blob: 29b14d7f72dcc1b61e5ac1123002c3f8de9b29e0 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700242 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
243 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700245 "src/f32-vbinary/gen/vadd-scalar-x1.c",
246 "src/f32-vbinary/gen/vadd-scalar-x2.c",
247 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700248 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
249 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
250 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700251 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700254 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
255 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
256 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700257 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
258 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
259 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700260 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
261 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
262 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700263 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700266 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
267 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
268 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700269 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
270 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
271 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700272 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
273 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
274 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800275 "src/f32-vbinary/gen/vmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vmax-scalar-x4.c",
278 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
279 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
280 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
281 "src/f32-vbinary/gen/vmin-scalar-x1.c",
282 "src/f32-vbinary/gen/vmin-scalar-x2.c",
283 "src/f32-vbinary/gen/vmin-scalar-x4.c",
284 "src/f32-vbinary/gen/vminc-scalar-x1.c",
285 "src/f32-vbinary/gen/vminc-scalar-x2.c",
286 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700287 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700290 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
291 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
292 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700293 "src/f32-vbinary/gen/vmul-scalar-x1.c",
294 "src/f32-vbinary/gen/vmul-scalar-x2.c",
295 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700296 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
297 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
298 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700299 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
300 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
301 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700302 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
303 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
304 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700305 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
306 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
307 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700308 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
309 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
310 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
312 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
313 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700314 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
315 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
316 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700317 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
318 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
319 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700320 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
321 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
322 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700323 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
324 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
325 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
326 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
327 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
328 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700329 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
330 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
331 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700332 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
333 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
334 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vsub-scalar-x1.c",
336 "src/f32-vbinary/gen/vsub-scalar-x2.c",
337 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700338 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
339 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
340 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700341 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
342 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
343 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700344 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
345 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
346 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700347 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
348 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
349 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700350 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
351 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
352 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700353 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
354 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
355 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
356 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
357 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
358 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
359 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
360 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
361 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700362 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
363 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
364 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700365 "src/f32-vunary/gen/vabs-scalar-x1.c",
366 "src/f32-vunary/gen/vabs-scalar-x2.c",
367 "src/f32-vunary/gen/vabs-scalar-x4.c",
368 "src/f32-vunary/gen/vneg-scalar-x1.c",
369 "src/f32-vunary/gen/vneg-scalar-x2.c",
370 "src/f32-vunary/gen/vneg-scalar-x4.c",
371 "src/f32-vunary/gen/vsqr-scalar-x1.c",
372 "src/f32-vunary/gen/vsqr-scalar-x2.c",
373 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800374 "src/math/expminus-scalar-lut2048-p1.c",
375 "src/math/expminus-scalar-lut64-p2.c",
376 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700377 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700378 "src/math/roundne-scalar-nearbyint.c",
379 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700380 "src/math/roundd-scalar-addsub.c",
381 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700382 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700383 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700384 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700385 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700386 "src/math/roundz-scalar-addsub.c",
387 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700388 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800389 "src/math/sigmoid-scalar-lut2048-p1-div.c",
390 "src/math/sigmoid-scalar-lut64-p2-div.c",
391 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700392 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
393 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700394 "src/qs8-requantization/fp32-scalar-lrintf.c",
395 "src/qs8-requantization/fp32-scalar-magic.c",
396 "src/qs8-requantization/precise-scalar-signed64.c",
397 "src/qs8-requantization/precise-scalar-unsigned32.c",
398 "src/qs8-requantization/precise-scalar-unsigned64.c",
399 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700400 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
401 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
402 "src/qu8-dwconv/up1x9-minmax-scalar.c",
403 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
404 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700405 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
406 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700407 "src/qu8-gemm/2x2-minmax-scalar.c",
408 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700409 "src/qu8-requantization/fp32-scalar-lrintf.c",
410 "src/qu8-requantization/fp32-scalar-magic.c",
411 "src/qu8-requantization/precise-scalar-signed64.c",
412 "src/qu8-requantization/precise-scalar-unsigned32.c",
413 "src/qu8-requantization/precise-scalar-unsigned64.c",
414 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700415 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700416 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700417 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700418 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700419 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700420 "src/x32-fill/scalar-float.c",
421 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700422 "src/x32-packx/x2-scalar.c",
423 "src/x32-packx/x3-scalar.c",
424 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700425 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700426 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700427 "src/x32-unpool/scalar.c",
428 "src/x32-zip/x2-scalar.c",
429 "src/x32-zip/x3-scalar.c",
430 "src/x32-zip/x4-scalar.c",
431 "src/x32-zip/xm-scalar.c",
432 "src/x8-lut/scalar.c",
433 "src/x8-zip/x2-scalar.c",
434 "src/x8-zip/x3-scalar.c",
435 "src/x8-zip/x4-scalar.c",
436 "src/x8-zip/xm-scalar.c",
437]
438
Marat Dukhan436ebe62019-12-04 15:10:12 -0800439WASM_UKERNELS = [
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",
Frank Barchard37297a62020-08-27 12:15:10 -0700509 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700510 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
511 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
512 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
513 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
514 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
515 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
516 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
517 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
518 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
519 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
520 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
521 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700522 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
523 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
524 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
525 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
526 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
527 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
528 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
529 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
530 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
531 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
532 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
533 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800534 "src/f32-vbinary/gen/vmax-wasm-x1.c",
535 "src/f32-vbinary/gen/vmax-wasm-x2.c",
536 "src/f32-vbinary/gen/vmax-wasm-x4.c",
537 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
538 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
539 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
540 "src/f32-vbinary/gen/vmin-wasm-x1.c",
541 "src/f32-vbinary/gen/vmin-wasm-x2.c",
542 "src/f32-vbinary/gen/vmin-wasm-x4.c",
543 "src/f32-vbinary/gen/vminc-wasm-x1.c",
544 "src/f32-vbinary/gen/vminc-wasm-x2.c",
545 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
549 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
550 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
551 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
552 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
553 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
554 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
555 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
556 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
557 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
558 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
561 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
562 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
563 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700564 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
565 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
566 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
567 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
568 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
569 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
570 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
573 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
574 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
575 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
576 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
577 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
578 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
579 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
580 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
581 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700582 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
583 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
584 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700585 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
586 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
587 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800588]
589
Marat Dukhan290055c2020-06-09 12:24:29 -0700590WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700591 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
592 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
593 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700594 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
595 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
596 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
597 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700598 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
599 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
600 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
601 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700602 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
603 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
604 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
605 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
606 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
607 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700608 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
609 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
610 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
611 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
612 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
613 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
614 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
615 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
616 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
617 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
618 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
619 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
620 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
621 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
622 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
623 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
624 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
625 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
626 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
627 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
628 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
629 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
630 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
631 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700632 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
633 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
634 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700635 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700636 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
637 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
638 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
639 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700640 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
641 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
642 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
643 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
644 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
645 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
646 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
647 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
648 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
649 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
650 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
651 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
652 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
653 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
654 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
655 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
656 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
657 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
658 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
659 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
660 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
661 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
662 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
663 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
664 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
665 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
666 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
667 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
668 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
669 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700670 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
671 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700672 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
673 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
674 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
675 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
676 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
677 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
678 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
679 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
680 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
681 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
682 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
683 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
684 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
685 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
686 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
687 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
688 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
689 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
690 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
691 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
692 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
693 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
694 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
695 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
696 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
697 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
698 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
699 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
700 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
701 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700702 "src/f32-ibilinear/gen/wasmsimd-c4.c",
703 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700704 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
705 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
706 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700707 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700708 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
709 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
710 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
711 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700712 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
716 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
717 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
718 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
719 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
720 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
721 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
722 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
723 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
724 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
725 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
726 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
727 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
730 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
731 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
732 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
733 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
734 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
735 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
736 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
737 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
738 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
739 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
740 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
741 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700742 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
743 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700744 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
745 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
746 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
747 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700748 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
749 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700750 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
751 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
752 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700753 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
754 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700755 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
756 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
757 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
758 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700759 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
760 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700761 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
762 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
763 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700764 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
765 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700766 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
767 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
768 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
769 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
770 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
771 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
772 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700773 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
774 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700775 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
776 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
777 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
778 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700779 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
780 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
781 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
782 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
783 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
784 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
785 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
786 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
787 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
788 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
789 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
790 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700791 "src/f32-relu/gen/wasmsimd-x4.c",
792 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700793 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700794 "src/f32-rmax/wasmsimd-arm.c",
795 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700796 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
797 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
798 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
799 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
800 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
801 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700802 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
803 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
804 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
805 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
806 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
807 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700808 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
809 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
810 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
811 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700812 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
813 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700814 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
815 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700816 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
817 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
818 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
819 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700820 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
821 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700822 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
823 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700824 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
825 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
826 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
827 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700828 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
829 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700830 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
831 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700832 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
833 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
834 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
835 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700836 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
837 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700838 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
839 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700840 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
841 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
842 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
843 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700844 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
845 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
846 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
847 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700848 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
849 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
850 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
851 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700852 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
853 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
854 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
855 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700856 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
857 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
858 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
859 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700860 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
861 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700862 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
863 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700864 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
865 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
866 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
867 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700868 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
869 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700870 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
871 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700872 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
873 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
874 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
875 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700876 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
877 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700878 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
879 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
882 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
