blob: 5f0283c89b8b628a7c1f9a32c2a557d288285729 [file] [log] [blame]
Marat Dukhan08c4a432019-10-03 09:29:21 -07001# Copyright 2019 Google LLC
2#
3# This source code is licensed under the BSD-style license found in the
4# LICENSE file in the root directory of this source tree.
5#
6# Description:
7# XNNPACK - optimized floating-point neural network operators library
8
Marat Dukhan10a38082020-04-17 03:58:35 -07009load(":build_defs.bzl", "xnnpack_aggregate_library", "xnnpack_benchmark", "xnnpack_binary", "xnnpack_cc_library", "xnnpack_gcc_std_copts", "xnnpack_min_size_copts", "xnnpack_msvc_std_copts", "xnnpack_optional_armcl_copts", "xnnpack_optional_armcl_deps", "xnnpack_optional_dnnl_copts", "xnnpack_optional_dnnl_deps", "xnnpack_optional_gemmlowp_copts", "xnnpack_optional_gemmlowp_deps", "xnnpack_optional_ruy_copts", "xnnpack_optional_ruy_deps", "xnnpack_optional_tflite_copts", "xnnpack_optional_tflite_deps", "xnnpack_std_cxxopts", "xnnpack_unit_test", "xnnpack_visibility")
Marat Dukhan69c3f2c2019-11-06 12:30:01 -080010
Marat Dukhan08c4a432019-10-03 09:29:21 -070011licenses(["notice"])
12
13exports_files(["LICENSE"])
14
Marat Dukhan1b354632020-03-23 12:50:22 -070015OPERATOR_BENCHMARK_DEPS = [
16 ":XNNPACK",
17 ":bench_utils",
18 "@cpuinfo",
Frank Barchard0c849732020-06-12 13:31:32 -070019 "@FP16",
Marat Dukhan1b354632020-03-23 12:50:22 -070020 "@pthreadpool",
21]
22
Marat Dukhan08c4a432019-10-03 09:29:21 -070023MICROKERNEL_BENCHMARK_DEPS = [
24 ":ukernels",
25 ":bench_utils",
Frank Barchard7e955972019-10-11 10:34:25 -070026 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070027 "@cpuinfo",
28 "@FP16",
29 "@pthreadpool",
30]
31
Marat Dukhan6adff4e2019-10-14 18:32:07 -070032ACCURACY_EVAL_DEPS = [
33 ":XNNPACK",
34 ":ukernels",
35 "@FP16",
36 "@pthreadpool",
37]
38
Marat Dukhan08c4a432019-10-03 09:29:21 -070039MICROKERNEL_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070040 ":ukernels_test_mode",
Frank Barchard7e955972019-10-11 10:34:25 -070041 ":enable_assembly",
Marat Dukhan08c4a432019-10-03 09:29:21 -070042 "@cpuinfo",
43 "@FP16",
44 "@pthreadpool",
45]
46
Marat Dukhan1b354632020-03-23 12:50:22 -070047OPERATOR_TEST_DEPS = [
Marat Dukhan33fcf782020-05-24 14:27:15 -070048 ":XNNPACK_test_mode",
Marat Dukhan1b354632020-03-23 12:50:22 -070049 "@pthreadpool",
50 "@FP16",
51]
52
Marat Dukhan08c4a432019-10-03 09:29:21 -070053OPERATOR_SRCS = [
Marat Dukhane8265432020-04-28 18:42:59 -070054 "src/operators/argmax-pooling-nhwc.c",
55 "src/operators/average-pooling-nhwc.c",
56 "src/operators/binary-elementwise-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070057 "src/operators/channel-shuffle-nc.c",
Marat Dukhan065b11e2020-05-22 09:49:41 -070058 "src/operators/constant-pad-nd.c",
Marat Dukhane8265432020-04-28 18:42:59 -070059 "src/operators/convolution-nchw.c",
60 "src/operators/convolution-nhwc.c",
61 "src/operators/deconvolution-nhwc.c",
62 "src/operators/fully-connected-nc.c",
63 "src/operators/global-average-pooling-ncw.c",
64 "src/operators/global-average-pooling-nwc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070065 "src/operators/leaky-relu-nc.c",
66 "src/operators/max-pooling-nhwc.c",
67 "src/operators/prelu-nc.c",
68 "src/operators/resize-bilinear-nhwc.c",
69 "src/operators/sigmoid-nc.c",
70 "src/operators/softmax-nc.c",
Marat Dukhanc3065f52020-06-04 13:33:32 -070071 "src/operators/unary-elementwise-nc.c",
Marat Dukhane8265432020-04-28 18:42:59 -070072 "src/operators/unpooling-nhwc.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -070073]
74
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070075SUBGRAPH_SRCS = [
Marat Dukhan5fab4092020-06-10 01:28:28 -070076 "src/subgraph/abs.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070077 "src/subgraph/add2.c",
78 "src/subgraph/argmax-pooling-2d.c",
79 "src/subgraph/average-pooling-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070080 "src/subgraph/bankers-rounding.c",
81 "src/subgraph/ceiling.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070082 "src/subgraph/clamp.c",
83 "src/subgraph/convolution-2d.c",
84 "src/subgraph/deconvolution-2d.c",
85 "src/subgraph/depthwise-convolution-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070086 "src/subgraph/divide.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070087 "src/subgraph/fully-connected.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070088 "src/subgraph/floor.c",
Marat Dukhana059b7d2020-06-11 11:41:27 -070089 "src/subgraph/global-average-pooling-2d.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070090 "src/subgraph/hardswish.c",
Marat Dukhan5bbebac2020-06-10 19:42:15 -070091 "src/subgraph/leaky-relu.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070092 "src/subgraph/max-pooling-2d.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -070093 "src/subgraph/maximum2.c",
94 "src/subgraph/minimum2.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070095 "src/subgraph/multiply2.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -070096 "src/subgraph/negate.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -070097 "src/subgraph/prelu.c",
98 "src/subgraph/sigmoid.c",
99 "src/subgraph/softmax.c",
100 "src/subgraph/static-constant-pad.c",
Marat Dukhand27202d2020-07-09 23:43:40 -0700101 "src/subgraph/static-reshape.c",
Marat Dukhanaff24e22020-07-23 01:43:58 -0700102 "src/subgraph/static-resize-bilinear-2d.c",
Marat Dukhan5fab4092020-06-10 01:28:28 -0700103 "src/subgraph/square.c",
Marat Dukhan51a01c62020-07-09 03:26:57 -0700104 "src/subgraph/square-root.c",
Marat Dukhan9d3a4592020-06-05 16:52:42 -0700105 "src/subgraph/squared-difference.c",
106 "src/subgraph/subtract.c",
Marat Dukhan95e8b7a2020-06-03 12:46:26 -0700107 "src/subgraph/unpooling-2d.c",
108]
109
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800110TABLE_SRCS = [
111 "src/tables/exp2-k-over-64.c",
112 "src/tables/exp2-k-over-2048.c",
113]
114
Marat Dukhan08c4a432019-10-03 09:29:21 -0700115SCALAR_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -0800116 "src/f32-argmaxpool/4x-scalar-c1.c",
Marat Dukhan1e782c42019-11-21 17:02:40 -0800117 "src/f32-argmaxpool/9p8x-scalar-c1.c",
Marat Dukhan329da642019-11-19 21:44:39 -0800118 "src/f32-argmaxpool/9x-scalar-c1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700119 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
120 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700121 "src/f32-clamp/gen/scalar-x1.c",
122 "src/f32-clamp/gen/scalar-x2.c",
123 "src/f32-clamp/gen/scalar-x4.c",
Marat Dukhan6b7dfae2019-12-04 16:00:52 -0800124 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700125 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700126 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
127 "src/f32-dwconv-chw/3x3p1-scalar.c",
128 "src/f32-dwconv-chw/3x3s2p1-scalar.c",
129 "src/f32-dwconv-chw/5x5p2-scalar.c",
130 "src/f32-dwconv-chw/5x5s2p2-scalar.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700131 "src/f32-dwconv/gen/up1x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700132 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700133 "src/f32-dwconv/gen/up1x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700134 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700135 "src/f32-dwconv/gen/up1x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700136 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700137 "src/f32-dwconv/gen/up2x4-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700138 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700139 "src/f32-dwconv/gen/up2x9-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700140 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700141 "src/f32-dwconv/gen/up2x25-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700142 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700143 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700144 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700145 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700146 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700147 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700148 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700149 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700150 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700151 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700152 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700153 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700154 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -0700155 "src/f32-gavgpool-cw/scalar-x1.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700156 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
157 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700158 "src/f32-gemm/gen/1x4-scalar.c",
159 "src/f32-gemm/gen/2x4-scalar.c",
160 "src/f32-gemm/gen/4x2-scalar.c",
161 "src/f32-gemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700162 "src/f32-gemm/gen/1x4-relu-scalar.c",
163 "src/f32-gemm/gen/2x4-relu-scalar.c",
164 "src/f32-gemm/gen/4x2-relu-scalar.c",
165 "src/f32-gemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700166 "src/f32-gemm/gen/1x4-minmax-scalar.c",
167 "src/f32-gemm/gen/2x4-minmax-scalar.c",
168 "src/f32-gemm/gen/4x2-minmax-scalar.c",
169 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700170 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
171 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
172 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700173 "src/f32-hswish/gen/hswish-scalar-x1.c",
174 "src/f32-hswish/gen/hswish-scalar-x2.c",
175 "src/f32-hswish/gen/hswish-scalar-x4.c",
Marat Dukhan660fd192020-03-10 04:55:30 -0700176 "src/f32-ibilinear/gen/scalar-c1.c",
177 "src/f32-ibilinear/gen/scalar-c2.c",
178 "src/f32-ibilinear/gen/scalar-c4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700179 "src/f32-igemm/gen/1x4-scalar.c",
180 "src/f32-igemm/gen/2x4-scalar.c",
181 "src/f32-igemm/gen/4x2-scalar.c",
182 "src/f32-igemm/gen/4x4-scalar.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700183 "src/f32-igemm/gen/1x4-relu-scalar.c",
184 "src/f32-igemm/gen/2x4-relu-scalar.c",
185 "src/f32-igemm/gen/4x2-relu-scalar.c",
186 "src/f32-igemm/gen/4x4-relu-scalar.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700187 "src/f32-igemm/gen/1x4-minmax-scalar.c",
188 "src/f32-igemm/gen/2x4-minmax-scalar.c",
189 "src/f32-igemm/gen/4x2-minmax-scalar.c",
190 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700191 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
192 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
193 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700194 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
195 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
196 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700197 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
198 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
199 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
200 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -0800201 "src/f32-prelu/gen/scalar-2x1.c",
202 "src/f32-prelu/gen/scalar-2x4.c",
Marat Dukhanf46f6752020-01-21 11:03:49 -0800203 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
204 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
205 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
206 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
207 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
208 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
209 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
210 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
211 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
212 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
213 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
214 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700215 "src/f32-rmax/scalar.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700216 "src/f32-relu/gen/scalar-x1.c",
217 "src/f32-relu/gen/scalar-x2.c",
218 "src/f32-relu/gen/scalar-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700219 "src/f32-relu/gen/scalar-x8.c",
Marat Dukhan3a77ea72019-12-23 12:10:24 -0800220 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
221 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
222 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
223 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
224 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
225 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
226 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
227 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
228 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Marat Dukhan355ab432020-04-09 19:01:52 -0700229 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
230 "src/f32-spmm/gen/1x1-minmax-scalar.c",
231 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
232 "src/f32-spmm/gen/2x1-minmax-scalar.c",
233 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
234 "src/f32-spmm/gen/4x1-minmax-scalar.c",
235 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
236 "src/f32-spmm/gen/8x1-minmax-scalar.c",
237 "src/f32-spmm/gen/8x2-minmax-scalar.c",
238 "src/f32-spmm/gen/8x4-minmax-scalar.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700239 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
240 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
241 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700242 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700243 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
244 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
245 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700246 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700247 "src/f32-vbinary/gen/vadd-scalar-x1.c",
248 "src/f32-vbinary/gen/vadd-scalar-x2.c",
249 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700250 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700251 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
252 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
253 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700254 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700255 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
256 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
257 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700258 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700259 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
260 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
261 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700262 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700263 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
264 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
265 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700266 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700267 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
268 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
269 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700270 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700271 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
272 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
273 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700274 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700275 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
276 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
277 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700278 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700279 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
280 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
281 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700282 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700283 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
284 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
285 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700286 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800287 "src/f32-vbinary/gen/vmax-scalar-x1.c",
288 "src/f32-vbinary/gen/vmax-scalar-x2.c",
289 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700290 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800291 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
292 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
293 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700294 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800295 "src/f32-vbinary/gen/vmin-scalar-x1.c",
296 "src/f32-vbinary/gen/vmin-scalar-x2.c",
297 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700298 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800299 "src/f32-vbinary/gen/vminc-scalar-x1.c",
300 "src/f32-vbinary/gen/vminc-scalar-x2.c",
301 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700302 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700303 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700306 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700307 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700310 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700311 "src/f32-vbinary/gen/vmul-scalar-x1.c",
312 "src/f32-vbinary/gen/vmul-scalar-x2.c",
313 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700314 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700315 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700318 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700319 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
320 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
321 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700322 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700323 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
324 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
325 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700326 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700327 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
328 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
329 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700330 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700331 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
332 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
333 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700334 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700335 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
336 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
337 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700338 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700339 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
340 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
341 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700342 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700343 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
344 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
345 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700346 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700347 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
348 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
349 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700350 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700351 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
352 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
353 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700354 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -0700355 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
356 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
357 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700358 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700359 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
360 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
361 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700362 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700363 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
364 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
365 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700366 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700367 "src/f32-vbinary/gen/vsub-scalar-x1.c",
368 "src/f32-vbinary/gen/vsub-scalar-x2.c",
369 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700370 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700371 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
372 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
373 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700374 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700375 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
376 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
377 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700378 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Frank Barchard8e229db2020-07-06 23:31:35 -0700379 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
380 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
381 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700382 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700383 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
384 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
385 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -0700386 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
387 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
388 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700389 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
390 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
391 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
392 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
393 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
394 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
395 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
396 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
397 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -0700398 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
399 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
400 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
Marat Dukhan5020b962020-06-08 13:30:10 -0700401 "src/f32-vunary/gen/vabs-scalar-x1.c",
402 "src/f32-vunary/gen/vabs-scalar-x2.c",
403 "src/f32-vunary/gen/vabs-scalar-x4.c",
404 "src/f32-vunary/gen/vneg-scalar-x1.c",
405 "src/f32-vunary/gen/vneg-scalar-x2.c",
406 "src/f32-vunary/gen/vneg-scalar-x4.c",
407 "src/f32-vunary/gen/vsqr-scalar-x1.c",
408 "src/f32-vunary/gen/vsqr-scalar-x2.c",
409 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800410 "src/math/expminus-scalar-lut2048-p1.c",
411 "src/math/expminus-scalar-lut64-p2.c",
412 "src/math/expminus-scalar-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -0700413 "src/math/roundne-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700414 "src/math/roundne-scalar-nearbyint.c",
415 "src/math/roundne-scalar-rint.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700416 "src/math/roundd-scalar-addsub.c",
417 "src/math/roundd-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700418 "src/math/roundd-scalar-floor.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700419 "src/math/roundu-scalar-addsub.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700420 "src/math/roundu-scalar-ceil.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -0700421 "src/math/roundu-scalar-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -0700422 "src/math/roundz-scalar-addsub.c",
423 "src/math/roundz-scalar-cvt.c",
Marat Dukhanffbf96a2020-05-14 02:59:08 -0700424 "src/math/roundz-scalar-trunc.c",
Marat Dukhan5739f702019-12-22 19:45:09 -0800425 "src/math/sigmoid-scalar-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-p5-div.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700428 "src/qs8-gemm/gen/8x8c4-minmax-scalar.c",
429 "src/qs8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -0700430 "src/qs8-requantization/fp32-scalar-lrintf.c",
431 "src/qs8-requantization/fp32-scalar-magic.c",
432 "src/qs8-requantization/precise-scalar-signed64.c",
433 "src/qs8-requantization/precise-scalar-unsigned32.c",
434 "src/qs8-requantization/precise-scalar-unsigned64.c",
435 "src/qs8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700436 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
437 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
438 "src/qu8-dwconv/up1x9-minmax-scalar.c",
439 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
440 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700441 "src/qu8-gemm/gen/8x8c4-minmax-scalar.c",
442 "src/qu8-gemm/gen/12x4c4-minmax-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700443 "src/qu8-gemm/2x2-minmax-scalar.c",
444 "src/qu8-igemm/2x2-minmax-scalar.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -0700445 "src/qu8-requantization/fp32-scalar-lrintf.c",
446 "src/qu8-requantization/fp32-scalar-magic.c",
447 "src/qu8-requantization/precise-scalar-signed64.c",
448 "src/qu8-requantization/precise-scalar-unsigned32.c",
449 "src/qu8-requantization/precise-scalar-unsigned64.c",
450 "src/qu8-requantization/q31-scalar.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -0700451 "src/qu8-vadd/minmax-scalar.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700452 "src/u8-clamp/scalar-x4.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700453 "src/u8-lut32norm/scalar.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700454 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700455 "src/u8-rmax/scalar.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -0700456 "src/x32-fill/scalar-float.c",
457 "src/x32-fill/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700458 "src/x32-packx/x2-scalar.c",
459 "src/x32-packx/x3-scalar.c",
460 "src/x32-packx/x4-scalar.c",
Marat Dukhan63523d42020-05-22 17:07:33 -0700461 "src/x32-pad/scalar-float.c",
Frank Barchardc9c320e2020-08-07 22:12:46 -0700462 "src/x32-pad/scalar-int.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -0700463 "src/x32-unpool/scalar.c",
464 "src/x32-zip/x2-scalar.c",
465 "src/x32-zip/x3-scalar.c",
466 "src/x32-zip/x4-scalar.c",
467 "src/x32-zip/xm-scalar.c",
468 "src/x8-lut/scalar.c",
469 "src/x8-zip/x2-scalar.c",
470 "src/x8-zip/x3-scalar.c",
471 "src/x8-zip/x4-scalar.c",
472 "src/x8-zip/xm-scalar.c",
473]
474
Marat Dukhan436ebe62019-12-04 15:10:12 -0800475WASM_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -0700476 "src/f32-avgpool/9p8x-minmax-wasm-c1.c",
477 "src/f32-avgpool/9x-minmax-wasm-c1.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -0700478 "src/f32-clamp/gen/wasm-x1.c",
479 "src/f32-clamp/gen/wasm-x2.c",
480 "src/f32-clamp/gen/wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700481 "src/f32-dwconv/gen/up1x4-wasm-acc2.c",
482 "src/f32-dwconv/gen/up1x4-wasm.c",
483 "src/f32-dwconv/gen/up1x9-wasm-acc2.c",
484 "src/f32-dwconv/gen/up1x9-wasm.c",
485 "src/f32-dwconv/gen/up1x25-wasm-acc2.c",
486 "src/f32-dwconv/gen/up1x25-wasm.c",
487 "src/f32-dwconv/gen/up2x4-wasm-acc2.c",
488 "src/f32-dwconv/gen/up2x4-wasm.c",
489 "src/f32-dwconv/gen/up2x9-wasm-acc2.c",
490 "src/f32-dwconv/gen/up2x9-wasm.c",
491 "src/f32-dwconv/gen/up2x25-wasm-acc2.c",
492 "src/f32-dwconv/gen/up2x25-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700493 "src/f32-dwconv/gen/up1x4-minmax-wasm-acc2.c",
494 "src/f32-dwconv/gen/up1x4-minmax-wasm.c",
495 "src/f32-dwconv/gen/up1x9-minmax-wasm-acc2.c",
496 "src/f32-dwconv/gen/up1x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700497 "src/f32-dwconv/gen/up1x25-minmax-wasm-acc2.c",
498 "src/f32-dwconv/gen/up1x25-minmax-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700499 "src/f32-dwconv/gen/up2x4-minmax-wasm-acc2.c",
500 "src/f32-dwconv/gen/up2x4-minmax-wasm.c",
501 "src/f32-dwconv/gen/up2x9-minmax-wasm-acc2.c",
502 "src/f32-dwconv/gen/up2x9-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700503 "src/f32-dwconv/gen/up2x25-minmax-wasm-acc2.c",
504 "src/f32-dwconv/gen/up2x25-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700505 "src/f32-gavgpool/7p7x-minmax-wasm-c1.c",
506 "src/f32-gavgpool/7x-minmax-wasm-c1.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700507 "src/f32-gemm/gen-inc/1x4inc-minmax-wasm.c",
508 "src/f32-gemm/gen-inc/2x4inc-minmax-wasm.c",
509 "src/f32-gemm/gen-inc/4x4inc-minmax-wasm.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700510 "src/f32-gemm/gen/1x4-wasm.c",
511 "src/f32-gemm/gen/2x4-wasm.c",
512 "src/f32-gemm/gen/4x2-wasm.c",
513 "src/f32-gemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700514 "src/f32-gemm/gen/1x4-relu-wasm.c",
515 "src/f32-gemm/gen/2x4-relu-wasm.c",
516 "src/f32-gemm/gen/4x2-relu-wasm.c",
517 "src/f32-gemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700518 "src/f32-gemm/gen/1x4-minmax-wasm.c",
519 "src/f32-gemm/gen/2x4-minmax-wasm.c",
520 "src/f32-gemm/gen/4x2-minmax-wasm.c",
521 "src/f32-gemm/gen/4x4-minmax-wasm.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700522 "src/f32-hswish/gen/hswish-wasm-x1.c",
523 "src/f32-hswish/gen/hswish-wasm-x2.c",
524 "src/f32-hswish/gen/hswish-wasm-x4.c",
Marat Dukhan163a7e62020-04-09 04:19:26 -0700525 "src/f32-igemm/gen/1x4-wasm.c",
526 "src/f32-igemm/gen/2x4-wasm.c",
527 "src/f32-igemm/gen/4x2-wasm.c",
528 "src/f32-igemm/gen/4x4-wasm.c",
Marat Dukhan467f6362020-05-22 23:21:55 -0700529 "src/f32-igemm/gen/1x4-relu-wasm.c",
530 "src/f32-igemm/gen/2x4-relu-wasm.c",
531 "src/f32-igemm/gen/4x2-relu-wasm.c",
532 "src/f32-igemm/gen/4x4-relu-wasm.c",
Marat Dukhan1c587112020-04-08 20:04:28 -0700533 "src/f32-igemm/gen/1x4-minmax-wasm.c",
534 "src/f32-igemm/gen/2x4-minmax-wasm.c",
535 "src/f32-igemm/gen/4x2-minmax-wasm.c",
536 "src/f32-igemm/gen/4x4-minmax-wasm.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700537 "src/f32-maxpool/9p8x-minmax-wasm-c1.c",
538 "src/f32-pavgpool/9p8x-minmax-wasm-c1.c",
539 "src/f32-pavgpool/9x-minmax-wasm-c1.c",
Marat Dukhan7c1f8082020-06-25 13:26:20 -0700540 "src/f32-prelu/gen/wasm-2x1.c",
541 "src/f32-prelu/gen/wasm-2x4.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700542 "src/f32-relu/gen/wasm-x1.c",
543 "src/f32-relu/gen/wasm-x2.c",
544 "src/f32-relu/gen/wasm-x4.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700545 "src/f32-relu/gen/wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700546 "src/f32-vbinary/gen/vadd-minmax-wasm-x1.c",
547 "src/f32-vbinary/gen/vadd-minmax-wasm-x2.c",
548 "src/f32-vbinary/gen/vadd-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700549 "src/f32-vbinary/gen/vadd-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700550 "src/f32-vbinary/gen/vaddc-minmax-wasm-x1.c",
551 "src/f32-vbinary/gen/vaddc-minmax-wasm-x2.c",
552 "src/f32-vbinary/gen/vaddc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700553 "src/f32-vbinary/gen/vaddc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700554 "src/f32-vbinary/gen/vdiv-minmax-wasm-x1.c",
555 "src/f32-vbinary/gen/vdiv-minmax-wasm-x2.c",
556 "src/f32-vbinary/gen/vdiv-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700557 "src/f32-vbinary/gen/vdiv-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700558 "src/f32-vbinary/gen/vdivc-minmax-wasm-x1.c",
559 "src/f32-vbinary/gen/vdivc-minmax-wasm-x2.c",
560 "src/f32-vbinary/gen/vdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700561 "src/f32-vbinary/gen/vdivc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700562 "src/f32-vbinary/gen/vadd-relu-wasm-x1.c",
563 "src/f32-vbinary/gen/vadd-relu-wasm-x2.c",
564 "src/f32-vbinary/gen/vadd-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700565 "src/f32-vbinary/gen/vadd-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700566 "src/f32-vbinary/gen/vaddc-relu-wasm-x1.c",
567 "src/f32-vbinary/gen/vaddc-relu-wasm-x2.c",
568 "src/f32-vbinary/gen/vaddc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700569 "src/f32-vbinary/gen/vaddc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700570 "src/f32-vbinary/gen/vdiv-relu-wasm-x1.c",
571 "src/f32-vbinary/gen/vdiv-relu-wasm-x2.c",
572 "src/f32-vbinary/gen/vdiv-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700573 "src/f32-vbinary/gen/vdiv-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700574 "src/f32-vbinary/gen/vdivc-relu-wasm-x1.c",
575 "src/f32-vbinary/gen/vdivc-relu-wasm-x2.c",
576 "src/f32-vbinary/gen/vdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700577 "src/f32-vbinary/gen/vdivc-relu-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800578 "src/f32-vbinary/gen/vmax-wasm-x1.c",
579 "src/f32-vbinary/gen/vmax-wasm-x2.c",
580 "src/f32-vbinary/gen/vmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700581 "src/f32-vbinary/gen/vmax-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800582 "src/f32-vbinary/gen/vmaxc-wasm-x1.c",
583 "src/f32-vbinary/gen/vmaxc-wasm-x2.c",
584 "src/f32-vbinary/gen/vmaxc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700585 "src/f32-vbinary/gen/vmaxc-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800586 "src/f32-vbinary/gen/vmin-wasm-x1.c",
587 "src/f32-vbinary/gen/vmin-wasm-x2.c",
588 "src/f32-vbinary/gen/vmin-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700589 "src/f32-vbinary/gen/vmin-wasm-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -0800590 "src/f32-vbinary/gen/vminc-wasm-x1.c",
591 "src/f32-vbinary/gen/vminc-wasm-x2.c",
592 "src/f32-vbinary/gen/vminc-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700593 "src/f32-vbinary/gen/vminc-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700594 "src/f32-vbinary/gen/vmul-minmax-wasm-x1.c",
595 "src/f32-vbinary/gen/vmul-minmax-wasm-x2.c",
596 "src/f32-vbinary/gen/vmul-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700597 "src/f32-vbinary/gen/vmul-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700598 "src/f32-vbinary/gen/vmulc-minmax-wasm-x1.c",
599 "src/f32-vbinary/gen/vmulc-minmax-wasm-x2.c",
600 "src/f32-vbinary/gen/vmulc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700601 "src/f32-vbinary/gen/vmulc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700602 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x1.c",
603 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x2.c",
604 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700605 "src/f32-vbinary/gen/vrdivc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700606 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x1.c",
607 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x2.c",
608 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700609 "src/f32-vbinary/gen/vrsubc-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700610 "src/f32-vbinary/gen/vsub-minmax-wasm-x1.c",
611 "src/f32-vbinary/gen/vsub-minmax-wasm-x2.c",
612 "src/f32-vbinary/gen/vsub-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700613 "src/f32-vbinary/gen/vsub-minmax-wasm-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -0700614 "src/f32-vbinary/gen/vsubc-minmax-wasm-x1.c",
615 "src/f32-vbinary/gen/vsubc-minmax-wasm-x2.c",
616 "src/f32-vbinary/gen/vsubc-minmax-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700617 "src/f32-vbinary/gen/vsubc-minmax-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700618 "src/f32-vbinary/gen/vmul-relu-wasm-x1.c",
