-tester.h",
4755 "src/xnnpack/AlignedAllocator.h",
4756 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004757 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004758)
4759
4760xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004761 name = "f16_dwconv_minmax_test",
4762 srcs = [
4763 "test/f16-dwconv-minmax.cc",
4764 "test/dwconv-microkernel-tester.h",
4765 "src/xnnpack/AlignedAllocator.h",
4766 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004767 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004768)
4769
4770xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004771 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004772 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004773 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004774 "test/dwconv-microkernel-tester.h",
4775 "src/xnnpack/AlignedAllocator.h",
4776 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004777 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004778)
4779
4780xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004781 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004783 "test/f32-dwconv-chw.cc",
4784 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004785 "src/xnnpack/AlignedAllocator.h",
4786 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004787 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004788)
4789
4790xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004791 name = "f16_gavgpool_minmax_test",
4792 srcs = [
4793 "test/f16-gavgpool-minmax.cc",
4794 "test/gavgpool-microkernel-tester.h",
4795 "src/xnnpack/AlignedAllocator.h",
4796 ] + MICROKERNEL_TEST_HDRS,
4797 deps = MICROKERNEL_TEST_DEPS,
4798)
4799
4800xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004801 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004802 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004803 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004804 "test/gavgpool-microkernel-tester.h",
4805 "src/xnnpack/AlignedAllocator.h",
4806 ] + MICROKERNEL_TEST_HDRS,
4807 deps = MICROKERNEL_TEST_DEPS,
4808)
4809
4810xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004811 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004813 "test/f32-gavgpool-cw.cc",
4814 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004815 "src/xnnpack/AlignedAllocator.h",
4816 ] + MICROKERNEL_TEST_HDRS,
4817 deps = MICROKERNEL_TEST_DEPS,
4818)
4819
4820xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004821 name = "f32_gemm_test",
4822 srcs = [
4823 "test/f32-gemm.cc",
4824 "test/gemm-microkernel-tester.h",
4825 "src/xnnpack/AlignedAllocator.h",
4826 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004827 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004828)
4829
4830xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004831 name = "f32_gemm_relu_test",
4832 srcs = [
4833 "test/f32-gemm-relu.cc",
4834 "test/gemm-microkernel-tester.h",
4835 "src/xnnpack/AlignedAllocator.h",
4836 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004837 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004838)
4839
4840xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004841 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004842 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004843 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004844 "test/gemm-microkernel-tester.h",
4845 "src/xnnpack/AlignedAllocator.h",
4846 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004847 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004848)
4849
4850xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004851 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004852 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004853 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004854 "test/gemm-microkernel-tester.h",
4855 "src/xnnpack/AlignedAllocator.h",
4856 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004857 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858)
4859
4860xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004861 name = "f16_hswish_test",
4862 srcs = [
4863 "test/f16-hswish.cc",
4864 "test/hswish-microkernel-tester.h",
4865 ] + MICROKERNEL_TEST_HDRS,
4866 deps = MICROKERNEL_TEST_DEPS,
4867)
4868
4869xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870 name = "f32_hswish_test",
4871 srcs = [
4872 "test/f32-hswish.cc",
4873 "test/hswish-microkernel-tester.h",
4874 ] + MICROKERNEL_TEST_HDRS,
4875 deps = MICROKERNEL_TEST_DEPS,
4876)
4877
4878xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004879 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004880 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004881 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004882 "test/maxpool-microkernel-tester.h",
4883 ] + MICROKERNEL_TEST_HDRS,
4884 deps = MICROKERNEL_TEST_DEPS,
4885)
4886
4887xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004888 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004889 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004890 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004891 "test/avgpool-microkernel-tester.h",
4892 "src/xnnpack/AlignedAllocator.h",
4893 ] + MICROKERNEL_TEST_HDRS,
4894 deps = MICROKERNEL_TEST_DEPS,
4895)
4896
4897xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004898 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004900 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 "test/gemm-microkernel-tester.h",
4902 "src/xnnpack/AlignedAllocator.h",
4903 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004904 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004905)
4906
4907xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004908 name = "f16_prelu_test",
4909 srcs = [
4910 "test/f16-prelu.cc",
4911 "test/prelu-microkernel-tester.h",
4912 "src/xnnpack/AlignedAllocator.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004918 name = "f32_prelu_test",
4919 srcs = [
4920 "test/f32-prelu.cc",
4921 "test/prelu-microkernel-tester.h",
4922 "src/xnnpack/AlignedAllocator.h",
4923 ] + MICROKERNEL_TEST_HDRS,
4924 deps = MICROKERNEL_TEST_DEPS,
4925)
4926
4927xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004928 name = "f32_raddexpminusmax_test",
4929 srcs = [
4930 "test/f32-raddexpminusmax.cc",
4931 "test/raddexpminusmax-microkernel-tester.h",
4932 ] + MICROKERNEL_TEST_HDRS,