883 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700884 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
885 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700886 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700888 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
889 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700890 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
891 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
894 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
895 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700896 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
897 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700898 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700900 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
901 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
902 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
903 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700904 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
905 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700906 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
907 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700908 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
909 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
910 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
911 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -0700912 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
913 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
914 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
915 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -0700916 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
917 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
918 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
919 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
920 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
921 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
922 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
923 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
924 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
925 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
926 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
927 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
928 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
929 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700930 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
931 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -0700932 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
933 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
934 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
935 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
936 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
937 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -0700938 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
939 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
940 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -0700941 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
942 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
943 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -0700944 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
945 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
946 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -0700947 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
948 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
949 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
950 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
951 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
952 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
953 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
954 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
955 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
956 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
957 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
958 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
959 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
960 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
961 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700962 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -0700963 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700964 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -0700965 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -0700966 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -0700967 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -0700968 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -0700969 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -0700970 "src/x32-zip/x2-wasmsimd.c",
971 "src/x32-zip/x3-wasmsimd.c",
972 "src/x32-zip/x4-wasmsimd.c",
973 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700974 "src/math/roundne-wasmsimd-addsub.c",
975 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700976 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700977 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700978 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -0700979 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -0700980 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -0700981 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
982 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -0700983]
984
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800985PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -0700986 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -0700987 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -0700988 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -0700989 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -0700990 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -0700991 "src/f32-spmm/gen/4x1-minmax-psimd.c",
992 "src/f32-spmm/gen/8x1-minmax-psimd.c",
993 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700994]
995
Marat Dukhan8d3c07e2020-01-02 01:20:59 -0800996PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700997 "src/qs8-requantization/precise-psimd.c",
998 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700999 "src/qu8-requantization/precise-psimd.c",
1000 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001001]
1002
Marat Dukhan08c4a432019-10-03 09:29:21 -07001003# ISA-specific micro-kernels
1004NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001005 "src/f32-argmaxpool/4x-neon-c4.c",
1006 "src/f32-argmaxpool/9p8x-neon-c4.c",
1007 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001008 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1009 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001010 "src/f32-clamp/gen/neon-x4.c",
1011 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001012 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1013 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001014 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1015 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001016 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1017 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001018 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1019 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001020 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1021 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1022 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1023 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001024 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1025 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1026 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1027 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001028 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1029 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1030 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1031 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001032 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001033 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1034 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001035 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1036 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1037 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1038 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1039 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1040 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1041 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1042 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1043 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1044 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1045 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1046 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1047 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1048 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1049 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1050 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1051 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1052 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1053 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1054 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1055 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1056 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1057 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1058 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1059 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1060 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1061 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1062 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1063 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1064 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1065 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001066 "src/f32-hswish/gen/hswish-neon-x4.c",
1067 "src/f32-hswish/gen/hswish-neon-x8.c",
1068 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001069 "src/f32-ibilinear/gen/neon-c4.c",
1070 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001071 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1072 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1073 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1074 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1075 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1076 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1077 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1078 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1079 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1080 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1081 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1082 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1083 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1084 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1085 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1086 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001087 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1088 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1089 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001090 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1091 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001092 "src/f32-prelu/gen/neon-1x4.c",
1093 "src/f32-prelu/gen/neon-1x8.c",
1094 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001095 "src/f32-prelu/gen/neon-2x4.c",
1096 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001097 "src/f32-prelu/gen/neon-2x16.c",
1098 "src/f32-prelu/gen/neon-4x4.c",
1099 "src/f32-prelu/gen/neon-4x8.c",
1100 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001101 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1102 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1103 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1104 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1105 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1106 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1107 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1108 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1109 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1110 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1111 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1112 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1113 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1114 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1115 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1116 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1117 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1118 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1119 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1120 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1121 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1122 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1123 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1124 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001125 "src/f32-relu/gen/neon-x4.c",
1126 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001127 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001128 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001129 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1130 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1131 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1132 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1133 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1134 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1135 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1136 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1137 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1138 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1139 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1140 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1141 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1142 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1143 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1144 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1145 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1146 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001147 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1148 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1149 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1150 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001151 "src/f32-vbinary/gen/vmax-neon-x4.c",
1152 "src/f32-vbinary/gen/vmax-neon-x8.c",
1153 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1154 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1155 "src/f32-vbinary/gen/vmin-neon-x4.c",
1156 "src/f32-vbinary/gen/vmin-neon-x8.c",
1157 "src/f32-vbinary/gen/vminc-neon-x4.c",
1158 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001159 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1160 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1161 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1162 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1163 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1164 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001165 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1166 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1167 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1168 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001169 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1170 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1171 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1172 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001173 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1174 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001175 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1176 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001177 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1178 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1179 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1180 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1181 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1182 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1183 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1184 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001185 "src/f32-vunary/gen/vabs-neon-x4.c",
1186 "src/f32-vunary/gen/vabs-neon-x8.c",
1187 "src/f32-vunary/gen/vneg-neon-x4.c",
1188 "src/f32-vunary/gen/vneg-neon-x8.c",
1189 "src/f32-vunary/gen/vsqr-neon-x4.c",
1190 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001191 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1192 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1193 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1194 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001195 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1196 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1197 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1198 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1199 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1200 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1201 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1202 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001203 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1204 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1205 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1206 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1207 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1208 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1209 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1210 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001211 "src/qs8-requantization/precise-neon.c",
1212 "src/qs8-requantization/fp32-neon.c",
1213 "src/qs8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001214 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1215 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1216 "src/qu8-dwconv/up8x9-minmax-neon.c",
1217 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1218 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1219 "src/qu8-gemm/4x8-minmax-neon.c",
1220 "src/qu8-gemm/8x8-minmax-neon.c",
1221 "src/qu8-igemm/4x8-minmax-neon.c",
1222 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001223 "src/qu8-requantization/precise-neon.c",
1224 "src/qu8-requantization/fp32-neon.c",
1225 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001226 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001227 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001228 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001229 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001230 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001231 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001232 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001233 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001234 "src/x32-zip/x2-neon.c",
1235 "src/x32-zip/x3-neon.c",
1236 "src/x32-zip/x4-neon.c",
1237 "src/x32-zip/xm-neon.c",
1238 "src/x8-zip/x2-neon.c",
1239 "src/x8-zip/x3-neon.c",
1240 "src/x8-zip/x4-neon.c",
1241 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001242 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001243 "src/math/roundd-neon-addsub.c",
1244 "src/math/roundd-neon-cvt.c",
1245 "src/math/roundu-neon-addsub.c",
1246 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001247 "src/math/roundz-neon-addsub.c",
1248 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001249 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001250 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1251 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1252 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1253 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1254 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1255 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001256 "src/math/sqrt-neon-nr1rsqrts.c",
1257 "src/math/sqrt-neon-nr2rsqrts.c",
1258 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001259]
1260
1261NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001262 "src/f32-ibilinear/gen/neonfma-c4.c",
1263 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001264 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1265 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1266 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1267 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1268 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1269 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1270 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1271 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1272 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001273 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1274 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1275 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1276 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001277 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1278 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1279 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1280 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001281 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1282 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1283 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1284 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001285 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1286 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1287 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1288 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1289 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1290 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1291 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1292 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1293 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1294 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1295 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1296 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1297 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1298 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1299 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1300 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1301 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1302 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001303 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1304 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001305 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1306 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1307 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1308 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1309 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1310 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1311 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1312 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1313 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1314 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1315 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1316 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1317 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1318 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1319 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1320 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1321 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1322 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1323 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1324 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1325 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1326 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1327 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1328 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001329 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1330 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1331 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1332 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1333 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1334 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1335 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1336 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1337 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1338 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1339 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1340 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1341 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1342 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1343 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1344 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1345 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1346 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1347 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1348 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1349 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1350 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1351 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1352 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1353 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1354 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1355 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1356 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1357 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1358 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1359 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1360 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1361 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1362 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1363 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1364 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1365 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1366 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1367 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1368 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1369 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1370 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1371 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1372 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1373 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1374 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1375 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1376 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1377 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1378 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1379 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1380 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1381 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1382 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001383 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1384 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001385 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1386 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1387 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1388 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1389 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1390 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1391 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1392 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1393 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1394 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1395 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1396 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1397 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1398 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1399 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1400 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1401 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1402 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1403 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1404 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001405 "src/math/exp-neonfma-lut64-p2.c",
1406 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001407 "src/math/expminus-neonfma-lut2048-p1.c",
1408 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001409 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001410 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1411 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1412 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1413 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1414 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1415 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1416 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1417 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1418 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1419 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1420 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1421 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1422 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1423 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1424 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1425 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1426 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1427 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001428 "src/math/sqrt-neonfma-nr1fma.c",
1429 "src/math/sqrt-neonfma-nr2fma.c",
1430 "src/math/sqrt-neonfma-nr3fma.c",
1431 "src/math/sqrt-neonfma-nr2fma1adj.c",
1432 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001433]
1434