619 "src/f32-vbinary/gen/vmul-relu-wasm-x2.c",
620 "src/f32-vbinary/gen/vmul-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700621 "src/f32-vbinary/gen/vmul-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700622 "src/f32-vbinary/gen/vmulc-relu-wasm-x1.c",
623 "src/f32-vbinary/gen/vmulc-relu-wasm-x2.c",
624 "src/f32-vbinary/gen/vmulc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700625 "src/f32-vbinary/gen/vmulc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700626 "src/f32-vbinary/gen/vrdivc-relu-wasm-x1.c",
627 "src/f32-vbinary/gen/vrdivc-relu-wasm-x2.c",
628 "src/f32-vbinary/gen/vrdivc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700629 "src/f32-vbinary/gen/vrdivc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700630 "src/f32-vbinary/gen/vrsubc-relu-wasm-x1.c",
631 "src/f32-vbinary/gen/vrsubc-relu-wasm-x2.c",
632 "src/f32-vbinary/gen/vrsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700633 "src/f32-vbinary/gen/vrsubc-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700634 "src/f32-vbinary/gen/vsub-relu-wasm-x1.c",
635 "src/f32-vbinary/gen/vsub-relu-wasm-x2.c",
636 "src/f32-vbinary/gen/vsub-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700637 "src/f32-vbinary/gen/vsub-relu-wasm-x8.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700638 "src/f32-vbinary/gen/vsubc-relu-wasm-x1.c",
639 "src/f32-vbinary/gen/vsubc-relu-wasm-x2.c",
640 "src/f32-vbinary/gen/vsubc-relu-wasm-x4.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700641 "src/f32-vbinary/gen/vsubc-relu-wasm-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -0700642 "src/f32-vlrelu/gen/vlrelu-wasm-x1.c",
643 "src/f32-vlrelu/gen/vlrelu-wasm-x2.c",
644 "src/f32-vlrelu/gen/vlrelu-wasm-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -0700645 "src/f32-vmulcaddc/gen/c1-minmax-wasm-2x.c",
646 "src/f32-vmulcaddc/gen/c2-minmax-wasm-2x.c",
647 "src/f32-vmulcaddc/gen/c4-minmax-wasm-2x.c",
Marat Dukhan436ebe62019-12-04 15:10:12 -0800648]
649
Marat Dukhan290055c2020-06-09 12:24:29 -0700650WASMSIMD_UKERNELS = [
Marat Dukhan40f05522020-07-16 22:33:12 -0700651 "src/f32-argmaxpool/4x-wasmsimd-c4.c",
652 "src/f32-argmaxpool/9p8x-wasmsimd-c4.c",
653 "src/f32-argmaxpool/9x-wasmsimd-c4.c",
Marat Dukhan3b7432d2020-07-16 17:46:32 -0700654 "src/f32-avgpool/9p8x-minmax-wasmsimd-arm-c4.c",
655 "src/f32-avgpool/9p8x-minmax-wasmsimd-x86-c4.c",
656 "src/f32-avgpool/9x-minmax-wasmsimd-arm-c4.c",
657 "src/f32-avgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan3fa52c82020-07-08 12:54:33 -0700658 "src/f32-clamp/gen/wasmsimd-arm-x4.c",
659 "src/f32-clamp/gen/wasmsimd-arm-x8.c",
660 "src/f32-clamp/gen/wasmsimd-x86-x4.c",
661 "src/f32-clamp/gen/wasmsimd-x86-x8.c",
Marat Dukhanb8e7b072020-06-16 12:34:23 -0700662 "src/f32-dwconv/gen/up4x25-wasmsimd.c",
663 "src/f32-dwconv/gen/up4x4-wasmsimd.c",
664 "src/f32-dwconv/gen/up4x9-wasmsimd.c",
665 "src/f32-dwconv/gen/up8x25-wasmsimd.c",
666 "src/f32-dwconv/gen/up8x4-wasmsimd.c",
667 "src/f32-dwconv/gen/up8x9-wasmsimd.c",
Marat Dukhanac014d72020-06-16 08:36:47 -0700668 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-arm.c",
669 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-arm.c",
670 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-arm.c",
671 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-arm.c",
672 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-arm.c",
673 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-arm.c",
674 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-arm.c",
675 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-arm.c",
676 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-arm.c",
677 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-arm.c",
678 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-arm.c",
679 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-arm.c",
680 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-acc2-x86.c",
681 "src/f32-dwconv/gen/up4x25-minmax-wasmsimd-x86.c",
682 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-acc2-x86.c",
683 "src/f32-dwconv/gen/up4x4-minmax-wasmsimd-x86.c",
684 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-acc2-x86.c",
685 "src/f32-dwconv/gen/up4x9-minmax-wasmsimd-x86.c",
686 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-acc2-x86.c",
687 "src/f32-dwconv/gen/up8x25-minmax-wasmsimd-x86.c",
688 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-acc2-x86.c",
689 "src/f32-dwconv/gen/up8x4-minmax-wasmsimd-x86.c",
690 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-acc2-x86.c",
691 "src/f32-dwconv/gen/up8x9-minmax-wasmsimd-x86.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700692 "src/f32-gemm/gen/1x8-wasmsimd-splat.c",
693 "src/f32-gemm/gen/4x8-wasmsimd-splat.c",
694 "src/f32-gemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700695 "src/f32-gemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700696 "src/f32-gemm/gen/1x8-relu-wasmsimd-splat.c",
697 "src/f32-gemm/gen/4x8-relu-wasmsimd-splat.c",
698 "src/f32-gemm/gen/5x8-relu-wasmsimd-splat.c",
699 "src/f32-gemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700700 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
701 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
702 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-arm.c",
703 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
704 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
705 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-arm.c",
706 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
707 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
708 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-arm.c",
709 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
710 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
711 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-arm.c",
712 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
713 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
714 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-arm.c",
715 "src/f32-gemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
716 "src/f32-gemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
717 "src/f32-gemm/gen/1x8s4-minmax-wasmsimd-x86.c",
718 "src/f32-gemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
719 "src/f32-gemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
720 "src/f32-gemm/gen/3x8s4-minmax-wasmsimd-x86.c",
721 "src/f32-gemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
722 "src/f32-gemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
723 "src/f32-gemm/gen/4x8s4-minmax-wasmsimd-x86.c",
724 "src/f32-gemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
725 "src/f32-gemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
726 "src/f32-gemm/gen/5x8s4-minmax-wasmsimd-x86.c",
727 "src/f32-gemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
728 "src/f32-gemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
729 "src/f32-gemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700730 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-arm.c",
731 "src/f32-gemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700732 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-arm.c",
733 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-arm.c",
734 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-arm.c",
735 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-arm.c",
736 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-arm.c",
737 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-arm.c",
738 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-arm.c",
739 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-arm.c",
740 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-arm.c",
741 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-arm.c",
742 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-arm.c",
743 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-arm.c",
744 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-arm.c",
745 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-arm.c",
746 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-arm.c",
747 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-loadsplat-x86.c",
748 "src/f32-gemm/gen-inc/1x8inc-minmax-wasmsimd-splat-x86.c",
749 "src/f32-gemm/gen-inc/1x8s4inc-minmax-wasmsimd-x86.c",
750 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-loadsplat-x86.c",
751 "src/f32-gemm/gen-inc/3x8inc-minmax-wasmsimd-splat-x86.c",
752 "src/f32-gemm/gen-inc/3x8s4inc-minmax-wasmsimd-x86.c",
753 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-loadsplat-x86.c",
754 "src/f32-gemm/gen-inc/4x8inc-minmax-wasmsimd-splat-x86.c",
755 "src/f32-gemm/gen-inc/4x8s4inc-minmax-wasmsimd-x86.c",
756 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-loadsplat-x86.c",
757 "src/f32-gemm/gen-inc/5x8inc-minmax-wasmsimd-splat-x86.c",
758 "src/f32-gemm/gen-inc/5x8s4inc-minmax-wasmsimd-x86.c",
759 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-loadsplat-x86.c",
760 "src/f32-gemm/gen-inc/6x8inc-minmax-wasmsimd-splat-x86.c",
761 "src/f32-gemm/gen-inc/6x8s4inc-minmax-wasmsimd-x86.c",
Marat Dukhan00d1d6e2020-07-09 01:37:27 -0700762 "src/f32-ibilinear/gen/wasmsimd-c4.c",
763 "src/f32-ibilinear/gen/wasmsimd-c8.c",
Marat Dukhan802808c2020-06-16 11:01:17 -0700764 "src/f32-igemm/gen/1x8-wasmsimd-splat.c",
765 "src/f32-igemm/gen/4x8-wasmsimd-splat.c",
766 "src/f32-igemm/gen/5x8-wasmsimd-splat.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700767 "src/f32-igemm/gen/4x2c4-wasmsimd.c",
Marat Dukhan688f6d82020-07-14 17:02:11 -0700768 "src/f32-igemm/gen/1x8-relu-wasmsimd-splat.c",
769 "src/f32-igemm/gen/4x8-relu-wasmsimd-splat.c",
770 "src/f32-igemm/gen/5x8-relu-wasmsimd-splat.c",
771 "src/f32-igemm/gen/4x2c4-relu-wasmsimd.c",
Marat Dukhan1bbf96b2020-06-15 23:01:20 -0700772 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-arm.c",
773 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-arm.c",
774 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-arm.c",
775 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-arm.c",
776 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-arm.c",
777 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-arm.c",
778 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-arm.c",
779 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-arm.c",
780 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-arm.c",
781 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-arm.c",
782 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-arm.c",
783 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-arm.c",
784 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-arm.c",
785 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-arm.c",
786 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-arm.c",
787 "src/f32-igemm/gen/1x8-minmax-wasmsimd-loadsplat-x86.c",
788 "src/f32-igemm/gen/1x8-minmax-wasmsimd-splat-x86.c",
789 "src/f32-igemm/gen/1x8s4-minmax-wasmsimd-x86.c",
790 "src/f32-igemm/gen/3x8-minmax-wasmsimd-loadsplat-x86.c",
791 "src/f32-igemm/gen/3x8-minmax-wasmsimd-splat-x86.c",
792 "src/f32-igemm/gen/3x8s4-minmax-wasmsimd-x86.c",
793 "src/f32-igemm/gen/4x8-minmax-wasmsimd-loadsplat-x86.c",
794 "src/f32-igemm/gen/4x8-minmax-wasmsimd-splat-x86.c",
795 "src/f32-igemm/gen/4x8s4-minmax-wasmsimd-x86.c",
796 "src/f32-igemm/gen/5x8-minmax-wasmsimd-loadsplat-x86.c",
797 "src/f32-igemm/gen/5x8-minmax-wasmsimd-splat-x86.c",
798 "src/f32-igemm/gen/5x8s4-minmax-wasmsimd-x86.c",
799 "src/f32-igemm/gen/6x8-minmax-wasmsimd-loadsplat-x86.c",
800 "src/f32-igemm/gen/6x8-minmax-wasmsimd-splat-x86.c",
801 "src/f32-igemm/gen/6x8s4-minmax-wasmsimd-x86.c",
Marat Dukhane39e6462020-07-09 01:33:36 -0700802 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-arm.c",
803 "src/f32-igemm/gen/4x2c4-minmax-wasmsimd-x86.c",
Marat Dukhanc6016802020-07-16 18:51:28 -0700804 "src/f32-gavgpool/7p7x-minmax-wasmsimd-arm-c4.c",
805 "src/f32-gavgpool/7p7x-minmax-wasmsimd-x86-c4.c",
806 "src/f32-gavgpool/7x-minmax-wasmsimd-arm-c4.c",
807 "src/f32-gavgpool/7x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan18493512020-07-24 14:13:48 -0700808 "src/f32-gavgpool-cw/wasmsimd-arm-x4.c",
809 "src/f32-gavgpool-cw/wasmsimd-x86-x4.c",
Marat Dukhanefaac272020-07-16 16:58:07 -0700810 "src/f32-hswish/gen/hswish-wasmsimd-x4.c",
811 "src/f32-hswish/gen/hswish-wasmsimd-x8.c",
812 "src/f32-hswish/gen/hswish-wasmsimd-x16.c",
Marat Dukhanf6e24802020-07-08 22:20:40 -0700813 "src/f32-maxpool/9p8x-minmax-wasmsimd-arm-c4.c",
814 "src/f32-maxpool/9p8x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan1483c532020-07-16 18:08:19 -0700815 "src/f32-pavgpool/9p8x-minmax-wasmsimd-arm-c4.c",
816 "src/f32-pavgpool/9p8x-minmax-wasmsimd-x86-c4.c",
817 "src/f32-pavgpool/9x-minmax-wasmsimd-arm-c4.c",
818 "src/f32-pavgpool/9x-minmax-wasmsimd-x86-c4.c",
Marat Dukhan0d0d8822020-07-23 23:37:56 -0700819 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-arm.c",
820 "src/f32-ppmm/gen/4x8-minmax-wasmsimd-splat-x86.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700821 "src/f32-prelu/gen/wasmsimd-bitselect-1x4.c",
822 "src/f32-prelu/gen/wasmsimd-bitselect-1x8.c",
823 "src/f32-prelu/gen/wasmsimd-bitselect-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700824 "src/f32-prelu/gen/wasmsimd-bitselect-2x4.c",
825 "src/f32-prelu/gen/wasmsimd-bitselect-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700826 "src/f32-prelu/gen/wasmsimd-bitselect-2x16.c",
827 "src/f32-prelu/gen/wasmsimd-bitselect-4x4.c",
828 "src/f32-prelu/gen/wasmsimd-bitselect-4x8.c",
829 "src/f32-prelu/gen/wasmsimd-bitselect-4x16.c",
830 "src/f32-prelu/gen/wasmsimd-minmax-1x4.c",
831 "src/f32-prelu/gen/wasmsimd-minmax-1x8.c",
832 "src/f32-prelu/gen/wasmsimd-minmax-1x16.c",
Marat Dukhan195f8eb2020-06-25 12:50:57 -0700833 "src/f32-prelu/gen/wasmsimd-minmax-2x4.c",
834 "src/f32-prelu/gen/wasmsimd-minmax-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -0700835 "src/f32-prelu/gen/wasmsimd-minmax-2x16.c",
836 "src/f32-prelu/gen/wasmsimd-minmax-4x4.c",
837 "src/f32-prelu/gen/wasmsimd-minmax-4x8.c",
838 "src/f32-prelu/gen/wasmsimd-minmax-4x16.c",
Marat Dukhan52238f02020-07-16 15:30:28 -0700839 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x4.c",
840 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8.c",
841 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x8-acc2.c",
842 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12.c",
843 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc2.c",
844 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x12-acc3.c",
845 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16.c",
846 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc2.c",
847 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x16-acc4.c",
848 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20.c",
849 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc2.c",
850 "src/f32-raddstoreexpminusmax/gen/wasmsimd-p5-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -0700851 "src/f32-relu/gen/wasmsimd-x4.c",
852 "src/f32-relu/gen/wasmsimd-x8.c",
Frank Barchard37297a62020-08-27 12:15:10 -0700853 "src/f32-relu/gen/wasmsimd-x16.c",
Marat Dukhan8c417962020-07-08 12:27:50 -0700854 "src/f32-rmax/wasmsimd-arm.c",
855 "src/f32-rmax/wasmsimd-x86.c",
Marat Dukhanb3635ed2020-07-16 12:36:28 -0700856 "src/f32-sigmoid/gen/wasmsimd-p5-div-x4.c",
857 "src/f32-sigmoid/gen/wasmsimd-p5-div-x8.c",
858 "src/f32-sigmoid/gen/wasmsimd-p5-div-x12.c",
859 "src/f32-sigmoid/gen/wasmsimd-p5-div-x16.c",
860 "src/f32-sigmoid/gen/wasmsimd-p5-div-x20.c",
861 "src/f32-sigmoid/gen/wasmsimd-p5-div-x24.c",
Marat Dukhand187a5b2020-07-20 01:07:17 -0700862 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x4.c",
863 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x8.c",
864 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x12.c",
865 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x16.c",
866 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x20.c",
867 "src/f32-sigmoid/gen/wasmsimd-lut64-p2-div-x24.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700868 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x4.c",
869 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700870 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700871 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x4.c",
872 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700873 "src/f32-vbinary/gen/vadd-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700874 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x4.c",
875 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700876 "src/f32-vbinary/gen/vadd-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700877 "src/f32-vbinary/gen/vadd-wasmsimd-x4.c",
878 "src/f32-vbinary/gen/vadd-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700879 "src/f32-vbinary/gen/vadd-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700880 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x4.c",
881 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700882 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700883 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x4.c",
884 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700885 "src/f32-vbinary/gen/vaddc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700886 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x4.c",
887 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700888 "src/f32-vbinary/gen/vaddc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700889 "src/f32-vbinary/gen/vaddc-wasmsimd-x4.c",
890 "src/f32-vbinary/gen/vaddc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700891 "src/f32-vbinary/gen/vaddc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700892 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x4.c",
893 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700894 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700895 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x4.c",
896 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700897 "src/f32-vbinary/gen/vdiv-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700898 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x4.c",
899 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700900 "src/f32-vbinary/gen/vdiv-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700901 "src/f32-vbinary/gen/vdiv-wasmsimd-x4.c",
902 "src/f32-vbinary/gen/vdiv-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700903 "src/f32-vbinary/gen/vdiv-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700904 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x4.c",
905 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700906 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700907 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x4.c",
908 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700909 "src/f32-vbinary/gen/vdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700910 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x4.c",
911 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700912 "src/f32-vbinary/gen/vdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700913 "src/f32-vbinary/gen/vdivc-wasmsimd-x4.c",
914 "src/f32-vbinary/gen/vdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700915 "src/f32-vbinary/gen/vdivc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700916 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x4.c",
917 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700918 "src/f32-vbinary/gen/vmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700919 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x4.c",
920 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700921 "src/f32-vbinary/gen/vmax-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700922 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x4.c",
923 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700924 "src/f32-vbinary/gen/vmaxc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700925 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x4.c",
926 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700927 "src/f32-vbinary/gen/vmaxc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700928 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x4.c",
929 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700930 "src/f32-vbinary/gen/vmin-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700931 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x4.c",
932 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700933 "src/f32-vbinary/gen/vmin-wasmsimd-x86-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700934 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x4.c",
935 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700936 "src/f32-vbinary/gen/vminc-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700937 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x4.c",
938 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700939 "src/f32-vbinary/gen/vminc-wasmsimd-x86-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700940 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x4.c",
941 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700942 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700943 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x4.c",
944 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700945 "src/f32-vbinary/gen/vmul-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700946 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x4.c",
947 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700948 "src/f32-vbinary/gen/vmul-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700949 "src/f32-vbinary/gen/vmul-wasmsimd-x4.c",
950 "src/f32-vbinary/gen/vmul-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700951 "src/f32-vbinary/gen/vmul-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700952 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x4.c",
953 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700954 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700955 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x4.c",
956 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700957 "src/f32-vbinary/gen/vmulc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700958 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x4.c",
959 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700960 "src/f32-vbinary/gen/vmulc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700961 "src/f32-vbinary/gen/vmulc-wasmsimd-x4.c",
962 "src/f32-vbinary/gen/vmulc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700963 "src/f32-vbinary/gen/vmulc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700964 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x4.c",
965 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700966 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700967 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x4.c",
968 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700969 "src/f32-vbinary/gen/vrdivc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700970 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x4.c",
971 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700972 "src/f32-vbinary/gen/vrdivc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700973 "src/f32-vbinary/gen/vrdivc-wasmsimd-x4.c",
974 "src/f32-vbinary/gen/vrdivc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700975 "src/f32-vbinary/gen/vrdivc-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700976 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x4.c",
977 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700978 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700979 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x4.c",
980 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700981 "src/f32-vbinary/gen/vrsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -0700982 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x4.c",
983 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700984 "src/f32-vbinary/gen/vrsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -0700985 "src/f32-vbinary/gen/vrsubc-wasmsimd-x4.c",
986 "src/f32-vbinary/gen/vrsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700987 "src/f32-vbinary/gen/vrsubc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700988 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x4.c",
989 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700990 "src/f32-vbinary/gen/vsqrdiff-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -0700991 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x4.c",
992 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700993 "src/f32-vbinary/gen/vsqrdiffc-wasmsimd-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700994 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x4.c",
995 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700996 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-arm-x16.c",
Marat Dukhan93d1ba12020-06-26 12:33:35 -0700997 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x4.c",
998 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -0700999 "src/f32-vbinary/gen/vsub-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001000 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x4.c",
1001 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001002 "src/f32-vbinary/gen/vsub-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001003 "src/f32-vbinary/gen/vsub-wasmsimd-x4.c",
1004 "src/f32-vbinary/gen/vsub-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001005 "src/f32-vbinary/gen/vsub-wasmsimd-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001006 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x4.c",
1007 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001008 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-arm-x16.c",
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07001009 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x4.c",
1010 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001011 "src/f32-vbinary/gen/vsubc-minmax-wasmsimd-x86-x16.c",
Frank Barchard674778d2020-08-08 10:17:25 -07001012 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x4.c",
1013 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001014 "src/f32-vbinary/gen/vsubc-relu-wasmsimd-x16.c",
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07001015 "src/f32-vbinary/gen/vsubc-wasmsimd-x4.c",
1016 "src/f32-vbinary/gen/vsubc-wasmsimd-x8.c",
Frank Barchard9c7308f2020-08-31 17:03:01 -07001017 "src/f32-vbinary/gen/vsubc-wasmsimd-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001018 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x4.c",
1019 "src/f32-vlrelu/gen/vlrelu-wasmsimd-bitselect-x8.c",
1020 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x4.c",
1021 "src/f32-vlrelu/gen/vlrelu-wasmsimd-minmax-x8.c",
Marat Dukhand816f622020-07-15 10:14:39 -07001022 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-arm-2x.c",
1023 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-arm-2x.c",
1024 "src/f32-vmulcaddc/gen/c4-minmax-wasmsimd-x86-2x.c",
1025 "src/f32-vmulcaddc/gen/c8-minmax-wasmsimd-x86-2x.c",
Marat Dukhanb82b2cd2020-07-16 02:23:42 -07001026 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x4.c",
1027 "src/f32-vrnd/gen/vrndne-wasmsimd-addsub-x8.c",
1028 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x4.c",
1029 "src/f32-vrnd/gen/vrndz-wasmsimd-addsub-x8.c",
1030 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x4.c",
1031 "src/f32-vrnd/gen/vrndz-wasmsimd-cvt-x8.c",
1032 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x4.c",
1033 "src/f32-vrnd/gen/vrndu-wasmsimd-addsub-x8.c",
1034 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x4.c",
1035 "src/f32-vrnd/gen/vrndu-wasmsimd-cvt-x8.c",
1036 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x4.c",
1037 "src/f32-vrnd/gen/vrndd-wasmsimd-addsub-x8.c",
1038 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x4.c",
1039 "src/f32-vrnd/gen/vrndd-wasmsimd-cvt-x8.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001040 "src/f32-vsqrt/gen/wasmsimd-sqrt-x4.c",
1041 "src/f32-vsqrt/gen/wasmsimd-sqrt-x8.c",
Marat Dukhan37c83512020-06-29 13:25:53 -07001042 "src/f32-vunary/gen/vabs-wasmsimd-x4.c",
1043 "src/f32-vunary/gen/vabs-wasmsimd-x8.c",
1044 "src/f32-vunary/gen/vneg-wasmsimd-x4.c",
1045 "src/f32-vunary/gen/vneg-wasmsimd-x8.c",
1046 "src/f32-vunary/gen/vsqr-wasmsimd-x4.c",
1047 "src/f32-vunary/gen/vsqr-wasmsimd-x8.c",
Marat Dukhancc8f34c2020-08-05 16:36:38 -07001048 "src/qs8-dwconv/gen/up8x9-minmax-wasmsimd-mul16.c",
1049 "src/qs8-dwconv/gen/up16x9-minmax-wasmsimd-mul16.c",
1050 "src/qs8-dwconv/gen/up24x9-minmax-wasmsimd-mul16.c",
Marat Dukhanef451802020-08-06 11:53:47 -07001051 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
1052 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
1053 "src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhanb5e3d172020-08-06 13:29:53 -07001054 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
1055 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
1056 "src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
Marat Dukhan27203da2020-08-05 15:19:03 -07001057 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1058 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1059 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1060 "src/qs8-gemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1061 "src/qs8-gemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1062 "src/qs8-gemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
1063 "src/qs8-gemm/gen/1x4c8-xw-minmax-wasmsimd.c",
1064 "src/qs8-gemm/gen/2x4c8-xw-minmax-wasmsimd.c",
1065 "src/qs8-gemm/gen/3x4c8-xw-minmax-wasmsimd.c",
1066 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld64.c",
1067 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld64.c",
1068 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld64.c",
1069 "src/qs8-igemm/gen/1x4c8-minmax-wasmsimd-ld128.c",
1070 "src/qs8-igemm/gen/2x4c8-minmax-wasmsimd-ld128.c",
1071 "src/qs8-igemm/gen/3x4c8-minmax-wasmsimd-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001072 "src/qs8-requantization/fp32-wasmsimd.c",
Marat Dukhan138560c2020-08-03 18:57:34 -07001073 "src/qs8-requantization/q31-wasmsimd.c",
Marat Dukhan5df27f82020-09-02 23:59:21 -07001074 "src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
1075 "src/qs8-vadd/gen/minmax-wasmsimd-x16.c",
1076 "src/qs8-vadd/gen/minmax-wasmsimd-x24.c",
1077 "src/qs8-vadd/gen/minmax-wasmsimd-x32.c",
1078 "src/qs8-vaddc/gen/minmax-wasmsimd-x8.c",
1079 "src/qs8-vaddc/gen/minmax-wasmsimd-x16.c",
1080 "src/qs8-vaddc/gen/minmax-wasmsimd-x24.c",
1081 "src/qs8-vaddc/gen/minmax-wasmsimd-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001082 "src/qu8-requantization/fp32-wasmsimd.c",
Marat Dukhan22de5e72020-08-03 21:53:18 -07001083 "src/qu8-requantization/q31-wasmsimd.c",
Marat Dukhan8ee37012020-07-16 13:17:13 -07001084 "src/x32-fill/wasmsimd.c",
Marat Dukhan66d99e92020-07-16 12:56:21 -07001085 "src/x32-packx/x4-wasmsimd.c",
Marat Dukhan9306ae02020-07-16 15:51:13 -07001086 "src/x32-pad/wasmsimd.c",
Marat Dukhan9d4bfa22020-07-16 19:07:04 -07001087 "src/x32-unpool/wasmsimd.c",
Marat Dukhane3b78762020-07-16 20:02:58 -07001088 "src/x32-zip/x2-wasmsimd.c",
1089 "src/x32-zip/x3-wasmsimd.c",
1090 "src/x32-zip/x4-wasmsimd.c",
1091 "src/x32-zip/xm-wasmsimd.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001092 "src/math/roundne-wasmsimd-addsub.c",
1093 "src/math/roundd-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001094 "src/math/roundd-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001095 "src/math/roundu-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001096 "src/math/roundu-wasmsimd-cvt.c",
Marat Dukhand3f3d872020-06-24 13:08:25 -07001097 "src/math/roundz-wasmsimd-addsub.c",
Marat Dukhan88da62c2020-07-15 17:56:09 -07001098 "src/math/roundz-wasmsimd-cvt.c",
Marat Dukhan773da432020-07-17 18:55:36 -07001099 "src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
1100 "src/math/sigmoid-wasmsimd-rr2-p5-div.c",
Marat Dukhan290055c2020-06-09 12:24:29 -07001101]
1102
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001103PSIMD_FASTMATH_UKERNELS = [
Erich Elsen0a1970e2020-06-10 09:24:59 -07001104 "src/f32-conv-hwc2chw/3x3s2p1c3x4-psimd-2x2.c",
Erich Elsene6214af2020-06-10 22:17:22 -07001105 "src/f32-dwconv-chw/3x3p1-psimd.c",
Erich Elsenfd7a6e32020-06-11 12:04:44 -07001106 "src/f32-dwconv-chw/3x3s2p1-psimd.c",
Erich Elsen28928892020-06-12 08:08:19 -07001107 "src/f32-dwconv-chw/5x5p2-psimd.c",
Erich Elsen7465a892020-06-13 14:02:04 -07001108 "src/f32-dwconv-chw/5x5s2p2-psimd.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001109 "src/f32-spmm/gen/4x1-minmax-psimd.c",
1110 "src/f32-spmm/gen/8x1-minmax-psimd.c",
1111 "src/f32-spmm/gen/16x1-minmax-psimd.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001112]
1113
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001114PSIMD_ACCMATH_UKERNELS = [
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001115 "src/qs8-requantization/precise-psimd.c",
1116 "src/qs8-requantization/fp32-psimd.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001117 "src/qu8-requantization/precise-psimd.c",
1118 "src/qu8-requantization/fp32-psimd.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001119]
1120
Marat Dukhan08c4a432019-10-03 09:29:21 -07001121# ISA-specific micro-kernels
1122NEON_UKERNELS = [
Marat Dukhanef25c6d2020-07-24 00:59:40 -07001123 "src/f32-argmaxpool/4x-neon-c4.c",
1124 "src/f32-argmaxpool/9p8x-neon-c4.c",
1125 "src/f32-argmaxpool/9x-neon-c4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001126 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
1127 "src/f32-avgpool/9x-minmax-neon-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001128 "src/f32-clamp/gen/neon-x4.c",
1129 "src/f32-clamp/gen/neon-x8.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001130 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
1131 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001132 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
1133 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001134 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
1135 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001136 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
1137 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001138 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
1139 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
1140 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
1141 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001142 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
1143 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
1144 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
1145 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001146 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
1147 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
1148 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
1149 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001150 "src/f32-gavgpool-cw/neon-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001151 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
1152 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001153 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