4933 deps = MICROKERNEL_TEST_DEPS,
4934)
4935
4936xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004937 name = "f32_raddextexp_test",
4938 srcs = [
4939 "test/f32-raddextexp.cc",
4940 "test/raddextexp-microkernel-tester.h",
4941 ] + MICROKERNEL_TEST_HDRS,
4942 deps = MICROKERNEL_TEST_DEPS,
4943)
4944
4945xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004946 name = "f32_raddstoreexpminusmax_test",
4947 srcs = [
4948 "test/f32-raddstoreexpminusmax.cc",
4949 "test/raddstoreexpminusmax-microkernel-tester.h",
4950 ] + MICROKERNEL_TEST_HDRS,
4951 deps = MICROKERNEL_TEST_DEPS,
4952)
4953
4954xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004955 name = "f32_relu_test",
4956 srcs = [
4957 "test/f32-relu.cc",
4958 "test/vunary-microkernel-tester.h",
4959 ] + MICROKERNEL_TEST_HDRS,
4960 deps = MICROKERNEL_TEST_DEPS,
4961)
4962
4963xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004964 name = "f32_rmax_test",
4965 srcs = [
4966 "test/f32-rmax.cc",
4967 "test/rmax-microkernel-tester.h",
4968 ] + MICROKERNEL_TEST_HDRS,
4969 deps = MICROKERNEL_TEST_DEPS,
4970)
4971
4972xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004973 name = "f32_sigmoid_test",
4974 srcs = [
4975 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004976 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004977 ] + MICROKERNEL_TEST_HDRS,
4978 deps = MICROKERNEL_TEST_DEPS,
4979)
4980
4981xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004982 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004983 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004984 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004985 "test/spmm-microkernel-tester.h",
4986 "src/xnnpack/AlignedAllocator.h",
4987 ] + MICROKERNEL_TEST_HDRS,
4988 deps = MICROKERNEL_TEST_DEPS,
4989)
4990
4991xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07004992 name = "f16_vadd_minmax_test",
4993 srcs = [
4994 "test/f16-vadd-minmax.cc",
4995 "test/vbinary-microkernel-tester.h",
4996 ] + MICROKERNEL_TEST_HDRS,
4997 deps = MICROKERNEL_TEST_DEPS,
4998)
4999
5000xnnpack_unit_test(
5001 name = "f16_vaddc_minmax_test",
5002 srcs = [
5003 "test/f16-vaddc-minmax.cc",
5004 "test/vbinaryc-microkernel-tester.h",
5005 ] + MICROKERNEL_TEST_HDRS,
5006 deps = MICROKERNEL_TEST_DEPS,
5007)
5008
5009xnnpack_unit_test(
5010 name = "f16_vdiv_minmax_test",
5011 srcs = [
5012 "test/f16-vdiv-minmax.cc",
5013 "test/vbinary-microkernel-tester.h",
5014 ] + MICROKERNEL_TEST_HDRS,
5015 deps = MICROKERNEL_TEST_DEPS,
5016)
5017
5018xnnpack_unit_test(
5019 name = "f16_vdivc_minmax_test",
5020 srcs = [
5021 "test/f16-vdivc-minmax.cc",
5022 "test/vbinaryc-microkernel-tester.h",
5023 ] + MICROKERNEL_TEST_HDRS,
5024 deps = MICROKERNEL_TEST_DEPS,
5025)
5026
5027xnnpack_unit_test(
5028 name = "f16_vrdivc_minmax_test",
5029 srcs = [
5030 "test/f16-vrdivc-minmax.cc",
5031 "test/vbinaryc-microkernel-tester.h",
5032 ] + MICROKERNEL_TEST_HDRS,
5033 deps = MICROKERNEL_TEST_DEPS,
5034)
5035
5036xnnpack_unit_test(
5037 name = "f16_vmax_test",
5038 srcs = [
5039 "test/f16-vmax.cc",
5040 "test/vbinary-microkernel-tester.h",
5041 ] + MICROKERNEL_TEST_HDRS,
5042 deps = MICROKERNEL_TEST_DEPS,
5043)
5044
5045xnnpack_unit_test(
5046 name = "f16_vmaxc_test",
5047 srcs = [
5048 "test/f16-vmaxc.cc",
5049 "test/vbinaryc-microkernel-tester.h",
5050 ] + MICROKERNEL_TEST_HDRS,
5051 deps = MICROKERNEL_TEST_DEPS,
5052)
5053
5054xnnpack_unit_test(
5055 name = "f16_vmin_test",
5056 srcs = [
5057 "test/f16-vmin.cc",
5058 "test/vbinary-microkernel-tester.h",
5059 ] + MICROKERNEL_TEST_HDRS,
5060 deps = MICROKERNEL_TEST_DEPS,
5061)
5062
5063xnnpack_unit_test(
5064 name = "f16_vminc_test",
5065 srcs = [
5066 "test/f16-vminc.cc",
5067 "test/vbinaryc-microkernel-tester.h",
5068 ] + MICROKERNEL_TEST_HDRS,
5069 deps = MICROKERNEL_TEST_DEPS,
5070)
5071
5072xnnpack_unit_test(
5073 name = "f16_vmul_minmax_test",
5074 srcs = [
5075 "test/f16-vmul-minmax.cc",
5076 "test/vbinary-microkernel-tester.h",
5077 ] + MICROKERNEL_TEST_HDRS,
5078 deps = MICROKERNEL_TEST_DEPS,
5079)
5080
5081xnnpack_unit_test(
5082 name = "f16_vmulc_minmax_test",
5083 srcs = [
5084 "test/f16-vmulc-minmax.cc",
5085 "test/vbinaryc-microkernel-tester.h",
5086 ] + MICROKERNEL_TEST_HDRS,
5087 deps = MICROKERNEL_TEST_DEPS,
5088)
5089
5090xnnpack_unit_test(
5091 name = "f16_vsub_minmax_test",
5092 srcs = [
5093 "test/f16-vsub-minmax.cc",
5094 "test/vbinary-microkernel-tester.h",
5095 ] + MICROKERNEL_TEST_HDRS,
5096 deps = MICROKERNEL_TEST_DEPS,
5097)
5098
5099xnnpack_unit_test(
5100 name = "f16_vsubc_minmax_test",
5101 srcs = [
5102 "test/f16-vsubc-minmax.cc",
5103 "test/vbinaryc-microkernel-tester.h",
5104 ] + MICROKERNEL_TEST_HDRS,
5105 deps = MICROKERNEL_TEST_DEPS,
5106)
5107
5108xnnpack_unit_test(
5109 name = "f16_vrsubc_minmax_test",
5110 srcs = [
5111 "test/f16-vrsubc-minmax.cc",
5112 "test/vbinaryc-microkernel-tester.h",
5113 ] + MICROKERNEL_TEST_HDRS,
5114 deps = MICROKERNEL_TEST_DEPS,
5115)
5116
5117xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005118 name = "f32_vabs_test",
5119 srcs = [
5120 "test/f32-vabs.cc",
5121 "test/vunary-microkernel-tester.h",
5122 ] + MICROKERNEL_TEST_HDRS,
5123 deps = MICROKERNEL_TEST_DEPS,
5124)
5125
5126xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005127 name = "f32_vadd_test",
5128 srcs = [
5129 "test/f32-vadd.cc",
5130 "test/vbinary-microkernel-tester.h",
5131 ] + MICROKERNEL_TEST_HDRS,
5132 deps = MICROKERNEL_TEST_DEPS,
5133)
5134
5135xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005136 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005137 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005138 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005139 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005140 ] + MICROKERNEL_TEST_HDRS,
5141 deps = MICROKERNEL_TEST_DEPS,
5142)
5143
5144xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005145 name = "f32_vadd_relu_test",
5146 srcs = [
5147 "test/f32-vadd-relu.cc",
5148 "test/vbinary-microkernel-tester.h",
5149 ] + MICROKERNEL_TEST_HDRS,
5150 deps = MICROKERNEL_TEST_DEPS,
5151)
5152
5153xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005154 name = "f32_vaddc_test",
5155 srcs = [
5156 "test/f32-vaddc.cc",
5157 "test/vbinaryc-microkernel-tester.h",
5158 ] + MICROKERNEL_TEST_HDRS,
5159 deps = MICROKERNEL_TEST_DEPS,
5160)
5161