1435AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001436 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1437 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1438 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1439 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1440 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1441 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001442 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1443 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1444 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1445 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1446 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1447 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1448 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1449 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1450 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1451 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1452 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1453 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1454 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1455 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1456 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1457 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1458 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1459 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1460 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1461 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001462 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1463 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001464 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1465 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001466 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1467 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001468 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1469 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001470 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1471 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1472 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1473 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1474 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001475 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001493 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1494 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1495 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1496 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1497 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1498 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1499 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1500 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1501 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1502 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1503 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1504 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1505 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1506 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1507 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1508 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1509 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1510 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1512 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001513 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1514 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1515 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1516 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1517 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1518 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001519]
1520
Marat Dukhan8853b822020-05-07 12:19:01 -07001521NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001522 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1523 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1524 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1525 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1526 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1527 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1528 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1529 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001530 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001531 "src/math/roundd-neonv8.c",
1532 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001533 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001534]
1535
Marat Dukhan08c4a432019-10-03 09:29:21 -07001536AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001537 "src/f16-clamp/gen/neonfp16arith-x16.c",
1538 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001539 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1540 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1541 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1542 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1543 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1544 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1545 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1546 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1547 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1548 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1549 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1550 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001551 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1552 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001553 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1554 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001555 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001556 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001557 "src/f16-relu/gen/neonfp16arith-x16.c",
1558 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001559 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1560 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1561 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1562 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1563 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1564 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1565 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1566 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1567 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1568 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1569 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1570 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1571 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1572 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1573 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1574 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1575 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1576 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1577 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1578 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1579 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1580 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1581 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1582 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1583 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1584 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1585 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1586 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001587 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1588 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001589 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1590 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1591 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1592 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1593 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1594 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1595 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1596 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001597 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1598 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1599 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1600 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001601 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1602 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1603 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1604 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1605 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1606 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1607 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1608 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1609 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1610 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1611 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1612 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001613 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1614 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1615 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1616 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1617 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1618 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1619 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1620 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001621]
1622
Benoit Jacoba9644732020-08-13 12:48:55 -07001623NEONDOT_UKERNELS = [
1624 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1625 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1626 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001627 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1628 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1629 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001630]
1631
Marat Dukhan08c4a432019-10-03 09:29:21 -07001632SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001633 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1634 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001635 "src/f32-clamp/gen/sse-x4.c",
1636 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001637 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1638 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001639 "src/f32-dwconv-chw/3x3p1-sse.c",
1640 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001641 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1642 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1643 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1644 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1645 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1646 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1647 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1648 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1649 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1650 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1651 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1652 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001653 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001654 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1655 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001656 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1657 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1658 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1659 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1660 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1661 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1662 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1663 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1664 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1665 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1666 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1667 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1668 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001669 "src/f32-hswish/gen/hswish-sse-x4.c",
1670 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001671 "src/f32-ibilinear/gen/sse-c4.c",
1672 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001673 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1674 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1675 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1676 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1677 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1678 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1679 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001680 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1681 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1682 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001683 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001684 "src/f32-prelu/gen/sse-2x4.c",
1685 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001686 "src/f32-relu/gen/sse-x4.c",
1687 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001688 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001689 "src/f32-spmm/gen/4x1-minmax-sse.c",
1690 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001691 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001692 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1693 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1694 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1695 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1696 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1697 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1698 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1699 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001700 "src/f32-vbinary/gen/vmax-sse-x4.c",
1701 "src/f32-vbinary/gen/vmax-sse-x8.c",
1702 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1703 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1704 "src/f32-vbinary/gen/vmin-sse-x4.c",
1705 "src/f32-vbinary/gen/vmin-sse-x8.c",
1706 "src/f32-vbinary/gen/vminc-sse-x4.c",
1707 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001708 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1709 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1710 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1711 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1712 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1713 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1714 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1715 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001716 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1717 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1718 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1719 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001720 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1721 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1722 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1723 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001724 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1725 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001726 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1727 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001728 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1729 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001730 "src/f32-vunary/gen/vabs-sse-x4.c",
1731 "src/f32-vunary/gen/vabs-sse-x8.c",
1732 "src/f32-vunary/gen/vneg-sse-x4.c",
1733 "src/f32-vunary/gen/vneg-sse-x8.c",
1734 "src/f32-vunary/gen/vsqr-sse-x4.c",
1735 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001736 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001737 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001738 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001739 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001740 "src/math/roundd-sse-addsub.c",
1741 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001742 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001743 "src/math/sqrt-sse-hh1mac.c",
1744 "src/math/sqrt-sse-nr1mac.c",
1745 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001746]
1747
1748SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001749 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1750 "src/f32-argmaxpool/4x-sse2-c4.c",
1751 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001752 "src/f32-prelu/gen/sse2-2x4.c",
1753 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001754 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1755 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1756 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1757 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1758 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1759 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1760 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1761 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1762 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1763 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1764 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1765 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001766 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001767 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001768 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001769 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001770 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1771 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001772 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1773 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001774 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1775 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1776 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1777 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1778 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1779 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1780 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1781 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001782 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1783 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1784 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001785 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1786 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1787 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001788 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1789 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1790 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001791 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1792 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001793 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1794 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001795 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1796 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001797 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1798 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001799 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001800 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1801 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001802 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001803 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1804 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1805 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001806 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1807 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001808 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1809 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001810 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1811 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001812 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001813 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1814 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001815 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001816 "src/qs8-requantization/precise-sse2.c",
1817 "src/qs8-requantization/fp32-sse2.c",
1818 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001819 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1820 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1821 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1822 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001823 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1824 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1825 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1826 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001827 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1828 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1829 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1830 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1831 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1832 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1833 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1834 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001835 "src/qu8-requantization/precise-sse2.c",
1836 "src/qu8-requantization/fp32-sse2.c",
1837 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001838 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001839 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001840 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001841 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001842 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001843 "src/x32-zip/x2-sse2.c",
1844 "src/x32-zip/x3-sse2.c",
1845 "src/x32-zip/x4-sse2.c",
1846 "src/x32-zip/xm-sse2.c",
1847 "src/x8-zip/x2-sse2.c",
1848 "src/x8-zip/x3-sse2.c",
1849 "src/x8-zip/x4-sse2.c",
1850 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001851 "src/math/exp-sse2-p5.c",
1852 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001853 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001854 "src/math/roundd-sse2-cvt.c",
1855 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001856 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001857 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001858]
1859
1860SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001861 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1862 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1863 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001864 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1865 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1866 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001867 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1868 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1869 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001870 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1871 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001872 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1873 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001874 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1875 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001876 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1877 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001878 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001879 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1880 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001881 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001882 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1883 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1884 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001885 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1886 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001887 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1888 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001889 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1890 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001891 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001892 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1893 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001894 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001895 "src/qs8-requantization/precise-ssse3.c",
1896 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001897 "src/qu8-requantization/precise-ssse3.c",
1898 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001899]
1900
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001901SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08001902 "src/f32-prelu/gen/sse41-2x4.c",
1903 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001904 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001905 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001906 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001907 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001908 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1909 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001910 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1911 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001912 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1913 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1914 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1915 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
1916 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1917 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1918 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1919 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001920 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1921 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1922 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001923 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1924 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1925 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001926 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1927 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1928 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001929 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001931 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1932 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001933 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1934 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001935 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1936 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001937 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001938 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1939 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001940 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001941 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1942 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1943 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001944 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1945 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001946 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1947 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001948 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1949 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001950 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001951 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1952 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001953 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001954 "src/qs8-requantization/fp32-sse4.c",
1955 "src/qs8-requantization/precise-sse4.c",
1956 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001957 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1958 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1959 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1960 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001961 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
1962 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
1963 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
1964 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001965 "src/qu8-requantization/precise-sse4.c",
1966 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001967 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001968 "src/math/roundd-sse41.c",
1969 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001970 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08001971]
1972
Marat Dukhan08c4a432019-10-03 09:29:21 -07001973AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001974 "src/f32-clamp/gen/avx-x8.c",
1975 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001976 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1977 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1978 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1979 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
1980 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1981 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1982 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1983 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
1984 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1985 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
1986 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1987 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
1988 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1989 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1990 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1991 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1992 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
1993 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1994 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1995 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1996 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1997 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
1998 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
1999 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2000 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2001 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2002 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2003 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2004 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2005 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002006 "src/f32-hswish/gen/hswish-avx-x8.c",
2007 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002008 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2009 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2010 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2011 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2012 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2013 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2014 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2015 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2016 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002017 "src/f32-prelu/gen/avx-2x8.c",
2018 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002019 "src/f32-relu/gen/avx-x16.c",