1154 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
1155 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
1156 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
1157 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
1158 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
1159 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
1160 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
1161 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
1162 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
1163 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
1164 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
1165 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
1166 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
1167 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
1168 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
1169 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
1170 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
1171 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
1172 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
1173 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
1174 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
1175 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
1176 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
1177 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
1178 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
1179 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
1180 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
1181 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
1182 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
1183 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001184 "src/f32-hswish/gen/hswish-neon-x4.c",
1185 "src/f32-hswish/gen/hswish-neon-x8.c",
1186 "src/f32-hswish/gen/hswish-neon-x16.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001187 "src/f32-ibilinear/gen/neon-c4.c",
1188 "src/f32-ibilinear/gen/neon-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001189 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
1190 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
1191 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
1192 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
1193 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
1194 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
1195 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
1196 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
1197 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
1198 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
1199 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
1200 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
1201 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
1202 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
1203 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
1204 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001205 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
1206 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
1207 "src/f32-pavgpool/9x-minmax-neon-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001208 "src/f32-ppmm/gen/4x8-minmax-neon.c",
1209 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001210 "src/f32-prelu/gen/neon-1x4.c",
1211 "src/f32-prelu/gen/neon-1x8.c",
1212 "src/f32-prelu/gen/neon-1x16.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001213 "src/f32-prelu/gen/neon-2x4.c",
1214 "src/f32-prelu/gen/neon-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001215 "src/f32-prelu/gen/neon-2x16.c",
1216 "src/f32-prelu/gen/neon-4x4.c",
1217 "src/f32-prelu/gen/neon-4x8.c",
1218 "src/f32-prelu/gen/neon-4x16.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001219 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
1220 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
1221 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
1222 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
1223 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
1224 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
1225 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
1226 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
1227 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
1228 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
1229 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
1230 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
1231 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
1232 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
1233 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
1234 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
1235 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
1236 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
1237 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
1238 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
1239 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
1240 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
1241 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
1242 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001243 "src/f32-relu/gen/neon-x4.c",
1244 "src/f32-relu/gen/neon-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001245 "src/f32-rmax/neon.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001246 "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001247 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
1248 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
1249 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
1250 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
1251 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
1252 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
1253 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
1254 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
1255 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
1256 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
1257 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
1258 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
1259 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
1260 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
1261 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
1262 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
1263 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
1264 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001265 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
1267 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
1268 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001269 "src/f32-vbinary/gen/vmax-neon-x4.c",
1270 "src/f32-vbinary/gen/vmax-neon-x8.c",
1271 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
1272 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
1273 "src/f32-vbinary/gen/vmin-neon-x4.c",
1274 "src/f32-vbinary/gen/vmin-neon-x8.c",
1275 "src/f32-vbinary/gen/vminc-neon-x4.c",
1276 "src/f32-vbinary/gen/vminc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001277 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
1278 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
1279 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
1280 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
1281 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
1282 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001283 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
1284 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
1285 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
1286 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001287 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
1288 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
1289 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
1290 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001291 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
1292 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001293 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
1294 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001295 "src/f32-vrnd/gen/vrndne-neon-x4.c",
1296 "src/f32-vrnd/gen/vrndne-neon-x8.c",
1297 "src/f32-vrnd/gen/vrndz-neon-x4.c",
1298 "src/f32-vrnd/gen/vrndz-neon-x8.c",
1299 "src/f32-vrnd/gen/vrndu-neon-x4.c",
1300 "src/f32-vrnd/gen/vrndu-neon-x8.c",
1301 "src/f32-vrnd/gen/vrndd-neon-x4.c",
1302 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001303 "src/f32-vunary/gen/vabs-neon-x4.c",
1304 "src/f32-vunary/gen/vabs-neon-x8.c",
1305 "src/f32-vunary/gen/vneg-neon-x4.c",
1306 "src/f32-vunary/gen/vneg-neon-x8.c",
1307 "src/f32-vunary/gen/vsqr-neon-x4.c",
1308 "src/f32-vunary/gen/vsqr-neon-x8.c",
Marat Dukhan023bcf92020-08-10 12:40:50 -07001309 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
1310 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
1311 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
1312 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
Marat Dukhan281262d2020-08-10 13:23:21 -07001313 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
1314 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
1315 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
1316 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
1317 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
1318 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
1319 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
1320 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
Marat Dukhan40bbafe2020-08-04 02:04:22 -07001321 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
1322 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
1323 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
1324 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
1325 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
1326 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
1327 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
1328 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07001329 "src/qs8-requantization/precise-neon.c",
1330 "src/qs8-requantization/fp32-neon.c",
1331 "src/qs8-requantization/q31-neon.c",
Marat Dukhanba7b2792020-09-02 14:26:45 -07001332 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
1333 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
1334 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
1335 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
1336 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
1337 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
1338 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
1339 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001340 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
1341 "src/qu8-avgpool/9x-minmax-neon-c8.c",
1342 "src/qu8-dwconv/up8x9-minmax-neon.c",
1343 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
1344 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
1345 "src/qu8-gemm/4x8-minmax-neon.c",
1346 "src/qu8-gemm/8x8-minmax-neon.c",
1347 "src/qu8-igemm/4x8-minmax-neon.c",
1348 "src/qu8-igemm/8x8-minmax-neon.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001349 "src/qu8-requantization/precise-neon.c",
1350 "src/qu8-requantization/fp32-neon.c",
1351 "src/qu8-requantization/q31-neon.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001352 "src/qu8-vadd/minmax-neon.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001353 "src/u8-clamp/neon-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001354 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001355 "src/u8-rmax/neon.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001356 "src/x32-fill/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001357 "src/x32-packx/x4-neon-st4.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001358 "src/x32-pad/neon.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001359 "src/x32-unpool/neon.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001360 "src/x32-zip/x2-neon.c",
1361 "src/x32-zip/x3-neon.c",
1362 "src/x32-zip/x4-neon.c",
1363 "src/x32-zip/xm-neon.c",
1364 "src/x8-zip/x2-neon.c",
1365 "src/x8-zip/x3-neon.c",
1366 "src/x8-zip/x4-neon.c",
1367 "src/x8-zip/xm-neon.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001368 "src/math/roundne-neon-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001369 "src/math/roundd-neon-addsub.c",
1370 "src/math/roundd-neon-cvt.c",
1371 "src/math/roundu-neon-addsub.c",
1372 "src/math/roundu-neon-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001373 "src/math/roundz-neon-addsub.c",
1374 "src/math/roundz-neon-cvt.c",
Marat Dukhan68b3b452020-01-02 10:11:15 -08001375 "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001376 "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c",
1377 "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c",
1378 "src/math/sigmoid-neon-rr1-p5-nr2recps.c",
1379 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
1380 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
1381 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001382 "src/math/sqrt-neon-nr1rsqrts.c",
1383 "src/math/sqrt-neon-nr2rsqrts.c",
1384 "src/math/sqrt-neon-nr3rsqrts.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001385]
1386
1387NEONFMA_UKERNELS = [
Marat Dukhan660fd192020-03-10 04:55:30 -07001388 "src/f32-ibilinear/gen/neonfma-c4.c",
1389 "src/f32-ibilinear/gen/neonfma-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001390 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1391 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1392 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1393 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1394 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1395 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
1396 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
1397 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1398 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001399 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
1400 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
1401 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
1402 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001403 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
1404 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
1405 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
1406 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
Marat Dukhanf5425ea2020-04-24 01:46:00 -07001407 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
1408 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
1409 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
1410 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001411 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
1412 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
1413 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
1414 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1415 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
1416 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
1417 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
1418 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
1419 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
1420 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
1421 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
1422 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
1423 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
1424 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
1425 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
1426 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
1427 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
1428 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001429 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1430 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Marat Dukhan8137e4c2020-01-25 12:56:58 -08001431 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1432 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1433 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1434 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1435 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1436 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1437 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1438 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1439 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1440 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
1441 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1442 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1443 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
1444 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
1445 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
1446 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
1447 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1448 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
1449 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
1450 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1451 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
1452 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1453 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1454 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001455 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1456 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1457 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1458 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1459 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1460 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1461 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1462 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1463 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1464 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1465 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1466 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1467 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1468 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1469 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1470 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1471 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1472 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
1473 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1474 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1475 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1476 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1477 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1478 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
1479 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1480 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1481 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1482 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1483 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1484 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
1485 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1486 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1487 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1488 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1489 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1490 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
1491 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1492 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1493 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1494 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1495 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1496 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
1497 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1498 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1499 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1500 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1501 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1502 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
1503 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1504 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1505 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1506 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1507 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1508 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001509 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1510 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001511 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1512 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1513 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1514 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1515 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1516 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1517 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1518 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1519 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1520 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1521 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1522 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1523 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1524 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1525 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1526 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1527 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1528 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1529 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1530 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Marat Dukhan797a8fe2019-11-14 20:21:57 -08001531 "src/math/exp-neonfma-lut64-p2.c",
1532 "src/math/exp-neonfma-p5.c",
Marat Dukhan189ae802019-11-26 11:28:44 -08001533 "src/math/expminus-neonfma-lut2048-p1.c",
1534 "src/math/expminus-neonfma-lut64-p2.c",
Marat Dukhan346a9e52019-11-15 09:06:30 -08001535 "src/math/expminus-neonfma-p5.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001536 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1537 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1538 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
1539 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1540 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1541 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
1542 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1543 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1544 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
1545 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1546 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1547 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
1548 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1549 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1550 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
1551 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1552 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1553 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001554 "src/math/sqrt-neonfma-nr1fma.c",
1555 "src/math/sqrt-neonfma-nr2fma.c",
1556 "src/math/sqrt-neonfma-nr3fma.c",
1557 "src/math/sqrt-neonfma-nr2fma1adj.c",
1558 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001559]
1560
1561AARCH64_NEONFMA_UKERNELS = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001562 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1563 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1564 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1565 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1566 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1567 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001568 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1569 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1570 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1571 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1572 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1573 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1574 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
1575 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
1576 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
1577 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
1578 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1579 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1580 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
1581 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1582 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1583 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
1584 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1585 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1586 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1587 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001588 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1589 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001590 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1591 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
Marat Dukhan56b10cd2020-05-18 09:35:49 -07001592 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1593 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
Marat Dukhance7a3f82020-05-17 21:46:44 -07001594 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1595 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001596 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1597 "src/f32-dwconv-chw/3x3p1-neonfma.c",
1598 "src/f32-dwconv-chw/5x5p2-neonfma.c",
1599 "src/f32-dwconv-chw/3x3s2p1-neonfma.c",
1600 "src/f32-dwconv-chw/5x5s2p2-neonfma.c",
Marat Dukhan4a24a582020-01-06 13:30:00 -08001601 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1602 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1603 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1604 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1605 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1606 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
1607 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1608 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1609 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1610 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1611 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1612 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1613 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1614 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1615 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1616 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1617 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1618 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001619 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1620 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1621 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
1622 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1623 "src/f32-spmm/gen/16x1-minmax-neonfma-unroll2.c",
1624 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1625 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1626 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
1627 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1628 "src/f32-spmm/gen/4x1-minmax-neonfma-unroll2.c",
1629 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1630 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1631 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
1632 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1633 "src/f32-spmm/gen/8x1-minmax-neonfma-unroll2.c",
1634 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1635 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1636 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001637 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1638 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Marat Dukhan77221d32020-01-06 10:04:39 -08001639 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
1640 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
1641 "src/math/sigmoid-neonfma-rr1-p5-div.c",
1642 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
1643 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
1644 "src/math/sigmoid-neonfma-rr2-p5-div.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001645]
1646
Marat Dukhan8853b822020-05-07 12:19:01 -07001647NEONV8_UKERNELS = [
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001648 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1649 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1650 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1651 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
1652 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1653 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1654 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1655 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001656 "src/math/roundne-neonv8.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001657 "src/math/roundd-neonv8.c",
1658 "src/math/roundu-neonv8.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001659 "src/math/roundz-neonv8.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07001660]
1661
Marat Dukhan08c4a432019-10-03 09:29:21 -07001662AARCH64_NEONFP16ARITH_UKERNELS = [
Frank Barchardb1966592020-05-12 13:47:06 -07001663 "src/f16-clamp/gen/neonfp16arith-x16.c",
1664 "src/f16-clamp/gen/neonfp16arith-x8.c",
Frank Barchard5a599a62020-06-04 20:12:44 -07001665 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1666 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
1667 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1668 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1669 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1670 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1671 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1672 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1673 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1674 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1675 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1676 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Frank Barchard0bb49a72020-06-04 11:35:11 -07001677 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1678 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001679 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1680 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Frank Barchardb1966592020-05-12 13:47:06 -07001681 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Frank Barcharda5316982020-07-23 13:19:28 -07001682 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001683 "src/f16-relu/gen/neonfp16arith-x16.c",
1684 "src/f16-relu/gen/neonfp16arith-x8.c",
Frank Barchardd793f6c2020-05-08 13:37:43 -07001685 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1686 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1687 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1688 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1689 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1690 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1691 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1692 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1693 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1694 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1695 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1696 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1697 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1698 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1699 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1700 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1701 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1702 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1703 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1704 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1705 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1706 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1707 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1708 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1709 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1710 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1711 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1712 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
Frank Barchard2a1049c2020-06-03 02:31:27 -07001713 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1714 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
Frank Barchard83ea9392020-04-20 20:15:14 -07001715 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1716 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1717 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1718 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1719 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1720 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1721 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1722 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
Frank Barchardb0e4fae2020-05-04 15:27:51 -07001723 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1724 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1725 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1726 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
Frank Barchard3f9f99f2020-05-06 01:12:04 -07001727 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1728 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1729 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1730 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
1731 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1732 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1733 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1734 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1735 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1736 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1737 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1738 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001739 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
1740 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-unroll2.c",
1741 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
1742 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-unroll2.c",
1743 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
1744 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-unroll2.c",
1745 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
1746 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-unroll2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001747]
1748
Benoit Jacoba9644732020-08-13 12:48:55 -07001749NEONDOT_UKERNELS = [
1750 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1751 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001753 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1754 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
1755 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001756]
1757