5162xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005163 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005164 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005165 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005166 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005167 ] + MICROKERNEL_TEST_HDRS,
5168 deps = MICROKERNEL_TEST_DEPS,
5169)
5170
5171xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005172 name = "f32_vaddc_relu_test",
5173 srcs = [
5174 "test/f32-vaddc-relu.cc",
5175 "test/vbinaryc-microkernel-tester.h",
5176 ] + MICROKERNEL_TEST_HDRS,
5177 deps = MICROKERNEL_TEST_DEPS,
5178)
5179
5180xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005181 name = "f32_vdiv_test",
5182 srcs = [
5183 "test/f32-vdiv.cc",
5184 "test/vbinary-microkernel-tester.h",
5185 ] + MICROKERNEL_TEST_HDRS,
5186 deps = MICROKERNEL_TEST_DEPS,
5187)
5188
5189xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005190 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005191 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005192 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005193 "test/vbinary-microkernel-tester.h",
5194 ] + MICROKERNEL_TEST_HDRS,
5195 deps = MICROKERNEL_TEST_DEPS,
5196)
5197
5198xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005199 name = "f32_vdiv_relu_test",
5200 srcs = [
5201 "test/f32-vdiv-relu.cc",
5202 "test/vbinary-microkernel-tester.h",
5203 ] + MICROKERNEL_TEST_HDRS,
5204 deps = MICROKERNEL_TEST_DEPS,
5205)
5206
5207xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005208 name = "f32_vdivc_test",
5209 srcs = [
5210 "test/f32-vdivc.cc",
5211 "test/vbinaryc-microkernel-tester.h",
5212 ] + MICROKERNEL_TEST_HDRS,
5213 deps = MICROKERNEL_TEST_DEPS,
5214)
5215
5216xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005217 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005218 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005219 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005220 "test/vbinaryc-microkernel-tester.h",
5221 ] + MICROKERNEL_TEST_HDRS,
5222 deps = MICROKERNEL_TEST_DEPS,
5223)
5224
5225xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005226 name = "f32_vdivc_relu_test",
5227 srcs = [
5228 "test/f32-vdivc-relu.cc",
5229 "test/vbinaryc-microkernel-tester.h",
5230 ] + MICROKERNEL_TEST_HDRS,
5231 deps = MICROKERNEL_TEST_DEPS,
5232)
5233
5234xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005235 name = "f32_vrdivc_test",
5236 srcs = [
5237 "test/f32-vrdivc.cc",
5238 "test/vbinaryc-microkernel-tester.h",
5239 ] + MICROKERNEL_TEST_HDRS,
5240 deps = MICROKERNEL_TEST_DEPS,
5241)
5242
5243xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005244 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005245 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005246 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005247 "test/vbinaryc-microkernel-tester.h",
5248 ] + MICROKERNEL_TEST_HDRS,
5249 deps = MICROKERNEL_TEST_DEPS,
5250)
5251
5252xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005253 name = "f32_vrdivc_relu_test",
5254 srcs = [
5255 "test/f32-vrdivc-relu.cc",
5256 "test/vbinaryc-microkernel-tester.h",
5257 ] + MICROKERNEL_TEST_HDRS,
5258 deps = MICROKERNEL_TEST_DEPS,
5259)
5260
5261xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005262 name = "f32_vmax_test",
5263 srcs = [
5264 "test/f32-vmax.cc",
5265 "test/vbinary-microkernel-tester.h",
5266 ] + MICROKERNEL_TEST_HDRS,
5267 deps = MICROKERNEL_TEST_DEPS,
5268)
5269
5270xnnpack_unit_test(
5271 name = "f32_vmaxc_test",
5272 srcs = [
5273 "test/f32-vmaxc.cc",
5274 "test/vbinaryc-microkernel-tester.h",
5275 ] + MICROKERNEL_TEST_HDRS,
5276 deps = MICROKERNEL_TEST_DEPS,
5277)
5278
5279xnnpack_unit_test(
5280 name = "f32_vmin_test",
5281 srcs = [
5282 "test/f32-vmin.cc",
5283 "test/vbinary-microkernel-tester.h",
5284 ] + MICROKERNEL_TEST_HDRS,
5285 deps = MICROKERNEL_TEST_DEPS,
5286)
5287
5288xnnpack_unit_test(
5289 name = "f32_vminc_test",
5290 srcs = [
5291 "test/f32-vminc.cc",
5292 "test/vbinaryc-microkernel-tester.h",
5293 ] + MICROKERNEL_TEST_HDRS,
5294 deps = MICROKERNEL_TEST_DEPS,
5295)
5296
5297xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005298 name = "f32_vmul_test",
5299 srcs = [
5300 "test/f32-vmul.cc",
5301 "test/vbinary-microkernel-tester.h",
5302 ] + MICROKERNEL_TEST_HDRS,
5303 deps = MICROKERNEL_TEST_DEPS,
5304)
5305
5306xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005307 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005308 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005309 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005310 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005311 ] + MICROKERNEL_TEST_HDRS,
5312 deps = MICROKERNEL_TEST_DEPS,
5313)
5314
5315xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005316 name = "f32_vmul_relu_test",
5317 srcs = [
5318 "test/f32-vmul-relu.cc",
5319 "test/vbinary-microkernel-tester.h",
5320 ] + MICROKERNEL_TEST_HDRS,
5321 deps = MICROKERNEL_TEST_DEPS,
5322)
5323
5324xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005325 name = "f32_vmulc_test",
5326 srcs = [
5327 "test/f32-vmulc.cc",
5328 "test/vbinaryc-microkernel-tester.h",
5329 ] + MICROKERNEL_TEST_HDRS,
5330 deps = MICROKERNEL_TEST_DEPS,
5331)
5332
5333xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005334 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005335 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005336 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005337 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005338 ] + MICROKERNEL_TEST_HDRS,
5339 deps = MICROKERNEL_TEST_DEPS,
5340)
5341
5342xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005343 name = "f32_vmulc_relu_test",
5344 srcs = [
5345 "test/f32-vmulc-relu.cc",
5346 "test/vbinaryc-microkernel-tester.h",
5347 ] + MICROKERNEL_TEST_HDRS,
5348 deps = MICROKERNEL_TEST_DEPS,
5349)
5350
5351xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005352 name = "f16_vmulcaddc_minmax_test",
5353 srcs = [
5354 "test/f16-vmulcaddc-minmax.cc",
5355 "test/vmulcaddc-microkernel-tester.h",
5356 "src/xnnpack/AlignedAllocator.h",
5357 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005358 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005359)
5360