2020 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002021 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002022 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2023 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2024 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2025 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2026 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2027 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2028 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2029 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002030 "src/f32-vbinary/gen/vmax-avx-x8.c",
2031 "src/f32-vbinary/gen/vmax-avx-x16.c",
2032 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2033 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2034 "src/f32-vbinary/gen/vmin-avx-x8.c",
2035 "src/f32-vbinary/gen/vmin-avx-x16.c",
2036 "src/f32-vbinary/gen/vminc-avx-x8.c",
2037 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002038 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2039 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2040 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2041 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2042 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2043 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2044 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2045 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002046 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2047 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2048 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2049 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002050 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2051 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2052 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2053 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002054 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2055 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002056 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002057 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2058 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2059 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2060 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2061 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2062 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2063 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2064 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002065 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2066 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002067 "src/f32-vunary/gen/vabs-avx-x8.c",
2068 "src/f32-vunary/gen/vabs-avx-x16.c",
2069 "src/f32-vunary/gen/vneg-avx-x8.c",
2070 "src/f32-vunary/gen/vneg-avx-x16.c",
2071 "src/f32-vunary/gen/vsqr-avx-x8.c",
2072 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002073]
2074
Marat Dukhan1566fee2020-08-02 21:55:41 -07002075XOP_UKERNELS = [
2076 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2077 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2078 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2079 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002080 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2081 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002082 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2083 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002084 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002085 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2086 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002087 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002088 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2089 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2090 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002091 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2092 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2093 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2094 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2095 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2096 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002097 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002098 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2099 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002100 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002101]
2102
Marat Dukhanfda12b82019-11-21 12:27:59 -08002103FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002104 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2105 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2106 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2107 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2108 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2109 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2110 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2111 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2112 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2113 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2114 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2115 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2116 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2117 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2118 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2119 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2120 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2121 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2122 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2123 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2124 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2125 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2126 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2127 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2128 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2129 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2130 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2131 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2132 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2133 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2134 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2135 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2136 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2137 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2138 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2139 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2140 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2141 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2142 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2143 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002144 "src/f32-hswish/gen/hswish-fma3-x8.c",
2145 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002146 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2147 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2148 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2149 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2150 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2151 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2152 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2153 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2154 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2155 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2156 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2157 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2158 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2159 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002160 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2161 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2162 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2163 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2164 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2165 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2166 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2167 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002168 "src/math/sqrt-fma3-nr1fma.c",
2169 "src/math/sqrt-fma3-nr2fma.c",
2170 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002171]
2172
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002173AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002174 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2175 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2176 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2177 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2178 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2179 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2180 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2181 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2182 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2183 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2184 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2185 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2186 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2187 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2188 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2189 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2190 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2191 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2192 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2193 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2194 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2195 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2196 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2197 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2198 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2199 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2200 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2201 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2202 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2203 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2204 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2205 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2206 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2207 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2208 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2209 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002210 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2211 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2212 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2213 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2214 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2215 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2216 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2217 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2218 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2219 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2220 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2221 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2222 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2223 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2224 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2225 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2226 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2227 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2228 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2229 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2230 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2231 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2232 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2233 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2234 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2235 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2236 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2237 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2238 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2239 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2244 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2245 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2246 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2247 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2248 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2249 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2250 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2251 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2252 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2253 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2254 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2255 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2256 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2257 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2258 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2259 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2260 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2261 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2262 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2263 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002264 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2265 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002266 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2267 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2268 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2269 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002270 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2271 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2272 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002273 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2274 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2275 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002276 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2277 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2278 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002279 "src/math/exp-avx2-p5.c",
2280 "src/math/exp-avx2-perm-p3.c",
2281 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002282 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002283 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002284 "src/math/sigmoid-avx2-rr2-p5-div.c",
2285 "src/math/sigmoid-avx2-rr1-p5-div.c",
2286 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2287 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2288 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2289 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002290]
2291
Marat Dukhan08c4a432019-10-03 09:29:21 -07002292AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002293 "src/f32-clamp/gen/avx512f-x16.c",
2294 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002295 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2296 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2297 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2298 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2299 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2300 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2301 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2302 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2303 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2304 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2305 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2306 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2307 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2308 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2309 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2310 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2311 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2312 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2313 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2314 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2315 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2316 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2317 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2318 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002319 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2320 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002321 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2322 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2323 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2324 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2325 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2326 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002327 "src/f32-prelu/gen/avx512f-2x16.c",
2328 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002329 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2330 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2331 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2332 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2333 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2334 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2335 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2336 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2337 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2338 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2339 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2340 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2341 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2342 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2343 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2344 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2345 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2346 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2347 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2348 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2349 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2350 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2351 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2352 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2353 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2354 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2355 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2356 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2357 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2358 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2359 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2360 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2361 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2362 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2363 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2364 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002365 "src/f32-relu/gen/avx512f-x16.c",
2366 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002367 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002368 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2369 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2370 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2371 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2372 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2373 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2374 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2375 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002376 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2377 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2378 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2379 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2380 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2381 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2382 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2383 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002384 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2385 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2386 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2387 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2388 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2389 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2390 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2391 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002392 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2393 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2394 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2395 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002396 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2397 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2398 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2399 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002400 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2401 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002402 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002403 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2404 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2405 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2406 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2407 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2408 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2409 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2410 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2411 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2412 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2413 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2414 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2415 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2416 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2417 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2418 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2419 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2420 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2421 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2422 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2423 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2424 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2425 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2426 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002427 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2428 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2429 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2430 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2431 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2432 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2433 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2434 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002435 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2436 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2437 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2438 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2439 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2440 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2441 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2442 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002443 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2444 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2445 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2446 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2447 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2448 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002449 "src/math/exp-avx512f-p5-scalef.c",
2450 "src/math/exp-avx512f-p5.c",
2451 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002452 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002453 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002454 "src/math/sqrt-avx512f-nr1fma.c",
2455 "src/math/sqrt-avx512f-nr2fma.c",
2456 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002457]
2458
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002459AVX512SKX_UKERNELS = [
2460 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2461 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2462 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2463 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2464 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2465 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2466 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2467 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2468]
2469
Frank Barchardbcedc082020-08-17 18:00:51 -07002470WASM32_ASM_UKERNELS = [
2471 "src/f32-relu/wasm_shr_x1.S",
2472 "src/f32-relu/wasm_shr_x2.S",
2473 "src/f32-relu/wasm_shr_x4.S",
2474]
2475
Marat Dukhan08c4a432019-10-03 09:29:21 -07002476AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002477 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002478 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002479 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2480 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002481 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002482 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002483 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002484 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002485 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2486 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002487 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2488 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2489 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2490 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002491]
2492
2493AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002494 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002495 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2496 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002497 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002498 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2499 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002500 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2501 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2502 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002503 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002504 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2505 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2506 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002507 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002508 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2509 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002510 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002511 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2512 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2513 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2514 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2515 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2516 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2517 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2518 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2519 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2520 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2521 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2522 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2523 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2524 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2525 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2526 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2527 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2528 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002529 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2530 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002531 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002532 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2533 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2534 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2535 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2536 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2537 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2538 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2539 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2540 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2541 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2542 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2543 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2544 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2545 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2546 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2547 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2548 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2549 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002550 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2551 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2552 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2553 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2554 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2555 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2556 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2557 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2558 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2559 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2560 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2561 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2562 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2563 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2564 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2565 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2566 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2567 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002568]
2569
Marat Dukhan1b354632020-03-23 12:50:22 -07002570INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002571 "src/xnnpack/argmaxpool.h",
2572 "src/xnnpack/avgpool.h",
2573 "src/xnnpack/clamp.h",
2574 "src/xnnpack/common.h",
2575 "src/xnnpack/conv.h",
2576 "src/xnnpack/dwconv.h",
2577 "src/xnnpack/gavgpool.h",
2578 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002579 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002580 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002581 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002582 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002583 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002584 "src/xnnpack/lut.h",