Marat Dukhan08c4a432019-10-03 09:29:21 -07001758SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001759 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1760 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001761 "src/f32-clamp/gen/sse-x4.c",
1762 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001763 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1764 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001765 "src/f32-dwconv-chw/3x3p1-sse.c",
1766 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001767 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1768 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1769 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1770 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1771 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1772 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1773 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1775 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1777 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001779 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001780 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1781 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001782 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1783 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1784 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1785 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1786 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1787 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1788 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1789 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1790 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1791 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1792 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1793 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1794 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001795 "src/f32-hswish/gen/hswish-sse-x4.c",
1796 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001797 "src/f32-ibilinear/gen/sse-c4.c",
1798 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001799 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1800 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1801 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1802 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1803 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1804 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1805 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001806 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1807 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1808 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001809 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001810 "src/f32-prelu/gen/sse-2x4.c",
1811 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001812 "src/f32-relu/gen/sse-x4.c",
1813 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001814 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001815 "src/f32-spmm/gen/4x1-minmax-sse.c",
1816 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001817 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001818 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1819 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1820 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1821 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1822 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001826 "src/f32-vbinary/gen/vmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1829 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1830 "src/f32-vbinary/gen/vmin-sse-x4.c",
1831 "src/f32-vbinary/gen/vmin-sse-x8.c",
1832 "src/f32-vbinary/gen/vminc-sse-x4.c",
1833 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001834 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1835 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1836 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1837 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1838 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001842 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1843 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1844 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1845 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001846 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1847 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1848 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1849 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001850 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1851 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001852 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1853 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001854 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1855 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001856 "src/f32-vunary/gen/vabs-sse-x4.c",
1857 "src/f32-vunary/gen/vabs-sse-x8.c",
1858 "src/f32-vunary/gen/vneg-sse-x4.c",
1859 "src/f32-vunary/gen/vneg-sse-x8.c",
1860 "src/f32-vunary/gen/vsqr-sse-x4.c",
1861 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001862 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001863 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001864 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001865 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001866 "src/math/roundd-sse-addsub.c",
1867 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001868 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001869 "src/math/sqrt-sse-hh1mac.c",
1870 "src/math/sqrt-sse-nr1mac.c",
1871 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001872]
1873
1874SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001875 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1876 "src/f32-argmaxpool/4x-sse2-c4.c",
1877 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001878 "src/f32-prelu/gen/sse2-2x4.c",
1879 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001880 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1881 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1882 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1883 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1884 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1885 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001892 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001893 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001894 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001895 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001896 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1897 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001898 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1899 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001900 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1901 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1902 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1903 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1904 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1905 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1906 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1907 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001908 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1909 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1910 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001911 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1912 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1913 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001914 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1915 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1916 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001917 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1918 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001919 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1920 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001921 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1922 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001923 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1924 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001925 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001926 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1927 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001928 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001929 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1930 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1931 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001932 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1933 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001934 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1935 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001936 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1937 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001938 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001939 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1940 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001941 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001942 "src/qs8-requantization/precise-sse2.c",
1943 "src/qs8-requantization/fp32-sse2.c",
1944 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001945 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1946 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1947 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1948 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001949 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1950 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1951 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1952 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001953 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1954 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1955 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1956 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1957 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1958 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1959 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1960 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001961 "src/qu8-requantization/precise-sse2.c",
1962 "src/qu8-requantization/fp32-sse2.c",
1963 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001964 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001965 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001966 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001967 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001968 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001969 "src/x32-zip/x2-sse2.c",
1970 "src/x32-zip/x3-sse2.c",
1971 "src/x32-zip/x4-sse2.c",
1972 "src/x32-zip/xm-sse2.c",
1973 "src/x8-zip/x2-sse2.c",
1974 "src/x8-zip/x3-sse2.c",
1975 "src/x8-zip/x4-sse2.c",
1976 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001977 "src/math/exp-sse2-p5.c",
1978 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001979 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001980 "src/math/roundd-sse2-cvt.c",
1981 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001982 "src/math/roundz-sse2-cvt.c",
Marat Dukhan80bafd22019-11-18 10:16:01 -08001983 "src/math/sigmoid-sse2-p5-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001984]
1985
1986SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001987 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1988 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1989 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001990 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1991 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1992 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001993 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1994 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1995 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001996 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1997 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001998 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1999 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002000 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2001 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002002 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2003 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002004 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002005 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2006 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002007 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002008 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2009 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2010 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002011 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2012 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002013 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2014 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002015 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2016 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002017 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002018 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2019 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002020 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002021 "src/qs8-requantization/precise-ssse3.c",
2022 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002023 "src/qu8-requantization/precise-ssse3.c",
2024 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002025]
2026
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002027SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002028 "src/f32-prelu/gen/sse41-2x4.c",
2029 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002030 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002031 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002032 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002033 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002034 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2035 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002036 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2037 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002038 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2039 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2040 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2041 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2042 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2043 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2044 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2045 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002046 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2047 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2048 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002049 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2050 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2051 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002052 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2053 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2054 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002055 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2056 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002057 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2058 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002059 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2060 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002061 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2062 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002063 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002064 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2065 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002066 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002067 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2068 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2069 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002070 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2071 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002072 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2073 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002074 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2075 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002076 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002077 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2078 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002079 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002080 "src/qs8-requantization/fp32-sse4.c",
2081 "src/qs8-requantization/precise-sse4.c",
2082 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002083 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2084 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2085 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2086 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002087 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2088 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2089 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2090 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002091 "src/qu8-requantization/precise-sse4.c",
2092 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002093 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002094 "src/math/roundd-sse41.c",
2095 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002096 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002097]
2098
Marat Dukhan08c4a432019-10-03 09:29:21 -07002099AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002100 "src/f32-clamp/gen/avx-x8.c",
2101 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002102 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2103 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2104 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2105 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2106 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2107 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2108 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2109 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2110 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2111 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2112 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2113 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2114 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2115 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2116 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2117 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2118 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2119 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2120 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2121 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2122 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2123 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2124 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2125 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2126 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2127 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2128 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2129 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2130 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2131 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002132 "src/f32-hswish/gen/hswish-avx-x8.c",
2133 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002134 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2135 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2136 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2137 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2138 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2139 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2140 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2141 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2142 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002143 "src/f32-prelu/gen/avx-2x8.c",
2144 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002145 "src/f32-relu/gen/avx-x16.c",
2146 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002147 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002148 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2149 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2150 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2151 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2152 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2153 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2154 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2155 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002156 "src/f32-vbinary/gen/vmax-avx-x8.c",
2157 "src/f32-vbinary/gen/vmax-avx-x16.c",
2158 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2159 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2160 "src/f32-vbinary/gen/vmin-avx-x8.c",
2161 "src/f32-vbinary/gen/vmin-avx-x16.c",
2162 "src/f32-vbinary/gen/vminc-avx-x8.c",
2163 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002164 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2165 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2166 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2167 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2168 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2169 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2170 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2171 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002172 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2173 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2174 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2175 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002176 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2177 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2178 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2179 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002180 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2181 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002182 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002183 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2184 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2185 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2186 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2187 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2188 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2189 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2190 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002191 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2192 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002193 "src/f32-vunary/gen/vabs-avx-x8.c",
2194 "src/f32-vunary/gen/vabs-avx-x16.c",
2195 "src/f32-vunary/gen/vneg-avx-x8.c",
2196 "src/f32-vunary/gen/vneg-avx-x16.c",
2197 "src/f32-vunary/gen/vsqr-avx-x8.c",
2198 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002199]
2200
Marat Dukhan1566fee2020-08-02 21:55:41 -07002201XOP_UKERNELS = [
2202 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2203 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2204 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2205 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002206 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2207 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002208 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2209 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002210 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002211 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2212 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002213 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002214 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2215 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2216 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002217 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2218 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2219 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2220 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2221 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2222 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002223 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002224 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2225 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002226 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002227]
2228
Marat Dukhanfda12b82019-11-21 12:27:59 -08002229FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002230 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2231 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2232 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2233 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2234 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2235 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2236 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2237 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2238 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2239 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2240 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2241 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2242 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2243 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2244 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2245 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2246 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2247 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2248 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2249 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2250 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2251 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2252 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2253 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2254 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2255 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2256 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2257 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2258 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2259 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2260 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2261 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2262 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2263 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2264 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2265 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2266 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2267 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2268 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2269 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002270 "src/f32-hswish/gen/hswish-fma3-x8.c",
2271 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002272 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2273 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2274 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2275 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2276 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2277 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2278 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2279 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2280 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2281 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2282 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2283 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2284 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2285 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002286 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2287 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2288 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2289 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2290 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2291 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2292 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2293 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002294 "src/math/sqrt-fma3-nr1fma.c",
2295 "src/math/sqrt-fma3-nr2fma.c",
2296 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002297]
2298
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002299AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002300 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2301 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2302 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2303 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2304 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2305 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2306 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2307 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2308 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2309 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2310 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2311 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2312 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2313 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2314 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2315 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2316 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2317 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2318 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2319 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2320 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2321 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2322 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2323 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2324 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2325 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2326 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2327 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2328 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2329 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2330 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2331 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2332 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2333 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2334 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2335 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002336 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2337 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2338 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2339 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2340 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2341 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2342 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2343 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2344 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2345 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2346 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2347 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2348 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2349 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2350 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2351 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2352 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2353 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2354 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2355 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2356 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2357 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2358 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2359 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2360 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2361 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2362 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2363 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2364 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2365 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002366 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2367 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2368 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2369 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2370 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2371 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2372 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2373 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2374 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2375 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2376 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2377 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2378 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2379 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2380 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2381 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2382 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2383 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2384 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2385 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2386 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2387 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2388 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2389 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002390 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2391 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002392 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2393 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2394 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2395 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002396 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2397 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2398 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002399 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2400 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2401 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002402 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2403 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2404 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002405 "src/math/exp-avx2-p5.c",