5361xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005362 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005363 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005364 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005365 "test/vmulcaddc-microkernel-tester.h",
5366 "src/xnnpack/AlignedAllocator.h",
5367 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005368 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005369)
5370
5371xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005372 name = "f32_vlrelu_test",
5373 srcs = [
5374 "test/f32-vlrelu.cc",
5375 "test/vunary-microkernel-tester.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005381 name = "f32_vneg_test",
5382 srcs = [
5383 "test/f32-vneg.cc",
5384 "test/vunary-microkernel-tester.h",
5385 ] + MICROKERNEL_TEST_HDRS,
5386 deps = MICROKERNEL_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005390 name = "f32_vrndne_test",
5391 srcs = [
5392 "test/f32-vrndne.cc",
5393 "test/vunary-microkernel-tester.h",
5394 ] + MICROKERNEL_TEST_HDRS,
5395 deps = MICROKERNEL_TEST_DEPS,
5396)
5397
5398xnnpack_unit_test(
5399 name = "f32_vrndz_test",
5400 srcs = [
5401 "test/f32-vrndz.cc",
5402 "test/vunary-microkernel-tester.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 deps = MICROKERNEL_TEST_DEPS,
5405)
5406
5407xnnpack_unit_test(
5408 name = "f32_vrndu_test",
5409 srcs = [
5410 "test/f32-vrndu.cc",
5411 "test/vunary-microkernel-tester.h",
5412 ] + MICROKERNEL_TEST_HDRS,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
5417 name = "f32_vrndd_test",
5418 srcs = [
5419 "test/f32-vrndd.cc",
5420 "test/vunary-microkernel-tester.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005426 name = "f32_vscale_test",
5427 srcs = [
5428 "test/f32-vscale.cc",
5429 "test/vscale-microkernel-tester.h",
5430 ] + MICROKERNEL_TEST_HDRS,
5431 deps = MICROKERNEL_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005435 name = "f32_vscaleexpminusmax_test",
5436 srcs = [
5437 "test/f32-vscaleexpminusmax.cc",
5438 "test/vscaleexpminusmax-microkernel-tester.h",
5439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005444 name = "f32_vscaleextexp_test",
5445 srcs = [
5446 "test/f32-vscaleextexp.cc",
5447 "test/vscaleextexp-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005453 name = "f32_vsqr_test",
5454 srcs = [
5455 "test/f32-vsqr.cc",
5456 "test/vunary-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005462 name = "f32_vsqrdiff_test",
5463 srcs = [
5464 "test/f32-vsqrdiff.cc",
5465 "test/vbinary-microkernel-tester.h",
5466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
5471 name = "f32_vsqrdiffc_test",
5472 srcs = [
5473 "test/f32-vsqrdiffc.cc",
5474 "test/vbinaryc-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005480 name = "f32_vsqrt_test",
5481 srcs = [
5482 "test/f32-vsqrt.cc",
5483 "test/vunary-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005489 name = "f32_vsub_test",
5490 srcs = [
5491 "test/f32-vsub.cc",
5492 "test/vbinary-microkernel-tester.h",
5493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005498 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005499 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005500 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005501 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005507 name = "f32_vsub_relu_test",
5508 srcs = [
5509 "test/f32-vsub-relu.cc",
5510 "test/vbinary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005516 name = "f32_vsubc_test",
5517 srcs = [
5518 "test/f32-vsubc.cc",
5519 "test/vbinaryc-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005525 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005526 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005527 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005528 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005534 name = "f32_vsubc_relu_test",
5535 srcs = [
5536 "test/f32-vsubc-relu.cc",
5537 "test/vbinaryc-microkernel-tester.h",
5538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005543 name = "f32_vrsubc_test",
5544 srcs = [
5545 "test/f32-vrsubc.cc",
5546 "test/vbinaryc-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005552 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005553 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005554 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005555 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005561 name = "f32_vrsubc_relu_test",
5562 srcs = [
5563 "test/f32-vrsubc-relu.cc",
5564 "test/vbinaryc-microkernel-tester.h",
5565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005570 name = "qs8_dwconv_minmax_test",
5571 srcs = [
5572 "test/qs8-dwconv-minmax.cc",
5573 "test/dwconv-microkernel-tester.h",
5574 "src/xnnpack/AlignedAllocator.h",
5575 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5576 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5577)
5578
5579xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005580 name = "qs8_gavgpool_minmax_test",
5581 srcs = [
5582 "test/qs8-gavgpool-minmax.cc",
5583 "test/gavgpool-microkernel-tester.h",
5584 "src/xnnpack/AlignedAllocator.h",
5585 ] + MICROKERNEL_TEST_HDRS,
5586 deps = MICROKERNEL_TEST_DEPS,
5587)
5588
5589xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005590 name = "qs8_gemm_minmax_test",
5591 srcs = [
5592 "test/qs8-gemm-minmax.cc",
5593 "test/gemm-microkernel-tester.h",
5594 "src/xnnpack/AlignedAllocator.h",
5595 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5597)
5598
5599xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005600 name = "qs8_igemm_minmax_test",
5601 srcs = [
5602 "test/qs8-igemm-minmax.cc",
5603 "test/gemm-microkernel-tester.h",
5604 "src/xnnpack/AlignedAllocator.h",
5605 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5606 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5607)
5608
5609xnnpack_unit_test(
5610 name = "qs8_requantization_test",
5611 srcs = [
5612 "src/xnnpack/requantization-stubs.h",
5613 "test/qs8-requantization.cc",
5614 "test/requantization-tester.h",
5615 ] + MICROKERNEL_TEST_HDRS,
5616 deps = MICROKERNEL_TEST_DEPS,
5617)
5618
5619xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005620 name = "qs8_vadd_minmax_test",
5621 srcs = [
5622 "test/qs8-vadd-minmax.cc",
5623 "test/vadd-microkernel-tester.h",
5624 ] + MICROKERNEL_TEST_HDRS,
5625 deps = MICROKERNEL_TEST_DEPS,
5626)
5627