2585 "src/xnnpack/math.h",
2586 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002587 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002588 "src/xnnpack/packx.h",
2589 "src/xnnpack/pad.h",
2590 "src/xnnpack/params.h",
2591 "src/xnnpack/pavgpool.h",
2592 "src/xnnpack/ppmm.h",
2593 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002594 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002595 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002596 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002597 "src/xnnpack/rmax.h",
2598 "src/xnnpack/scalar-utils.h",
2599 "src/xnnpack/spmm.h",
2600 "src/xnnpack/unpool.h",
2601 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002602 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002603 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002604 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002605 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002606 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002607 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002608 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002609]
2610
2611INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002612 "include/xnnpack.h",
2613 "src/xnnpack/allocator.h",
2614 "src/xnnpack/compute.h",
2615 "src/xnnpack/im2col.h",
2616 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002617 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002618 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002619 "src/xnnpack/operator.h",
2620 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002621 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002622 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002623 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002624 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002625]
2626
Marat Dukhan1b354632020-03-23 12:50:22 -07002627ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002628 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002629]
2630
Marat Dukhan1b354632020-03-23 12:50:22 -07002631MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002632 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002633 "include/xnnpack.h",
2634]
2635
Marat Dukhan1b354632020-03-23 12:50:22 -07002636MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002637 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002638 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002639 "src/xnnpack/requantization.h",
2640 "include/xnnpack.h",
2641]
2642
2643OPERATOR_TEST_PARAMS_HDRS = [
2644 "src/xnnpack/params.h",
2645 "src/xnnpack/common.h",
2646]
2647
2648WEIGHTS_PACK_HDRS = [
2649 "src/xnnpack/pack.h",
2650 "src/xnnpack/operator.h",
2651 "src/xnnpack/compute.h",
2652]
2653
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002654LOGGING_COPTS = select({
2655 # No logging in optimized mode
2656 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2657 # Full logging in debug mode
2658 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2659 # Error-only logging in default (fastbuild) mode
2660 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2661})
2662
Marat Dukhan3b59de22020-06-03 20:15:19 -07002663LOGGING_SRCS = select({
2664 # No logging in optimized mode
2665 ":optimized_build": [],
2666 "//conditions:default": [
2667 "src/operator-strings.c",
2668 "src/subgraph-strings.c",
2669 ],
2670})
2671
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002672LOGGING_HDRS = [
2673 "src/xnnpack/log.h",
2674]
2675
Marat Dukhan08c4a432019-10-03 09:29:21 -07002676xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002677 name = "tables",
2678 srcs = TABLE_SRCS,
2679 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002680 gcc_copts = xnnpack_gcc_std_copts(),
2681 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002682)
2683
2684xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002685 name = "scalar_ukernels",
2686 srcs = SCALAR_UKERNELS,
2687 hdrs = INTERNAL_HDRS,
2688 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002689 gcc_copts = xnnpack_gcc_std_copts(),
2690 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002691 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002692 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002693 "@FP16",
2694 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002695 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002696 ],
2697)
2698
2699xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002700 name = "scalar_ukernels_test_mode",
2701 srcs = SCALAR_UKERNELS,
2702 hdrs = INTERNAL_HDRS,
2703 aarch32_copts = ["-marm"],
2704 copts = [
2705 "-UNDEBUG",
2706 "-DXNN_TEST_MODE=1",
2707 ],
2708 gcc_copts = xnnpack_gcc_std_copts(),
2709 msvc_copts = xnnpack_msvc_std_copts(),
2710 deps = [
2711 ":tables",
2712 "@FP16",
2713 "@FXdiv",
2714 "@pthreadpool",
2715 ],
2716)
2717
2718xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002719 name = "wasm_ukernels",
2720 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002721 gcc_copts = xnnpack_gcc_std_copts(),
2722 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002723 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002724 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002725 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002726 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002727 "@FP16",
2728 "@FXdiv",
2729 "@pthreadpool",
2730 ],
2731)
2732
2733xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002734 name = "wasm_ukernels_test_mode",
2735 hdrs = INTERNAL_HDRS,
2736 copts = [
2737 "-UNDEBUG",
2738 "-DXNN_TEST_MODE=1",
2739 ],
2740 gcc_copts = xnnpack_gcc_std_copts(),
2741 msvc_copts = xnnpack_msvc_std_copts(),
2742 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002743 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002744 deps = [
2745 ":tables",
2746 "@FP16",
2747 "@FXdiv",
2748 "@pthreadpool",
2749 ],
2750)
2751
2752xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002753 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002754 hdrs = INTERNAL_HDRS,
2755 aarch32_copts = [
2756 "-marm",
2757 "-mfpu=neon",
2758 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002759 gcc_copts = xnnpack_gcc_std_copts(),
2760 gcc_x86_copts = ["-msse2"],
2761 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002762 optimized_copts = [
2763 "-O3",
2764 "-ffast-math",
2765 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002766 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002767 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002768 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002769 "@FP16",
2770 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002771 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002772 ],
2773)
2774
2775xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002776 name = "psimd_fastmath_ukernels_test_mode",
2777 hdrs = INTERNAL_HDRS,
2778 aarch32_copts = [
2779 "-marm",
2780 "-mfpu=neon",
2781 ],
2782 copts = [
2783 "-UNDEBUG",
2784 "-DXNN_TEST_MODE=1",
2785 ],
2786 gcc_copts = xnnpack_gcc_std_copts(),
2787 gcc_x86_copts = ["-msse2"],
2788 msvc_copts = xnnpack_msvc_std_copts(),
2789 optimized_copts = [
2790 "-O3",
2791 "-ffast-math",
2792 ],
2793 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2794 deps = [
2795 ":tables",
2796 "@FP16",
2797 "@psimd",
2798 "@pthreadpool",
2799 ],
2800)
2801
2802xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002803 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002804 hdrs = INTERNAL_HDRS,
2805 aarch32_copts = [
2806 "-marm",
2807 "-mfpu=neon",
2808 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002809 gcc_copts = xnnpack_gcc_std_copts(),
2810 gcc_x86_copts = ["-msse2"],
2811 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002812 optimized_copts = [
2813 "-O3",
2814 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002815 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002816 deps = [
2817 ":tables",
2818 "@FP16",
2819 "@psimd",
2820 "@pthreadpool",
2821 ],
2822)
2823
2824xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002825 name = "psimd_accmath_ukernels_test_mode",
2826 hdrs = INTERNAL_HDRS,
2827 aarch32_copts = [
2828 "-marm",
2829 "-mfpu=neon",
2830 ],
2831 copts = [
2832 "-UNDEBUG",
2833 "-DXNN_TEST_MODE=1",
2834 ],
2835 gcc_copts = xnnpack_gcc_std_copts(),
2836 gcc_x86_copts = ["-msse2"],
2837 msvc_copts = xnnpack_msvc_std_copts(),
2838 optimized_copts = [
2839 "-O3",
2840 ],
2841 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2842 deps = [
2843 ":tables",
2844 "@FP16",
2845 "@psimd",
2846 "@pthreadpool",
2847 ],
2848)
2849
2850xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002851 name = "neon_ukernels",
2852 hdrs = INTERNAL_HDRS,
2853 aarch32_copts = [
2854 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002855 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002856 "-mfpu=neon",
2857 ],
2858 aarch32_srcs = NEON_UKERNELS,
2859 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002860 gcc_copts = xnnpack_gcc_std_copts(),
2861 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002862 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002863 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002864 "@FP16",
2865 "@pthreadpool",
2866 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002867)
2868
2869xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002870 name = "neon_ukernels_test_mode",
2871 hdrs = INTERNAL_HDRS,
2872 aarch32_copts = [
2873 "-marm",
2874 "-march=armv7-a",
2875 "-mfpu=neon",
2876 ],
2877 aarch32_srcs = NEON_UKERNELS,
2878 aarch64_srcs = NEON_UKERNELS,
2879 copts = [
2880 "-UNDEBUG",
2881 "-DXNN_TEST_MODE=1",
2882 ],
2883 gcc_copts = xnnpack_gcc_std_copts(),
2884 msvc_copts = xnnpack_msvc_std_copts(),
2885 deps = [
2886 ":tables",
2887 "@FP16",
2888 "@pthreadpool",
2889 ],
2890)
2891
2892xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002893 name = "neonfma_ukernels",
2894 hdrs = INTERNAL_HDRS,
2895 aarch32_copts = [
2896 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002897 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 "-mfpu=neon-vfpv4",
2899 ],
2900 aarch32_srcs = NEONFMA_UKERNELS,
2901 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002902 apple_aarch32_copts = [
2903 "-mcpu=swift",
2904 "-mtune=generic",
2905 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002906 gcc_copts = xnnpack_gcc_std_copts(),
2907 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002908 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002909 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002910 "@FP16",
2911 "@pthreadpool",
2912 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002913)
2914
2915xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002916 name = "neonfma_ukernels_test_mode",
2917 hdrs = INTERNAL_HDRS,
2918 aarch32_copts = [
2919 "-marm",
2920 "-march=armv7-a",
2921 "-mfpu=neon-vfpv4",
2922 ],
2923 aarch32_srcs = NEONFMA_UKERNELS,
2924 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002925 apple_aarch32_copts = [
2926 "-mcpu=swift",
2927 "-mtune=generic",
2928 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002929 copts = [
2930 "-UNDEBUG",
2931 "-DXNN_TEST_MODE=1",
2932 ],
2933 gcc_copts = xnnpack_gcc_std_copts(),
2934 msvc_copts = xnnpack_msvc_std_copts(),
2935 deps = [
2936 ":tables",
2937 "@FP16",
2938 "@pthreadpool",
2939 ],
2940)
2941
2942xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07002943 name = "neonv8_ukernels",
2944 hdrs = INTERNAL_HDRS,
2945 aarch32_copts = [
2946 "-marm",
2947 "-march=armv8-a",
2948 "-mfpu=neon-fp-armv8",
2949 ],
2950 aarch32_srcs = NEONV8_UKERNELS,
2951 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002952 apple_aarch32_copts = [
2953 "-mcpu=cyclone",
2954 "-mtune=generic",
2955 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07002956 gcc_copts = xnnpack_gcc_std_copts(),
2957 msvc_copts = xnnpack_msvc_std_copts(),
2958 deps = [
2959 ":tables",
2960 "@FP16",
2961 "@pthreadpool",
2962 ],
2963)
2964
2965xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002966 name = "neonv8_ukernels_test_mode",
2967 hdrs = INTERNAL_HDRS,
2968 aarch32_copts = [
2969 "-marm",
2970 "-march=armv8-a",
2971 "-mfpu=neon-fp-armv8",
2972 ],
2973 aarch32_srcs = NEONV8_UKERNELS,
2974 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07002975 apple_aarch32_copts = [
2976 "-mcpu=cyclone",
2977 "-mtune=generic",
2978 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07002979 copts = [
2980 "-UNDEBUG",
2981 "-DXNN_TEST_MODE=1",
2982 ],
2983 gcc_copts = xnnpack_gcc_std_copts(),
2984 msvc_copts = xnnpack_msvc_std_copts(),
2985 deps = [
2986 ":tables",
2987 "@FP16",
2988 "@pthreadpool",
2989 ],
2990)
2991
2992xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993 name = "neonfp16arith_ukernels",
2994 hdrs = INTERNAL_HDRS,
2995 aarch64_copts = ["-march=armv8.2-a+fp16"],
2996 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002997 gcc_copts = xnnpack_gcc_std_copts(),
2998 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002999 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003000 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003001 "@FP16",
3002 "@pthreadpool",
3003 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003004)
3005
3006xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003007 name = "neonfp16arith_ukernels_test_mode",
3008 hdrs = INTERNAL_HDRS,
3009 aarch64_copts = ["-march=armv8.2-a+fp16"],
3010 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3011 copts = [
3012 "-UNDEBUG",
3013 "-DXNN_TEST_MODE=1",
3014 ],
3015 gcc_copts = xnnpack_gcc_std_copts(),
3016 msvc_copts = xnnpack_msvc_std_copts(),
3017 deps = [
3018 ":tables",
3019 "@FP16",
3020 "@pthreadpool",
3021 ],
3022)
3023
3024xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003025 name = "neondot_ukernels",
3026 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003027 aarch32_copts = [
3028 "-marm",
3029 "-march=armv8.2-a+dotprod",
3030 "-mfpu=neon-fp-armv8",
3031 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003032 aarch32_srcs = NEONDOT_UKERNELS,
3033 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3034 aarch64_srcs = NEONDOT_UKERNELS,
3035 gcc_copts = xnnpack_gcc_std_copts(),
3036 msvc_copts = xnnpack_msvc_std_copts(),
3037 deps = [
3038 ":tables",
3039 "@FP16",
3040 "@pthreadpool",
3041 ],
3042)
3043
3044xnnpack_cc_library(
3045 name = "neondot_ukernels_test_mode",
3046 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003047 aarch32_copts = [
3048 "-marm",
3049 "-march=armv8.2-a+dotprod",
3050 "-mfpu=neon-fp-armv8",
3051 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003052 aarch32_srcs = NEONDOT_UKERNELS,
3053 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3054 aarch64_srcs = NEONDOT_UKERNELS,
3055 copts = [
3056 "-UNDEBUG",
3057 "-DXNN_TEST_MODE=1",
3058 ],
3059 gcc_copts = xnnpack_gcc_std_copts(),
3060 msvc_copts = xnnpack_msvc_std_copts(),
3061 deps = [
3062 ":tables",
3063 "@FP16",
3064 "@pthreadpool",
3065 ],
3066)
3067
3068xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003069 name = "sse2_ukernels",
3070 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003071 gcc_copts = xnnpack_gcc_std_copts(),
3072 gcc_x86_copts = ["-msse2"],
3073 msvc_copts = xnnpack_msvc_std_copts(),
3074 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003075 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003076 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003077 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003078 "@FP16",
3079 "@pthreadpool",
3080 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003081)
3082
3083xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003084 name = "sse2_ukernels_test_mode",
3085 hdrs = INTERNAL_HDRS,
3086 copts = [
3087 "-UNDEBUG",
3088 "-DXNN_TEST_MODE=1",
3089 ],
3090 gcc_copts = xnnpack_gcc_std_copts(),
3091 gcc_x86_copts = ["-msse2"],
3092 msvc_copts = xnnpack_msvc_std_copts(),
3093 msvc_x86_32_copts = ["/arch:SSE2"],
3094 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3095 deps = [
3096 ":tables",
3097 "@FP16",
3098 "@pthreadpool",
3099 ],
3100)
3101
3102xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003103 name = "ssse3_ukernels",
3104 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003105 gcc_copts = xnnpack_gcc_std_copts(),
3106 gcc_x86_copts = ["-mssse3"],
3107 msvc_copts = xnnpack_msvc_std_copts(),
3108 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003109 x86_srcs = SSSE3_UKERNELS,
3110 deps = [
3111 ":tables",
3112 "@FP16",
3113 "@pthreadpool",
3114 ],
3115)
3116
3117xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003118 name = "ssse3_ukernels_test_mode",
3119 hdrs = INTERNAL_HDRS,
3120 copts = [
3121 "-UNDEBUG",
3122 "-DXNN_TEST_MODE=1",
3123 ],
3124 gcc_copts = xnnpack_gcc_std_copts(),
3125 gcc_x86_copts = ["-mssse3"],
3126 msvc_copts = xnnpack_msvc_std_copts(),
3127 msvc_x86_32_copts = ["/arch:SSE2"],
3128 x86_srcs = SSSE3_UKERNELS,
3129 deps = [
3130 ":tables",
3131 "@FP16",
3132 "@pthreadpool",
3133 ],
3134)
3135
3136xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003137 name = "sse41_ukernels",
3138 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003139 gcc_copts = xnnpack_gcc_std_copts(),
3140 gcc_x86_copts = ["-msse4.1"],
3141 msvc_copts = xnnpack_msvc_std_copts(),
3142 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003143 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003144 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003145 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003146 "@FP16",
3147 "@pthreadpool",
3148 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003149)
3150
3151xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003152 name = "sse41_ukernels_test_mode",
3153 hdrs = INTERNAL_HDRS,
3154 copts = [
3155 "-UNDEBUG",
3156 "-DXNN_TEST_MODE=1",
3157 ],
3158 gcc_copts = xnnpack_gcc_std_copts(),
3159 gcc_x86_copts = ["-msse4.1"],
3160 msvc_copts = xnnpack_msvc_std_copts(),
3161 msvc_x86_32_copts = ["/arch:SSE2"],
3162 x86_srcs = SSE41_UKERNELS,
3163 deps = [
3164 ":tables",
3165 "@FP16",
3166 "@pthreadpool",
3167 ],
3168)
3169
3170xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003171 name = "avx_ukernels",
3172 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003173 gcc_copts = xnnpack_gcc_std_copts(),
3174 gcc_x86_copts = ["-mavx"],
3175 msvc_copts = xnnpack_msvc_std_copts(),
3176 msvc_x86_32_copts = ["/arch:AVX"],
3177 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003178 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003179 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003180 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003181 "@FP16",
3182 "@pthreadpool",
3183 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003184)
3185
3186xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003187 name = "avx_ukernels_test_mode",
3188 hdrs = INTERNAL_HDRS,
3189 copts = [
3190 "-UNDEBUG",
3191 "-DXNN_TEST_MODE=1",
3192 ],
3193 gcc_copts = xnnpack_gcc_std_copts(),
3194 gcc_x86_copts = ["-mavx"],
3195 msvc_copts = xnnpack_msvc_std_copts(),
3196 msvc_x86_32_copts = ["/arch:AVX"],
3197 msvc_x86_64_copts = ["/arch:AVX"],
3198 x86_srcs = AVX_UKERNELS,
3199 deps = [
3200 ":tables",
3201 "@FP16",
3202 "@pthreadpool",
3203 ],
3204)
3205
3206xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003207 name = "xop_ukernels",
3208 hdrs = INTERNAL_HDRS,
3209 gcc_copts = xnnpack_gcc_std_copts(),
3210 gcc_x86_copts = ["-mxop"],
3211 msvc_copts = xnnpack_msvc_std_copts(),
3212 msvc_x86_32_copts = ["/arch:AVX"],
3213 msvc_x86_64_copts = ["/arch:AVX"],
3214 x86_srcs = XOP_UKERNELS,
3215 deps = [
3216 ":tables",
3217 "@FP16",
3218 "@pthreadpool",
3219 ],
3220)
3221
3222xnnpack_cc_library(
3223 name = "xop_ukernels_test_mode",
3224 hdrs = INTERNAL_HDRS,
3225 copts = [
3226 "-UNDEBUG",
3227 "-DXNN_TEST_MODE=1",
3228 ],
3229 gcc_copts = xnnpack_gcc_std_copts(),
3230 gcc_x86_copts = ["-mxop"],
3231 msvc_copts = xnnpack_msvc_std_copts(),
3232 msvc_x86_32_copts = ["/arch:AVX"],
3233 msvc_x86_64_copts = ["/arch:AVX"],
3234 x86_srcs = XOP_UKERNELS,
3235 deps = [
3236 ":tables",
3237 "@FP16",
3238 "@pthreadpool",
3239 ],
3240)
3241
3242xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003243 name = "fma3_ukernels",
3244 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003245 gcc_copts = xnnpack_gcc_std_copts(),
3246 gcc_x86_copts = ["-mfma"],
3247 msvc_copts = xnnpack_msvc_std_copts(),
3248 msvc_x86_32_copts = ["/arch:AVX"],
3249 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003250 x86_srcs = FMA3_UKERNELS,
3251 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003252 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003253 "@FP16",
3254 "@pthreadpool",
3255 ],
3256)
3257
3258xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003259 name = "fma3_ukernels_test_mode",
3260 hdrs = INTERNAL_HDRS,
3261 copts = [
3262 "-UNDEBUG",
3263 "-DXNN_TEST_MODE=1",
3264 ],
3265 gcc_copts = xnnpack_gcc_std_copts(),
3266 gcc_x86_copts = ["-mfma"],
3267 msvc_copts = xnnpack_msvc_std_copts(),
3268 msvc_x86_32_copts = ["/arch:AVX"],
3269 msvc_x86_64_copts = ["/arch:AVX"],
3270 x86_srcs = FMA3_UKERNELS,
3271 deps = [
3272 ":tables",
3273 "@FP16",
3274 "@pthreadpool",
3275 ],
3276)
3277
3278xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003279 name = "avx2_ukernels",
3280 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003281 gcc_copts = xnnpack_gcc_std_copts(),
3282 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003283 "-mfma",
3284 "-mavx2",
3285 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003286 msvc_copts = xnnpack_msvc_std_copts(),
3287 msvc_x86_32_copts = ["/arch:AVX2"],
3288 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003289 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003290 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003291 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003292 "@FP16",
3293 "@pthreadpool",
3294 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003295)
3296
3297xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003298 name = "avx2_ukernels_test_mode",
3299 hdrs = INTERNAL_HDRS,
3300 copts = [
3301 "-UNDEBUG",
3302 "-DXNN_TEST_MODE=1",
3303 ],
3304 gcc_copts = xnnpack_gcc_std_copts(),
3305 gcc_x86_copts = [
3306 "-mfma",
3307 "-mavx2",
3308 ],
3309 msvc_copts = xnnpack_msvc_std_copts(),
3310 msvc_x86_32_copts = ["/arch:AVX2"],
3311 msvc_x86_64_copts = ["/arch:AVX2"],
3312 x86_srcs = AVX2_UKERNELS,
3313 deps = [
3314 ":tables",
3315 "@FP16",
3316 "@pthreadpool",
3317 ],
3318)
3319
3320xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003321 name = "avx512f_ukernels",
3322 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003323 gcc_copts = xnnpack_gcc_std_copts(),
3324 gcc_x86_copts = ["-mavx512f"],
3325 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3326 msvc_copts = xnnpack_msvc_std_copts(),
3327 msvc_x86_32_copts = ["/arch:AVX512"],
3328 msvc_x86_64_copts = ["/arch:AVX512"],
3329 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003330 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003331 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003332 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003333 "@FP16",
3334 "@pthreadpool",
3335 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003336)
3337
3338xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003339 name = "avx512f_ukernels_test_mode",
3340 hdrs = INTERNAL_HDRS,
3341 copts = [
3342 "-UNDEBUG",
3343 "-DXNN_TEST_MODE=1",
3344 ],
3345 gcc_copts = xnnpack_gcc_std_copts(),
3346 gcc_x86_copts = ["-mavx512f"],
3347 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3348 msvc_copts = xnnpack_msvc_std_copts(),
3349 msvc_x86_32_copts = ["/arch:AVX512"],
3350 msvc_x86_64_copts = ["/arch:AVX512"],
3351 msys_copts = ["-fno-asynchronous-unwind-tables"],
3352 x86_srcs = AVX512F_UKERNELS,
3353 deps = [
3354 ":tables",
3355 "@FP16",
3356 "@pthreadpool",
3357 ],
3358)
3359
3360xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003361 name = "avx512skx_ukernels",
3362 hdrs = INTERNAL_HDRS,
3363 gcc_copts = xnnpack_gcc_std_copts(),
3364 gcc_x86_copts = [
3365 "-mavx512f",
3366 "-mavx512cd",
3367 "-mavx512bw",
3368 "-mavx512dq",
3369 "-mavx512vl",
3370 ],
3371 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3372 msvc_copts = xnnpack_msvc_std_copts(),
3373 msvc_x86_32_copts = ["/arch:AVX512"],
3374 msvc_x86_64_copts = ["/arch:AVX512"],
3375 msys_copts = ["-fno-asynchronous-unwind-tables"],
3376 x86_srcs = AVX512SKX_UKERNELS,
3377 deps = [
3378 ":tables",
3379 "@FP16",
3380 "@pthreadpool",
3381 ],
3382)
3383
3384xnnpack_cc_library(
3385 name = "avx512skx_ukernels_test_mode",
3386 hdrs = INTERNAL_HDRS,
3387 copts = [
3388 "-UNDEBUG",
3389 "-DXNN_TEST_MODE=1",
3390 ],
3391 gcc_copts = xnnpack_gcc_std_copts(),
3392 gcc_x86_copts = [
3393 "-mavx512f",
3394 "-mavx512cd",
3395 "-mavx512bw",
3396 "-mavx512dq",
3397 "-mavx512vl",
3398 ],
3399 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3400 msvc_copts = xnnpack_msvc_std_copts(),
3401 msvc_x86_32_copts = ["/arch:AVX512"],
3402 msvc_x86_64_copts = ["/arch:AVX512"],
3403 msys_copts = ["-fno-asynchronous-unwind-tables"],
3404 x86_srcs = AVX512SKX_UKERNELS,
3405 deps = [
3406 ":tables",
3407 "@FP16",
3408 "@pthreadpool",
3409 ],
3410)
3411
3412xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003413 name = "asm_ukernels",
3414 hdrs = ["src/xnnpack/assembly.h"],
3415 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003416 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003417 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003418 wasm_srcs = WASM32_ASM_UKERNELS,
3419 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003420)
3421
Marat Dukhan3b59de22020-06-03 20:15:19 -07003422xnnpack_cc_library(
3423 name = "logging_utils",
3424 srcs = LOGGING_SRCS,
3425 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3426 copts = LOGGING_COPTS + [
3427 "-Isrc",
3428 "-Iinclude",
3429 ] + select({
3430 ":debug_build": [],
3431 "//conditions:default": xnnpack_min_size_copts(),
3432 }),
3433 gcc_copts = xnnpack_gcc_std_copts(),
3434 msvc_copts = xnnpack_msvc_std_copts(),
3435 visibility = xnnpack_visibility(),
3436 deps = [
3437 "@FP16",
3438 "@clog",
3439 "@pthreadpool",
3440 ],
3441)
3442
Marat Dukhan08c4a432019-10-03 09:29:21 -07003443xnnpack_aggregate_library(
3444 name = "ukernels",
3445 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003446 ":neon_ukernels",
3447 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003448 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003449 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003450 ":asm_ukernels",
3451 ],
3452 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003453 ":neon_ukernels",
3454 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003455 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003456 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003457 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003458 ":asm_ukernels",
3459 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003460 generic_deps = [
3461 ":scalar_ukernels",
3462 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003463 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003464 ":psimd_fastmath_ukernels",
3465 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003466 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 wasm_deps = [
3468 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003469 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003470 ],
3471 wasmsimd_deps = [
3472 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003473 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003474 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003475 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003476 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003477 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003478 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003479 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003480 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003481 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003482 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003483 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003484 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003485 ],
3486)
3487
Marat Dukhan33fcf782020-05-24 14:27:15 -07003488xnnpack_aggregate_library(
3489 name = "ukernels_test_mode",
3490 aarch32_deps = [
3491 ":neon_ukernels_test_mode",
3492 ":neonfma_ukernels_test_mode",
3493 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003494 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003495 ":asm_ukernels",
3496 ],
3497 aarch64_deps = [
3498 ":neon_ukernels_test_mode",
3499 ":neonfma_ukernels_test_mode",
3500 ":neonv8_ukernels_test_mode",
3501 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003502 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003503 ":asm_ukernels",
3504 ],
3505 generic_deps = [
3506 ":scalar_ukernels_test_mode",
3507 ],
3508 psimd_deps = [
3509 ":psimd_fastmath_ukernels_test_mode",
3510 ":psimd_accmath_ukernels_test_mode",
3511 ],
3512 wasm_deps = [
3513 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003514 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003515 ],
3516 wasmsimd_deps = [
3517 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003518 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003519 ],
3520 x86_deps = [
3521 ":sse2_ukernels_test_mode",
3522 ":ssse3_ukernels_test_mode",
3523 ":sse41_ukernels_test_mode",
3524 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003525 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003526 ":fma3_ukernels_test_mode",
3527 ":avx2_ukernels_test_mode",
3528 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003529 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003530 ],
3531)
3532
Marat Dukhan86df3842020-06-17 14:03:31 -07003533# This target must be compatible with Android NDK r17 and compile with Clang 6.
3534# It is similar to :ukernels target, but excludes microkernels which require
3535# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003536xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003537 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003538 aarch32_deps = [
3539 ":neon_ukernels",
3540 ":neonfma_ukernels",
3541 ":neonv8_ukernels",
3542 ":asm_ukernels",
3543 ],
3544 aarch64_deps = [
3545 ":neon_ukernels",
3546 ":neonfma_ukernels",
3547 ":neonv8_ukernels",
3548 ":asm_ukernels",
3549 ],
3550 generic_deps = [
3551 ":scalar_ukernels",
3552 ],
3553 psimd_deps = [
3554 ":psimd_fastmath_ukernels",
3555 ":psimd_accmath_ukernels",
3556 ],
3557 wasm_deps = [
3558 ":wasm_ukernels",
3559 ],
3560 wasmsimd_deps = [
3561 ":wasm_ukernels",
3562 ],
3563 x86_deps = [
3564 ":sse2_ukernels",
3565 ":ssse3_ukernels",
3566 ":sse41_ukernels",
3567 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003568 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003569 ":fma3_ukernels",