2406 "src/math/exp-avx2-perm-p3.c",
2407 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002408 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002409 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002410 "src/math/sigmoid-avx2-rr2-p5-div.c",
2411 "src/math/sigmoid-avx2-rr1-p5-div.c",
2412 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2413 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2414 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2415 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002416]
2417
Marat Dukhan08c4a432019-10-03 09:29:21 -07002418AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002419 "src/f32-clamp/gen/avx512f-x16.c",
2420 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002421 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2422 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2423 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2424 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2425 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2426 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2427 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2428 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2429 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2430 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2431 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2432 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2433 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2434 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2435 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2436 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2437 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2438 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2439 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2440 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2441 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2442 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2443 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2444 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002445 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2446 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002447 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2448 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2449 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2450 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2451 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2452 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002453 "src/f32-prelu/gen/avx512f-2x16.c",
2454 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002455 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2456 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2457 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2458 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2459 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2460 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2461 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2462 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2463 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2464 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2465 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2466 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2467 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2468 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2469 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2470 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2471 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2472 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2473 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2474 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2475 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2476 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2477 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2478 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2479 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2480 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2481 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2482 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2483 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2484 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2485 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2486 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2487 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2488 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2489 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2490 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002491 "src/f32-relu/gen/avx512f-x16.c",
2492 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002493 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002494 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2495 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2496 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2497 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2498 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2499 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2500 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2501 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002502 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2503 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2504 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2505 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2506 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2507 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2508 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2509 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002510 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2511 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2512 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2513 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2514 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2515 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2516 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2517 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002518 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2519 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2520 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2521 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002522 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2523 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2524 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2525 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002526 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2527 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002528 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002529 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2530 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2531 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2532 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2533 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2534 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2535 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2536 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2537 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2538 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2539 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2540 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2541 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2542 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2543 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2544 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2545 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2546 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2547 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2548 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2549 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2550 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2551 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2552 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002553 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2554 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2555 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2556 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2557 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2558 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2559 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2560 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002561 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2562 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2563 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2564 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2565 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2566 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2567 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2568 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002569 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2570 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2571 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2572 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2573 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2574 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002575 "src/math/exp-avx512f-p5-scalef.c",
2576 "src/math/exp-avx512f-p5.c",
2577 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002578 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002579 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002580 "src/math/sqrt-avx512f-nr1fma.c",
2581 "src/math/sqrt-avx512f-nr2fma.c",
2582 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002583]
2584
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002585AVX512SKX_UKERNELS = [
2586 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2587 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2588 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2589 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2590 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2591 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2592 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2593 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2594]
2595
Frank Barchardbcedc082020-08-17 18:00:51 -07002596WASM32_ASM_UKERNELS = [
2597 "src/f32-relu/wasm_shr_x1.S",
2598 "src/f32-relu/wasm_shr_x2.S",
2599 "src/f32-relu/wasm_shr_x4.S",
2600]
2601
Marat Dukhan08c4a432019-10-03 09:29:21 -07002602AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002603 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002604 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002605 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2606 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002607 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002608 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002609 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002610 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002611 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2612 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002613 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2614 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2615 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2616 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002617]
2618
2619AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002620 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002621 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2622 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002623 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002624 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2625 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002626 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2627 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2628 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002629 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002630 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2631 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2632 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002633 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002634 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2635 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002636 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002637 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2638 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2639 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2640 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2641 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2642 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2643 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2644 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2645 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2646 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2647 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2648 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2649 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2650 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2651 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2652 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2653 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2654 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002655 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2656 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002657 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002658 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2659 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2660 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2661 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2662 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2663 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2664 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2665 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2666 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2667 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2668 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2669 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2670 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2671 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2672 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2673 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2674 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2675 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002676 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2677 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2678 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2679 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2680 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2681 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2682 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2683 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2684 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2685 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2686 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2687 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2688 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2689 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2690 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2691 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2692 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2693 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002694]
2695
Marat Dukhan1b354632020-03-23 12:50:22 -07002696INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002697 "src/xnnpack/argmaxpool.h",
2698 "src/xnnpack/avgpool.h",
2699 "src/xnnpack/clamp.h",
2700 "src/xnnpack/common.h",
2701 "src/xnnpack/conv.h",
2702 "src/xnnpack/dwconv.h",
2703 "src/xnnpack/gavgpool.h",
2704 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002705 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002706 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002707 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002708 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002709 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002710 "src/xnnpack/lut.h",
2711 "src/xnnpack/math.h",
2712 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002713 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002714 "src/xnnpack/packx.h",
2715 "src/xnnpack/pad.h",
2716 "src/xnnpack/params.h",
2717 "src/xnnpack/pavgpool.h",
2718 "src/xnnpack/ppmm.h",
2719 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002720 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002721 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002722 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002723 "src/xnnpack/rmax.h",
2724 "src/xnnpack/scalar-utils.h",
2725 "src/xnnpack/spmm.h",
2726 "src/xnnpack/unpool.h",
2727 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002728 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002729 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002730 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002731 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002732 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002733 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002734 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002735]
2736
2737INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002738 "include/xnnpack.h",
2739 "src/xnnpack/allocator.h",
2740 "src/xnnpack/compute.h",
2741 "src/xnnpack/im2col.h",
2742 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002743 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002744 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002745 "src/xnnpack/operator.h",
2746 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002747 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002748 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002749 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002750 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002751]
2752
Marat Dukhan1b354632020-03-23 12:50:22 -07002753ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002754 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002755]
2756
Marat Dukhan1b354632020-03-23 12:50:22 -07002757MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002758 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002759 "include/xnnpack.h",
2760]
2761
Marat Dukhan1b354632020-03-23 12:50:22 -07002762MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002763 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002764 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002765 "src/xnnpack/requantization.h",
2766 "include/xnnpack.h",
2767]
2768
2769OPERATOR_TEST_PARAMS_HDRS = [
2770 "src/xnnpack/params.h",
2771 "src/xnnpack/common.h",
2772]
2773
2774WEIGHTS_PACK_HDRS = [
2775 "src/xnnpack/pack.h",
2776 "src/xnnpack/operator.h",
2777 "src/xnnpack/compute.h",
2778]
2779
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002780LOGGING_COPTS = select({
2781 # No logging in optimized mode
2782 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2783 # Full logging in debug mode
2784 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2785 # Error-only logging in default (fastbuild) mode
2786 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2787})
2788
Marat Dukhan3b59de22020-06-03 20:15:19 -07002789LOGGING_SRCS = select({
2790 # No logging in optimized mode
2791 ":optimized_build": [],
2792 "//conditions:default": [
2793 "src/operator-strings.c",
2794 "src/subgraph-strings.c",
2795 ],
2796})
2797
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002798LOGGING_HDRS = [
2799 "src/xnnpack/log.h",
2800]
2801
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002803 name = "tables",
2804 srcs = TABLE_SRCS,
2805 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002806 gcc_copts = xnnpack_gcc_std_copts(),
2807 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002808)
2809
2810xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002811 name = "scalar_ukernels",
2812 srcs = SCALAR_UKERNELS,
2813 hdrs = INTERNAL_HDRS,
2814 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002815 gcc_copts = xnnpack_gcc_std_copts(),
2816 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002817 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002818 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002819 "@FP16",
2820 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002821 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002822 ],
2823)
2824
2825xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002826 name = "scalar_ukernels_test_mode",
2827 srcs = SCALAR_UKERNELS,
2828 hdrs = INTERNAL_HDRS,
2829 aarch32_copts = ["-marm"],
2830 copts = [
2831 "-UNDEBUG",
2832 "-DXNN_TEST_MODE=1",
2833 ],
2834 gcc_copts = xnnpack_gcc_std_copts(),
2835 msvc_copts = xnnpack_msvc_std_copts(),
2836 deps = [
2837 ":tables",
2838 "@FP16",
2839 "@FXdiv",
2840 "@pthreadpool",
2841 ],
2842)
2843
2844xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002845 name = "wasm_ukernels",
2846 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002847 gcc_copts = xnnpack_gcc_std_copts(),
2848 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002849 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002850 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002851 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002852 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002853 "@FP16",
2854 "@FXdiv",
2855 "@pthreadpool",
2856 ],
2857)
2858
2859xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002860 name = "wasm_ukernels_test_mode",
2861 hdrs = INTERNAL_HDRS,
2862 copts = [
2863 "-UNDEBUG",
2864 "-DXNN_TEST_MODE=1",
2865 ],
2866 gcc_copts = xnnpack_gcc_std_copts(),
2867 msvc_copts = xnnpack_msvc_std_copts(),
2868 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002869 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002870 deps = [
2871 ":tables",
2872 "@FP16",
2873 "@FXdiv",
2874 "@pthreadpool",
2875 ],
2876)
2877
2878xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002879 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002880 hdrs = INTERNAL_HDRS,
2881 aarch32_copts = [
2882 "-marm",
2883 "-mfpu=neon",
2884 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002885 gcc_copts = xnnpack_gcc_std_copts(),
2886 gcc_x86_copts = ["-msse2"],
2887 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002888 optimized_copts = [
2889 "-O3",
2890 "-ffast-math",
2891 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002892 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002893 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002894 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002895 "@FP16",
2896 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002897 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002898 ],
2899)
2900
2901xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002902 name = "psimd_fastmath_ukernels_test_mode",
2903 hdrs = INTERNAL_HDRS,
2904 aarch32_copts = [
2905 "-marm",
2906 "-mfpu=neon",
2907 ],
2908 copts = [
2909 "-UNDEBUG",
2910 "-DXNN_TEST_MODE=1",
2911 ],
2912 gcc_copts = xnnpack_gcc_std_copts(),
2913 gcc_x86_copts = ["-msse2"],
2914 msvc_copts = xnnpack_msvc_std_copts(),
2915 optimized_copts = [
2916 "-O3",
2917 "-ffast-math",
2918 ],
2919 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2920 deps = [
2921 ":tables",
2922 "@FP16",
2923 "@psimd",
2924 "@pthreadpool",
2925 ],
2926)
2927
2928xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002929 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002930 hdrs = INTERNAL_HDRS,
2931 aarch32_copts = [
2932 "-marm",
2933 "-mfpu=neon",
2934 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002935 gcc_copts = xnnpack_gcc_std_copts(),
2936 gcc_x86_copts = ["-msse2"],
2937 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002938 optimized_copts = [
2939 "-O3",
2940 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002941 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002942 deps = [
2943 ":tables",
2944 "@FP16",
2945 "@psimd",
2946 "@pthreadpool",
2947 ],
2948)
2949
2950xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002951 name = "psimd_accmath_ukernels_test_mode",
2952 hdrs = INTERNAL_HDRS,
2953 aarch32_copts = [
2954 "-marm",
2955 "-mfpu=neon",
2956 ],
2957 copts = [
2958 "-UNDEBUG",
2959 "-DXNN_TEST_MODE=1",
2960 ],
2961 gcc_copts = xnnpack_gcc_std_copts(),
2962 gcc_x86_copts = ["-msse2"],
2963 msvc_copts = xnnpack_msvc_std_copts(),
2964 optimized_copts = [
2965 "-O3",
2966 ],
2967 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
2968 deps = [
2969 ":tables",
2970 "@FP16",
2971 "@psimd",
2972 "@pthreadpool",
2973 ],
2974)
2975
2976xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002977 name = "neon_ukernels",
2978 hdrs = INTERNAL_HDRS,
2979 aarch32_copts = [
2980 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07002981 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002982 "-mfpu=neon",
2983 ],
2984 aarch32_srcs = NEON_UKERNELS,
2985 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002986 gcc_copts = xnnpack_gcc_std_copts(),
2987 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08002988 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002989 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002990 "@FP16",
2991 "@pthreadpool",
2992 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07002993)
2994
2995xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002996 name = "neon_ukernels_test_mode",
2997 hdrs = INTERNAL_HDRS,
2998 aarch32_copts = [
2999 "-marm",
3000 "-march=armv7-a",
3001 "-mfpu=neon",
3002 ],
3003 aarch32_srcs = NEON_UKERNELS,
3004 aarch64_srcs = NEON_UKERNELS,
3005 copts = [
3006 "-UNDEBUG",
3007 "-DXNN_TEST_MODE=1",
3008 ],
3009 gcc_copts = xnnpack_gcc_std_copts(),
3010 msvc_copts = xnnpack_msvc_std_copts(),
3011 deps = [
3012 ":tables",
3013 "@FP16",
3014 "@pthreadpool",
3015 ],
3016)
3017
3018xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003019 name = "neonfma_ukernels",
3020 hdrs = INTERNAL_HDRS,
3021 aarch32_copts = [
3022 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003023 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003024 "-mfpu=neon-vfpv4",
3025 ],
3026 aarch32_srcs = NEONFMA_UKERNELS,
3027 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003028 apple_aarch32_copts = [
3029 "-mcpu=swift",
3030 "-mtune=generic",
3031 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003032 gcc_copts = xnnpack_gcc_std_copts(),
3033 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003034 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003035 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003036 "@FP16",
3037 "@pthreadpool",
3038 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003039)
3040
3041xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003042 name = "neonfma_ukernels_test_mode",
3043 hdrs = INTERNAL_HDRS,
3044 aarch32_copts = [
3045 "-marm",
3046 "-march=armv7-a",
3047 "-mfpu=neon-vfpv4",
3048 ],
3049 aarch32_srcs = NEONFMA_UKERNELS,
3050 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003051 apple_aarch32_copts = [
3052 "-mcpu=swift",
3053 "-mtune=generic",
3054 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003055 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 Dukhan8853b822020-05-07 12:19:01 -07003069 name = "neonv8_ukernels",
3070 hdrs = INTERNAL_HDRS,
3071 aarch32_copts = [
3072 "-marm",
3073 "-march=armv8-a",
3074 "-mfpu=neon-fp-armv8",
3075 ],
3076 aarch32_srcs = NEONV8_UKERNELS,
3077 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003078 apple_aarch32_copts = [
3079 "-mcpu=cyclone",
3080 "-mtune=generic",
3081 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003082 gcc_copts = xnnpack_gcc_std_copts(),
3083 msvc_copts = xnnpack_msvc_std_copts(),
3084 deps = [
3085 ":tables",
3086 "@FP16",
3087 "@pthreadpool",
3088 ],
3089)
3090
3091xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003092 name = "neonv8_ukernels_test_mode",
3093 hdrs = INTERNAL_HDRS,
3094 aarch32_copts = [
3095 "-marm",
3096 "-march=armv8-a",
3097 "-mfpu=neon-fp-armv8",
3098 ],
3099 aarch32_srcs = NEONV8_UKERNELS,
3100 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003101 apple_aarch32_copts = [
3102 "-mcpu=cyclone",
3103 "-mtune=generic",
3104 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003105 copts = [
3106 "-UNDEBUG",
3107 "-DXNN_TEST_MODE=1",
3108 ],
3109 gcc_copts = xnnpack_gcc_std_copts(),
3110 msvc_copts = xnnpack_msvc_std_copts(),
3111 deps = [
3112 ":tables",
3113 "@FP16",
3114 "@pthreadpool",
3115 ],
3116)
3117
3118xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003119 name = "neonfp16arith_ukernels",
3120 hdrs = INTERNAL_HDRS,
3121 aarch64_copts = ["-march=armv8.2-a+fp16"],
3122 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003123 gcc_copts = xnnpack_gcc_std_copts(),
3124 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003125 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003126 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003127 "@FP16",
3128 "@pthreadpool",
3129 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003130)
3131
3132xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003133 name = "neonfp16arith_ukernels_test_mode",
3134 hdrs = INTERNAL_HDRS,
3135 aarch64_copts = ["-march=armv8.2-a+fp16"],
3136 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3137 copts = [
3138 "-UNDEBUG",
3139 "-DXNN_TEST_MODE=1",
3140 ],
3141 gcc_copts = xnnpack_gcc_std_copts(),
3142 msvc_copts = xnnpack_msvc_std_copts(),
3143 deps = [
3144 ":tables",
3145 "@FP16",
3146 "@pthreadpool",
3147 ],
3148)
3149
3150xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003151 name = "neondot_ukernels",
3152 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003153 aarch32_copts = [
3154 "-marm",
3155 "-march=armv8.2-a+dotprod",
3156 "-mfpu=neon-fp-armv8",
3157 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003158 aarch32_srcs = NEONDOT_UKERNELS,
3159 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3160 aarch64_srcs = NEONDOT_UKERNELS,
3161 gcc_copts = xnnpack_gcc_std_copts(),
3162 msvc_copts = xnnpack_msvc_std_copts(),
3163 deps = [
3164 ":tables",
3165 "@FP16",
3166 "@pthreadpool",
3167 ],
3168)
3169
3170xnnpack_cc_library(
3171 name = "neondot_ukernels_test_mode",
3172 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003173 aarch32_copts = [
3174 "-marm",
3175 "-march=armv8.2-a+dotprod",
3176 "-mfpu=neon-fp-armv8",
3177 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003178 aarch32_srcs = NEONDOT_UKERNELS,
3179 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3180 aarch64_srcs = NEONDOT_UKERNELS,
3181 copts = [
3182 "-UNDEBUG",
3183 "-DXNN_TEST_MODE=1",
3184 ],
3185 gcc_copts = xnnpack_gcc_std_copts(),
3186 msvc_copts = xnnpack_msvc_std_copts(),
3187 deps = [
3188 ":tables",
3189 "@FP16",
3190 "@pthreadpool",
3191 ],
3192)
3193
3194xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003195 name = "sse2_ukernels",
3196 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003197 gcc_copts = xnnpack_gcc_std_copts(),
3198 gcc_x86_copts = ["-msse2"],
3199 msvc_copts = xnnpack_msvc_std_copts(),
3200 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003201 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003202 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003203 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003204 "@FP16",
3205 "@pthreadpool",
3206 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003207)
3208
3209xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003210 name = "sse2_ukernels_test_mode",
3211 hdrs = INTERNAL_HDRS,
3212 copts = [
3213 "-UNDEBUG",
3214 "-DXNN_TEST_MODE=1",
3215 ],
3216 gcc_copts = xnnpack_gcc_std_copts(),
3217 gcc_x86_copts = ["-msse2"],
3218 msvc_copts = xnnpack_msvc_std_copts(),
3219 msvc_x86_32_copts = ["/arch:SSE2"],
3220 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3221 deps = [
3222 ":tables",
3223 "@FP16",
3224 "@pthreadpool",
3225 ],
3226)
3227
3228xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003229 name = "ssse3_ukernels",
3230 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003231 gcc_copts = xnnpack_gcc_std_copts(),
3232 gcc_x86_copts = ["-mssse3"],
3233 msvc_copts = xnnpack_msvc_std_copts(),
3234 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003235 x86_srcs = SSSE3_UKERNELS,
3236 deps = [
3237 ":tables",
3238 "@FP16",
3239 "@pthreadpool",
3240 ],
3241)
3242
3243xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003244 name = "ssse3_ukernels_test_mode",
3245 hdrs = INTERNAL_HDRS,
3246 copts = [
3247 "-UNDEBUG",
3248 "-DXNN_TEST_MODE=1",
3249 ],
3250 gcc_copts = xnnpack_gcc_std_copts(),
3251 gcc_x86_copts = ["-mssse3"],
3252 msvc_copts = xnnpack_msvc_std_copts(),
3253 msvc_x86_32_copts = ["/arch:SSE2"],
3254 x86_srcs = SSSE3_UKERNELS,
3255 deps = [
3256 ":tables",
3257 "@FP16",
3258 "@pthreadpool",
3259 ],
3260)
3261
3262xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003263 name = "sse41_ukernels",
3264 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003265 gcc_copts = xnnpack_gcc_std_copts(),
3266 gcc_x86_copts = ["-msse4.1"],
3267 msvc_copts = xnnpack_msvc_std_copts(),
3268 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003269 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003270 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003271 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003272 "@FP16",
3273 "@pthreadpool",
3274 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003275)
3276
3277xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003278 name = "sse41_ukernels_test_mode",
3279 hdrs = INTERNAL_HDRS,
3280 copts = [
3281 "-UNDEBUG",
3282 "-DXNN_TEST_MODE=1",
3283 ],
3284 gcc_copts = xnnpack_gcc_std_copts(),
3285 gcc_x86_copts = ["-msse4.1"],
3286 msvc_copts = xnnpack_msvc_std_copts(),
3287 msvc_x86_32_copts = ["/arch:SSE2"],
3288 x86_srcs = SSE41_UKERNELS,
3289 deps = [
3290 ":tables",
3291 "@FP16",
3292 "@pthreadpool",
3293 ],
3294)
3295
3296xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003297 name = "avx_ukernels",
3298 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003299 gcc_copts = xnnpack_gcc_std_copts(),
3300 gcc_x86_copts = ["-mavx"],
3301 msvc_copts = xnnpack_msvc_std_copts(),
3302 msvc_x86_32_copts = ["/arch:AVX"],
3303 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003304 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003305 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003306 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003307 "@FP16",
3308 "@pthreadpool",
3309 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003310)
3311