5628xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005629 name = "qs8_vaddc_minmax_test",
5630 srcs = [
5631 "test/qs8-vaddc-minmax.cc",
5632 "test/vaddc-microkernel-tester.h",
5633 ] + MICROKERNEL_TEST_HDRS,
5634 deps = MICROKERNEL_TEST_DEPS,
5635)
5636
5637xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005638 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005639 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005640 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005641 "test/avgpool-microkernel-tester.h",
5642 "src/xnnpack/AlignedAllocator.h",
5643 ] + MICROKERNEL_TEST_HDRS,
5644 deps = MICROKERNEL_TEST_DEPS,
5645)
5646
5647xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005648 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005649 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005650 "test/qu8-dwconv-minmax.cc",
5651 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005652 "src/xnnpack/AlignedAllocator.h",
5653 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005654 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005655)
5656
5657xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005658 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005659 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005660 "test/qu8-igemm-minmax.cc",
5661 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005662 "src/xnnpack/AlignedAllocator.h",
5663 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005664 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005665)
5666
5667xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005668 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005669 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005670 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005671 "test/gavgpool-microkernel-tester.h",
5672 "src/xnnpack/AlignedAllocator.h",
5673 ] + MICROKERNEL_TEST_HDRS,
5674 deps = MICROKERNEL_TEST_DEPS,
5675)
5676
5677xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005678 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005679 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005680 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005681 "test/gemm-microkernel-tester.h",
5682 "src/xnnpack/AlignedAllocator.h",
5683 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005684 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005685)
5686
5687xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005688 name = "qu8_requantization_test",
5689 srcs = [
5690 "src/xnnpack/requantization-stubs.h",
5691 "test/qu8-requantization.cc",
5692 "test/requantization-tester.h",
5693 ] + MICROKERNEL_TEST_HDRS,
5694 deps = MICROKERNEL_TEST_DEPS,
5695)
5696
5697xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005698 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005699 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005700 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005701 "test/vadd-microkernel-tester.h",
5702 ] + MICROKERNEL_TEST_HDRS,
5703 deps = MICROKERNEL_TEST_DEPS,
5704)
5705
5706xnnpack_unit_test(
5707 name = "u8_clamp_test",
5708 srcs = [
5709 "test/u8-clamp.cc",
5710 "test/clamp-microkernel-tester.h",
5711 ] + MICROKERNEL_TEST_HDRS,
5712 deps = MICROKERNEL_TEST_DEPS,
5713)
5714
5715xnnpack_unit_test(
5716 name = "u8_lut32norm_test",
5717 srcs = [
5718 "test/u8-lut32norm.cc",
5719 "test/lut-norm-microkernel-tester.h",
5720 ] + MICROKERNEL_TEST_HDRS,
5721 deps = MICROKERNEL_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005725 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005726 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005727 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005728 "test/maxpool-microkernel-tester.h",
5729 ] + MICROKERNEL_TEST_HDRS,
5730 deps = MICROKERNEL_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
5734 name = "u8_rmax_test",
5735 srcs = [
5736 "test/u8-rmax.cc",
5737 "test/rmax-microkernel-tester.h",
5738 ] + MICROKERNEL_TEST_HDRS,
5739 deps = MICROKERNEL_TEST_DEPS,
5740)
5741
5742xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005743 name = "x32_fill_test",
5744 srcs = [
5745 "test/x32-fill.cc",
5746 "test/fill-microkernel-tester.h",
5747 ] + MICROKERNEL_TEST_HDRS,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005752 name = "x32_packx_test",
5753 srcs = [
5754 "test/x32-packx.cc",
5755 "test/pack-microkernel-tester.h",
5756 "src/xnnpack/AlignedAllocator.h",
5757 ] + MICROKERNEL_TEST_HDRS,
5758 deps = MICROKERNEL_TEST_DEPS,
5759)
5760
5761xnnpack_unit_test(
5762 name = "x32_pad_test",
5763 srcs = [
5764 "test/x32-pad.cc",
5765 "test/pad-microkernel-tester.h",
5766 ] + MICROKERNEL_TEST_HDRS,
5767 deps = MICROKERNEL_TEST_DEPS,
5768)
5769
5770xnnpack_unit_test(
5771 name = "x32_unpool_test",
5772 srcs = [
5773 "test/x32-unpool.cc",
5774 "test/unpool-microkernel-tester.h",
5775 ] + MICROKERNEL_TEST_HDRS,
5776 deps = MICROKERNEL_TEST_DEPS,
5777)
5778
5779xnnpack_unit_test(
5780 name = "x32_zip_test",
5781 srcs = [
5782 "test/x32-zip.cc",
5783 "test/zip-microkernel-tester.h",
5784 ] + MICROKERNEL_TEST_HDRS,
5785 deps = MICROKERNEL_TEST_DEPS,
5786)
5787
5788xnnpack_unit_test(
5789 name = "x8_lut_test",
5790 srcs = [
5791 "test/x8-lut.cc",
5792 "test/lut-microkernel-tester.h",
5793 ] + MICROKERNEL_TEST_HDRS,
5794 deps = MICROKERNEL_TEST_DEPS,
5795)
5796
5797xnnpack_unit_test(
5798 name = "x8_zip_test",
5799 srcs = [
5800 "test/x8-zip.cc",
5801 "test/zip-microkernel-tester.h",
5802 ] + MICROKERNEL_TEST_HDRS,
5803 deps = MICROKERNEL_TEST_DEPS,
5804)
5805
Marat Dukhan20c3b922020-03-10 03:45:06 -07005806########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005807
5808xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005809 name = "operator_size_test",
5810 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005811 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005812)
5813
Marat Dukhan20c3b922020-03-10 03:45:06 -07005814xnnpack_binary(
5815 name = "subgraph_size_test",
5816 srcs = ["test/subgraph-size.c"],
5817 deps = [":XNNPACK"],
5818)
5819
5820########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005821
5822xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005823 name = "abs_nc_test",
5824 srcs = [
5825 "test/abs-nc.cc",
5826 "test/abs-operator-tester.h",
5827 ],
5828 deps = OPERATOR_TEST_DEPS,
5829)
5830
5831xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005832 name = "add_nd_test",