3570 ":avx2_ukernels",
3571 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003572 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003573 ],
3574)
3575
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576xnnpack_cc_library(
3577 name = "im2col",
3578 srcs = ["src/im2col.c"],
3579 hdrs = [
3580 "src/xnnpack/common.h",
3581 "src/xnnpack/im2col.h",
3582 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003583 gcc_copts = xnnpack_gcc_std_copts(),
3584 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003585)
3586
3587xnnpack_cc_library(
3588 name = "indirection",
3589 srcs = ["src/indirection.c"],
3590 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003591 gcc_copts = xnnpack_gcc_std_copts(),
3592 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003593 deps = [
3594 "@FP16",
3595 "@FXdiv",
3596 "@pthreadpool",
3597 ],
3598)
3599
3600xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003601 name = "indirection_test_mode",
3602 srcs = ["src/indirection.c"],
3603 hdrs = INTERNAL_HDRS,
3604 copts = [
3605 "-UNDEBUG",
3606 "-DXNN_TEST_MODE=1",
3607 ],
3608 gcc_copts = xnnpack_gcc_std_copts(),
3609 msvc_copts = xnnpack_msvc_std_copts(),
3610 deps = [
3611 "@FP16",
3612 "@FXdiv",
3613 "@pthreadpool",
3614 ],
3615)
3616
3617xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003618 name = "packing",
3619 srcs = ["src/packing.c"],
3620 hdrs = INTERNAL_HDRS,
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(
3631 name = "packing_test_mode",
3632 srcs = ["src/packing.c"],
3633 hdrs = INTERNAL_HDRS,
3634 copts = [
3635 "-UNDEBUG",
3636 "-DXNN_TEST_MODE=1",
3637 ],
3638 gcc_copts = xnnpack_gcc_std_copts(),
3639 msvc_copts = xnnpack_msvc_std_copts(),
3640 deps = [
3641 "@FP16",
3642 "@FXdiv",
3643 "@pthreadpool",
3644 ],
3645)
3646
3647xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 name = "operator_run",
3649 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003650 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003651 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003652 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3653 "//conditions:default": [],
3654 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003655 gcc_copts = xnnpack_gcc_std_copts(),
3656 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003657 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003658 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003659 "@FP16",
3660 "@FXdiv",
3661 "@clog",
3662 "@pthreadpool",
3663 ],
3664)
3665
Chao Mei6ddfc602020-05-13 22:29:36 -07003666xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003667 name = "operator_run_test_mode",
3668 srcs = ["src/operator-run.c"],
3669 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3670 copts = LOGGING_COPTS + [
3671 "-UNDEBUG",
3672 "-DXNN_TEST_MODE=1",
3673 ] + select({
3674 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3675 "//conditions:default": [],
3676 }),
3677 gcc_copts = xnnpack_gcc_std_copts(),
3678 msvc_copts = xnnpack_msvc_std_copts(),
3679 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003680 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003681 "@FP16",
3682 "@FXdiv",
3683 "@clog",
3684 "@pthreadpool",
3685 ],
3686)
3687
3688xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003689 name = "memory_planner",
3690 srcs = ["src/memory-planner.c"],
3691 hdrs = INTERNAL_HDRS,
3692 defines = select({
3693 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3694 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3695 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3696 }),
3697 gcc_copts = xnnpack_gcc_std_copts(),
3698 msvc_copts = xnnpack_msvc_std_copts(),
3699 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003700 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003701 "@pthreadpool",
3702 ],
3703)
3704
Marat Dukhan33fcf782020-05-24 14:27:15 -07003705xnnpack_cc_library(
3706 name = "memory_planner_test_mode",
3707 srcs = ["src/memory-planner.c"],
3708 hdrs = INTERNAL_HDRS,
3709 copts = [
3710 "-UNDEBUG",
3711 "-DXNN_TEST_MODE=1",
3712 ],
3713 defines = select({
3714 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3715 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3716 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3717 }),
3718 gcc_copts = xnnpack_gcc_std_copts(),
3719 msvc_copts = xnnpack_msvc_std_copts(),
3720 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003721 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003722 "@pthreadpool",
3723 ],
3724)
3725
Marat Dukhan08c4a432019-10-03 09:29:21 -07003726cc_library(
3727 name = "enable_assembly",
3728 defines = select({
3729 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3730 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003731 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003732 }),
3733)
3734
Marat Dukhan9de90e02020-06-18 16:04:12 -07003735cc_library(
3736 name = "enable_sparse",
3737 defines = select({
3738 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3739 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003740 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003741 }),
3742)
3743
Marat Dukhancf056b22019-10-07 10:26:29 -07003744xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745 name = "operators",
3746 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003747 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003748 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003749 ],
3750 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003751 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003752 "-Isrc",
3753 "-Iinclude",
3754 ] + select({
3755 ":debug_build": [],
3756 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003757 }) + select({
3758 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3759 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003760 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003761 gcc_copts = xnnpack_gcc_std_copts(),
3762 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003763 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003764 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003765 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003766 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003767 "@FP16",
3768 "@FXdiv",
3769 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003770 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003771 ],
3772)
3773
Marat Dukhan10a38082020-04-17 03:58:35 -07003774xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003775 name = "operators_test_mode",
3776 srcs = OPERATOR_SRCS + [
3777 "src/memory.c",
3778 "src/operator-delete.c",
3779 ],
3780 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3781 copts = LOGGING_COPTS + [
3782 "-Isrc",
3783 "-Iinclude",
3784 "-UNDEBUG",
3785 "-DXNN_TEST_MODE=1",
3786 ] + select({
3787 ":debug_build": [],
3788 "//conditions:default": xnnpack_min_size_copts(),
3789 }) + select({
3790 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3791 "//conditions:default": [],
3792 }),
3793 gcc_copts = xnnpack_gcc_std_copts(),
3794 msvc_copts = xnnpack_msvc_std_copts(),
3795 deps = [
3796 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003797 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003798 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003799 "@FP16",
3800 "@FXdiv",
3801 "@clog",
3802 "@pthreadpool",
3803 ],
3804)
3805
3806xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003807 name = "XNNPACK",
3808 srcs = [
3809 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003810 "src/runtime.c",
3811 "src/subgraph.c",
3812 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003813 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003814 hdrs = ["include/xnnpack.h"],
3815 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003816 "-Isrc",
3817 "-Iinclude",
3818 ] + select({
3819 ":debug_build": [],
3820 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003821 }) + select({
3822 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3823 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003824 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003825 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003826 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003827 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003828 visibility = xnnpack_visibility(),
3829 deps = [
3830 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003831 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003832 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003833 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003834 ":operator_run",
3835 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003836 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003837 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003838 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003839 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003840 ] + select({
3841 ":emscripten": [],
3842 "//conditions:default": ["@cpuinfo"],
3843 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003844)
3845
Marat Dukhan10a38082020-04-17 03:58:35 -07003846xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003847 name = "XNNPACK_test_mode",
3848 srcs = [
3849 "src/init.c",
3850 "src/runtime.c",
3851 "src/subgraph.c",
3852 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003853 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003854 hdrs = ["include/xnnpack.h"],
3855 copts = LOGGING_COPTS + [
3856 "-Isrc",
3857 "-Iinclude",
3858 "-UNDEBUG",
3859 "-DXNN_TEST_MODE=1",
3860 ] + select({
3861 ":debug_build": [],
3862 "//conditions:default": xnnpack_min_size_copts(),
3863 }) + select({
3864 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3865 "//conditions:default": [],
3866 }),
3867 gcc_copts = xnnpack_gcc_std_copts(),
3868 includes = ["include"],
3869 msvc_copts = xnnpack_msvc_std_copts(),
3870 visibility = xnnpack_visibility(),
3871 deps = [
3872 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003873 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003874 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003875 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003876 ":operator_run_test_mode",
3877 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003878 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003879 "@clog",
3880 "@FP16",
3881 "@pthreadpool",
3882 ] + select({
3883 ":emscripten": [],
3884 "//conditions:default": ["@cpuinfo"],
3885 }),
3886)
3887
3888xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07003889 name = "xnnpack_f32",
3890 srcs = [
3891 "src/init.c",
3892 "src/runtime.c",
3893 "src/subgraph.c",
3894 "src/tensor.c",
3895 ] + SUBGRAPH_SRCS,
3896 hdrs = ["include/xnnpack.h"],
3897 copts = LOGGING_COPTS + [
3898 "-Isrc",
3899 "-Iinclude",
3900 ] + select({
3901 ":debug_build": [],
3902 "//conditions:default": xnnpack_min_size_copts(),
3903 }) + select({
3904 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3905 "//conditions:default": [],
3906 }),
3907 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003908 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003909 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07003910 "XNN_NO_U8_OPERATORS",
3911 "XNN_NO_X8_OPERATORS",
3912 "XNN_NO_F16_OPERATORS",
3913 "XNN_NO_X16_OPERATORS",
3914 ],
3915 gcc_copts = xnnpack_gcc_std_copts(),
3916 includes = ["include"],
3917 msvc_copts = xnnpack_msvc_std_copts(),
3918 visibility = xnnpack_visibility(),
3919 deps = [
3920 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003921 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07003922 ":logging_utils",
3923 ":memory_planner",
3924 ":operator_run",
3925 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07003926 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003927 "@clog",
3928 "@FP16",
3929 "@pthreadpool",
3930 ] + select({
3931 ":emscripten": [],
3932 "//conditions:default": ["@cpuinfo"],
3933 }),
3934)
3935
3936xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003937 name = "xnnpack_operators_nhwc_f32",
3938 srcs = [
3939 "src/init.c",
3940 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003941 hdrs = ["include/xnnpack.h"],
3942 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003943 "-Isrc",
3944 "-Iinclude",
3945 ] + select({
3946 ":debug_build": [],
3947 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003948 }) + select({
3949 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3950 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003951 }),
3952 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07003953 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07003954 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003955 "XNN_NO_U8_OPERATORS",
3956 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08003957 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003958 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003959 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003960 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003961 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003962 visibility = xnnpack_visibility(),
3963 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003964 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003965 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003966 ":operator_run",
3967 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003968 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003969 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003971 ] + select({
3972 ":emscripten": [],
3973 "//conditions:default": ["@cpuinfo"],
3974 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003975)
3976
Marat Dukhancf056b22019-10-07 10:26:29 -07003977xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003978 name = "bench_utils",
3979 srcs = ["bench/utils.cc"],
3980 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08003981 deps = [
3982 "@com_google_benchmark//:benchmark",
3983 "@cpuinfo",
3984 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003985)
3986
Frank Barchard7e955972019-10-11 10:34:25 -07003987######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07003988
3989xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07003990 name = "qs8_gemm_bench",
3991 srcs = [
3992 "bench/gemm.h",
3993 "bench/qs8-gemm.cc",
3994 "src/xnnpack/AlignedAllocator.h",
3995 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
3996 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
3997)
3998
3999xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004000 name = "qs8_requantization_bench",
4001 srcs = [
4002 "bench/qs8-requantization.cc",
4003 "src/xnnpack/requantization-stubs.h",
4004 "src/xnnpack/AlignedAllocator.h",
4005 ] + MICROKERNEL_BENCHMARK_HDRS,
4006 deps = MICROKERNEL_BENCHMARK_DEPS,
4007)
4008
4009xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004010 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004011 srcs = [
4012 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004013 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004014 "src/xnnpack/AlignedAllocator.h",
4015 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004016 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004017 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004018)
4019
4020xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004021 name = "qu8_requantization_bench",
4022 srcs = [
4023 "bench/qu8-requantization.cc",
4024 "src/xnnpack/requantization-stubs.h",
4025 "src/xnnpack/AlignedAllocator.h",
4026 ] + MICROKERNEL_BENCHMARK_HDRS,
4027 deps = MICROKERNEL_BENCHMARK_DEPS,
4028)
4029
4030xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004031 name = "f16_igemm_bench",
4032 srcs = [
4033 "bench/f16-igemm.cc",
4034 "bench/conv.h",
4035 "bench/google/conv.h",
4036 "src/xnnpack/AlignedAllocator.h",
4037 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004038 deps = MICROKERNEL_BENCHMARK_DEPS + [
4039 ":indirection",
4040 ":packing",
4041 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004042)
4043
4044xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004045 name = "f16_gemm_bench",
4046 srcs = [
4047 "bench/f16-gemm.cc",
4048 "bench/gemm.h",
4049 "src/xnnpack/AlignedAllocator.h",
4050 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004051 deps = MICROKERNEL_BENCHMARK_DEPS + [
4052 ":packing",
4053 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004054)
4055
4056xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004057 name = "f16_spmm_bench",
4058 srcs = [
4059 "bench/f16-spmm.cc",
4060 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004061 "src/xnnpack/AlignedAllocator.h",
4062 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004063 deps = MICROKERNEL_BENCHMARK_DEPS,
4064)
4065
4066xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004067 name = "f32_igemm_bench",
4068 srcs = [
4069 "bench/f32-igemm.cc",
4070 "bench/conv.h",
4071 "src/xnnpack/AlignedAllocator.h",
4072 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004073 deps = MICROKERNEL_BENCHMARK_DEPS + [
4074 ":indirection",
4075 ":packing",
4076 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004077)
4078
4079xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004080 name = "f16_relu_bench",
4081 srcs = [
4082 "bench/f16-relu.cc",
4083 "src/xnnpack/AlignedAllocator.h",
4084 ] + MICROKERNEL_BENCHMARK_HDRS,
4085 deps = MICROKERNEL_BENCHMARK_DEPS,
4086)
4087
4088xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004089 name = "f32_conv_hwc_bench",
4090 srcs = [
4091 "bench/f32-conv-hwc.cc",
4092 "bench/dconv.h",
4093 "src/xnnpack/AlignedAllocator.h",
4094 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004095 deps = MICROKERNEL_BENCHMARK_DEPS + [
4096 ":packing",
4097 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004098)
4099
4100xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004101 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004102 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004103 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004104 "bench/dconv.h",
4105 "src/xnnpack/AlignedAllocator.h",
4106 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004107 deps = MICROKERNEL_BENCHMARK_DEPS + [
4108 ":packing",
4109 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004110)
4111
4112xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004113 name = "f16_dwconv_bench",
4114 srcs = [
4115 "bench/f16-dwconv.cc",
4116 "bench/dwconv.h",
4117 "bench/google/dwconv.h",
4118 "src/xnnpack/AlignedAllocator.h",
4119 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004120 deps = MICROKERNEL_BENCHMARK_DEPS + [
4121 ":indirection",
4122 ":packing",
4123 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004124)
4125
4126xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004127 name = "f32_dwconv_bench",
4128 srcs = [
4129 "bench/f32-dwconv.cc",
4130 "bench/dwconv.h",
4131 "src/xnnpack/AlignedAllocator.h",
4132 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004133 deps = MICROKERNEL_BENCHMARK_DEPS + [
4134 ":indirection",
4135 ":packing",
4136 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137)
4138
4139xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004140 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004141 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004142 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004143 "bench/dwconv.h",
4144 "src/xnnpack/AlignedAllocator.h",
4145 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004146 deps = MICROKERNEL_BENCHMARK_DEPS + [
4147 ":indirection",
4148 ":packing",
4149 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004150)
4151
4152xnnpack_benchmark(
4153 name = "f32_gemm_bench",
4154 srcs = [
4155 "bench/f32-gemm.cc",
4156 "bench/gemm.h",
4157 "src/xnnpack/AlignedAllocator.h",
4158 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004159 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004160 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004161)
4162
4163xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004164 name = "f32_hswish_bench",
4165 srcs = [
4166 "bench/f32-hswish.cc",
4167 "src/xnnpack/AlignedAllocator.h",
4168 ] + MICROKERNEL_BENCHMARK_HDRS,
4169 deps = MICROKERNEL_BENCHMARK_DEPS,
4170)
4171
4172xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004173 name = "f32_raddexpminusmax_bench",
4174 srcs = [
4175 "bench/f32-raddexpminusmax.cc",
4176 "src/xnnpack/AlignedAllocator.h",
4177 ] + MICROKERNEL_BENCHMARK_HDRS,
4178 deps = MICROKERNEL_BENCHMARK_DEPS,
4179)
4180
4181xnnpack_benchmark(
4182 name = "f32_raddextexp_bench",
4183 srcs = [
4184 "bench/f32-raddextexp.cc",
4185 "src/xnnpack/AlignedAllocator.h",
4186 ] + MICROKERNEL_BENCHMARK_HDRS,
4187 deps = MICROKERNEL_BENCHMARK_DEPS,
4188)
4189
4190xnnpack_benchmark(
4191 name = "f32_raddstoreexpminusmax_bench",
4192 srcs = [
4193 "bench/f32-raddstoreexpminusmax.cc",
4194 "src/xnnpack/AlignedAllocator.h",
4195 ] + MICROKERNEL_BENCHMARK_HDRS,
4196 deps = MICROKERNEL_BENCHMARK_DEPS,
4197)
4198
4199xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004200 name = "f32_relu_bench",
4201 srcs = [
4202 "bench/f32-relu.cc",
4203 "src/xnnpack/AlignedAllocator.h",
4204 ] + MICROKERNEL_BENCHMARK_HDRS,
4205 deps = MICROKERNEL_BENCHMARK_DEPS,
4206)
4207
4208xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004209 name = "f32_rmax_bench",
4210 srcs = [
4211 "bench/f32-rmax.cc",
4212 "src/xnnpack/AlignedAllocator.h",
4213 ] + MICROKERNEL_BENCHMARK_HDRS,
4214 deps = MICROKERNEL_BENCHMARK_DEPS,
4215)
4216
4217xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004218 name = "f32_sigmoid_bench",
4219 srcs = [
4220 "bench/f32-sigmoid.cc",
4221 "src/xnnpack/AlignedAllocator.h",
4222 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004223 deps = MICROKERNEL_BENCHMARK_DEPS,
4224)
4225
4226xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004227 name = "f32_spmm_bench",
4228 srcs = [
4229 "bench/f32-spmm.cc",
4230 "bench/gemm.h",
4231 "src/xnnpack/AlignedAllocator.h",
4232 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004233 deps = MICROKERNEL_BENCHMARK_DEPS,
4234)
4235
4236xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004237 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004238 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004239 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004240 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004241 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004242 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004243)
4244
4245xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004246 name = "f32_vscaleexpminusmax_bench",
4247 srcs = [
4248 "bench/f32-vscaleexpminusmax.cc",
4249 "src/xnnpack/AlignedAllocator.h",
4250 ] + MICROKERNEL_BENCHMARK_HDRS,
4251 deps = MICROKERNEL_BENCHMARK_DEPS,
4252)
4253
4254xnnpack_benchmark(
4255 name = "f32_vscaleextexp_bench",
4256 srcs = [
4257 "bench/f32-vscaleextexp.cc",
4258 "src/xnnpack/AlignedAllocator.h",
4259 ] + MICROKERNEL_BENCHMARK_HDRS,
4260 deps = MICROKERNEL_BENCHMARK_DEPS,
4261)
4262
4263xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004264 name = "f32_vsqrt_bench",
4265 srcs = [
4266 "bench/f32-vsqrt.cc",
4267 "src/xnnpack/AlignedAllocator.h",
4268 ] + MICROKERNEL_BENCHMARK_HDRS,
4269 deps = MICROKERNEL_BENCHMARK_DEPS,
4270)
4271
4272xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004273 name = "f32_im2col_gemm_bench",
4274 srcs = [
4275 "bench/f32-im2col-gemm.cc",
4276 "bench/conv.h",
4277 "src/xnnpack/AlignedAllocator.h",
4278 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004279 deps = MICROKERNEL_BENCHMARK_DEPS + [
4280 ":im2col",
4281 ":packing",
4282 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004283)
4284
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004285xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004286 name = "rounding_bench",
4287 srcs = [
4288 "bench/rounding.cc",
4289 "src/xnnpack/math-stubs.h",
4290 "src/xnnpack/AlignedAllocator.h",
4291 ] + MICROKERNEL_BENCHMARK_HDRS,
4292 deps = MICROKERNEL_BENCHMARK_DEPS,
4293)
4294
Marat Dukhan08c4a432019-10-03 09:29:21 -07004295########################### Benchmarks for operators ###########################
4296
4297xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004298 name = "average_pooling_bench",
4299 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004300 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004301 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004302 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004303)
4304
4305xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004306 name = "bankers_rounding_bench",
4307 srcs = ["bench/bankers-rounding.cc"],
4308 copts = xnnpack_optional_tflite_copts(),
4309 tags = ["nowin32"],
4310 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4311)
4312
4313xnnpack_benchmark(
4314 name = "ceiling_bench",
4315 srcs = ["bench/ceiling.cc"],
4316 copts = xnnpack_optional_tflite_copts(),
4317 tags = ["nowin32"],
4318 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4319)
4320
4321xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004322 name = "channel_shuffle_bench",
4323 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004324 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004325)
4326
4327xnnpack_benchmark(
4328 name = "convolution_bench",
4329 srcs = ["bench/convolution.cc"],
4330 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004331 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004332 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004333)
4334
4335xnnpack_benchmark(
4336 name = "deconvolution_bench",
4337 srcs = ["bench/deconvolution.cc"],
4338 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004339 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004340 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004341)
4342
4343xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004344 name = "floor_bench",
4345 srcs = ["bench/floor.cc"],
4346 copts = xnnpack_optional_tflite_copts(),
4347 tags = ["nowin32"],
4348 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4349)
4350
4351xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004352 name = "global_average_pooling_bench",
4353 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004354 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004355)
4356
4357xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004358 name = "hardswish_bench",
4359 srcs = ["bench/hardswish.cc"],
4360 copts = xnnpack_optional_tflite_copts(),
4361 tags = ["nowin32"],
4362 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4363)
4364
4365xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004366 name = "max_pooling_bench",
4367 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004368 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004369)
4370
4371xnnpack_benchmark(
4372 name = "sigmoid_bench",
4373 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004374 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004375 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004376 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004377)
4378
4379xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004380 name = "prelu_bench",
4381 srcs = ["bench/prelu.cc"],
4382 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004383 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004384 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004385)
4386
4387xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004388 name = "softmax_bench",
4389 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004390 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004391 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004392 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004393)
4394
Marat Dukhan87727142020-06-24 15:24:10 -07004395xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004396 name = "square_root_bench",
4397 srcs = ["bench/square-root.cc"],
4398 copts = xnnpack_optional_tflite_copts(),
4399 tags = ["nowin32"],
4400 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4401)
4402
4403xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004404 name = "truncation_bench",
4405 srcs = ["bench/truncation.cc"],
4406 deps = OPERATOR_BENCHMARK_DEPS,
4407)
4408
Marat Dukhanc068bb62019-10-04 13:24:39 -07004409############################# End-to-end benchmarks ############################
4410
4411cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004412 name = "fp32_mobilenet_v1",
4413 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004414 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004415 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004416 linkstatic = True,
4417 deps = [
4418 ":XNNPACK",
4419 "@pthreadpool",
4420 ],
4421)
4422
4423cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004424 name = "fp16_mobilenet_v1",
4425 srcs = ["models/fp16-mobilenet-v1.cc"],
4426 hdrs = ["models/models.h"],
4427 copts = xnnpack_std_cxxopts(),
4428 linkstatic = True,
4429 deps = [
4430 ":XNNPACK",
4431 "@FP16",
4432 "@pthreadpool",
4433 ],
4434)
4435
4436cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004437 name = "qs8_mobilenet_v1",
4438 srcs = ["models/qs8-mobilenet-v1.cc"],
4439 hdrs = ["models/models.h"],
4440 copts = xnnpack_std_cxxopts(),
4441 linkstatic = True,
4442 deps = [
4443 ":XNNPACK",
4444 "@pthreadpool",
4445 ],
4446)
4447
4448cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004449 name = "fp32_mobilenet_v2",
4450 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004451 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004452 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004453 linkstatic = True,
4454 deps = [
4455 ":XNNPACK",
4456 "@pthreadpool",
4457 ],
4458)
4459
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004460cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004461 name = "fp16_mobilenet_v2",
4462 srcs = ["models/fp16-mobilenet-v2.cc"],
4463 hdrs = ["models/models.h"],
4464 copts = xnnpack_std_cxxopts(),
4465 linkstatic = True,
4466 deps = [
4467 ":XNNPACK",
4468 "@FP16",
4469 "@pthreadpool",
4470 ],
4471)
4472
4473cc_library(
4474 name = "fp32_mobilenet_v3_large",
4475 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004476 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004477 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004478 linkstatic = True,
4479 deps = [
4480 ":XNNPACK",
4481 "@pthreadpool",
4482 ],
4483)
4484
4485cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004486 name = "fp32_mobilenet_v3_small",
4487 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004488 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004489 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004490 linkstatic = True,
4491 deps = [
4492 ":XNNPACK",
4493 "@pthreadpool",
4494 ],
4495)
4496
Marat Dukhanc068bb62019-10-04 13:24:39 -07004497xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004498 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004499 srcs = [