3312xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003313 name = "avx_ukernels_test_mode",
3314 hdrs = INTERNAL_HDRS,
3315 copts = [
3316 "-UNDEBUG",
3317 "-DXNN_TEST_MODE=1",
3318 ],
3319 gcc_copts = xnnpack_gcc_std_copts(),
3320 gcc_x86_copts = ["-mavx"],
3321 msvc_copts = xnnpack_msvc_std_copts(),
3322 msvc_x86_32_copts = ["/arch:AVX"],
3323 msvc_x86_64_copts = ["/arch:AVX"],
3324 x86_srcs = AVX_UKERNELS,
3325 deps = [
3326 ":tables",
3327 "@FP16",
3328 "@pthreadpool",
3329 ],
3330)
3331
3332xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003333 name = "xop_ukernels",
3334 hdrs = INTERNAL_HDRS,
3335 gcc_copts = xnnpack_gcc_std_copts(),
3336 gcc_x86_copts = ["-mxop"],
3337 msvc_copts = xnnpack_msvc_std_copts(),
3338 msvc_x86_32_copts = ["/arch:AVX"],
3339 msvc_x86_64_copts = ["/arch:AVX"],
3340 x86_srcs = XOP_UKERNELS,
3341 deps = [
3342 ":tables",
3343 "@FP16",
3344 "@pthreadpool",
3345 ],
3346)
3347
3348xnnpack_cc_library(
3349 name = "xop_ukernels_test_mode",
3350 hdrs = INTERNAL_HDRS,
3351 copts = [
3352 "-UNDEBUG",
3353 "-DXNN_TEST_MODE=1",
3354 ],
3355 gcc_copts = xnnpack_gcc_std_copts(),
3356 gcc_x86_copts = ["-mxop"],
3357 msvc_copts = xnnpack_msvc_std_copts(),
3358 msvc_x86_32_copts = ["/arch:AVX"],
3359 msvc_x86_64_copts = ["/arch:AVX"],
3360 x86_srcs = XOP_UKERNELS,
3361 deps = [
3362 ":tables",
3363 "@FP16",
3364 "@pthreadpool",
3365 ],
3366)
3367
3368xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003369 name = "fma3_ukernels",
3370 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003371 gcc_copts = xnnpack_gcc_std_copts(),
3372 gcc_x86_copts = ["-mfma"],
3373 msvc_copts = xnnpack_msvc_std_copts(),
3374 msvc_x86_32_copts = ["/arch:AVX"],
3375 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003376 x86_srcs = FMA3_UKERNELS,
3377 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003378 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003379 "@FP16",
3380 "@pthreadpool",
3381 ],
3382)
3383
3384xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003385 name = "fma3_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 = ["-mfma"],
3393 msvc_copts = xnnpack_msvc_std_copts(),
3394 msvc_x86_32_copts = ["/arch:AVX"],
3395 msvc_x86_64_copts = ["/arch:AVX"],
3396 x86_srcs = FMA3_UKERNELS,
3397 deps = [
3398 ":tables",
3399 "@FP16",
3400 "@pthreadpool",
3401 ],
3402)
3403
3404xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003405 name = "avx2_ukernels",
3406 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003407 gcc_copts = xnnpack_gcc_std_copts(),
3408 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003409 "-mfma",
3410 "-mavx2",
3411 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003412 msvc_copts = xnnpack_msvc_std_copts(),
3413 msvc_x86_32_copts = ["/arch:AVX2"],
3414 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003415 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003416 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003417 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003418 "@FP16",
3419 "@pthreadpool",
3420 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003421)
3422
3423xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003424 name = "avx2_ukernels_test_mode",
3425 hdrs = INTERNAL_HDRS,
3426 copts = [
3427 "-UNDEBUG",
3428 "-DXNN_TEST_MODE=1",
3429 ],
3430 gcc_copts = xnnpack_gcc_std_copts(),
3431 gcc_x86_copts = [
3432 "-mfma",
3433 "-mavx2",
3434 ],
3435 msvc_copts = xnnpack_msvc_std_copts(),
3436 msvc_x86_32_copts = ["/arch:AVX2"],
3437 msvc_x86_64_copts = ["/arch:AVX2"],
3438 x86_srcs = AVX2_UKERNELS,
3439 deps = [
3440 ":tables",
3441 "@FP16",
3442 "@pthreadpool",
3443 ],
3444)
3445
3446xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003447 name = "avx512f_ukernels",
3448 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003449 gcc_copts = xnnpack_gcc_std_copts(),
3450 gcc_x86_copts = ["-mavx512f"],
3451 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3452 msvc_copts = xnnpack_msvc_std_copts(),
3453 msvc_x86_32_copts = ["/arch:AVX512"],
3454 msvc_x86_64_copts = ["/arch:AVX512"],
3455 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003456 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003457 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003458 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003459 "@FP16",
3460 "@pthreadpool",
3461 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003462)
3463
3464xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003465 name = "avx512f_ukernels_test_mode",
3466 hdrs = INTERNAL_HDRS,
3467 copts = [
3468 "-UNDEBUG",
3469 "-DXNN_TEST_MODE=1",
3470 ],
3471 gcc_copts = xnnpack_gcc_std_copts(),
3472 gcc_x86_copts = ["-mavx512f"],
3473 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3474 msvc_copts = xnnpack_msvc_std_copts(),
3475 msvc_x86_32_copts = ["/arch:AVX512"],
3476 msvc_x86_64_copts = ["/arch:AVX512"],
3477 msys_copts = ["-fno-asynchronous-unwind-tables"],
3478 x86_srcs = AVX512F_UKERNELS,
3479 deps = [
3480 ":tables",
3481 "@FP16",
3482 "@pthreadpool",
3483 ],
3484)
3485
3486xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003487 name = "avx512skx_ukernels",
3488 hdrs = INTERNAL_HDRS,
3489 gcc_copts = xnnpack_gcc_std_copts(),
3490 gcc_x86_copts = [
3491 "-mavx512f",
3492 "-mavx512cd",
3493 "-mavx512bw",
3494 "-mavx512dq",
3495 "-mavx512vl",
3496 ],
3497 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3498 msvc_copts = xnnpack_msvc_std_copts(),
3499 msvc_x86_32_copts = ["/arch:AVX512"],
3500 msvc_x86_64_copts = ["/arch:AVX512"],
3501 msys_copts = ["-fno-asynchronous-unwind-tables"],
3502 x86_srcs = AVX512SKX_UKERNELS,
3503 deps = [
3504 ":tables",
3505 "@FP16",
3506 "@pthreadpool",
3507 ],
3508)
3509
3510xnnpack_cc_library(
3511 name = "avx512skx_ukernels_test_mode",
3512 hdrs = INTERNAL_HDRS,
3513 copts = [
3514 "-UNDEBUG",
3515 "-DXNN_TEST_MODE=1",
3516 ],
3517 gcc_copts = xnnpack_gcc_std_copts(),
3518 gcc_x86_copts = [
3519 "-mavx512f",
3520 "-mavx512cd",
3521 "-mavx512bw",
3522 "-mavx512dq",
3523 "-mavx512vl",
3524 ],
3525 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3526 msvc_copts = xnnpack_msvc_std_copts(),
3527 msvc_x86_32_copts = ["/arch:AVX512"],
3528 msvc_x86_64_copts = ["/arch:AVX512"],
3529 msys_copts = ["-fno-asynchronous-unwind-tables"],
3530 x86_srcs = AVX512SKX_UKERNELS,
3531 deps = [
3532 ":tables",
3533 "@FP16",
3534 "@pthreadpool",
3535 ],
3536)
3537
3538xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003539 name = "asm_ukernels",
3540 hdrs = ["src/xnnpack/assembly.h"],
3541 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003542 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003543 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003544 wasm_srcs = WASM32_ASM_UKERNELS,
3545 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003546)
3547
Marat Dukhan3b59de22020-06-03 20:15:19 -07003548xnnpack_cc_library(
3549 name = "logging_utils",
3550 srcs = LOGGING_SRCS,
3551 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3552 copts = LOGGING_COPTS + [
3553 "-Isrc",
3554 "-Iinclude",
3555 ] + select({
3556 ":debug_build": [],
3557 "//conditions:default": xnnpack_min_size_copts(),
3558 }),
3559 gcc_copts = xnnpack_gcc_std_copts(),
3560 msvc_copts = xnnpack_msvc_std_copts(),
3561 visibility = xnnpack_visibility(),
3562 deps = [
3563 "@FP16",
3564 "@clog",
3565 "@pthreadpool",
3566 ],
3567)
3568
Marat Dukhan08c4a432019-10-03 09:29:21 -07003569xnnpack_aggregate_library(
3570 name = "ukernels",
3571 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003572 ":neon_ukernels",
3573 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003574 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003575 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003576 ":asm_ukernels",
3577 ],
3578 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003579 ":neon_ukernels",
3580 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003581 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003583 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003584 ":asm_ukernels",
3585 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003586 generic_deps = [
3587 ":scalar_ukernels",
3588 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003589 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003590 ":psimd_fastmath_ukernels",
3591 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003592 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003593 wasm_deps = [
3594 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003595 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003596 ],
3597 wasmsimd_deps = [
3598 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003599 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003600 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003601 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003602 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003603 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003604 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003605 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003606 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003607 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003608 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003609 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003610 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003611 ],
3612)
3613
Marat Dukhan33fcf782020-05-24 14:27:15 -07003614xnnpack_aggregate_library(
3615 name = "ukernels_test_mode",
3616 aarch32_deps = [
3617 ":neon_ukernels_test_mode",
3618 ":neonfma_ukernels_test_mode",
3619 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003620 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003621 ":asm_ukernels",
3622 ],
3623 aarch64_deps = [
3624 ":neon_ukernels_test_mode",
3625 ":neonfma_ukernels_test_mode",
3626 ":neonv8_ukernels_test_mode",
3627 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003628 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003629 ":asm_ukernels",
3630 ],
3631 generic_deps = [
3632 ":scalar_ukernels_test_mode",
3633 ],
3634 psimd_deps = [
3635 ":psimd_fastmath_ukernels_test_mode",
3636 ":psimd_accmath_ukernels_test_mode",
3637 ],
3638 wasm_deps = [
3639 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003640 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003641 ],
3642 wasmsimd_deps = [
3643 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003644 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003645 ],
3646 x86_deps = [
3647 ":sse2_ukernels_test_mode",
3648 ":ssse3_ukernels_test_mode",
3649 ":sse41_ukernels_test_mode",
3650 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003651 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003652 ":fma3_ukernels_test_mode",
3653 ":avx2_ukernels_test_mode",
3654 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003655 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003656 ],
3657)
3658
Marat Dukhan86df3842020-06-17 14:03:31 -07003659# This target must be compatible with Android NDK r17 and compile with Clang 6.
3660# It is similar to :ukernels target, but excludes microkernels which require
3661# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003662xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003663 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003664 aarch32_deps = [
3665 ":neon_ukernels",
3666 ":neonfma_ukernels",
3667 ":neonv8_ukernels",
3668 ":asm_ukernels",
3669 ],
3670 aarch64_deps = [
3671 ":neon_ukernels",
3672 ":neonfma_ukernels",
3673 ":neonv8_ukernels",
3674 ":asm_ukernels",
3675 ],
3676 generic_deps = [
3677 ":scalar_ukernels",
3678 ],
3679 psimd_deps = [
3680 ":psimd_fastmath_ukernels",
3681 ":psimd_accmath_ukernels",
3682 ],
3683 wasm_deps = [
3684 ":wasm_ukernels",
3685 ],
3686 wasmsimd_deps = [
3687 ":wasm_ukernels",
3688 ],
3689 x86_deps = [
3690 ":sse2_ukernels",
3691 ":ssse3_ukernels",
3692 ":sse41_ukernels",
3693 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003694 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003695 ":fma3_ukernels",
3696 ":avx2_ukernels",
3697 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003698 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003699 ],
3700)
3701
Marat Dukhan08c4a432019-10-03 09:29:21 -07003702xnnpack_cc_library(
3703 name = "im2col",
3704 srcs = ["src/im2col.c"],
3705 hdrs = [
3706 "src/xnnpack/common.h",
3707 "src/xnnpack/im2col.h",
3708 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003709 gcc_copts = xnnpack_gcc_std_copts(),
3710 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003711)
3712
3713xnnpack_cc_library(
3714 name = "indirection",
3715 srcs = ["src/indirection.c"],
3716 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003717 gcc_copts = xnnpack_gcc_std_copts(),
3718 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003719 deps = [
3720 "@FP16",
3721 "@FXdiv",
3722 "@pthreadpool",
3723 ],
3724)
3725
3726xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003727 name = "indirection_test_mode",
3728 srcs = ["src/indirection.c"],
3729 hdrs = INTERNAL_HDRS,
3730 copts = [
3731 "-UNDEBUG",
3732 "-DXNN_TEST_MODE=1",
3733 ],
3734 gcc_copts = xnnpack_gcc_std_copts(),
3735 msvc_copts = xnnpack_msvc_std_copts(),
3736 deps = [
3737 "@FP16",
3738 "@FXdiv",
3739 "@pthreadpool",
3740 ],
3741)
3742
3743xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003744 name = "packing",
3745 srcs = ["src/packing.c"],
3746 hdrs = INTERNAL_HDRS,
3747 gcc_copts = xnnpack_gcc_std_copts(),
3748 msvc_copts = xnnpack_msvc_std_copts(),
3749 deps = [
3750 "@FP16",
3751 "@FXdiv",
3752 "@pthreadpool",
3753 ],
3754)
3755
3756xnnpack_cc_library(
3757 name = "packing_test_mode",
3758 srcs = ["src/packing.c"],
3759 hdrs = INTERNAL_HDRS,
3760 copts = [
3761 "-UNDEBUG",
3762 "-DXNN_TEST_MODE=1",
3763 ],
3764 gcc_copts = xnnpack_gcc_std_copts(),
3765 msvc_copts = xnnpack_msvc_std_copts(),
3766 deps = [
3767 "@FP16",
3768 "@FXdiv",
3769 "@pthreadpool",
3770 ],
3771)
3772
3773xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003774 name = "operator_run",
3775 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003776 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003777 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003778 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3779 "//conditions:default": [],
3780 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003781 gcc_copts = xnnpack_gcc_std_copts(),
3782 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003783 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003784 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003785 "@FP16",
3786 "@FXdiv",
3787 "@clog",
3788 "@pthreadpool",
3789 ],
3790)
3791
Chao Mei6ddfc602020-05-13 22:29:36 -07003792xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003793 name = "operator_run_test_mode",
3794 srcs = ["src/operator-run.c"],
3795 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3796 copts = LOGGING_COPTS + [
3797 "-UNDEBUG",
3798 "-DXNN_TEST_MODE=1",
3799 ] + select({
3800 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3801 "//conditions:default": [],
3802 }),
3803 gcc_copts = xnnpack_gcc_std_copts(),
3804 msvc_copts = xnnpack_msvc_std_copts(),
3805 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003806 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003807 "@FP16",
3808 "@FXdiv",
3809 "@clog",
3810 "@pthreadpool",
3811 ],
3812)
3813
3814xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003815 name = "memory_planner",
3816 srcs = ["src/memory-planner.c"],
3817 hdrs = INTERNAL_HDRS,
3818 defines = select({
3819 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3820 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3821 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3822 }),
3823 gcc_copts = xnnpack_gcc_std_copts(),
3824 msvc_copts = xnnpack_msvc_std_copts(),
3825 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003826 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003827 "@pthreadpool",
3828 ],
3829)
3830
Marat Dukhan33fcf782020-05-24 14:27:15 -07003831xnnpack_cc_library(
3832 name = "memory_planner_test_mode",
3833 srcs = ["src/memory-planner.c"],
3834 hdrs = INTERNAL_HDRS,
3835 copts = [
3836 "-UNDEBUG",
3837 "-DXNN_TEST_MODE=1",
3838 ],
3839 defines = select({
3840 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3841 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3842 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3843 }),
3844 gcc_copts = xnnpack_gcc_std_copts(),
3845 msvc_copts = xnnpack_msvc_std_copts(),
3846 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003847 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003848 "@pthreadpool",
3849 ],
3850)
3851
Marat Dukhan08c4a432019-10-03 09:29:21 -07003852cc_library(
3853 name = "enable_assembly",
3854 defines = select({
3855 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3856 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003857 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003858 }),
3859)
3860
Marat Dukhan9de90e02020-06-18 16:04:12 -07003861cc_library(
3862 name = "enable_sparse",
3863 defines = select({
3864 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3865 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003866 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003867 }),
3868)
3869
Marat Dukhancf056b22019-10-07 10:26:29 -07003870xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003871 name = "operators",
3872 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003873 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003874 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003875 ],
3876 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003877 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003878 "-Isrc",
3879 "-Iinclude",
3880 ] + select({
3881 ":debug_build": [],
3882 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003883 }) + select({
3884 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3885 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003886 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003887 gcc_copts = xnnpack_gcc_std_copts(),
3888 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003889 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003890 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003891 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003892 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003893 "@FP16",
3894 "@FXdiv",
3895 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003896 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003897 ],
3898)
3899
Marat Dukhan10a38082020-04-17 03:58:35 -07003900xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003901 name = "operators_test_mode",
3902 srcs = OPERATOR_SRCS + [
3903 "src/memory.c",
3904 "src/operator-delete.c",
3905 ],
3906 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3907 copts = LOGGING_COPTS + [
3908 "-Isrc",
3909 "-Iinclude",
3910 "-UNDEBUG",
3911 "-DXNN_TEST_MODE=1",
3912 ] + select({
3913 ":debug_build": [],
3914 "//conditions:default": xnnpack_min_size_copts(),
3915 }) + select({
3916 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3917 "//conditions:default": [],
3918 }),
3919 gcc_copts = xnnpack_gcc_std_copts(),
3920 msvc_copts = xnnpack_msvc_std_copts(),
3921 deps = [
3922 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003923 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003924 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003925 "@FP16",
3926 "@FXdiv",
3927 "@clog",
3928 "@pthreadpool",
3929 ],
3930)
3931
3932xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003933 name = "XNNPACK",
3934 srcs = [
3935 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003936 "src/runtime.c",
3937 "src/subgraph.c",
3938 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003939 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003940 hdrs = ["include/xnnpack.h"],
3941 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003942 "-Isrc",
3943 "-Iinclude",
3944 ] + select({
3945 ":debug_build": [],
3946 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003947 }) + select({
3948 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3949 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003950 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003951 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003952 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003953 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003954 visibility = xnnpack_visibility(),
3955 deps = [
3956 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003957 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003958 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003959 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003960 ":operator_run",
3961 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003962 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003963 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07003964 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003965 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07003966 ] + select({
3967 ":emscripten": [],
3968 "//conditions:default": ["@cpuinfo"],
3969 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003970)
3971
Marat Dukhan10a38082020-04-17 03:58:35 -07003972xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003973 name = "XNNPACK_test_mode",
3974 srcs = [
3975 "src/init.c",
3976 "src/runtime.c",
3977 "src/subgraph.c",
3978 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003979 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07003980 hdrs = ["include/xnnpack.h"],
3981 copts = LOGGING_COPTS + [
3982 "-Isrc",
3983 "-Iinclude",
3984 "-UNDEBUG",
3985 "-DXNN_TEST_MODE=1",
3986 ] + select({
3987 ":debug_build": [],
3988 "//conditions:default": xnnpack_min_size_copts(),
3989 }) + select({
3990 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3991 "//conditions:default": [],
3992 }),
3993 gcc_copts = xnnpack_gcc_std_copts(),
3994 includes = ["include"],
3995 msvc_copts = xnnpack_msvc_std_copts(),
3996 visibility = xnnpack_visibility(),
3997 deps = [
3998 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07003999 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004000 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004001 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004002 ":operator_run_test_mode",
4003 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004004 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004005 "@clog",
4006 "@FP16",
4007 "@pthreadpool",
4008 ] + select({
4009 ":emscripten": [],
4010 "//conditions:default": ["@cpuinfo"],
4011 }),
4012)
4013
4014xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004015 name = "xnnpack_f32",
4016 srcs = [
4017 "src/init.c",
4018 "src/runtime.c",
4019 "src/subgraph.c",
4020 "src/tensor.c",
4021 ] + SUBGRAPH_SRCS,
4022 hdrs = ["include/xnnpack.h"],
4023 copts = LOGGING_COPTS + [
4024 "-Isrc",
4025 "-Iinclude",
4026 ] + select({
4027 ":debug_build": [],
4028 "//conditions:default": xnnpack_min_size_copts(),
4029 }) + select({
4030 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4031 "//conditions:default": [],
4032 }),
4033 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004034 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004035 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004036 "XNN_NO_U8_OPERATORS",
4037 "XNN_NO_X8_OPERATORS",
4038 "XNN_NO_F16_OPERATORS",
4039 "XNN_NO_X16_OPERATORS",
4040 ],
4041 gcc_copts = xnnpack_gcc_std_copts(),
4042 includes = ["include"],
4043 msvc_copts = xnnpack_msvc_std_copts(),
4044 visibility = xnnpack_visibility(),
4045 deps = [
4046 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004047 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004048 ":logging_utils",
4049 ":memory_planner",
4050 ":operator_run",
4051 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004052 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004053 "@clog",
4054 "@FP16",
4055 "@pthreadpool",
4056 ] + select({
4057 ":emscripten": [],
4058 "//conditions:default": ["@cpuinfo"],
4059 }),
4060)
4061
4062xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004063 name = "xnnpack_operators_nhwc_f32",
4064 srcs = [
4065 "src/init.c",
4066 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004067 hdrs = ["include/xnnpack.h"],
4068 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004069 "-Isrc",
4070 "-Iinclude",
4071 ] + select({
4072 ":debug_build": [],
4073 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004074 }) + select({
4075 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4076 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004077 }),
4078 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004079 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004080 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004081 "XNN_NO_U8_OPERATORS",
4082 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004083 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004084 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004085 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004086 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004087 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004088 visibility = xnnpack_visibility(),
4089 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004090 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004091 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004092 ":operator_run",
4093 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004094 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004095 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004096 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004097 ] + select({
4098 ":emscripten": [],
4099 "//conditions:default": ["@cpuinfo"],
4100 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004101)
4102
Marat Dukhancf056b22019-10-07 10:26:29 -07004103xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004104 name = "bench_utils",
4105 srcs = ["bench/utils.cc"],
4106 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004107 deps = [
4108 "@com_google_benchmark//:benchmark",
4109 "@cpuinfo",
4110 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004111)
4112
Frank Barchard7e955972019-10-11 10:34:25 -07004113######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004114
4115xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004116 name = "qs8_gemm_bench",
4117 srcs = [
4118 "bench/gemm.h",
4119 "bench/qs8-gemm.cc",
4120 "src/xnnpack/AlignedAllocator.h",
4121 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4122 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4123)
4124
4125xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004126 name = "qs8_requantization_bench",
4127 srcs = [
4128 "bench/qs8-requantization.cc",
4129 "src/xnnpack/requantization-stubs.h",
4130 "src/xnnpack/AlignedAllocator.h",
4131 ] + MICROKERNEL_BENCHMARK_HDRS,
4132 deps = MICROKERNEL_BENCHMARK_DEPS,
4133)
4134
4135xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004136 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004137 srcs = [
4138 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004139 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004140 "src/xnnpack/AlignedAllocator.h",
4141 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004142 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004143 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144)
4145
4146xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004147 name = "qu8_requantization_bench",
4148 srcs = [
4149 "bench/qu8-requantization.cc",
4150 "src/xnnpack/requantization-stubs.h",
4151 "src/xnnpack/AlignedAllocator.h",
4152 ] + MICROKERNEL_BENCHMARK_HDRS,
4153 deps = MICROKERNEL_BENCHMARK_DEPS,
4154)
4155
4156xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004157 name = "f16_igemm_bench",
4158 srcs = [
4159 "bench/f16-igemm.cc",
4160 "bench/conv.h",
4161 "bench/google/conv.h",
4162 "src/xnnpack/AlignedAllocator.h",
4163 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004164 deps = MICROKERNEL_BENCHMARK_DEPS + [
4165 ":indirection",
4166 ":packing",
4167 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004168)
4169
4170xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004171 name = "f16_gemm_bench",
4172 srcs = [
4173 "bench/f16-gemm.cc",
4174 "bench/gemm.h",
4175 "src/xnnpack/AlignedAllocator.h",
4176 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004177 deps = MICROKERNEL_BENCHMARK_DEPS + [
4178 ":packing",
4179 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004180)
4181
4182xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004183 name = "f16_spmm_bench",
4184 srcs = [
4185 "bench/f16-spmm.cc",
4186 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004187 "src/xnnpack/AlignedAllocator.h",
4188 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004189 deps = MICROKERNEL_BENCHMARK_DEPS,
4190)
4191
4192xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004193 name = "f32_igemm_bench",
4194 srcs = [
4195 "bench/f32-igemm.cc",
4196 "bench/conv.h",
4197 "src/xnnpack/AlignedAllocator.h",
4198 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004199 deps = MICROKERNEL_BENCHMARK_DEPS + [
4200 ":indirection",
4201 ":packing",
4202 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004203)
4204
4205xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004206 name = "f16_relu_bench",
4207 srcs = [
4208 "bench/f16-relu.cc",
4209 "src/xnnpack/AlignedAllocator.h",
4210 ] + MICROKERNEL_BENCHMARK_HDRS,
4211 deps = MICROKERNEL_BENCHMARK_DEPS,
4212)
4213
4214xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004215 name = "f32_conv_hwc_bench",
4216 srcs = [
4217 "bench/f32-conv-hwc.cc",
4218 "bench/dconv.h",
4219 "src/xnnpack/AlignedAllocator.h",
4220 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004221 deps = MICROKERNEL_BENCHMARK_DEPS + [
4222 ":packing",
4223 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004224)
4225
4226xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004227 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004228 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004229 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004230 "bench/dconv.h",
4231 "src/xnnpack/AlignedAllocator.h",
4232 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004233 deps = MICROKERNEL_BENCHMARK_DEPS + [
4234 ":packing",
4235 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004236)
4237
4238xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004239 name = "f16_dwconv_bench",
4240 srcs = [
4241 "bench/f16-dwconv.cc",
4242 "bench/dwconv.h",
4243 "bench/google/dwconv.h",
4244 "src/xnnpack/AlignedAllocator.h",
4245 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004246 deps = MICROKERNEL_BENCHMARK_DEPS + [
4247 ":indirection",
4248 ":packing",
4249 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004250)
4251
4252xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004253 name = "f32_dwconv_bench",
4254 srcs = [
4255 "bench/f32-dwconv.cc",
4256 "bench/dwconv.h",
4257 "src/xnnpack/AlignedAllocator.h",
4258 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004259 deps = MICROKERNEL_BENCHMARK_DEPS + [
4260 ":indirection",
4261 ":packing",
4262 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004263)
4264
4265xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004266 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004268 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004269 "bench/dwconv.h",
4270 "src/xnnpack/AlignedAllocator.h",
4271 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004272 deps = MICROKERNEL_BENCHMARK_DEPS + [
4273 ":indirection",
4274 ":packing",
4275 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004276)
4277
4278xnnpack_benchmark(
4279 name = "f32_gemm_bench",
4280 srcs = [
4281 "bench/f32-gemm.cc",
4282 "bench/gemm.h",
4283 "src/xnnpack/AlignedAllocator.h",
4284 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004285 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004286 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004287)
4288
4289xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004290 name = "f32_hswish_bench",
4291 srcs = [
4292 "bench/f32-hswish.cc",
4293 "src/xnnpack/AlignedAllocator.h",
4294 ] + MICROKERNEL_BENCHMARK_HDRS,
4295 deps = MICROKERNEL_BENCHMARK_DEPS,
4296)
4297
4298xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004299 name = "f32_raddexpminusmax_bench",
4300 srcs = [
4301 "bench/f32-raddexpminusmax.cc",
4302 "src/xnnpack/AlignedAllocator.h",
4303 ] + MICROKERNEL_BENCHMARK_HDRS,
4304 deps = MICROKERNEL_BENCHMARK_DEPS,
4305)
4306
4307xnnpack_benchmark(
4308 name = "f32_raddextexp_bench",
4309 srcs = [
4310 "bench/f32-raddextexp.cc",
4311 "src/xnnpack/AlignedAllocator.h",
4312 ] + MICROKERNEL_BENCHMARK_HDRS,
4313 deps = MICROKERNEL_BENCHMARK_DEPS,
4314)
4315
4316xnnpack_benchmark(
4317 name = "f32_raddstoreexpminusmax_bench",
4318 srcs = [
4319 "bench/f32-raddstoreexpminusmax.cc",
4320 "src/xnnpack/AlignedAllocator.h",
4321 ] + MICROKERNEL_BENCHMARK_HDRS,
4322 deps = MICROKERNEL_BENCHMARK_DEPS,
4323)
4324
4325xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004326 name = "f32_relu_bench",
4327 srcs = [
4328 "bench/f32-relu.cc",
4329 "src/xnnpack/AlignedAllocator.h",
4330 ] + MICROKERNEL_BENCHMARK_HDRS,
4331 deps = MICROKERNEL_BENCHMARK_DEPS,
4332)
4333
4334xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004335 name = "f32_rmax_bench",
4336 srcs = [
4337 "bench/f32-rmax.cc",
4338 "src/xnnpack/AlignedAllocator.h",
4339 ] + MICROKERNEL_BENCHMARK_HDRS,
4340 deps = MICROKERNEL_BENCHMARK_DEPS,
4341)
4342
4343xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004344 name = "f32_sigmoid_bench",
4345 srcs = [
4346 "bench/f32-sigmoid.cc",
4347 "src/xnnpack/AlignedAllocator.h",
4348 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004349 deps = MICROKERNEL_BENCHMARK_DEPS,
4350)
4351
4352xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004353 name = "f32_spmm_bench",
4354 srcs = [
4355 "bench/f32-spmm.cc",
4356 "bench/gemm.h",
4357 "src/xnnpack/AlignedAllocator.h",
4358 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004359 deps = MICROKERNEL_BENCHMARK_DEPS,
4360)
4361
4362xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004363 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004364 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004365 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004366 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004367 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004368 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004369)
4370
4371xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004372 name = "f32_vscaleexpminusmax_bench",
4373 srcs = [
4374 "bench/f32-vscaleexpminusmax.cc",
4375 "src/xnnpack/AlignedAllocator.h",
4376 ] + MICROKERNEL_BENCHMARK_HDRS,
4377 deps = MICROKERNEL_BENCHMARK_DEPS,
4378)
4379
4380xnnpack_benchmark(
4381 name = "f32_vscaleextexp_bench",
4382 srcs = [
4383 "bench/f32-vscaleextexp.cc",
4384 "src/xnnpack/AlignedAllocator.h",
4385 ] + MICROKERNEL_BENCHMARK_HDRS,
4386 deps = MICROKERNEL_BENCHMARK_DEPS,
4387)
4388
4389xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004390 name = "f32_vsqrt_bench",
4391 srcs = [
4392 "bench/f32-vsqrt.cc",
4393 "src/xnnpack/AlignedAllocator.h",