5833 srcs = [
5834 "test/add-nd.cc",
5835 "test/binary-elementwise-operator-tester.h",
5836 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005837 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005838)
5839
5840xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005841 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005842 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005843 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005844 "test/argmax-pooling-operator-tester.h",
5845 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005846 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005847)
5848
5849xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005850 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005851 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005852 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005853 "test/average-pooling-operator-tester.h",
5854 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005855 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005856)
5857
5858xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005859 name = "bankers_rounding_nc_test",
5860 srcs = [
5861 "test/bankers-rounding-nc.cc",
5862 "test/bankers-rounding-operator-tester.h",
5863 ],
5864 deps = OPERATOR_TEST_DEPS,
5865)
5866
5867xnnpack_unit_test(
5868 name = "ceiling_nc_test",
5869 srcs = [
5870 "test/ceiling-nc.cc",
5871 "test/ceiling-operator-tester.h",
5872 ],
5873 deps = OPERATOR_TEST_DEPS,
5874)
5875
5876xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005877 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005878 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005879 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005880 "test/channel-shuffle-operator-tester.h",
5881 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005882 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005883)
5884
5885xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005886 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005887 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005888 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005889 "test/clamp-operator-tester.h",
5890 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005891 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005892)
5893
5894xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005895 name = "constant_pad_nd_test",
5896 srcs = [
5897 "test/constant-pad-nd.cc",
5898 "test/constant-pad-operator-tester.h",
5899 ],
5900 deps = OPERATOR_TEST_DEPS,
5901)
5902
5903xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005904 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005905 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005906 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005907 "test/convolution-operator-tester.h",
5908 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005909 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005910)
5911
5912xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005913 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005914 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005915 "test/convolution-nchw.cc",
5916 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005917 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005918 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005919)
5920
5921xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005922 name = "copy_nc_test",
5923 srcs = [
5924 "test/copy-nc.cc",
5925 "test/copy-operator-tester.h",
5926 ],
5927 deps = OPERATOR_TEST_DEPS,
5928)
5929
5930xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005931 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005932 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005933 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934 "test/deconvolution-operator-tester.h",
5935 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005936 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005937)
5938
5939xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005940 name = "divide_nd_test",
5941 srcs = [
5942 "test/binary-elementwise-operator-tester.h",
5943 "test/divide-nd.cc",
5944 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005945 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005946)
5947
5948xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005949 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005950 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005951 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005952 "test/fully-connected-operator-tester.h",
5953 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005954 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005955)
5956
5957xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005958 name = "floor_nc_test",
5959 srcs = [
5960 "test/floor-nc.cc",
5961 "test/floor-operator-tester.h",
5962 ],
5963 deps = OPERATOR_TEST_DEPS,
5964)
5965
5966xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005967 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005968 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005969 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005970 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005971 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005972 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005973)
5974
5975xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005976 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005977 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005978 "test/global-average-pooling-ncw.cc",
5979 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005980 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005981 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005982)
5983
5984xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005985 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005986 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005987 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005988 "test/hardswish-operator-tester.h",
5989 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005990 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005991)
5992