4500 "bench/f32-dwconv-e2e.cc",
4501 "bench/end2end.h",
4502 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004503 deps = MICROKERNEL_BENCHMARK_DEPS + [
4504 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004505 ":fp32_mobilenet_v1",
4506 ":fp32_mobilenet_v2",
4507 ":fp32_mobilenet_v3_large",
4508 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004509 ],
4510)
4511
4512xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004513 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004514 srcs = [
4515 "bench/f32-gemm-e2e.cc",
4516 "bench/end2end.h",
4517 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004518 deps = MICROKERNEL_BENCHMARK_DEPS + [
4519 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004520 ":fp32_mobilenet_v1",
4521 ":fp32_mobilenet_v2",
4522 ":fp32_mobilenet_v3_large",
4523 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004524 ],
4525)
4526
4527xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004528 name = "end2end_bench",
4529 srcs = ["bench/end2end.cc"],
4530 deps = [
4531 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004532 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004533 ":fp16_mobilenet_v1",
4534 ":fp16_mobilenet_v2",
4535 ":fp32_mobilenet_v1",
4536 ":fp32_mobilenet_v2",
4537 ":fp32_mobilenet_v3_large",
4538 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004539 ":qs8_mobilenet_v1",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004540 "@pthreadpool",
4541 ],
4542)
4543
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004544#################### Accuracy evaluation for math functions ####################
4545
4546xnnpack_benchmark(
4547 name = "f32_exp_eval",
4548 srcs = [
4549 "eval/f32-exp.cc",
4550 "src/xnnpack/AlignedAllocator.h",
4551 ] + ACCURACY_EVAL_HDRS,
4552 deps = ACCURACY_EVAL_DEPS,
4553)
4554
Marat Dukhan515c9772019-10-17 18:07:57 -07004555xnnpack_benchmark(
4556 name = "f32_expminus_eval",
4557 srcs = [
4558 "eval/f32-expminus.cc",
4559 "src/xnnpack/AlignedAllocator.h",
4560 ] + ACCURACY_EVAL_HDRS,
4561 deps = ACCURACY_EVAL_DEPS,
4562)
4563
Marat Dukhan98ba4412019-10-23 02:14:28 -07004564xnnpack_benchmark(
4565 name = "f32_extexp_eval",
4566 srcs = [
4567 "eval/f32-extexp.cc",
4568 "src/xnnpack/AlignedAllocator.h",
4569 ] + ACCURACY_EVAL_HDRS,
4570 deps = ACCURACY_EVAL_DEPS,
4571)
4572
Marat Dukhan8853b822020-05-07 12:19:01 -07004573xnnpack_unit_test(
4574 name = "f32_roundne_eval",
4575 srcs = [
4576 "eval/f32-roundne.cc",
4577 "src/xnnpack/AlignedAllocator.h",
4578 "src/xnnpack/math-stubs.h",
4579 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004580 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004581 deps = MICROKERNEL_TEST_DEPS,
4582)
4583
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004584xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004585 name = "f32_roundd_eval",
4586 srcs = [
4587 "eval/f32-roundd.cc",
4588 "src/xnnpack/AlignedAllocator.h",
4589 "src/xnnpack/math-stubs.h",
4590 ] + MICROKERNEL_TEST_HDRS,
4591 automatic = False,
4592 deps = MICROKERNEL_TEST_DEPS,
4593)
4594
4595xnnpack_unit_test(
4596 name = "f32_roundu_eval",
4597 srcs = [
4598 "eval/f32-roundu.cc",
4599 "src/xnnpack/AlignedAllocator.h",
4600 "src/xnnpack/math-stubs.h",
4601 ] + MICROKERNEL_TEST_HDRS,
4602 automatic = False,
4603 deps = MICROKERNEL_TEST_DEPS,
4604)
4605
4606xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004607 name = "f32_roundz_eval",
4608 srcs = [
4609 "eval/f32-roundz.cc",
4610 "src/xnnpack/AlignedAllocator.h",
4611 "src/xnnpack/math-stubs.h",
4612 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004613 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004614 deps = MICROKERNEL_TEST_DEPS,
4615)
4616
Marat Dukhan346a9e52019-11-15 09:06:30 -08004617xnnpack_benchmark(
4618 name = "f32_sigmoid_eval",
4619 srcs = [
4620 "eval/f32-sigmoid.cc",
4621 "src/xnnpack/AlignedAllocator.h",
4622 ] + ACCURACY_EVAL_HDRS,
4623 deps = ACCURACY_EVAL_DEPS,
4624)
4625
Marat Dukhan84000762020-06-29 18:38:43 -07004626xnnpack_benchmark(
4627 name = "f32_sqrt_eval",
4628 srcs = [
4629 "eval/f32-sqrt.cc",
4630 "src/xnnpack/AlignedAllocator.h",
4631 ] + ACCURACY_EVAL_HDRS,
4632 deps = ACCURACY_EVAL_DEPS + [
4633 ":bench_utils",
4634 ],
4635)
4636
Marat Dukhan08c4a432019-10-03 09:29:21 -07004637######################### Unit tests for micro-kernels #########################
4638
4639xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004640 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004641 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004642 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004643 "test/gemm-microkernel-tester.h",
4644 "src/xnnpack/AlignedAllocator.h",
4645 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004646 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004647)
4648
4649xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004650 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004651 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004652 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004653 "test/spmm-microkernel-tester.h",
4654 "src/xnnpack/AlignedAllocator.h",
4655 ] + MICROKERNEL_TEST_HDRS,
4656 deps = MICROKERNEL_TEST_DEPS,
4657)
4658
4659xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004660 name = "f32_argmaxpool_test",
4661 srcs = [
4662 "test/f32-argmaxpool.cc",
4663 "test/argmaxpool-microkernel-tester.h",
4664 "src/xnnpack/AlignedAllocator.h",
4665 ] + MICROKERNEL_TEST_HDRS,
4666 deps = MICROKERNEL_TEST_DEPS,
4667)
4668
4669xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004670 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004671 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004672 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004673 "test/avgpool-microkernel-tester.h",
4674 "src/xnnpack/AlignedAllocator.h",
4675 ] + MICROKERNEL_TEST_HDRS,
4676 deps = MICROKERNEL_TEST_DEPS,
4677)
4678
4679xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004680 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004681 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004682 "test/f32-ibilinear.cc",
4683 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004684 "src/xnnpack/AlignedAllocator.h",
4685 ] + MICROKERNEL_TEST_HDRS,
4686 deps = MICROKERNEL_TEST_DEPS,
4687)
4688
4689xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004690 name = "f16_clamp_test",
4691 srcs = [
4692 "test/f16-clamp.cc",
4693 "test/clamp-microkernel-tester.h",
4694 ] + MICROKERNEL_TEST_HDRS,
4695 deps = MICROKERNEL_TEST_DEPS,
4696)
4697
4698xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004699 name = "f32_clamp_test",
4700 srcs = [
4701 "test/f32-clamp.cc",
4702 "test/clamp-microkernel-tester.h",
4703 ] + MICROKERNEL_TEST_HDRS,
4704 deps = MICROKERNEL_TEST_DEPS,
4705)
4706
4707xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004708 name = "f32_igemm_test",
4709 srcs = [
4710 "test/f32-igemm.cc",
4711 "test/gemm-microkernel-tester.h",
4712 "src/xnnpack/AlignedAllocator.h",
4713 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004714 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004715)
4716
4717xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004718 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004719 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004720 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004721 "test/gemm-microkernel-tester.h",
4722 "src/xnnpack/AlignedAllocator.h",
4723 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004724 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004725)
4726
4727xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004728 name = "f32_igemm_minmax_test",
4729 srcs = [
4730 "test/f32-igemm-minmax.cc",
4731 "test/gemm-microkernel-tester.h",
4732 "src/xnnpack/AlignedAllocator.h",
4733 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004734 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004735)
4736
4737xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004738 name = "f16_igemm_minmax_test",
4739 srcs = [
4740 "test/f16-igemm-minmax.cc",
4741 "test/gemm-microkernel-tester.h",
4742 "src/xnnpack/AlignedAllocator.h",
4743 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004744 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004745)
4746
4747xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004748 name = "f32_conv_hwc_test",
4749 srcs = [
4750 "test/f32-conv-hwc.cc",
4751 "test/conv-hwc-microkernel-tester.h",
4752 "src/xnnpack/AlignedAllocator.h",
4753 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004754 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004755)
4756
4757xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004758 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004759 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004760 "test/f32-conv-hwc2chw.cc",
4761 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004762 "src/xnnpack/AlignedAllocator.h",
4763 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004764 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004765)
4766
4767xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004768 name = "f32_dwconv_test",
4769 srcs = [
4770 "test/f32-dwconv.cc",
4771 "test/dwconv-microkernel-tester.h",
4772 "src/xnnpack/AlignedAllocator.h",
4773 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004774 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004775)
4776
4777xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004778 name = "f16_dwconv_minmax_test",
4779 srcs = [
4780 "test/f16-dwconv-minmax.cc",
4781 "test/dwconv-microkernel-tester.h",
4782 "src/xnnpack/AlignedAllocator.h",
4783 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004784 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004785)
4786
4787xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004788 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004789 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004790 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004791 "test/dwconv-microkernel-tester.h",
4792 "src/xnnpack/AlignedAllocator.h",
4793 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004794 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004795)
4796
4797xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004798 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004800 "test/f32-dwconv-chw.cc",
4801 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004802 "src/xnnpack/AlignedAllocator.h",
4803 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004804 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004805)
4806
4807xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004808 name = "f16_gavgpool_minmax_test",
4809 srcs = [
4810 "test/f16-gavgpool-minmax.cc",
4811 "test/gavgpool-microkernel-tester.h",
4812 "src/xnnpack/AlignedAllocator.h",
4813 ] + MICROKERNEL_TEST_HDRS,
4814 deps = MICROKERNEL_TEST_DEPS,
4815)
4816
4817xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004818 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004819 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004820 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004821 "test/gavgpool-microkernel-tester.h",
4822 "src/xnnpack/AlignedAllocator.h",
4823 ] + MICROKERNEL_TEST_HDRS,
4824 deps = MICROKERNEL_TEST_DEPS,
4825)
4826
4827xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004828 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004829 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004830 "test/f32-gavgpool-cw.cc",
4831 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004832 "src/xnnpack/AlignedAllocator.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004838 name = "f32_gemm_test",
4839 srcs = [
4840 "test/f32-gemm.cc",
4841 "test/gemm-microkernel-tester.h",
4842 "src/xnnpack/AlignedAllocator.h",
4843 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004844 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004845)
4846
4847xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004848 name = "f32_gemm_relu_test",
4849 srcs = [
4850 "test/f32-gemm-relu.cc",
4851 "test/gemm-microkernel-tester.h",
4852 "src/xnnpack/AlignedAllocator.h",
4853 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004854 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004855)
4856
4857xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004858 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004859 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004860 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004861 "test/gemm-microkernel-tester.h",
4862 "src/xnnpack/AlignedAllocator.h",
4863 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004864 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004865)
4866
4867xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004868 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004869 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004870 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004871 "test/gemm-microkernel-tester.h",
4872 "src/xnnpack/AlignedAllocator.h",
4873 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004874 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004875)
4876
4877xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004878 name = "f16_hswish_test",
4879 srcs = [
4880 "test/f16-hswish.cc",
4881 "test/hswish-microkernel-tester.h",
4882 ] + MICROKERNEL_TEST_HDRS,
4883 deps = MICROKERNEL_TEST_DEPS,
4884)
4885
4886xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004887 name = "f32_hswish_test",
4888 srcs = [
4889 "test/f32-hswish.cc",
4890 "test/hswish-microkernel-tester.h",
4891 ] + MICROKERNEL_TEST_HDRS,
4892 deps = MICROKERNEL_TEST_DEPS,
4893)
4894
4895xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004896 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004897 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004898 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004899 "test/maxpool-microkernel-tester.h",
4900 ] + MICROKERNEL_TEST_HDRS,
4901 deps = MICROKERNEL_TEST_DEPS,
4902)
4903
4904xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004905 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004906 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004907 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004908 "test/avgpool-microkernel-tester.h",
4909 "src/xnnpack/AlignedAllocator.h",
4910 ] + MICROKERNEL_TEST_HDRS,
4911 deps = MICROKERNEL_TEST_DEPS,
4912)
4913
4914xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004915 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004916 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004917 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004918 "test/gemm-microkernel-tester.h",
4919 "src/xnnpack/AlignedAllocator.h",
4920 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004921 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004922)
4923
4924xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004925 name = "f16_prelu_test",
4926 srcs = [
4927 "test/f16-prelu.cc",
4928 "test/prelu-microkernel-tester.h",
4929 "src/xnnpack/AlignedAllocator.h",
4930 ] + MICROKERNEL_TEST_HDRS,
4931 deps = MICROKERNEL_TEST_DEPS,
4932)
4933
4934xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004935 name = "f32_prelu_test",
4936 srcs = [
4937 "test/f32-prelu.cc",
4938 "test/prelu-microkernel-tester.h",
4939 "src/xnnpack/AlignedAllocator.h",
4940 ] + MICROKERNEL_TEST_HDRS,
4941 deps = MICROKERNEL_TEST_DEPS,
4942)
4943
4944xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004945 name = "f32_raddexpminusmax_test",
4946 srcs = [
4947 "test/f32-raddexpminusmax.cc",
4948 "test/raddexpminusmax-microkernel-tester.h",
4949 ] + MICROKERNEL_TEST_HDRS,
4950 deps = MICROKERNEL_TEST_DEPS,
4951)
4952
4953xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07004954 name = "f32_raddextexp_test",
4955 srcs = [
4956 "test/f32-raddextexp.cc",
4957 "test/raddextexp-microkernel-tester.h",
4958 ] + MICROKERNEL_TEST_HDRS,
4959 deps = MICROKERNEL_TEST_DEPS,
4960)
4961
4962xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07004963 name = "f32_raddstoreexpminusmax_test",
4964 srcs = [
4965 "test/f32-raddstoreexpminusmax.cc",
4966 "test/raddstoreexpminusmax-microkernel-tester.h",
4967 ] + MICROKERNEL_TEST_HDRS,
4968 deps = MICROKERNEL_TEST_DEPS,
4969)
4970
4971xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07004972 name = "f32_relu_test",
4973 srcs = [
4974 "test/f32-relu.cc",
4975 "test/vunary-microkernel-tester.h",
4976 ] + MICROKERNEL_TEST_HDRS,
4977 deps = MICROKERNEL_TEST_DEPS,
4978)
4979
4980xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004981 name = "f32_rmax_test",
4982 srcs = [
4983 "test/f32-rmax.cc",
4984 "test/rmax-microkernel-tester.h",
4985 ] + MICROKERNEL_TEST_HDRS,
4986 deps = MICROKERNEL_TEST_DEPS,
4987)
4988
4989xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08004990 name = "f32_sigmoid_test",
4991 srcs = [
4992 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08004993 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08004994 ] + MICROKERNEL_TEST_HDRS,
4995 deps = MICROKERNEL_TEST_DEPS,
4996)
4997
4998xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004999 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005001 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005002 "test/spmm-microkernel-tester.h",
5003 "src/xnnpack/AlignedAllocator.h",
5004 ] + MICROKERNEL_TEST_HDRS,
5005 deps = MICROKERNEL_TEST_DEPS,
5006)
5007
5008xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005009 name = "f16_vadd_minmax_test",
5010 srcs = [
5011 "test/f16-vadd-minmax.cc",
5012 "test/vbinary-microkernel-tester.h",
5013 ] + MICROKERNEL_TEST_HDRS,
5014 deps = MICROKERNEL_TEST_DEPS,
5015)
5016
5017xnnpack_unit_test(
5018 name = "f16_vaddc_minmax_test",
5019 srcs = [
5020 "test/f16-vaddc-minmax.cc",
5021 "test/vbinaryc-microkernel-tester.h",
5022 ] + MICROKERNEL_TEST_HDRS,
5023 deps = MICROKERNEL_TEST_DEPS,
5024)
5025
5026xnnpack_unit_test(
5027 name = "f16_vdiv_minmax_test",
5028 srcs = [
5029 "test/f16-vdiv-minmax.cc",
5030 "test/vbinary-microkernel-tester.h",
5031 ] + MICROKERNEL_TEST_HDRS,
5032 deps = MICROKERNEL_TEST_DEPS,
5033)
5034
5035xnnpack_unit_test(
5036 name = "f16_vdivc_minmax_test",
5037 srcs = [
5038 "test/f16-vdivc-minmax.cc",
5039 "test/vbinaryc-microkernel-tester.h",
5040 ] + MICROKERNEL_TEST_HDRS,
5041 deps = MICROKERNEL_TEST_DEPS,
5042)
5043
5044xnnpack_unit_test(
5045 name = "f16_vrdivc_minmax_test",
5046 srcs = [
5047 "test/f16-vrdivc-minmax.cc",
5048 "test/vbinaryc-microkernel-tester.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
5054 name = "f16_vmax_test",
5055 srcs = [
5056 "test/f16-vmax.cc",
5057 "test/vbinary-microkernel-tester.h",
5058 ] + MICROKERNEL_TEST_HDRS,
5059 deps = MICROKERNEL_TEST_DEPS,
5060)
5061
5062xnnpack_unit_test(
5063 name = "f16_vmaxc_test",
5064 srcs = [
5065 "test/f16-vmaxc.cc",
5066 "test/vbinaryc-microkernel-tester.h",
5067 ] + MICROKERNEL_TEST_HDRS,
5068 deps = MICROKERNEL_TEST_DEPS,
5069)
5070
5071xnnpack_unit_test(
5072 name = "f16_vmin_test",
5073 srcs = [
5074 "test/f16-vmin.cc",
5075 "test/vbinary-microkernel-tester.h",
5076 ] + MICROKERNEL_TEST_HDRS,
5077 deps = MICROKERNEL_TEST_DEPS,
5078)
5079
5080xnnpack_unit_test(
5081 name = "f16_vminc_test",
5082 srcs = [
5083 "test/f16-vminc.cc",
5084 "test/vbinaryc-microkernel-tester.h",
5085 ] + MICROKERNEL_TEST_HDRS,
5086 deps = MICROKERNEL_TEST_DEPS,
5087)
5088
5089xnnpack_unit_test(
5090 name = "f16_vmul_minmax_test",
5091 srcs = [
5092 "test/f16-vmul-minmax.cc",
5093 "test/vbinary-microkernel-tester.h",
5094 ] + MICROKERNEL_TEST_HDRS,
5095 deps = MICROKERNEL_TEST_DEPS,
5096)
5097
5098xnnpack_unit_test(
5099 name = "f16_vmulc_minmax_test",
5100 srcs = [
5101 "test/f16-vmulc-minmax.cc",
5102 "test/vbinaryc-microkernel-tester.h",
5103 ] + MICROKERNEL_TEST_HDRS,
5104 deps = MICROKERNEL_TEST_DEPS,
5105)
5106
5107xnnpack_unit_test(
5108 name = "f16_vsub_minmax_test",
5109 srcs = [
5110 "test/f16-vsub-minmax.cc",
5111 "test/vbinary-microkernel-tester.h",
5112 ] + MICROKERNEL_TEST_HDRS,
5113 deps = MICROKERNEL_TEST_DEPS,
5114)
5115
5116xnnpack_unit_test(
5117 name = "f16_vsubc_minmax_test",
5118 srcs = [
5119 "test/f16-vsubc-minmax.cc",
5120 "test/vbinaryc-microkernel-tester.h",
5121 ] + MICROKERNEL_TEST_HDRS,
5122 deps = MICROKERNEL_TEST_DEPS,
5123)
5124
5125xnnpack_unit_test(
5126 name = "f16_vrsubc_minmax_test",
5127 srcs = [
5128 "test/f16-vrsubc-minmax.cc",
5129 "test/vbinaryc-microkernel-tester.h",
5130 ] + MICROKERNEL_TEST_HDRS,
5131 deps = MICROKERNEL_TEST_DEPS,
5132)
5133
5134xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005135 name = "f32_vabs_test",
5136 srcs = [
5137 "test/f32-vabs.cc",
5138 "test/vunary-microkernel-tester.h",
5139 ] + MICROKERNEL_TEST_HDRS,
5140 deps = MICROKERNEL_TEST_DEPS,
5141)
5142
5143xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005144 name = "f32_vadd_test",
5145 srcs = [
5146 "test/f32-vadd.cc",
5147 "test/vbinary-microkernel-tester.h",
5148 ] + MICROKERNEL_TEST_HDRS,
5149 deps = MICROKERNEL_TEST_DEPS,
5150)
5151
5152xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005153 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005154 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005155 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005156 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005157 ] + MICROKERNEL_TEST_HDRS,
5158 deps = MICROKERNEL_TEST_DEPS,
5159)
5160
5161xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005162 name = "f32_vadd_relu_test",
5163 srcs = [
5164 "test/f32-vadd-relu.cc",
5165 "test/vbinary-microkernel-tester.h",
5166 ] + MICROKERNEL_TEST_HDRS,
5167 deps = MICROKERNEL_TEST_DEPS,
5168)
5169
5170xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005171 name = "f32_vaddc_test",
5172 srcs = [
5173 "test/f32-vaddc.cc",
5174 "test/vbinaryc-microkernel-tester.h",
5175 ] + MICROKERNEL_TEST_HDRS,
5176 deps = MICROKERNEL_TEST_DEPS,
5177)
5178
5179xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005180 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005181 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005182 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005183 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005184 ] + MICROKERNEL_TEST_HDRS,
5185 deps = MICROKERNEL_TEST_DEPS,
5186)
5187
5188xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005189 name = "f32_vaddc_relu_test",
5190 srcs = [
5191 "test/f32-vaddc-relu.cc",
5192 "test/vbinaryc-microkernel-tester.h",
5193 ] + MICROKERNEL_TEST_HDRS,
5194 deps = MICROKERNEL_TEST_DEPS,
5195)
5196
5197xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005198 name = "f32_vdiv_test",
5199 srcs = [
5200 "test/f32-vdiv.cc",
5201 "test/vbinary-microkernel-tester.h",
5202 ] + MICROKERNEL_TEST_HDRS,
5203 deps = MICROKERNEL_TEST_DEPS,
5204)
5205
5206xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005207 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005208 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005209 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005210 "test/vbinary-microkernel-tester.h",
5211 ] + MICROKERNEL_TEST_HDRS,
5212 deps = MICROKERNEL_TEST_DEPS,
5213)
5214
5215xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005216 name = "f32_vdiv_relu_test",
5217 srcs = [
5218 "test/f32-vdiv-relu.cc",
5219 "test/vbinary-microkernel-tester.h",
5220 ] + MICROKERNEL_TEST_HDRS,
5221 deps = MICROKERNEL_TEST_DEPS,
5222)
5223
5224xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005225 name = "f32_vdivc_test",
5226 srcs = [
5227 "test/f32-vdivc.cc",
5228 "test/vbinaryc-microkernel-tester.h",
5229 ] + MICROKERNEL_TEST_HDRS,
5230 deps = MICROKERNEL_TEST_DEPS,
5231)
5232
5233xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005234 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005235 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005236 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005237 "test/vbinaryc-microkernel-tester.h",
5238 ] + MICROKERNEL_TEST_HDRS,
5239 deps = MICROKERNEL_TEST_DEPS,
5240)
5241
5242xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005243 name = "f32_vdivc_relu_test",
5244 srcs = [
5245 "test/f32-vdivc-relu.cc",
5246 "test/vbinaryc-microkernel-tester.h",
5247 ] + MICROKERNEL_TEST_HDRS,
5248 deps = MICROKERNEL_TEST_DEPS,
5249)
5250
5251xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005252 name = "f32_vrdivc_test",
5253 srcs = [
5254 "test/f32-vrdivc.cc",
5255 "test/vbinaryc-microkernel-tester.h",
5256 ] + MICROKERNEL_TEST_HDRS,
5257 deps = MICROKERNEL_TEST_DEPS,
5258)
5259
5260xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005261 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005262 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005263 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005264 "test/vbinaryc-microkernel-tester.h",
5265 ] + MICROKERNEL_TEST_HDRS,
5266 deps = MICROKERNEL_TEST_DEPS,
5267)
5268
5269xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005270 name = "f32_vrdivc_relu_test",
5271 srcs = [
5272 "test/f32-vrdivc-relu.cc",
5273 "test/vbinaryc-microkernel-tester.h",
5274 ] + MICROKERNEL_TEST_HDRS,
5275 deps = MICROKERNEL_TEST_DEPS,
5276)
5277
5278xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005279 name = "f32_vmax_test",
5280 srcs = [
5281 "test/f32-vmax.cc",
5282 "test/vbinary-microkernel-tester.h",
5283 ] + MICROKERNEL_TEST_HDRS,
5284 deps = MICROKERNEL_TEST_DEPS,
5285)
5286
5287xnnpack_unit_test(
5288 name = "f32_vmaxc_test",
5289 srcs = [
5290 "test/f32-vmaxc.cc",
5291 "test/vbinaryc-microkernel-tester.h",
5292 ] + MICROKERNEL_TEST_HDRS,
5293 deps = MICROKERNEL_TEST_DEPS,
5294)
5295
5296xnnpack_unit_test(
5297 name = "f32_vmin_test",
5298 srcs = [
5299 "test/f32-vmin.cc",
5300 "test/vbinary-microkernel-tester.h",
5301 ] + MICROKERNEL_TEST_HDRS,
5302 deps = MICROKERNEL_TEST_DEPS,
5303)
5304
5305xnnpack_unit_test(
5306 name = "f32_vminc_test",
5307 srcs = [
5308 "test/f32-vminc.cc",
5309 "test/vbinaryc-microkernel-tester.h",
5310 ] + MICROKERNEL_TEST_HDRS,
5311 deps = MICROKERNEL_TEST_DEPS,
5312)
5313
5314xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005315 name = "f32_vmul_test",
5316 srcs = [
5317 "test/f32-vmul.cc",
5318 "test/vbinary-microkernel-tester.h",
5319 ] + MICROKERNEL_TEST_HDRS,
5320 deps = MICROKERNEL_TEST_DEPS,
5321)
5322
5323xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005324 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005325 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005326 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005327 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005328 ] + MICROKERNEL_TEST_HDRS,
5329 deps = MICROKERNEL_TEST_DEPS,
5330)
5331
5332xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005333 name = "f32_vmul_relu_test",
5334 srcs = [
5335 "test/f32-vmul-relu.cc",
5336 "test/vbinary-microkernel-tester.h",
5337 ] + MICROKERNEL_TEST_HDRS,
5338 deps = MICROKERNEL_TEST_DEPS,
5339)
5340
5341xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005342 name = "f32_vmulc_test",
5343 srcs = [
5344 "test/f32-vmulc.cc",
5345 "test/vbinaryc-microkernel-tester.h",
5346 ] + MICROKERNEL_TEST_HDRS,
5347 deps = MICROKERNEL_TEST_DEPS,
5348)
5349
5350xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005351 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005352 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005353 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005354 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005355 ] + MICROKERNEL_TEST_HDRS,
5356 deps = MICROKERNEL_TEST_DEPS,
5357)
5358
5359xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005360 name = "f32_vmulc_relu_test",
5361 srcs = [
5362 "test/f32-vmulc-relu.cc",
5363 "test/vbinaryc-microkernel-tester.h",
5364 ] + MICROKERNEL_TEST_HDRS,
5365 deps = MICROKERNEL_TEST_DEPS,
5366)
5367
5368xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005369 name = "f16_vmulcaddc_minmax_test",
5370 srcs = [
5371 "test/f16-vmulcaddc-minmax.cc",
5372 "test/vmulcaddc-microkernel-tester.h",
5373 "src/xnnpack/AlignedAllocator.h",
5374 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005375 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005376)
5377
5378xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005379 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005380 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005381 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005382 "test/vmulcaddc-microkernel-tester.h",
5383 "src/xnnpack/AlignedAllocator.h",
5384 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005385 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005386)
5387
5388xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005389 name = "f32_vlrelu_test",
5390 srcs = [
5391 "test/f32-vlrelu.cc",
5392 "test/vunary-microkernel-tester.h",
5393 ] + MICROKERNEL_TEST_HDRS,
5394 deps = MICROKERNEL_TEST_DEPS,
5395)
5396
5397xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005398 name = "f32_vneg_test",
5399 srcs = [
5400 "test/f32-vneg.cc",
5401 "test/vunary-microkernel-tester.h",
5402 ] + MICROKERNEL_TEST_HDRS,
5403 deps = MICROKERNEL_TEST_DEPS,
5404)
5405
5406xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005407 name = "f32_vrndne_test",
5408 srcs = [
5409 "test/f32-vrndne.cc",
5410 "test/vunary-microkernel-tester.h",
5411 ] + MICROKERNEL_TEST_HDRS,
5412 deps = MICROKERNEL_TEST_DEPS,
5413)
5414
5415xnnpack_unit_test(
5416 name = "f32_vrndz_test",
5417 srcs = [
5418 "test/f32-vrndz.cc",
5419 "test/vunary-microkernel-tester.h",
5420 ] + MICROKERNEL_TEST_HDRS,
5421 deps = MICROKERNEL_TEST_DEPS,
5422)
5423
5424xnnpack_unit_test(
5425 name = "f32_vrndu_test",
5426 srcs = [
5427 "test/f32-vrndu.cc",
5428 "test/vunary-microkernel-tester.h",
5429 ] + MICROKERNEL_TEST_HDRS,
5430 deps = MICROKERNEL_TEST_DEPS,
5431)
5432
5433xnnpack_unit_test(
5434 name = "f32_vrndd_test",
5435 srcs = [
5436 "test/f32-vrndd.cc",
5437 "test/vunary-microkernel-tester.h",
5438 ] + MICROKERNEL_TEST_HDRS,
5439 deps = MICROKERNEL_TEST_DEPS,
5440)
5441
5442xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005443 name = "f32_vscale_test",
5444 srcs = [
5445 "test/f32-vscale.cc",
5446 "test/vscale-microkernel-tester.h",
5447 ] + MICROKERNEL_TEST_HDRS,
5448 deps = MICROKERNEL_TEST_DEPS,
5449)
5450
5451xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005452 name = "f32_vscaleexpminusmax_test",
5453 srcs = [
5454 "test/f32-vscaleexpminusmax.cc",
5455 "test/vscaleexpminusmax-microkernel-tester.h",
5456 ] + MICROKERNEL_TEST_HDRS,
5457 deps = MICROKERNEL_TEST_DEPS,
5458)
5459
5460xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005461 name = "f32_vscaleextexp_test",
5462 srcs = [
5463 "test/f32-vscaleextexp.cc",
5464 "test/vscaleextexp-microkernel-tester.h",
5465 ] + MICROKERNEL_TEST_HDRS,
5466 deps = MICROKERNEL_TEST_DEPS,
5467)
5468
5469xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005470 name = "f32_vsqr_test",
5471 srcs = [
5472 "test/f32-vsqr.cc",
5473 "test/vunary-microkernel-tester.h",
5474 ] + MICROKERNEL_TEST_HDRS,