4394 ] + MICROKERNEL_BENCHMARK_HDRS,
4395 deps = MICROKERNEL_BENCHMARK_DEPS,
4396)
4397
4398xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004399 name = "f32_im2col_gemm_bench",
4400 srcs = [
4401 "bench/f32-im2col-gemm.cc",
4402 "bench/conv.h",
4403 "src/xnnpack/AlignedAllocator.h",
4404 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004405 deps = MICROKERNEL_BENCHMARK_DEPS + [
4406 ":im2col",
4407 ":packing",
4408 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004409)
4410
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004411xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004412 name = "rounding_bench",
4413 srcs = [
4414 "bench/rounding.cc",
4415 "src/xnnpack/math-stubs.h",
4416 "src/xnnpack/AlignedAllocator.h",
4417 ] + MICROKERNEL_BENCHMARK_HDRS,
4418 deps = MICROKERNEL_BENCHMARK_DEPS,
4419)
4420
Marat Dukhan08c4a432019-10-03 09:29:21 -07004421########################### Benchmarks for operators ###########################
4422
4423xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004424 name = "average_pooling_bench",
4425 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004426 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004427 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004428 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004429)
4430
4431xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004432 name = "bankers_rounding_bench",
4433 srcs = ["bench/bankers-rounding.cc"],
4434 copts = xnnpack_optional_tflite_copts(),
4435 tags = ["nowin32"],
4436 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4437)
4438
4439xnnpack_benchmark(
4440 name = "ceiling_bench",
4441 srcs = ["bench/ceiling.cc"],
4442 copts = xnnpack_optional_tflite_copts(),
4443 tags = ["nowin32"],
4444 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4445)
4446
4447xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004448 name = "channel_shuffle_bench",
4449 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004450 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004451)
4452
4453xnnpack_benchmark(
4454 name = "convolution_bench",
4455 srcs = ["bench/convolution.cc"],
4456 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004457 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004458 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004459)
4460
4461xnnpack_benchmark(
4462 name = "deconvolution_bench",
4463 srcs = ["bench/deconvolution.cc"],
4464 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004465 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004466 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467)
4468
4469xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004470 name = "floor_bench",
4471 srcs = ["bench/floor.cc"],
4472 copts = xnnpack_optional_tflite_copts(),
4473 tags = ["nowin32"],
4474 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4475)
4476
4477xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004478 name = "global_average_pooling_bench",
4479 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004480 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004481)
4482
4483xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004484 name = "hardswish_bench",
4485 srcs = ["bench/hardswish.cc"],
4486 copts = xnnpack_optional_tflite_copts(),
4487 tags = ["nowin32"],
4488 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4489)
4490
4491xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004492 name = "max_pooling_bench",
4493 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004494 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004495)
4496
4497xnnpack_benchmark(
4498 name = "sigmoid_bench",
4499 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004500 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004501 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004502 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004503)
4504
4505xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004506 name = "prelu_bench",
4507 srcs = ["bench/prelu.cc"],
4508 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004509 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004510 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004511)
4512
4513xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004514 name = "softmax_bench",
4515 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004516 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004517 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004518 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004519)
4520
Marat Dukhan87727142020-06-24 15:24:10 -07004521xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004522 name = "square_root_bench",
4523 srcs = ["bench/square-root.cc"],
4524 copts = xnnpack_optional_tflite_copts(),
4525 tags = ["nowin32"],
4526 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4527)
4528
4529xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004530 name = "truncation_bench",
4531 srcs = ["bench/truncation.cc"],
4532 deps = OPERATOR_BENCHMARK_DEPS,
4533)
4534
Marat Dukhanc068bb62019-10-04 13:24:39 -07004535############################# End-to-end benchmarks ############################
4536
4537cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004538 name = "fp32_mobilenet_v1",
4539 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004540 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004541 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004542 linkstatic = True,
4543 deps = [
4544 ":XNNPACK",
4545 "@pthreadpool",
4546 ],
4547)
4548
4549cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004550 name = "fp16_mobilenet_v1",
4551 srcs = ["models/fp16-mobilenet-v1.cc"],
4552 hdrs = ["models/models.h"],
4553 copts = xnnpack_std_cxxopts(),
4554 linkstatic = True,
4555 deps = [
4556 ":XNNPACK",
4557 "@FP16",
4558 "@pthreadpool",
4559 ],
4560)
4561
4562cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004563 name = "qs8_mobilenet_v1",
4564 srcs = ["models/qs8-mobilenet-v1.cc"],
4565 hdrs = ["models/models.h"],
4566 copts = xnnpack_std_cxxopts(),
4567 linkstatic = True,
4568 deps = [
4569 ":XNNPACK",
4570 "@pthreadpool",
4571 ],
4572)
4573
4574cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004575 name = "qs8_mobilenet_v2",
4576 srcs = ["models/qs8-mobilenet-v2.cc"],
4577 hdrs = ["models/models.h"],
4578 copts = xnnpack_std_cxxopts(),
4579 linkstatic = True,
4580 deps = [
4581 ":XNNPACK",
4582 "@pthreadpool",
4583 ],
4584)
4585
4586cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004587 name = "fp32_mobilenet_v2",
4588 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004589 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004590 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004591 linkstatic = True,
4592 deps = [
4593 ":XNNPACK",
4594 "@pthreadpool",
4595 ],
4596)
4597
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004598cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004599 name = "fp16_mobilenet_v2",
4600 srcs = ["models/fp16-mobilenet-v2.cc"],
4601 hdrs = ["models/models.h"],
4602 copts = xnnpack_std_cxxopts(),
4603 linkstatic = True,
4604 deps = [
4605 ":XNNPACK",
4606 "@FP16",
4607 "@pthreadpool",
4608 ],
4609)
4610
4611cc_library(
4612 name = "fp32_mobilenet_v3_large",
4613 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004614 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004615 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004616 linkstatic = True,
4617 deps = [
4618 ":XNNPACK",
4619 "@pthreadpool",
4620 ],
4621)
4622
4623cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004624 name = "fp32_mobilenet_v3_small",
4625 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004626 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004627 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004628 linkstatic = True,
4629 deps = [
4630 ":XNNPACK",
4631 "@pthreadpool",
4632 ],
4633)
4634
Marat Dukhanc068bb62019-10-04 13:24:39 -07004635xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004636 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004637 srcs = [
4638 "bench/f32-dwconv-e2e.cc",
4639 "bench/end2end.h",
4640 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004641 deps = MICROKERNEL_BENCHMARK_DEPS + [
4642 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004643 ":fp32_mobilenet_v1",
4644 ":fp32_mobilenet_v2",
4645 ":fp32_mobilenet_v3_large",
4646 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004647 ],
4648)
4649
4650xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004651 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004652 srcs = [
4653 "bench/f32-gemm-e2e.cc",
4654 "bench/end2end.h",
4655 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004656 deps = MICROKERNEL_BENCHMARK_DEPS + [
4657 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004658 ":fp32_mobilenet_v1",
4659 ":fp32_mobilenet_v2",
4660 ":fp32_mobilenet_v3_large",
4661 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004662 ],
4663)
4664
4665xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004666 name = "end2end_bench",
4667 srcs = ["bench/end2end.cc"],
4668 deps = [
4669 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004670 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004671 ":fp16_mobilenet_v1",
4672 ":fp16_mobilenet_v2",
4673 ":fp32_mobilenet_v1",
4674 ":fp32_mobilenet_v2",
4675 ":fp32_mobilenet_v3_large",
4676 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004677 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004678 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004679 "@pthreadpool",
4680 ],
4681)
4682
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004683#################### Accuracy evaluation for math functions ####################
4684
4685xnnpack_benchmark(
4686 name = "f32_exp_eval",
4687 srcs = [
4688 "eval/f32-exp.cc",
4689 "src/xnnpack/AlignedAllocator.h",
4690 ] + ACCURACY_EVAL_HDRS,
4691 deps = ACCURACY_EVAL_DEPS,
4692)
4693
Marat Dukhan515c9772019-10-17 18:07:57 -07004694xnnpack_benchmark(
4695 name = "f32_expminus_eval",
4696 srcs = [
4697 "eval/f32-expminus.cc",
4698 "src/xnnpack/AlignedAllocator.h",
4699 ] + ACCURACY_EVAL_HDRS,
4700 deps = ACCURACY_EVAL_DEPS,
4701)
4702
Marat Dukhan98ba4412019-10-23 02:14:28 -07004703xnnpack_benchmark(
4704 name = "f32_extexp_eval",
4705 srcs = [
4706 "eval/f32-extexp.cc",
4707 "src/xnnpack/AlignedAllocator.h",
4708 ] + ACCURACY_EVAL_HDRS,
4709 deps = ACCURACY_EVAL_DEPS,
4710)
4711
Marat Dukhan8853b822020-05-07 12:19:01 -07004712xnnpack_unit_test(
4713 name = "f32_roundne_eval",
4714 srcs = [
4715 "eval/f32-roundne.cc",
4716 "src/xnnpack/AlignedAllocator.h",
4717 "src/xnnpack/math-stubs.h",
4718 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004719 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004720 deps = MICROKERNEL_TEST_DEPS,
4721)
4722
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004723xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004724 name = "f32_roundd_eval",
4725 srcs = [
4726 "eval/f32-roundd.cc",
4727 "src/xnnpack/AlignedAllocator.h",
4728 "src/xnnpack/math-stubs.h",
4729 ] + MICROKERNEL_TEST_HDRS,
4730 automatic = False,
4731 deps = MICROKERNEL_TEST_DEPS,
4732)
4733
4734xnnpack_unit_test(
4735 name = "f32_roundu_eval",
4736 srcs = [
4737 "eval/f32-roundu.cc",
4738 "src/xnnpack/AlignedAllocator.h",
4739 "src/xnnpack/math-stubs.h",
4740 ] + MICROKERNEL_TEST_HDRS,
4741 automatic = False,
4742 deps = MICROKERNEL_TEST_DEPS,
4743)
4744
4745xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004746 name = "f32_roundz_eval",
4747 srcs = [
4748 "eval/f32-roundz.cc",
4749 "src/xnnpack/AlignedAllocator.h",
4750 "src/xnnpack/math-stubs.h",
4751 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004752 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004753 deps = MICROKERNEL_TEST_DEPS,
4754)
4755
Marat Dukhan346a9e52019-11-15 09:06:30 -08004756xnnpack_benchmark(
4757 name = "f32_sigmoid_eval",
4758 srcs = [
4759 "eval/f32-sigmoid.cc",
4760 "src/xnnpack/AlignedAllocator.h",
4761 ] + ACCURACY_EVAL_HDRS,
4762 deps = ACCURACY_EVAL_DEPS,
4763)
4764
Marat Dukhan84000762020-06-29 18:38:43 -07004765xnnpack_benchmark(
4766 name = "f32_sqrt_eval",
4767 srcs = [
4768 "eval/f32-sqrt.cc",
4769 "src/xnnpack/AlignedAllocator.h",
4770 ] + ACCURACY_EVAL_HDRS,
4771 deps = ACCURACY_EVAL_DEPS + [
4772 ":bench_utils",
4773 ],
4774)
4775
Marat Dukhan08c4a432019-10-03 09:29:21 -07004776######################### Unit tests for micro-kernels #########################
4777
4778xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004779 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004780 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004781 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004782 "test/gemm-microkernel-tester.h",
4783 "src/xnnpack/AlignedAllocator.h",
4784 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004785 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004786)
4787
4788xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004789 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004790 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004791 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004792 "test/spmm-microkernel-tester.h",
4793 "src/xnnpack/AlignedAllocator.h",
4794 ] + MICROKERNEL_TEST_HDRS,
4795 deps = MICROKERNEL_TEST_DEPS,
4796)
4797
4798xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004799 name = "f32_argmaxpool_test",
4800 srcs = [
4801 "test/f32-argmaxpool.cc",
4802 "test/argmaxpool-microkernel-tester.h",
4803 "src/xnnpack/AlignedAllocator.h",
4804 ] + MICROKERNEL_TEST_HDRS,
4805 deps = MICROKERNEL_TEST_DEPS,
4806)
4807
4808xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004809 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004810 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004811 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004812 "test/avgpool-microkernel-tester.h",
4813 "src/xnnpack/AlignedAllocator.h",
4814 ] + MICROKERNEL_TEST_HDRS,
4815 deps = MICROKERNEL_TEST_DEPS,
4816)
4817
4818xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004819 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004820 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004821 "test/f32-ibilinear.cc",
4822 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004823 "src/xnnpack/AlignedAllocator.h",
4824 ] + MICROKERNEL_TEST_HDRS,
4825 deps = MICROKERNEL_TEST_DEPS,
4826)
4827
4828xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004829 name = "f16_clamp_test",
4830 srcs = [
4831 "test/f16-clamp.cc",
4832 "test/clamp-microkernel-tester.h",
4833 ] + MICROKERNEL_TEST_HDRS,
4834 deps = MICROKERNEL_TEST_DEPS,
4835)
4836
4837xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004838 name = "f32_clamp_test",
4839 srcs = [
4840 "test/f32-clamp.cc",
4841 "test/clamp-microkernel-tester.h",
4842 ] + MICROKERNEL_TEST_HDRS,
4843 deps = MICROKERNEL_TEST_DEPS,
4844)
4845
4846xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004847 name = "f32_igemm_test",
4848 srcs = [
4849 "test/f32-igemm.cc",
4850 "test/gemm-microkernel-tester.h",
4851 "src/xnnpack/AlignedAllocator.h",
4852 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004853 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004854)
4855
4856xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004857 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004858 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004859 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004860 "test/gemm-microkernel-tester.h",
4861 "src/xnnpack/AlignedAllocator.h",
4862 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004863 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004864)
4865
4866xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004867 name = "f32_igemm_minmax_test",
4868 srcs = [
4869 "test/f32-igemm-minmax.cc",
4870 "test/gemm-microkernel-tester.h",
4871 "src/xnnpack/AlignedAllocator.h",
4872 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004873 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004874)
4875
4876xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004877 name = "f16_igemm_minmax_test",
4878 srcs = [
4879 "test/f16-igemm-minmax.cc",
4880 "test/gemm-microkernel-tester.h",
4881 "src/xnnpack/AlignedAllocator.h",
4882 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004883 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004884)
4885
4886xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004887 name = "f32_conv_hwc_test",
4888 srcs = [
4889 "test/f32-conv-hwc.cc",
4890 "test/conv-hwc-microkernel-tester.h",
4891 "src/xnnpack/AlignedAllocator.h",
4892 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004893 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004894)
4895
4896xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004897 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004898 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004899 "test/f32-conv-hwc2chw.cc",
4900 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004901 "src/xnnpack/AlignedAllocator.h",
4902 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004903 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004904)
4905
4906xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004907 name = "f32_dwconv_test",
4908 srcs = [
4909 "test/f32-dwconv.cc",
4910 "test/dwconv-microkernel-tester.h",
4911 "src/xnnpack/AlignedAllocator.h",
4912 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004913 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004914)
4915
4916xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004917 name = "f16_dwconv_minmax_test",
4918 srcs = [
4919 "test/f16-dwconv-minmax.cc",
4920 "test/dwconv-microkernel-tester.h",
4921 "src/xnnpack/AlignedAllocator.h",
4922 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004923 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004924)
4925
4926xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004927 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004928 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004929 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004930 "test/dwconv-microkernel-tester.h",
4931 "src/xnnpack/AlignedAllocator.h",
4932 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004933 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004934)
4935
4936xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004937 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004938 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004939 "test/f32-dwconv-chw.cc",
4940 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004941 "src/xnnpack/AlignedAllocator.h",
4942 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004943 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004944)
4945
4946xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07004947 name = "f16_gavgpool_minmax_test",
4948 srcs = [
4949 "test/f16-gavgpool-minmax.cc",
4950 "test/gavgpool-microkernel-tester.h",
4951 "src/xnnpack/AlignedAllocator.h",
4952 ] + MICROKERNEL_TEST_HDRS,
4953 deps = MICROKERNEL_TEST_DEPS,
4954)
4955
4956xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004957 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004959 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004960 "test/gavgpool-microkernel-tester.h",
4961 "src/xnnpack/AlignedAllocator.h",
4962 ] + MICROKERNEL_TEST_HDRS,
4963 deps = MICROKERNEL_TEST_DEPS,
4964)
4965
4966xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004967 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004968 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004969 "test/f32-gavgpool-cw.cc",
4970 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004971 "src/xnnpack/AlignedAllocator.h",
4972 ] + MICROKERNEL_TEST_HDRS,
4973 deps = MICROKERNEL_TEST_DEPS,
4974)
4975
4976xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004977 name = "f32_gemm_test",
4978 srcs = [
4979 "test/f32-gemm.cc",
4980 "test/gemm-microkernel-tester.h",
4981 "src/xnnpack/AlignedAllocator.h",
4982 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004983 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004984)
4985
4986xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004987 name = "f32_gemm_relu_test",
4988 srcs = [
4989 "test/f32-gemm-relu.cc",
4990 "test/gemm-microkernel-tester.h",
4991 "src/xnnpack/AlignedAllocator.h",
4992 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004993 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07004994)
4995
4996xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004997 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004998 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07004999 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005000 "test/gemm-microkernel-tester.h",
5001 "src/xnnpack/AlignedAllocator.h",
5002 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005003 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005004)
5005
5006xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005007 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005008 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005009 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005010 "test/gemm-microkernel-tester.h",
5011 "src/xnnpack/AlignedAllocator.h",
5012 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005013 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005014)
5015
5016xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005017 name = "f16_hswish_test",
5018 srcs = [
5019 "test/f16-hswish.cc",
5020 "test/hswish-microkernel-tester.h",
5021 ] + MICROKERNEL_TEST_HDRS,
5022 deps = MICROKERNEL_TEST_DEPS,
5023)
5024
5025xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005026 name = "f32_hswish_test",
5027 srcs = [
5028 "test/f32-hswish.cc",
5029 "test/hswish-microkernel-tester.h",
5030 ] + MICROKERNEL_TEST_HDRS,
5031 deps = MICROKERNEL_TEST_DEPS,
5032)
5033
5034xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005035 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005036 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005037 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005038 "test/maxpool-microkernel-tester.h",
5039 ] + MICROKERNEL_TEST_HDRS,
5040 deps = MICROKERNEL_TEST_DEPS,
5041)
5042
5043xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005044 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005045 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005046 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005047 "test/avgpool-microkernel-tester.h",
5048 "src/xnnpack/AlignedAllocator.h",
5049 ] + MICROKERNEL_TEST_HDRS,
5050 deps = MICROKERNEL_TEST_DEPS,
5051)
5052
5053xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005054 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005055 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005056 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005057 "test/gemm-microkernel-tester.h",
5058 "src/xnnpack/AlignedAllocator.h",
5059 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005060 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005061)
5062
5063xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005064 name = "f16_prelu_test",
5065 srcs = [
5066 "test/f16-prelu.cc",
5067 "test/prelu-microkernel-tester.h",
5068 "src/xnnpack/AlignedAllocator.h",
5069 ] + MICROKERNEL_TEST_HDRS,
5070 deps = MICROKERNEL_TEST_DEPS,
5071)
5072
5073xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005074 name = "f32_prelu_test",
5075 srcs = [
5076 "test/f32-prelu.cc",
5077 "test/prelu-microkernel-tester.h",
5078 "src/xnnpack/AlignedAllocator.h",
5079 ] + MICROKERNEL_TEST_HDRS,
5080 deps = MICROKERNEL_TEST_DEPS,
5081)
5082
5083xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005084 name = "f32_raddexpminusmax_test",
5085 srcs = [
5086 "test/f32-raddexpminusmax.cc",
5087 "test/raddexpminusmax-microkernel-tester.h",
5088 ] + MICROKERNEL_TEST_HDRS,
5089 deps = MICROKERNEL_TEST_DEPS,
5090)
5091
5092xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005093 name = "f32_raddextexp_test",
5094 srcs = [
5095 "test/f32-raddextexp.cc",
5096 "test/raddextexp-microkernel-tester.h",
5097 ] + MICROKERNEL_TEST_HDRS,
5098 deps = MICROKERNEL_TEST_DEPS,
5099)
5100
5101xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005102 name = "f32_raddstoreexpminusmax_test",
5103 srcs = [
5104 "test/f32-raddstoreexpminusmax.cc",
5105 "test/raddstoreexpminusmax-microkernel-tester.h",
5106 ] + MICROKERNEL_TEST_HDRS,
5107 deps = MICROKERNEL_TEST_DEPS,
5108)
5109
5110xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005111 name = "f32_relu_test",
5112 srcs = [
5113 "test/f32-relu.cc",
5114 "test/vunary-microkernel-tester.h",
5115 ] + MICROKERNEL_TEST_HDRS,
5116 deps = MICROKERNEL_TEST_DEPS,
5117)
5118
5119xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005120 name = "f32_rmax_test",
5121 srcs = [
5122 "test/f32-rmax.cc",
5123 "test/rmax-microkernel-tester.h",
5124 ] + MICROKERNEL_TEST_HDRS,
5125 deps = MICROKERNEL_TEST_DEPS,
5126)
5127
5128xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005129 name = "f32_sigmoid_test",
5130 srcs = [
5131 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005132 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005133 ] + MICROKERNEL_TEST_HDRS,
5134 deps = MICROKERNEL_TEST_DEPS,
5135)
5136
5137xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005138 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005139 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005140 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005141 "test/spmm-microkernel-tester.h",
5142 "src/xnnpack/AlignedAllocator.h",
5143 ] + MICROKERNEL_TEST_HDRS,
5144 deps = MICROKERNEL_TEST_DEPS,
5145)
5146
5147xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005148 name = "f16_vadd_minmax_test",
5149 srcs = [
5150 "test/f16-vadd-minmax.cc",
5151 "test/vbinary-microkernel-tester.h",
5152 ] + MICROKERNEL_TEST_HDRS,
5153 deps = MICROKERNEL_TEST_DEPS,
5154)
5155
5156xnnpack_unit_test(
5157 name = "f16_vaddc_minmax_test",
5158 srcs = [
5159 "test/f16-vaddc-minmax.cc",
5160 "test/vbinaryc-microkernel-tester.h",
5161 ] + MICROKERNEL_TEST_HDRS,
5162 deps = MICROKERNEL_TEST_DEPS,
5163)
5164
5165xnnpack_unit_test(
5166 name = "f16_vdiv_minmax_test",
5167 srcs = [
5168 "test/f16-vdiv-minmax.cc",
5169 "test/vbinary-microkernel-tester.h",
5170 ] + MICROKERNEL_TEST_HDRS,
5171 deps = MICROKERNEL_TEST_DEPS,
5172)
5173
5174xnnpack_unit_test(
5175 name = "f16_vdivc_minmax_test",
5176 srcs = [
5177 "test/f16-vdivc-minmax.cc",
5178 "test/vbinaryc-microkernel-tester.h",
5179 ] + MICROKERNEL_TEST_HDRS,
5180 deps = MICROKERNEL_TEST_DEPS,
5181)
5182
5183xnnpack_unit_test(
5184 name = "f16_vrdivc_minmax_test",
5185 srcs = [
5186 "test/f16-vrdivc-minmax.cc",
5187 "test/vbinaryc-microkernel-tester.h",
5188 ] + MICROKERNEL_TEST_HDRS,
5189 deps = MICROKERNEL_TEST_DEPS,
5190)
5191
5192xnnpack_unit_test(
5193 name = "f16_vmax_test",
5194 srcs = [
5195 "test/f16-vmax.cc",
5196 "test/vbinary-microkernel-tester.h",
5197 ] + MICROKERNEL_TEST_HDRS,
5198 deps = MICROKERNEL_TEST_DEPS,
5199)
5200
5201xnnpack_unit_test(
5202 name = "f16_vmaxc_test",
5203 srcs = [
5204 "test/f16-vmaxc.cc",
5205 "test/vbinaryc-microkernel-tester.h",
5206 ] + MICROKERNEL_TEST_HDRS,
5207 deps = MICROKERNEL_TEST_DEPS,
5208)
5209
5210xnnpack_unit_test(
5211 name = "f16_vmin_test",
5212 srcs = [
5213 "test/f16-vmin.cc",
5214 "test/vbinary-microkernel-tester.h",
5215 ] + MICROKERNEL_TEST_HDRS,
5216 deps = MICROKERNEL_TEST_DEPS,
5217)
5218
5219xnnpack_unit_test(
5220 name = "f16_vminc_test",
5221 srcs = [
5222 "test/f16-vminc.cc",
5223 "test/vbinaryc-microkernel-tester.h",
5224 ] + MICROKERNEL_TEST_HDRS,
5225 deps = MICROKERNEL_TEST_DEPS,
5226)
5227
5228xnnpack_unit_test(
5229 name = "f16_vmul_minmax_test",
5230 srcs = [
5231 "test/f16-vmul-minmax.cc",
5232 "test/vbinary-microkernel-tester.h",
5233 ] + MICROKERNEL_TEST_HDRS,
5234 deps = MICROKERNEL_TEST_DEPS,
5235)
5236
5237xnnpack_unit_test(
5238 name = "f16_vmulc_minmax_test",
5239 srcs = [
5240 "test/f16-vmulc-minmax.cc",
5241 "test/vbinaryc-microkernel-tester.h",
5242 ] + MICROKERNEL_TEST_HDRS,
5243 deps = MICROKERNEL_TEST_DEPS,
5244)
5245
5246xnnpack_unit_test(
5247 name = "f16_vsub_minmax_test",
5248 srcs = [
5249 "test/f16-vsub-minmax.cc",
5250 "test/vbinary-microkernel-tester.h",
5251 ] + MICROKERNEL_TEST_HDRS,
5252 deps = MICROKERNEL_TEST_DEPS,
5253)
5254
5255xnnpack_unit_test(
5256 name = "f16_vsubc_minmax_test",
5257 srcs = [
5258 "test/f16-vsubc-minmax.cc",
5259 "test/vbinaryc-microkernel-tester.h",
5260 ] + MICROKERNEL_TEST_HDRS,
5261 deps = MICROKERNEL_TEST_DEPS,
5262)
5263
5264xnnpack_unit_test(
5265 name = "f16_vrsubc_minmax_test",
5266 srcs = [
5267 "test/f16-vrsubc-minmax.cc",
5268 "test/vbinaryc-microkernel-tester.h",
5269 ] + MICROKERNEL_TEST_HDRS,
5270 deps = MICROKERNEL_TEST_DEPS,
5271)
5272
5273xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005274 name = "f32_vabs_test",
5275 srcs = [
5276 "test/f32-vabs.cc",
5277 "test/vunary-microkernel-tester.h",
5278 ] + MICROKERNEL_TEST_HDRS,
5279 deps = MICROKERNEL_TEST_DEPS,
5280)
5281
5282xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005283 name = "f32_vadd_test",
5284 srcs = [
5285 "test/f32-vadd.cc",
5286 "test/vbinary-microkernel-tester.h",
5287 ] + MICROKERNEL_TEST_HDRS,
5288 deps = MICROKERNEL_TEST_DEPS,
5289)
5290
5291xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005292 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005293 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005294 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005295 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005296 ] + MICROKERNEL_TEST_HDRS,
5297 deps = MICROKERNEL_TEST_DEPS,
5298)
5299
5300xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005301 name = "f32_vadd_relu_test",
5302 srcs = [
5303 "test/f32-vadd-relu.cc",
5304 "test/vbinary-microkernel-tester.h",
5305 ] + MICROKERNEL_TEST_HDRS,
5306 deps = MICROKERNEL_TEST_DEPS,
5307)
5308
5309xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005310 name = "f32_vaddc_test",
5311 srcs = [
5312 "test/f32-vaddc.cc",
5313 "test/vbinaryc-microkernel-tester.h",
5314 ] + MICROKERNEL_TEST_HDRS,
5315 deps = MICROKERNEL_TEST_DEPS,
5316)
5317
5318xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005319 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005320 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005321 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005322 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005323 ] + MICROKERNEL_TEST_HDRS,
5324 deps = MICROKERNEL_TEST_DEPS,
5325)
5326
5327xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005328 name = "f32_vaddc_relu_test",
5329 srcs = [
5330 "test/f32-vaddc-relu.cc",
5331 "test/vbinaryc-microkernel-tester.h",
5332 ] + MICROKERNEL_TEST_HDRS,
5333 deps = MICROKERNEL_TEST_DEPS,
5334)
5335
5336xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005337 name = "f32_vdiv_test",
5338 srcs = [
5339 "test/f32-vdiv.cc",
5340 "test/vbinary-microkernel-tester.h",
5341 ] + MICROKERNEL_TEST_HDRS,
5342 deps = MICROKERNEL_TEST_DEPS,
5343)
5344
5345xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005346 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005347 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005348 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005349 "test/vbinary-microkernel-tester.h",
5350 ] + MICROKERNEL_TEST_HDRS,
5351 deps = MICROKERNEL_TEST_DEPS,
5352)
5353
5354xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005355 name = "f32_vdiv_relu_test",
5356 srcs = [
5357 "test/f32-vdiv-relu.cc",
5358 "test/vbinary-microkernel-tester.h",
5359 ] + MICROKERNEL_TEST_HDRS,
5360 deps = MICROKERNEL_TEST_DEPS,
5361)
5362
5363xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005364 name = "f32_vdivc_test",
5365 srcs = [
5366 "test/f32-vdivc.cc",
5367 "test/vbinaryc-microkernel-tester.h",
5368 ] + MICROKERNEL_TEST_HDRS,
5369 deps = MICROKERNEL_TEST_DEPS,
5370)
5371
5372xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005373 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005374 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005375 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005376 "test/vbinaryc-microkernel-tester.h",
5377 ] + MICROKERNEL_TEST_HDRS,
5378 deps = MICROKERNEL_TEST_DEPS,
5379)
5380
5381xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005382 name = "f32_vdivc_relu_test",
5383 srcs = [
5384 "test/f32-vdivc-relu.cc",
5385 "test/vbinaryc-microkernel-tester.h",
5386 ] + MICROKERNEL_TEST_HDRS,
5387 deps = MICROKERNEL_TEST_DEPS,
5388)
5389
5390xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005391 name = "f32_vrdivc_test",
5392 srcs = [
5393 "test/f32-vrdivc.cc",
5394 "test/vbinaryc-microkernel-tester.h",
5395 ] + MICROKERNEL_TEST_HDRS,
5396 deps = MICROKERNEL_TEST_DEPS,
5397)
5398
5399xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005400 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005401 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005402 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005403 "test/vbinaryc-microkernel-tester.h",
5404 ] + MICROKERNEL_TEST_HDRS,
5405 deps = MICROKERNEL_TEST_DEPS,
5406)
5407
5408xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005409 name = "f32_vrdivc_relu_test",
5410 srcs = [
5411 "test/f32-vrdivc-relu.cc",
5412 "test/vbinaryc-microkernel-tester.h",
5413 ] + MICROKERNEL_TEST_HDRS,
5414 deps = MICROKERNEL_TEST_DEPS,
5415)
5416
5417xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005418 name = "f32_vmax_test",
5419 srcs = [
5420 "test/f32-vmax.cc",
5421 "test/vbinary-microkernel-tester.h",
5422 ] + MICROKERNEL_TEST_HDRS,
5423 deps = MICROKERNEL_TEST_DEPS,
5424)
5425
5426xnnpack_unit_test(
5427 name = "f32_vmaxc_test",
5428 srcs = [
5429 "test/f32-vmaxc.cc",
5430 "test/vbinaryc-microkernel-tester.h",
5431 ] + MICROKERNEL_TEST_HDRS,
5432 deps = MICROKERNEL_TEST_DEPS,
5433)
5434
5435xnnpack_unit_test(
5436 name = "f32_vmin_test",
5437 srcs = [
5438 "test/f32-vmin.cc",
5439 "test/vbinary-microkernel-tester.h",
5440 ] + MICROKERNEL_TEST_HDRS,
5441 deps = MICROKERNEL_TEST_DEPS,
5442)
5443
5444xnnpack_unit_test(
5445 name = "f32_vminc_test",
5446 srcs = [
5447 "test/f32-vminc.cc",
5448 "test/vbinaryc-microkernel-tester.h",
5449 ] + MICROKERNEL_TEST_HDRS,
5450 deps = MICROKERNEL_TEST_DEPS,
5451)
5452
5453xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005454 name = "f32_vmul_test",
5455 srcs = [
5456 "test/f32-vmul.cc",
5457 "test/vbinary-microkernel-tester.h",
5458 ] + MICROKERNEL_TEST_HDRS,
5459 deps = MICROKERNEL_TEST_DEPS,
5460)
5461
5462xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005463 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005464 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005465 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005466 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005467 ] + MICROKERNEL_TEST_HDRS,
5468 deps = MICROKERNEL_TEST_DEPS,
5469)
5470
5471xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005472 name = "f32_vmul_relu_test",
5473 srcs = [
5474 "test/f32-vmul-relu.cc",
5475 "test/vbinary-microkernel-tester.h",
5476 ] + MICROKERNEL_TEST_HDRS,
5477 deps = MICROKERNEL_TEST_DEPS,
5478)
5479
5480xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005481 name = "f32_vmulc_test",