5993xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005994 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005995 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005996 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005997 "test/leaky-relu-operator-tester.h",
5998 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005999 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000)
6001
6002xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006003 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006004 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006005 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006006 "test/max-pooling-operator-tester.h",
6007 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006008 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009)
6010
6011xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006012 name = "maximum_nd_test",
6013 srcs = [
6014 "test/binary-elementwise-operator-tester.h",
6015 "test/maximum-nd.cc",
6016 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006017 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006018)
6019
6020xnnpack_unit_test(
6021 name = "minimum_nd_test",
6022 srcs = [
6023 "test/binary-elementwise-operator-tester.h",
6024 "test/minimum-nd.cc",
6025 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006026 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006027)
6028
6029xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006030 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006031 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006032 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006033 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006034 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006035 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006036)
6037
6038xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006039 name = "negate_nc_test",
6040 srcs = [
6041 "test/negate-nc.cc",
6042 "test/negate-operator-tester.h",
6043 ],
6044 deps = OPERATOR_TEST_DEPS,
6045)
6046
6047xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006048 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006049 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006050 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006051 "test/prelu-operator-tester.h",
6052 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006053 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006054)
6055
6056xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006057 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006058 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006059 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006060 "test/resize-bilinear-operator-tester.h",
6061 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006062 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006063)
6064
6065xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006066 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006067 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006068 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006069 "test/sigmoid-operator-tester.h",
6070 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006071 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006072)
6073
6074xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006075 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006076 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006077 "test/softmax-nc.cc",
6078 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006079 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006080 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006081)
6082
6083xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006084 name = "square_nc_test",
6085 srcs = [
6086 "test/square-nc.cc",
6087 "test/square-operator-tester.h",
6088 ],
6089 deps = OPERATOR_TEST_DEPS,
6090)
6091
6092xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006093 name = "square_root_nc_test",
6094 srcs = [
6095 "test/square-root-nc.cc",
6096 "test/square-root-operator-tester.h",
6097 ],
6098 deps = OPERATOR_TEST_DEPS,
6099)
6100
6101xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006102 name = "squared_difference_nd_test",
6103 srcs = [
6104 "test/binary-elementwise-operator-tester.h",
6105 "test/squared-difference-nd.cc",
6106 ],
6107 deps = OPERATOR_TEST_DEPS,
6108)
6109
6110xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006111 name = "subtract_nd_test",
6112 srcs = [
6113 "test/binary-elementwise-operator-tester.h",
6114 "test/subtract-nd.cc",
6115 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006116 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006117)
6118
6119xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006120 name = "truncation_nc_test",
6121 srcs = [
6122 "test/truncation-nc.cc",
6123 "test/truncation-operator-tester.h",
6124 ],
6125 deps = OPERATOR_TEST_DEPS,
6126)
6127
6128xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006129 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006130 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006131 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 "test/unpooling-operator-tester.h",
6133 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006134 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006135)
6136
Chao Mei6ddfc602020-05-13 22:29:36 -07006137############################### Misc unit tests ###############################
6138
6139xnnpack_unit_test(
6140 name = "memory_planner_test",
6141 srcs = [
6142 "test/memory-planner-test.cc",
6143 ],
6144 deps = [
6145 ":XNNPACK",
6146 ":memory_planner",
6147 ],
6148)
6149
Marat Dukhan08c4a432019-10-03 09:29:21 -07006150############################# Build configurations #############################
6151
Marat Dukhanb8642352019-10-30 15:43:02 -07006152# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006154 name = "xnn_enable_assembly_explicit_true",
6155 define_values = {"xnn_enable_assembly": "true"},
6156)
6157
6158# Disables usage of assembly kernels.
6159config_setting(
6160 name = "xnn_enable_assembly_explicit_false",
6161 define_values = {"xnn_enable_assembly": "false"},
6162)
6163
Marat Dukhan9de90e02020-06-18 16:04:12 -07006164# Enables usage of sparse inference.
6165config_setting(
6166 name = "xnn_enable_sparse_explicit_true",
6167 define_values = {"xnn_enable_sparse": "true"},
6168)
6169
6170# Disables usage of sparse inference.