5475 deps = MICROKERNEL_TEST_DEPS,
5476)
5477
5478xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005479 name = "f32_vsqrdiff_test",
5480 srcs = [
5481 "test/f32-vsqrdiff.cc",
5482 "test/vbinary-microkernel-tester.h",
5483 ] + MICROKERNEL_TEST_HDRS,
5484 deps = MICROKERNEL_TEST_DEPS,
5485)
5486
5487xnnpack_unit_test(
5488 name = "f32_vsqrdiffc_test",
5489 srcs = [
5490 "test/f32-vsqrdiffc.cc",
5491 "test/vbinaryc-microkernel-tester.h",
5492 ] + MICROKERNEL_TEST_HDRS,
5493 deps = MICROKERNEL_TEST_DEPS,
5494)
5495
5496xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005497 name = "f32_vsqrt_test",
5498 srcs = [
5499 "test/f32-vsqrt.cc",
5500 "test/vunary-microkernel-tester.h",
5501 ] + MICROKERNEL_TEST_HDRS,
5502 deps = MICROKERNEL_TEST_DEPS,
5503)
5504
5505xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005506 name = "f32_vsub_test",
5507 srcs = [
5508 "test/f32-vsub.cc",
5509 "test/vbinary-microkernel-tester.h",
5510 ] + MICROKERNEL_TEST_HDRS,
5511 deps = MICROKERNEL_TEST_DEPS,
5512)
5513
5514xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005515 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005516 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005517 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005518 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005519 ] + MICROKERNEL_TEST_HDRS,
5520 deps = MICROKERNEL_TEST_DEPS,
5521)
5522
5523xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005524 name = "f32_vsub_relu_test",
5525 srcs = [
5526 "test/f32-vsub-relu.cc",
5527 "test/vbinary-microkernel-tester.h",
5528 ] + MICROKERNEL_TEST_HDRS,
5529 deps = MICROKERNEL_TEST_DEPS,
5530)
5531
5532xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005533 name = "f32_vsubc_test",
5534 srcs = [
5535 "test/f32-vsubc.cc",
5536 "test/vbinaryc-microkernel-tester.h",
5537 ] + MICROKERNEL_TEST_HDRS,
5538 deps = MICROKERNEL_TEST_DEPS,
5539)
5540
5541xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005542 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005543 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005544 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005545 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005546 ] + MICROKERNEL_TEST_HDRS,
5547 deps = MICROKERNEL_TEST_DEPS,
5548)
5549
5550xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005551 name = "f32_vsubc_relu_test",
5552 srcs = [
5553 "test/f32-vsubc-relu.cc",
5554 "test/vbinaryc-microkernel-tester.h",
5555 ] + MICROKERNEL_TEST_HDRS,
5556 deps = MICROKERNEL_TEST_DEPS,
5557)
5558
5559xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005560 name = "f32_vrsubc_test",
5561 srcs = [
5562 "test/f32-vrsubc.cc",
5563 "test/vbinaryc-microkernel-tester.h",
5564 ] + MICROKERNEL_TEST_HDRS,
5565 deps = MICROKERNEL_TEST_DEPS,
5566)
5567
5568xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005569 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005570 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005571 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005572 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005573 ] + MICROKERNEL_TEST_HDRS,
5574 deps = MICROKERNEL_TEST_DEPS,
5575)
5576
5577xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005578 name = "f32_vrsubc_relu_test",
5579 srcs = [
5580 "test/f32-vrsubc-relu.cc",
5581 "test/vbinaryc-microkernel-tester.h",
5582 ] + MICROKERNEL_TEST_HDRS,
5583 deps = MICROKERNEL_TEST_DEPS,
5584)
5585
5586xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005587 name = "qs8_dwconv_minmax_test",
5588 srcs = [
5589 "test/qs8-dwconv-minmax.cc",
5590 "test/dwconv-microkernel-tester.h",
5591 "src/xnnpack/AlignedAllocator.h",
5592 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5593 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5594)
5595
5596xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005597 name = "qs8_gavgpool_minmax_test",
5598 srcs = [
5599 "test/qs8-gavgpool-minmax.cc",
5600 "test/gavgpool-microkernel-tester.h",
5601 "src/xnnpack/AlignedAllocator.h",
5602 ] + MICROKERNEL_TEST_HDRS,
5603 deps = MICROKERNEL_TEST_DEPS,
5604)
5605
5606xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005607 name = "qs8_gemm_minmax_test",
5608 srcs = [
5609 "test/qs8-gemm-minmax.cc",
5610 "test/gemm-microkernel-tester.h",
5611 "src/xnnpack/AlignedAllocator.h",
5612 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5613 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5614)
5615
5616xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005617 name = "qs8_igemm_minmax_test",
5618 srcs = [
5619 "test/qs8-igemm-minmax.cc",
5620 "test/gemm-microkernel-tester.h",
5621 "src/xnnpack/AlignedAllocator.h",
5622 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5624)
5625
5626xnnpack_unit_test(
5627 name = "qs8_requantization_test",
5628 srcs = [
5629 "src/xnnpack/requantization-stubs.h",
5630 "test/qs8-requantization.cc",
5631 "test/requantization-tester.h",
5632 ] + MICROKERNEL_TEST_HDRS,
5633 deps = MICROKERNEL_TEST_DEPS,
5634)
5635
5636xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005637 name = "qs8_vadd_minmax_test",
5638 srcs = [
5639 "test/qs8-vadd-minmax.cc",
5640 "test/vadd-microkernel-tester.h",
5641 ] + MICROKERNEL_TEST_HDRS,
5642 deps = MICROKERNEL_TEST_DEPS,
5643)
5644
5645xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005646 name = "qs8_vaddc_minmax_test",
5647 srcs = [
5648 "test/qs8-vaddc-minmax.cc",
5649 "test/vaddc-microkernel-tester.h",
5650 ] + MICROKERNEL_TEST_HDRS,
5651 deps = MICROKERNEL_TEST_DEPS,
5652)
5653
5654xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005655 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005656 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005657 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005658 "test/avgpool-microkernel-tester.h",
5659 "src/xnnpack/AlignedAllocator.h",
5660 ] + MICROKERNEL_TEST_HDRS,
5661 deps = MICROKERNEL_TEST_DEPS,
5662)
5663
5664xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005665 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005666 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005667 "test/qu8-dwconv-minmax.cc",
5668 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005669 "src/xnnpack/AlignedAllocator.h",
5670 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005671 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005672)
5673
5674xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005675 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005676 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005677 "test/qu8-igemm-minmax.cc",
5678 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005679 "src/xnnpack/AlignedAllocator.h",
5680 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005681 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005682)
5683
5684xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005685 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005686 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005687 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005688 "test/gavgpool-microkernel-tester.h",
5689 "src/xnnpack/AlignedAllocator.h",
5690 ] + MICROKERNEL_TEST_HDRS,
5691 deps = MICROKERNEL_TEST_DEPS,
5692)
5693
5694xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005695 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005696 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005697 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005698 "test/gemm-microkernel-tester.h",
5699 "src/xnnpack/AlignedAllocator.h",
5700 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005701 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005702)
5703
5704xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005705 name = "qu8_requantization_test",
5706 srcs = [
5707 "src/xnnpack/requantization-stubs.h",
5708 "test/qu8-requantization.cc",
5709 "test/requantization-tester.h",
5710 ] + MICROKERNEL_TEST_HDRS,
5711 deps = MICROKERNEL_TEST_DEPS,
5712)
5713
5714xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005715 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005716 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005717 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005718 "test/vadd-microkernel-tester.h",
5719 ] + MICROKERNEL_TEST_HDRS,
5720 deps = MICROKERNEL_TEST_DEPS,
5721)
5722
5723xnnpack_unit_test(
5724 name = "u8_clamp_test",
5725 srcs = [
5726 "test/u8-clamp.cc",
5727 "test/clamp-microkernel-tester.h",
5728 ] + MICROKERNEL_TEST_HDRS,
5729 deps = MICROKERNEL_TEST_DEPS,
5730)
5731
5732xnnpack_unit_test(
5733 name = "u8_lut32norm_test",
5734 srcs = [
5735 "test/u8-lut32norm.cc",
5736 "test/lut-norm-microkernel-tester.h",
5737 ] + MICROKERNEL_TEST_HDRS,
5738 deps = MICROKERNEL_TEST_DEPS,
5739)
5740
5741xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005742 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005743 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005744 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005745 "test/maxpool-microkernel-tester.h",
5746 ] + MICROKERNEL_TEST_HDRS,
5747 deps = MICROKERNEL_TEST_DEPS,
5748)
5749
5750xnnpack_unit_test(
5751 name = "u8_rmax_test",
5752 srcs = [
5753 "test/u8-rmax.cc",
5754 "test/rmax-microkernel-tester.h",
5755 ] + MICROKERNEL_TEST_HDRS,
5756 deps = MICROKERNEL_TEST_DEPS,
5757)
5758
5759xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005760 name = "x32_fill_test",
5761 srcs = [
5762 "test/x32-fill.cc",
5763 "test/fill-microkernel-tester.h",
5764 ] + MICROKERNEL_TEST_HDRS,
5765 deps = MICROKERNEL_TEST_DEPS,
5766)
5767
5768xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005769 name = "x32_packx_test",
5770 srcs = [
5771 "test/x32-packx.cc",
5772 "test/pack-microkernel-tester.h",
5773 "src/xnnpack/AlignedAllocator.h",
5774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
5779 name = "x32_pad_test",
5780 srcs = [
5781 "test/x32-pad.cc",
5782 "test/pad-microkernel-tester.h",
5783 ] + MICROKERNEL_TEST_HDRS,
5784 deps = MICROKERNEL_TEST_DEPS,
5785)
5786
5787xnnpack_unit_test(
5788 name = "x32_unpool_test",
5789 srcs = [
5790 "test/x32-unpool.cc",
5791 "test/unpool-microkernel-tester.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
5797 name = "x32_zip_test",
5798 srcs = [
5799 "test/x32-zip.cc",
5800 "test/zip-microkernel-tester.h",
5801 ] + MICROKERNEL_TEST_HDRS,
5802 deps = MICROKERNEL_TEST_DEPS,
5803)
5804
5805xnnpack_unit_test(
5806 name = "x8_lut_test",
5807 srcs = [
5808 "test/x8-lut.cc",
5809 "test/lut-microkernel-tester.h",
5810 ] + MICROKERNEL_TEST_HDRS,
5811 deps = MICROKERNEL_TEST_DEPS,
5812)
5813
5814xnnpack_unit_test(
5815 name = "x8_zip_test",
5816 srcs = [
5817 "test/x8-zip.cc",
5818 "test/zip-microkernel-tester.h",
5819 ] + MICROKERNEL_TEST_HDRS,
5820 deps = MICROKERNEL_TEST_DEPS,
5821)
5822
Marat Dukhan20c3b922020-03-10 03:45:06 -07005823########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005824
5825xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005826 name = "operator_size_test",
5827 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005828 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005829)
5830
Marat Dukhan20c3b922020-03-10 03:45:06 -07005831xnnpack_binary(
5832 name = "subgraph_size_test",
5833 srcs = ["test/subgraph-size.c"],
5834 deps = [":XNNPACK"],
5835)
5836
5837########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005838
5839xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005840 name = "abs_nc_test",
5841 srcs = [
5842 "test/abs-nc.cc",
5843 "test/abs-operator-tester.h",
5844 ],
5845 deps = OPERATOR_TEST_DEPS,
5846)
5847
5848xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005849 name = "add_nd_test",
5850 srcs = [
5851 "test/add-nd.cc",
5852 "test/binary-elementwise-operator-tester.h",
5853 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005854 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005855)
5856
5857xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005858 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005859 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005860 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005861 "test/argmax-pooling-operator-tester.h",
5862 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005863 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005864)
5865
5866xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005867 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005868 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005869 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005870 "test/average-pooling-operator-tester.h",
5871 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005872 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005873)
5874
5875xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005876 name = "bankers_rounding_nc_test",
5877 srcs = [
5878 "test/bankers-rounding-nc.cc",
5879 "test/bankers-rounding-operator-tester.h",
5880 ],
5881 deps = OPERATOR_TEST_DEPS,
5882)
5883
5884xnnpack_unit_test(
5885 name = "ceiling_nc_test",
5886 srcs = [
5887 "test/ceiling-nc.cc",
5888 "test/ceiling-operator-tester.h",
5889 ],
5890 deps = OPERATOR_TEST_DEPS,
5891)
5892
5893xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005894 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005895 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005896 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005897 "test/channel-shuffle-operator-tester.h",
5898 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005899 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005900)
5901
5902xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005903 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005904 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005905 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906 "test/clamp-operator-tester.h",
5907 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005908 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005909)
5910
5911xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07005912 name = "constant_pad_nd_test",
5913 srcs = [
5914 "test/constant-pad-nd.cc",
5915 "test/constant-pad-operator-tester.h",
5916 ],
5917 deps = OPERATOR_TEST_DEPS,
5918)
5919
5920xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005921 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005922 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005923 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005924 "test/convolution-operator-tester.h",
5925 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005926 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005927)
5928
5929xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005930 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005931 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005932 "test/convolution-nchw.cc",
5933 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005934 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005935 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005936)
5937
5938xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07005939 name = "copy_nc_test",
5940 srcs = [
5941 "test/copy-nc.cc",
5942 "test/copy-operator-tester.h",
5943 ],
5944 deps = OPERATOR_TEST_DEPS,
5945)
5946
5947xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005948 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005949 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005950 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005951 "test/deconvolution-operator-tester.h",
5952 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005953 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005954)
5955
5956xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08005957 name = "divide_nd_test",
5958 srcs = [
5959 "test/binary-elementwise-operator-tester.h",
5960 "test/divide-nd.cc",
5961 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005962 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08005963)
5964
5965xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005966 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005967 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005968 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005969 "test/fully-connected-operator-tester.h",
5970 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005971 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005972)
5973
5974xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07005975 name = "floor_nc_test",
5976 srcs = [
5977 "test/floor-nc.cc",
5978 "test/floor-operator-tester.h",
5979 ],
5980 deps = OPERATOR_TEST_DEPS,
5981)
5982
5983xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005984 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005985 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005986 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005987 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07005988 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07005989 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005990)
5991
5992xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005993 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005994 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005995 "test/global-average-pooling-ncw.cc",
5996 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005997 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005998 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07005999)
6000
6001xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006002 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006003 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006004 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006005 "test/hardswish-operator-tester.h",
6006 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006007 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006008)
6009
6010xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006011 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006012 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006013 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006014 "test/leaky-relu-operator-tester.h",
6015 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006016 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006017)
6018
6019xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006020 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006021 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006022 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006023 "test/max-pooling-operator-tester.h",
6024 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006025 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006026)
6027
6028xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006029 name = "maximum_nd_test",
6030 srcs = [
6031 "test/binary-elementwise-operator-tester.h",
6032 "test/maximum-nd.cc",
6033 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006034 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006035)
6036
6037xnnpack_unit_test(
6038 name = "minimum_nd_test",
6039 srcs = [
6040 "test/binary-elementwise-operator-tester.h",
6041 "test/minimum-nd.cc",
6042 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006043 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006044)
6045
6046xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006047 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006048 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006049 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006050 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006051 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006052 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006053)
6054
6055xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006056 name = "negate_nc_test",
6057 srcs = [
6058 "test/negate-nc.cc",
6059 "test/negate-operator-tester.h",
6060 ],
6061 deps = OPERATOR_TEST_DEPS,
6062)
6063
6064xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006065 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006067 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006068 "test/prelu-operator-tester.h",
6069 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006070 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071)
6072
6073xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006074 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006075 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006076 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006077 "test/resize-bilinear-operator-tester.h",
6078 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006079 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006080)
6081
6082xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006083 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006084 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006085 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006086 "test/sigmoid-operator-tester.h",
6087 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006088 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006089)
6090
6091xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006092 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006094 "test/softmax-nc.cc",
6095 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006096 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006097 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006098)
6099
6100xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006101 name = "square_nc_test",
6102 srcs = [
6103 "test/square-nc.cc",
6104 "test/square-operator-tester.h",
6105 ],
6106 deps = OPERATOR_TEST_DEPS,
6107)
6108
6109xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006110 name = "square_root_nc_test",
6111 srcs = [
6112 "test/square-root-nc.cc",
6113 "test/square-root-operator-tester.h",
6114 ],
6115 deps = OPERATOR_TEST_DEPS,
6116)
6117
6118xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006119 name = "squared_difference_nd_test",
6120 srcs = [
6121 "test/binary-elementwise-operator-tester.h",
6122 "test/squared-difference-nd.cc",
6123 ],
6124 deps = OPERATOR_TEST_DEPS,
6125)
6126
6127xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006128 name = "subtract_nd_test",
6129 srcs = [
6130 "test/binary-elementwise-operator-tester.h",
6131 "test/subtract-nd.cc",
6132 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006133 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006134)
6135
6136xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006137 name = "truncation_nc_test",
6138 srcs = [
6139 "test/truncation-nc.cc",
6140 "test/truncation-operator-tester.h",
6141 ],
6142 deps = OPERATOR_TEST_DEPS,
6143)
6144
6145xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006146 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006147 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006148 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006149 "test/unpooling-operator-tester.h",
6150 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006151 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006152)
6153
Chao Mei6ddfc602020-05-13 22:29:36 -07006154############################### Misc unit tests ###############################
6155
6156xnnpack_unit_test(
6157 name = "memory_planner_test",
6158 srcs = [
6159 "test/memory-planner-test.cc",
6160 ],
6161 deps = [
6162 ":XNNPACK",
6163 ":memory_planner",
6164 ],
6165)
6166
Marat Dukhan08c4a432019-10-03 09:29:21 -07006167############################# Build configurations #############################
6168
Marat Dukhanb8642352019-10-30 15:43:02 -07006169# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006170config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006171 name = "xnn_enable_assembly_explicit_true",
6172 define_values = {"xnn_enable_assembly": "true"},
6173)
6174
6175# Disables usage of assembly kernels.
6176config_setting(
6177 name = "xnn_enable_assembly_explicit_false",
6178 define_values = {"xnn_enable_assembly": "false"},
6179)
6180
Marat Dukhan9de90e02020-06-18 16:04:12 -07006181# Enables usage of sparse inference.
6182config_setting(
6183 name = "xnn_enable_sparse_explicit_true",
6184 define_values = {"xnn_enable_sparse": "true"},
6185)
6186
6187# Disables usage of sparse inference.
6188config_setting(
6189 name = "xnn_enable_sparse_explicit_false",
6190 define_values = {"xnn_enable_sparse": "false"},
6191)
6192
Marat Dukhan05702cf2020-03-26 15:41:33 -07006193# Disables usage of HMP-aware optimizations.
6194config_setting(
6195 name = "xnn_enable_hmp_explicit_false",
6196 define_values = {"xnn_enable_hmp": "false"},
6197)
6198
Chao Mei6ddfc602020-05-13 22:29:36 -07006199# Enable usage of optimized memory allocation
6200config_setting(
6201 name = "xnn_enable_memopt_explicit_true",
6202 define_values = {"xnnpack_enable_memopt": "true"},
6203)
6204
6205# Disable usage of optimized memory allocation
6206config_setting(
6207 name = "xnn_enable_memopt_explicit_false",
6208 define_values = {"xnnpack_enable_memopt": "false"},
6209)
6210
Marat Dukhanb8642352019-10-30 15:43:02 -07006211# Builds with -c dbg
6212config_setting(
6213 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006214 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006215 "compilation_mode": "dbg",
6216 },
6217)
6218
6219# Builds with -c opt
6220config_setting(
6221 name = "optimized_build",
6222 values = {
6223 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 },
6225)
6226
6227config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006228 name = "linux_k8",
6229 values = {"cpu": "k8"},
6230)
6231
6232config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006233 name = "linux_arm",
6234 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006235)
6236
6237config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006238 name = "linux_armeabi",
6239 values = {"cpu": "armeabi"},
6240)
6241
6242config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006243 name = "linux_armhf",
6244 values = {"cpu": "armhf"},
6245)
6246
6247config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006248 name = "linux_armv7a",
6249 values = {"cpu": "armv7a"},
6250)
6251
6252config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006253 name = "linux_aarch64",
6254 values = {"cpu": "aarch64"},
6255)
6256
6257config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006258 name = "android",
6259 values = {"crosstool_top": "//external:android/crosstool"},
6260)
6261
6262config_setting(
6263 name = "android_armv7",
6264 values = {
6265 "crosstool_top": "//external:android/crosstool",
6266 "cpu": "armeabi-v7a",
6267 },
6268)
6269
6270config_setting(
6271 name = "android_arm64",
6272 values = {
6273 "crosstool_top": "//external:android/crosstool",
6274 "cpu": "arm64-v8a",
6275 },
6276)
6277
6278config_setting(
6279 name = "android_x86",
6280 values = {
6281 "crosstool_top": "//external:android/crosstool",
6282 "cpu": "x86",
6283 },
6284)
6285
6286config_setting(
6287 name = "android_x86_64",
6288 values = {
6289 "crosstool_top": "//external:android/crosstool",
6290 "cpu": "x86_64",
6291 },
6292)
6293
6294config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006295 name = "windows_x86_64",
6296 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006297)
6298
6299config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006300 name = "windows_x86_64_clang",
6301 values = {
6302 "compiler": "clang-cl",
6303 "cpu": "x64_windows",
6304 },
6305)
6306
6307config_setting(
6308 name = "windows_x86_64_mingw",
6309 values = {
6310 "compiler": "mingw-gcc",
6311 "cpu": "x64_windows",
6312 },
6313)
6314
6315config_setting(
6316 name = "windows_x86_64_msys",
6317 values = {
6318 "compiler": "msys-gcc",
6319 "cpu": "x64_windows",
6320 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006321)
6322
6323config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006324 name = "macos_x86_64",
6325 values = {
6326 "apple_platform_type": "macos",
6327 "cpu": "darwin",
6328 },
6329)
6330
6331config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006332 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006333 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006334)
6335
6336config_setting(
6337 name = "emscripten_wasm",
6338 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006339 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006340 "cpu": "wasm",
6341 },
6342)
6343
6344config_setting(
6345 name = "emscripten_wasmsimd",
6346 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006347 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006348 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006349 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006350 },
6351)
6352
6353config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006354 name = "ios_armv7",
6355 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006356 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006357 "cpu": "ios_armv7",
6358 },
6359)
6360
6361config_setting(
6362 name = "ios_arm64",
6363 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006364 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006365 "cpu": "ios_arm64",
6366 },
6367)
6368
6369config_setting(
6370 name = "ios_arm64e",
6371 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006372 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006373 "cpu": "ios_arm64e",
6374 },
6375)
6376
6377config_setting(
6378 name = "ios_x86",
6379 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006380 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006381 "cpu": "ios_i386",
6382 },
6383)
6384
6385config_setting(
6386 name = "ios_x86_64",
6387 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006388 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006389 "cpu": "ios_x86_64",
6390 },
6391)
6392
6393config_setting(
6394 name = "watchos_armv7k",
6395 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006396 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006397 "cpu": "watchos_armv7k",
6398 },
6399)
6400
6401config_setting(
6402 name = "watchos_arm64_32",
6403 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006404 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006405 "cpu": "watchos_arm64_32",
6406 },
6407)
6408
6409config_setting(
6410 name = "watchos_x86",
6411 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006412 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006413 "cpu": "watchos_i386",
6414 },
6415)
6416
6417config_setting(
6418 name = "watchos_x86_64",
6419 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006420 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006421 "cpu": "watchos_x86_64",
6422 },
6423)
6424
6425config_setting(
6426 name = "tvos_arm64",
6427 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006428 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006429 "cpu": "tvos_arm64",
6430 },
6431)
6432
6433config_setting(
6434 name = "tvos_x86_64",
6435 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006436 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006437 "cpu": "tvos_x86_64",
6438 },
6439)