5482 srcs = [
5483 "test/f32-vmulc.cc",
5484 "test/vbinaryc-microkernel-tester.h",
5485 ] + MICROKERNEL_TEST_HDRS,
5486 deps = MICROKERNEL_TEST_DEPS,
5487)
5488
5489xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005490 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005491 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005492 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005493 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005494 ] + MICROKERNEL_TEST_HDRS,
5495 deps = MICROKERNEL_TEST_DEPS,
5496)
5497
5498xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005499 name = "f32_vmulc_relu_test",
5500 srcs = [
5501 "test/f32-vmulc-relu.cc",
5502 "test/vbinaryc-microkernel-tester.h",
5503 ] + MICROKERNEL_TEST_HDRS,
5504 deps = MICROKERNEL_TEST_DEPS,
5505)
5506
5507xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005508 name = "f16_vmulcaddc_minmax_test",
5509 srcs = [
5510 "test/f16-vmulcaddc-minmax.cc",
5511 "test/vmulcaddc-microkernel-tester.h",
5512 "src/xnnpack/AlignedAllocator.h",
5513 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005514 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005515)
5516
5517xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005518 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005519 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005520 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005521 "test/vmulcaddc-microkernel-tester.h",
5522 "src/xnnpack/AlignedAllocator.h",
5523 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005524 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005525)
5526
5527xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005528 name = "f32_vlrelu_test",
5529 srcs = [
5530 "test/f32-vlrelu.cc",
5531 "test/vunary-microkernel-tester.h",
5532 ] + MICROKERNEL_TEST_HDRS,
5533 deps = MICROKERNEL_TEST_DEPS,
5534)
5535
5536xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005537 name = "f32_vneg_test",
5538 srcs = [
5539 "test/f32-vneg.cc",
5540 "test/vunary-microkernel-tester.h",
5541 ] + MICROKERNEL_TEST_HDRS,
5542 deps = MICROKERNEL_TEST_DEPS,
5543)
5544
5545xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005546 name = "f32_vrndne_test",
5547 srcs = [
5548 "test/f32-vrndne.cc",
5549 "test/vunary-microkernel-tester.h",
5550 ] + MICROKERNEL_TEST_HDRS,
5551 deps = MICROKERNEL_TEST_DEPS,
5552)
5553
5554xnnpack_unit_test(
5555 name = "f32_vrndz_test",
5556 srcs = [
5557 "test/f32-vrndz.cc",
5558 "test/vunary-microkernel-tester.h",
5559 ] + MICROKERNEL_TEST_HDRS,
5560 deps = MICROKERNEL_TEST_DEPS,
5561)
5562
5563xnnpack_unit_test(
5564 name = "f32_vrndu_test",
5565 srcs = [
5566 "test/f32-vrndu.cc",
5567 "test/vunary-microkernel-tester.h",
5568 ] + MICROKERNEL_TEST_HDRS,
5569 deps = MICROKERNEL_TEST_DEPS,
5570)
5571
5572xnnpack_unit_test(
5573 name = "f32_vrndd_test",
5574 srcs = [
5575 "test/f32-vrndd.cc",
5576 "test/vunary-microkernel-tester.h",
5577 ] + MICROKERNEL_TEST_HDRS,
5578 deps = MICROKERNEL_TEST_DEPS,
5579)
5580
5581xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005582 name = "f32_vscale_test",
5583 srcs = [
5584 "test/f32-vscale.cc",
5585 "test/vscale-microkernel-tester.h",
5586 ] + MICROKERNEL_TEST_HDRS,
5587 deps = MICROKERNEL_TEST_DEPS,
5588)
5589
5590xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005591 name = "f32_vscaleexpminusmax_test",
5592 srcs = [
5593 "test/f32-vscaleexpminusmax.cc",
5594 "test/vscaleexpminusmax-microkernel-tester.h",
5595 ] + MICROKERNEL_TEST_HDRS,
5596 deps = MICROKERNEL_TEST_DEPS,
5597)
5598
5599xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005600 name = "f32_vscaleextexp_test",
5601 srcs = [
5602 "test/f32-vscaleextexp.cc",
5603 "test/vscaleextexp-microkernel-tester.h",
5604 ] + MICROKERNEL_TEST_HDRS,
5605 deps = MICROKERNEL_TEST_DEPS,
5606)
5607
5608xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005609 name = "f32_vsqr_test",
5610 srcs = [
5611 "test/f32-vsqr.cc",
5612 "test/vunary-microkernel-tester.h",
5613 ] + MICROKERNEL_TEST_HDRS,
5614 deps = MICROKERNEL_TEST_DEPS,
5615)
5616
5617xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005618 name = "f32_vsqrdiff_test",
5619 srcs = [
5620 "test/f32-vsqrdiff.cc",
5621 "test/vbinary-microkernel-tester.h",
5622 ] + MICROKERNEL_TEST_HDRS,
5623 deps = MICROKERNEL_TEST_DEPS,
5624)
5625
5626xnnpack_unit_test(
5627 name = "f32_vsqrdiffc_test",
5628 srcs = [
5629 "test/f32-vsqrdiffc.cc",
5630 "test/vbinaryc-microkernel-tester.h",
5631 ] + MICROKERNEL_TEST_HDRS,
5632 deps = MICROKERNEL_TEST_DEPS,
5633)
5634
5635xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005636 name = "f32_vsqrt_test",
5637 srcs = [
5638 "test/f32-vsqrt.cc",
5639 "test/vunary-microkernel-tester.h",
5640 ] + MICROKERNEL_TEST_HDRS,
5641 deps = MICROKERNEL_TEST_DEPS,
5642)
5643
5644xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005645 name = "f32_vsub_test",
5646 srcs = [
5647 "test/f32-vsub.cc",
5648 "test/vbinary-microkernel-tester.h",
5649 ] + MICROKERNEL_TEST_HDRS,
5650 deps = MICROKERNEL_TEST_DEPS,
5651)
5652
5653xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005654 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005655 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005656 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005657 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005658 ] + MICROKERNEL_TEST_HDRS,
5659 deps = MICROKERNEL_TEST_DEPS,
5660)
5661
5662xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005663 name = "f32_vsub_relu_test",
5664 srcs = [
5665 "test/f32-vsub-relu.cc",
5666 "test/vbinary-microkernel-tester.h",
5667 ] + MICROKERNEL_TEST_HDRS,
5668 deps = MICROKERNEL_TEST_DEPS,
5669)
5670
5671xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005672 name = "f32_vsubc_test",
5673 srcs = [
5674 "test/f32-vsubc.cc",
5675 "test/vbinaryc-microkernel-tester.h",
5676 ] + MICROKERNEL_TEST_HDRS,
5677 deps = MICROKERNEL_TEST_DEPS,
5678)
5679
5680xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005681 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005682 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005683 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005684 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005685 ] + MICROKERNEL_TEST_HDRS,
5686 deps = MICROKERNEL_TEST_DEPS,
5687)
5688
5689xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005690 name = "f32_vsubc_relu_test",
5691 srcs = [
5692 "test/f32-vsubc-relu.cc",
5693 "test/vbinaryc-microkernel-tester.h",
5694 ] + MICROKERNEL_TEST_HDRS,
5695 deps = MICROKERNEL_TEST_DEPS,
5696)
5697
5698xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005699 name = "f32_vrsubc_test",
5700 srcs = [
5701 "test/f32-vrsubc.cc",
5702 "test/vbinaryc-microkernel-tester.h",
5703 ] + MICROKERNEL_TEST_HDRS,
5704 deps = MICROKERNEL_TEST_DEPS,
5705)
5706
5707xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005708 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005709 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005710 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005711 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005712 ] + MICROKERNEL_TEST_HDRS,
5713 deps = MICROKERNEL_TEST_DEPS,
5714)
5715
5716xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005717 name = "f32_vrsubc_relu_test",
5718 srcs = [
5719 "test/f32-vrsubc-relu.cc",
5720 "test/vbinaryc-microkernel-tester.h",
5721 ] + MICROKERNEL_TEST_HDRS,
5722 deps = MICROKERNEL_TEST_DEPS,
5723)
5724
5725xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005726 name = "qs8_dwconv_minmax_test",
5727 srcs = [
5728 "test/qs8-dwconv-minmax.cc",
5729 "test/dwconv-microkernel-tester.h",
5730 "src/xnnpack/AlignedAllocator.h",
5731 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5732 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5733)
5734
5735xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005736 name = "qs8_gavgpool_minmax_test",
5737 srcs = [
5738 "test/qs8-gavgpool-minmax.cc",
5739 "test/gavgpool-microkernel-tester.h",
5740 "src/xnnpack/AlignedAllocator.h",
5741 ] + MICROKERNEL_TEST_HDRS,
5742 deps = MICROKERNEL_TEST_DEPS,
5743)
5744
5745xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005746 name = "qs8_gemm_minmax_test",
5747 srcs = [
5748 "test/qs8-gemm-minmax.cc",
5749 "test/gemm-microkernel-tester.h",
5750 "src/xnnpack/AlignedAllocator.h",
5751 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5752 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5753)
5754
5755xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005756 name = "qs8_igemm_minmax_test",
5757 srcs = [
5758 "test/qs8-igemm-minmax.cc",
5759 "test/gemm-microkernel-tester.h",
5760 "src/xnnpack/AlignedAllocator.h",
5761 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5762 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5763)
5764
5765xnnpack_unit_test(
5766 name = "qs8_requantization_test",
5767 srcs = [
5768 "src/xnnpack/requantization-stubs.h",
5769 "test/qs8-requantization.cc",
5770 "test/requantization-tester.h",
5771 ] + MICROKERNEL_TEST_HDRS,
5772 deps = MICROKERNEL_TEST_DEPS,
5773)
5774
5775xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005776 name = "qs8_vadd_minmax_test",
5777 srcs = [
5778 "test/qs8-vadd-minmax.cc",
5779 "test/vadd-microkernel-tester.h",
5780 ] + MICROKERNEL_TEST_HDRS,
5781 deps = MICROKERNEL_TEST_DEPS,
5782)
5783
5784xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005785 name = "qs8_vaddc_minmax_test",
5786 srcs = [
5787 "test/qs8-vaddc-minmax.cc",
5788 "test/vaddc-microkernel-tester.h",
5789 ] + MICROKERNEL_TEST_HDRS,
5790 deps = MICROKERNEL_TEST_DEPS,
5791)
5792
5793xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005794 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005795 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005796 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005797 "test/avgpool-microkernel-tester.h",
5798 "src/xnnpack/AlignedAllocator.h",
5799 ] + MICROKERNEL_TEST_HDRS,
5800 deps = MICROKERNEL_TEST_DEPS,
5801)
5802
5803xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005804 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005805 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005806 "test/qu8-dwconv-minmax.cc",
5807 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005808 "src/xnnpack/AlignedAllocator.h",
5809 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005810 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005811)
5812
5813xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005814 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005815 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005816 "test/qu8-igemm-minmax.cc",
5817 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005818 "src/xnnpack/AlignedAllocator.h",
5819 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005820 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005821)
5822
5823xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005824 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005825 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005826 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005827 "test/gavgpool-microkernel-tester.h",
5828 "src/xnnpack/AlignedAllocator.h",
5829 ] + MICROKERNEL_TEST_HDRS,
5830 deps = MICROKERNEL_TEST_DEPS,
5831)
5832
5833xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005834 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005835 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005836 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005837 "test/gemm-microkernel-tester.h",
5838 "src/xnnpack/AlignedAllocator.h",
5839 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005840 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005841)
5842
5843xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005844 name = "qu8_requantization_test",
5845 srcs = [
5846 "src/xnnpack/requantization-stubs.h",
5847 "test/qu8-requantization.cc",
5848 "test/requantization-tester.h",
5849 ] + MICROKERNEL_TEST_HDRS,
5850 deps = MICROKERNEL_TEST_DEPS,
5851)
5852
5853xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005854 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005855 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005856 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005857 "test/vadd-microkernel-tester.h",
5858 ] + MICROKERNEL_TEST_HDRS,
5859 deps = MICROKERNEL_TEST_DEPS,
5860)
5861
5862xnnpack_unit_test(
5863 name = "u8_clamp_test",
5864 srcs = [
5865 "test/u8-clamp.cc",
5866 "test/clamp-microkernel-tester.h",
5867 ] + MICROKERNEL_TEST_HDRS,
5868 deps = MICROKERNEL_TEST_DEPS,
5869)
5870
5871xnnpack_unit_test(
5872 name = "u8_lut32norm_test",
5873 srcs = [
5874 "test/u8-lut32norm.cc",
5875 "test/lut-norm-microkernel-tester.h",
5876 ] + MICROKERNEL_TEST_HDRS,
5877 deps = MICROKERNEL_TEST_DEPS,
5878)
5879
5880xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005881 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005882 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005883 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005884 "test/maxpool-microkernel-tester.h",
5885 ] + MICROKERNEL_TEST_HDRS,
5886 deps = MICROKERNEL_TEST_DEPS,
5887)
5888
5889xnnpack_unit_test(
5890 name = "u8_rmax_test",
5891 srcs = [
5892 "test/u8-rmax.cc",
5893 "test/rmax-microkernel-tester.h",
5894 ] + MICROKERNEL_TEST_HDRS,
5895 deps = MICROKERNEL_TEST_DEPS,
5896)
5897
5898xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005899 name = "x32_fill_test",
5900 srcs = [
5901 "test/x32-fill.cc",
5902 "test/fill-microkernel-tester.h",
5903 ] + MICROKERNEL_TEST_HDRS,
5904 deps = MICROKERNEL_TEST_DEPS,
5905)
5906
5907xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908 name = "x32_packx_test",
5909 srcs = [
5910 "test/x32-packx.cc",
5911 "test/pack-microkernel-tester.h",
5912 "src/xnnpack/AlignedAllocator.h",
5913 ] + MICROKERNEL_TEST_HDRS,
5914 deps = MICROKERNEL_TEST_DEPS,
5915)
5916
5917xnnpack_unit_test(
5918 name = "x32_pad_test",
5919 srcs = [
5920 "test/x32-pad.cc",
5921 "test/pad-microkernel-tester.h",
5922 ] + MICROKERNEL_TEST_HDRS,
5923 deps = MICROKERNEL_TEST_DEPS,
5924)
5925
5926xnnpack_unit_test(
5927 name = "x32_unpool_test",
5928 srcs = [
5929 "test/x32-unpool.cc",
5930 "test/unpool-microkernel-tester.h",
5931 ] + MICROKERNEL_TEST_HDRS,
5932 deps = MICROKERNEL_TEST_DEPS,
5933)
5934
5935xnnpack_unit_test(
5936 name = "x32_zip_test",
5937 srcs = [
5938 "test/x32-zip.cc",
5939 "test/zip-microkernel-tester.h",
5940 ] + MICROKERNEL_TEST_HDRS,
5941 deps = MICROKERNEL_TEST_DEPS,
5942)
5943
5944xnnpack_unit_test(
5945 name = "x8_lut_test",
5946 srcs = [
5947 "test/x8-lut.cc",
5948 "test/lut-microkernel-tester.h",
5949 ] + MICROKERNEL_TEST_HDRS,
5950 deps = MICROKERNEL_TEST_DEPS,
5951)
5952
5953xnnpack_unit_test(
5954 name = "x8_zip_test",
5955 srcs = [
5956 "test/x8-zip.cc",
5957 "test/zip-microkernel-tester.h",
5958 ] + MICROKERNEL_TEST_HDRS,
5959 deps = MICROKERNEL_TEST_DEPS,
5960)
5961
Marat Dukhan20c3b922020-03-10 03:45:06 -07005962########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005963
5964xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07005965 name = "operator_size_test",
5966 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07005967 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005968)
5969
Marat Dukhan20c3b922020-03-10 03:45:06 -07005970xnnpack_binary(
5971 name = "subgraph_size_test",
5972 srcs = ["test/subgraph-size.c"],
5973 deps = [":XNNPACK"],
5974)
5975
5976########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07005977
5978xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005979 name = "abs_nc_test",
5980 srcs = [
5981 "test/abs-nc.cc",
5982 "test/abs-operator-tester.h",
5983 ],
5984 deps = OPERATOR_TEST_DEPS,
5985)
5986
5987xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005988 name = "add_nd_test",
5989 srcs = [
5990 "test/add-nd.cc",
5991 "test/binary-elementwise-operator-tester.h",
5992 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07005993 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08005994)
5995
5996xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08005997 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005998 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08005999 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006000 "test/argmax-pooling-operator-tester.h",
6001 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006002 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006003)
6004
6005xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006006 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006007 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006008 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006009 "test/average-pooling-operator-tester.h",
6010 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006011 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006012)
6013
6014xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006015 name = "bankers_rounding_nc_test",
6016 srcs = [
6017 "test/bankers-rounding-nc.cc",
6018 "test/bankers-rounding-operator-tester.h",
6019 ],
6020 deps = OPERATOR_TEST_DEPS,
6021)
6022
6023xnnpack_unit_test(
6024 name = "ceiling_nc_test",
6025 srcs = [
6026 "test/ceiling-nc.cc",
6027 "test/ceiling-operator-tester.h",
6028 ],
6029 deps = OPERATOR_TEST_DEPS,
6030)
6031
6032xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006033 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006035 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006036 "test/channel-shuffle-operator-tester.h",
6037 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006038 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006039)
6040
6041xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006042 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006043 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006044 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006045 "test/clamp-operator-tester.h",
6046 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006047 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006048)
6049
6050xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006051 name = "constant_pad_nd_test",
6052 srcs = [
6053 "test/constant-pad-nd.cc",
6054 "test/constant-pad-operator-tester.h",
6055 ],
6056 deps = OPERATOR_TEST_DEPS,
6057)
6058
6059xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006060 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006061 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006062 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006063 "test/convolution-operator-tester.h",
6064 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006065 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006066)
6067
6068xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006069 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006070 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006071 "test/convolution-nchw.cc",
6072 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006073 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006074 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006075)
6076
6077xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006078 name = "copy_nc_test",
6079 srcs = [
6080 "test/copy-nc.cc",
6081 "test/copy-operator-tester.h",
6082 ],
6083 deps = OPERATOR_TEST_DEPS,
6084)
6085
6086xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006087 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006088 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006089 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006090 "test/deconvolution-operator-tester.h",
6091 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006092 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006093)
6094
6095xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006096 name = "divide_nd_test",
6097 srcs = [
6098 "test/binary-elementwise-operator-tester.h",
6099 "test/divide-nd.cc",
6100 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006101 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006102)
6103
6104xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006105 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006106 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006107 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006108 "test/fully-connected-operator-tester.h",
6109 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006110 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006111)
6112
6113xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006114 name = "floor_nc_test",
6115 srcs = [
6116 "test/floor-nc.cc",
6117 "test/floor-operator-tester.h",
6118 ],
6119 deps = OPERATOR_TEST_DEPS,
6120)
6121
6122xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006123 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006124 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006125 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006126 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006127 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006128 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006129)
6130
6131xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006132 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006133 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006134 "test/global-average-pooling-ncw.cc",
6135 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006136 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006137 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006138)
6139
6140xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006141 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006142 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006143 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 "test/hardswish-operator-tester.h",
6145 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006146 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006147)
6148
6149xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006150 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006151 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006152 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006153 "test/leaky-relu-operator-tester.h",
6154 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006155 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006156)
6157
6158xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006159 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006160 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006161 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006162 "test/max-pooling-operator-tester.h",
6163 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006164 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006165)
6166
6167xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006168 name = "maximum_nd_test",
6169 srcs = [
6170 "test/binary-elementwise-operator-tester.h",
6171 "test/maximum-nd.cc",
6172 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006173 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006174)
6175
6176xnnpack_unit_test(
6177 name = "minimum_nd_test",
6178 srcs = [
6179 "test/binary-elementwise-operator-tester.h",
6180 "test/minimum-nd.cc",
6181 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006182 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006183)
6184
6185xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006186 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006187 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006188 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006189 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006190 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006191 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006192)
6193
6194xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006195 name = "negate_nc_test",
6196 srcs = [
6197 "test/negate-nc.cc",
6198 "test/negate-operator-tester.h",
6199 ],
6200 deps = OPERATOR_TEST_DEPS,
6201)
6202
6203xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006204 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006205 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006206 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207 "test/prelu-operator-tester.h",
6208 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006209 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006210)
6211
6212xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006213 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006214 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006215 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006216 "test/resize-bilinear-operator-tester.h",
6217 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006218 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006219)
6220
6221xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006222 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006223 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006224 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006225 "test/sigmoid-operator-tester.h",
6226 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006227 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006228)
6229
6230xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006231 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006232 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006233 "test/softmax-nc.cc",
6234 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006235 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006236 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006237)
6238
6239xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006240 name = "square_nc_test",
6241 srcs = [
6242 "test/square-nc.cc",
6243 "test/square-operator-tester.h",
6244 ],
6245 deps = OPERATOR_TEST_DEPS,
6246)
6247
6248xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006249 name = "square_root_nc_test",
6250 srcs = [
6251 "test/square-root-nc.cc",
6252 "test/square-root-operator-tester.h",
6253 ],
6254 deps = OPERATOR_TEST_DEPS,
6255)
6256
6257xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006258 name = "squared_difference_nd_test",
6259 srcs = [
6260 "test/binary-elementwise-operator-tester.h",
6261 "test/squared-difference-nd.cc",
6262 ],
6263 deps = OPERATOR_TEST_DEPS,
6264)
6265
6266xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006267 name = "subtract_nd_test",
6268 srcs = [
6269 "test/binary-elementwise-operator-tester.h",
6270 "test/subtract-nd.cc",
6271 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006272 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006273)
6274
6275xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006276 name = "truncation_nc_test",
6277 srcs = [
6278 "test/truncation-nc.cc",
6279 "test/truncation-operator-tester.h",
6280 ],
6281 deps = OPERATOR_TEST_DEPS,
6282)
6283
6284xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006285 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006286 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006287 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006288 "test/unpooling-operator-tester.h",
6289 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006290 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006291)
6292
Chao Mei6ddfc602020-05-13 22:29:36 -07006293############################### Misc unit tests ###############################
6294
6295xnnpack_unit_test(
6296 name = "memory_planner_test",
6297 srcs = [
6298 "test/memory-planner-test.cc",
6299 ],
6300 deps = [
6301 ":XNNPACK",
6302 ":memory_planner",
6303 ],
6304)
6305
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306############################# Build configurations #############################
6307
Marat Dukhanb8642352019-10-30 15:43:02 -07006308# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006309config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006310 name = "xnn_enable_assembly_explicit_true",
6311 define_values = {"xnn_enable_assembly": "true"},
6312)
6313
6314# Disables usage of assembly kernels.
6315config_setting(
6316 name = "xnn_enable_assembly_explicit_false",
6317 define_values = {"xnn_enable_assembly": "false"},
6318)
6319
Marat Dukhan9de90e02020-06-18 16:04:12 -07006320# Enables usage of sparse inference.
6321config_setting(
6322 name = "xnn_enable_sparse_explicit_true",
6323 define_values = {"xnn_enable_sparse": "true"},
6324)
6325
6326# Disables usage of sparse inference.
6327config_setting(
6328 name = "xnn_enable_sparse_explicit_false",
6329 define_values = {"xnn_enable_sparse": "false"},
6330)
6331
Marat Dukhan05702cf2020-03-26 15:41:33 -07006332# Disables usage of HMP-aware optimizations.
6333config_setting(
6334 name = "xnn_enable_hmp_explicit_false",
6335 define_values = {"xnn_enable_hmp": "false"},
6336)
6337
Chao Mei6ddfc602020-05-13 22:29:36 -07006338# Enable usage of optimized memory allocation
6339config_setting(
6340 name = "xnn_enable_memopt_explicit_true",
6341 define_values = {"xnnpack_enable_memopt": "true"},
6342)
6343
6344# Disable usage of optimized memory allocation
6345config_setting(
6346 name = "xnn_enable_memopt_explicit_false",
6347 define_values = {"xnnpack_enable_memopt": "false"},
6348)
6349
Marat Dukhanb8642352019-10-30 15:43:02 -07006350# Builds with -c dbg
6351config_setting(
6352 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006353 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006354 "compilation_mode": "dbg",
6355 },
6356)
6357
6358# Builds with -c opt
6359config_setting(
6360 name = "optimized_build",
6361 values = {
6362 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006363 },
6364)
6365
6366config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006367 name = "linux_k8",
6368 values = {"cpu": "k8"},
6369)
6370
6371config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006372 name = "linux_arm",
6373 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006374)
6375
6376config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006377 name = "linux_armeabi",
6378 values = {"cpu": "armeabi"},
6379)
6380
6381config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006382 name = "linux_armhf",
6383 values = {"cpu": "armhf"},
6384)
6385
6386config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006387 name = "linux_armv7a",
6388 values = {"cpu": "armv7a"},
6389)
6390
6391config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006392 name = "linux_aarch64",
6393 values = {"cpu": "aarch64"},
6394)
6395
6396config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006397 name = "android",
6398 values = {"crosstool_top": "//external:android/crosstool"},
6399)
6400
6401config_setting(
6402 name = "android_armv7",
6403 values = {
6404 "crosstool_top": "//external:android/crosstool",
6405 "cpu": "armeabi-v7a",
6406 },
6407)
6408
6409config_setting(
6410 name = "android_arm64",
6411 values = {
6412 "crosstool_top": "//external:android/crosstool",
6413 "cpu": "arm64-v8a",
6414 },
6415)
6416
6417config_setting(
6418 name = "android_x86",
6419 values = {
6420 "crosstool_top": "//external:android/crosstool",
6421 "cpu": "x86",
6422 },
6423)
6424
6425config_setting(
6426 name = "android_x86_64",
6427 values = {
6428 "crosstool_top": "//external:android/crosstool",
6429 "cpu": "x86_64",
6430 },
6431)
6432
6433config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006434 name = "windows_x86_64",
6435 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006436)
6437
6438config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006439 name = "windows_x86_64_clang",
6440 values = {
6441 "compiler": "clang-cl",
6442 "cpu": "x64_windows",
6443 },
6444)
6445
6446config_setting(
6447 name = "windows_x86_64_mingw",
6448 values = {
6449 "compiler": "mingw-gcc",
6450 "cpu": "x64_windows",
6451 },
6452)
6453
6454config_setting(
6455 name = "windows_x86_64_msys",
6456 values = {
6457 "compiler": "msys-gcc",
6458 "cpu": "x64_windows",
6459 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006460)
6461
6462config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006463 name = "macos_x86_64",
6464 values = {
6465 "apple_platform_type": "macos",
6466 "cpu": "darwin",
6467 },
6468)
6469
6470config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006471 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006472 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006473)
6474
6475config_setting(
6476 name = "emscripten_wasm",
6477 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006478 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006479 "cpu": "wasm",
6480 },
6481)
6482
6483config_setting(
6484 name = "emscripten_wasmsimd",
6485 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006486 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006487 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006488 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006489 },
6490)
6491
6492config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006493 name = "ios_armv7",
6494 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006495 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006496 "cpu": "ios_armv7",
6497 },
6498)
6499
6500config_setting(
6501 name = "ios_arm64",
6502 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006503 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006504 "cpu": "ios_arm64",
6505 },
6506)
6507
6508config_setting(
6509 name = "ios_arm64e",
6510 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006511 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006512 "cpu": "ios_arm64e",
6513 },
6514)
6515
6516config_setting(
6517 name = "ios_x86",
6518 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006519 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006520 "cpu": "ios_i386",
6521 },
6522)
6523
6524config_setting(
6525 name = "ios_x86_64",
6526 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006527 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006528 "cpu": "ios_x86_64",
6529 },
6530)
6531
6532config_setting(
6533 name = "watchos_armv7k",
6534 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006535 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006536 "cpu": "watchos_armv7k",
6537 },
6538)
6539
6540config_setting(
6541 name = "watchos_arm64_32",
6542 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006543 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006544 "cpu": "watchos_arm64_32",
6545 },
6546)
6547
6548config_setting(
6549 name = "watchos_x86",
6550 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006551 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006552 "cpu": "watchos_i386",
6553 },
6554)
6555
6556config_setting(
6557 name = "watchos_x86_64",
6558 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006559 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006560 "cpu": "watchos_x86_64",
6561 },
6562)
6563
6564config_setting(
6565 name = "tvos_arm64",
6566 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006567 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006568 "cpu": "tvos_arm64",
6569 },
6570)
6571
6572config_setting(
6573 name = "tvos_x86_64",
6574 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006575 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006576 "cpu": "tvos_x86_64",
6577 },
6578)