6171config_setting(
6172 name = "xnn_enable_sparse_explicit_false",
6173 define_values = {"xnn_enable_sparse": "false"},
6174)
6175
Marat Dukhan05702cf2020-03-26 15:41:33 -07006176# Disables usage of HMP-aware optimizations.
6177config_setting(
6178 name = "xnn_enable_hmp_explicit_false",
6179 define_values = {"xnn_enable_hmp": "false"},
6180)
6181
Chao Mei6ddfc602020-05-13 22:29:36 -07006182# Enable usage of optimized memory allocation
6183config_setting(
6184 name = "xnn_enable_memopt_explicit_true",
6185 define_values = {"xnnpack_enable_memopt": "true"},
6186)
6187
6188# Disable usage of optimized memory allocation
6189config_setting(
6190 name = "xnn_enable_memopt_explicit_false",
6191 define_values = {"xnnpack_enable_memopt": "false"},
6192)
6193
Marat Dukhanb8642352019-10-30 15:43:02 -07006194# Builds with -c dbg
6195config_setting(
6196 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006198 "compilation_mode": "dbg",
6199 },
6200)
6201
6202# Builds with -c opt
6203config_setting(
6204 name = "optimized_build",
6205 values = {
6206 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207 },
6208)
6209
6210config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006211 name = "linux_k8",
6212 values = {"cpu": "k8"},
6213)
6214
6215config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006216 name = "linux_arm",
6217 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006218)
6219
6220config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006221 name = "linux_armeabi",
6222 values = {"cpu": "armeabi"},
6223)
6224
6225config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006226 name = "linux_armhf",
6227 values = {"cpu": "armhf"},
6228)
6229
6230config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006231 name = "linux_armv7a",
6232 values = {"cpu": "armv7a"},
6233)
6234
6235config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006236 name = "linux_aarch64",
6237 values = {"cpu": "aarch64"},
6238)
6239
6240config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006241 name = "android",
6242 values = {"crosstool_top": "//external:android/crosstool"},
6243)
6244
6245config_setting(
6246 name = "android_armv7",
6247 values = {
6248 "crosstool_top": "//external:android/crosstool",
6249 "cpu": "armeabi-v7a",
6250 },
6251)
6252
6253config_setting(
6254 name = "android_arm64",
6255 values = {
6256 "crosstool_top": "//external:android/crosstool",
6257 "cpu": "arm64-v8a",
6258 },
6259)
6260
6261config_setting(
6262 name = "android_x86",
6263 values = {
6264 "crosstool_top": "//external:android/crosstool",
6265 "cpu": "x86",
6266 },
6267)
6268
6269config_setting(
6270 name = "android_x86_64",
6271 values = {
6272 "crosstool_top": "//external:android/crosstool",
6273 "cpu": "x86_64",
6274 },
6275)
6276
6277config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006278 name = "windows_x86_64",
6279 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006280)
6281
6282config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006283 name = "windows_x86_64_clang",
6284 values = {
6285 "compiler": "clang-cl",
6286 "cpu": "x64_windows",
6287 },
6288)
6289
6290config_setting(
6291 name = "windows_x86_64_mingw",
6292 values = {
6293 "compiler": "mingw-gcc",
6294 "cpu": "x64_windows",
6295 },
6296)
6297
6298config_setting(
6299 name = "windows_x86_64_msys",
6300 values = {
6301 "compiler": "msys-gcc",
6302 "cpu": "x64_windows",
6303 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006304)
6305
6306config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006307 name = "macos_x86_64",
6308 values = {
6309 "apple_platform_type": "macos",
6310 "cpu": "darwin",
6311 },
6312)
6313
6314config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006315 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006316 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006317)
6318
6319config_setting(
6320 name = "emscripten_wasm",
6321 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006322 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006323 "cpu": "wasm",
6324 },
6325)
6326
6327config_setting(
6328 name = "emscripten_wasmsimd",
6329 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006330 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006331 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006332 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006333 },
6334)
6335
6336config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006337 name = "ios_armv7",
6338 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006339 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006340 "cpu": "ios_armv7",
6341 },
6342)
6343
6344config_setting(
6345 name = "ios_arm64",
6346 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006347 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006348 "cpu": "ios_arm64",
6349 },
6350)
6351
6352config_setting(
6353 name = "ios_arm64e",
6354 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006355 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006356 "cpu": "ios_arm64e",
6357 },
6358)
6359
6360config_setting(
6361 name = "ios_x86",
6362 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006363 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006364 "cpu": "ios_i386",
6365 },
6366)
6367
6368config_setting(
6369 name = "ios_x86_64",
6370 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006371 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006372 "cpu": "ios_x86_64",
6373 },
6374)
6375
6376config_setting(
6377 name = "watchos_armv7k",
6378 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006379 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006380 "cpu": "watchos_armv7k",
6381 },
6382)
6383
6384config_setting(
6385 name = "watchos_arm64_32",
6386 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006387 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006388 "cpu": "watchos_arm64_32",
6389 },
6390)
6391
6392config_setting(
6393 name = "watchos_x86",
6394 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006395 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006396 "cpu": "watchos_i386",
6397 },
6398)
6399
6400config_setting(
6401 name = "watchos_x86_64",
6402 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006403 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006404 "cpu": "watchos_x86_64",
6405 },
6406)
6407
6408config_setting(
6409 name = "tvos_arm64",
6410 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006411 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006412 "cpu": "tvos_arm64",
6413 },
6414)
6415
6416config_setting(
6417 name = "tvos_x86_64",
6418 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006419 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006420 "cpu": "tvos_x86_64",
6421 },
6422)