blob: 80fc8412ccdb741e8815393944f42280dc797fc9 [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 Dukhanf8475d62020-09-17 15:01:43 -0700425 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
426 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
427 "src/math/sigmoid-scalar-rr2-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 = [
Marat Dukhand4c83032020-09-09 23:20:12 -07001750 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001751 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1752 "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001753 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001754 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001755 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001756 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1757 "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
Marat Dukhand4c83032020-09-09 23:20:12 -07001758 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
Benoit Jacoba05487f2020-08-14 13:37:21 -07001759 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
Benoit Jacoba9644732020-08-13 12:48:55 -07001760]
1761
Marat Dukhan08c4a432019-10-03 09:29:21 -07001762SSE_UKERNELS = [
Marat Dukhan99936602020-04-11 16:47:01 -07001763 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1764 "src/f32-avgpool/9x-minmax-sse-c4.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001765 "src/f32-clamp/gen/sse-x4.c",
1766 "src/f32-clamp/gen/sse-x8.c",
Erich Elsenb1233402020-06-08 15:53:15 -07001767 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1768 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001769 "src/f32-dwconv-chw/3x3p1-sse.c",
1770 "src/f32-dwconv-chw/3x3s2p1-sse.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001771 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1772 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
1773 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1774 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1775 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1776 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
1777 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1778 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1779 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1780 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1781 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1782 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Marat Dukhan1f29b802020-05-15 23:46:39 -07001783 "src/f32-gavgpool-cw/sse-x4.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001784 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1785 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001786 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1787 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1788 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
1789 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1790 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1791 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1792 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
1793 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1794 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1795 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
1796 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1797 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1798 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07001799 "src/f32-hswish/gen/hswish-sse-x4.c",
1800 "src/f32-hswish/gen/hswish-sse-x8.c",
Marat Dukhan660fd192020-03-10 04:55:30 -07001801 "src/f32-ibilinear/gen/sse-c4.c",
1802 "src/f32-ibilinear/gen/sse-c8.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001803 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1804 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1805 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
1806 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1807 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1808 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1809 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001810 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1811 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1812 "src/f32-pavgpool/9x-minmax-sse-c4.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07001813 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Marat Dukhan39b5e942020-06-24 15:03:48 -07001814 "src/f32-prelu/gen/sse-2x4.c",
1815 "src/f32-prelu/gen/sse-2x8.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07001816 "src/f32-relu/gen/sse-x4.c",
1817 "src/f32-relu/gen/sse-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001818 "src/f32-rmax/sse.c",
Marat Dukhan355ab432020-04-09 19:01:52 -07001819 "src/f32-spmm/gen/4x1-minmax-sse.c",
1820 "src/f32-spmm/gen/8x1-minmax-sse.c",
Erich Elsen6e80fdc2020-06-09 15:35:37 -07001821 "src/f32-spmm/gen/16x1-minmax-sse.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001822 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1823 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1824 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1825 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1826 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1827 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1828 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1829 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Marat Dukhan403b7d42019-12-05 12:49:11 -08001830 "src/f32-vbinary/gen/vmax-sse-x4.c",
1831 "src/f32-vbinary/gen/vmax-sse-x8.c",
1832 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1833 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1834 "src/f32-vbinary/gen/vmin-sse-x4.c",
1835 "src/f32-vbinary/gen/vmin-sse-x8.c",
1836 "src/f32-vbinary/gen/vminc-sse-x4.c",
1837 "src/f32-vbinary/gen/vminc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001838 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1839 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1840 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1841 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1842 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1843 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1844 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1845 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07001846 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1847 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1848 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1849 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07001850 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1851 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1852 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1853 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001854 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1855 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001856 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1857 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07001858 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1859 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07001860 "src/f32-vunary/gen/vabs-sse-x4.c",
1861 "src/f32-vunary/gen/vabs-sse-x8.c",
1862 "src/f32-vunary/gen/vneg-sse-x4.c",
1863 "src/f32-vunary/gen/vneg-sse-x8.c",
1864 "src/f32-vunary/gen/vsqr-sse-x4.c",
1865 "src/f32-vunary/gen/vsqr-sse-x8.c",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07001866 "src/x32-fill/sse.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001867 "src/x32-packx/x4-sse.c",
Marat Dukhan63523d42020-05-22 17:07:33 -07001868 "src/x32-pad/sse.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001869 "src/math/roundne-sse-addsub.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001870 "src/math/roundd-sse-addsub.c",
1871 "src/math/roundu-sse-addsub.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001872 "src/math/roundz-sse-addsub.c",
Marat Dukhan84000762020-06-29 18:38:43 -07001873 "src/math/sqrt-sse-hh1mac.c",
1874 "src/math/sqrt-sse-nr1mac.c",
1875 "src/math/sqrt-sse-nr2mac.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001876]
1877
1878SSE2_UKERNELS = [
Marat Dukhan329da642019-11-19 21:44:39 -08001879 "src/f32-argmaxpool/9p8x-sse2-c4.c",
1880 "src/f32-argmaxpool/4x-sse2-c4.c",
1881 "src/f32-argmaxpool/9x-sse2-c4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001882 "src/f32-prelu/gen/sse2-2x4.c",
1883 "src/f32-prelu/gen/sse2-2x8.c",
Marat Dukhanb39689d2020-01-24 13:32:20 -08001884 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
1885 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
1886 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
1887 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
1888 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1889 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
1890 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
1891 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1892 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
1893 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1894 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1895 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07001896 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1897 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1898 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1899 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1900 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1901 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001902 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001903 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001904 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08001905 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08001906 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1907 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07001908 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1909 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07001910 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1911 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1912 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1913 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1914 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1915 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1916 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1917 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001918 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1919 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1920 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07001921 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1922 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1923 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07001924 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1925 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1926 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001927 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1928 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001929 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1930 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001931 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1932 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001933 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1934 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001935 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001936 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1937 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001938 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07001939 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1940 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1941 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07001942 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1943 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07001944 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1945 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001946 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1947 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001948 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07001949 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1950 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07001951 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001952 "src/qs8-requantization/precise-sse2.c",
1953 "src/qs8-requantization/fp32-sse2.c",
1954 "src/qs8-requantization/q31-sse2.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07001955 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1956 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1957 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1958 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07001959 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1960 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1961 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1962 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001963 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1964 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
1965 "src/qu8-igemm/4x4c2-minmax-sse2.c",
1966 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1967 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1968 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1969 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1970 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07001971 "src/qu8-requantization/precise-sse2.c",
1972 "src/qu8-requantization/fp32-sse2.c",
1973 "src/qu8-requantization/q31-sse2.c",
Marat Dukhan08b7a972020-07-14 18:17:29 -07001974 "src/qu8-vadd/minmax-sse2.c",
Marat Dukhan5c5fa962020-03-10 18:38:33 -07001975 "src/u8-clamp/sse2-x64.c",
Marat Dukhan99936602020-04-11 16:47:01 -07001976 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001977 "src/u8-rmax/sse2.c",
Marat Dukhan57dccd82020-04-14 00:53:10 -07001978 "src/x32-unpool/sse2.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07001979 "src/x32-zip/x2-sse2.c",
1980 "src/x32-zip/x3-sse2.c",
1981 "src/x32-zip/x4-sse2.c",
1982 "src/x32-zip/xm-sse2.c",
1983 "src/x8-zip/x2-sse2.c",
1984 "src/x8-zip/x3-sse2.c",
1985 "src/x8-zip/x4-sse2.c",
1986 "src/x8-zip/xm-sse2.c",
Marat Dukhanffd68402019-11-15 15:19:11 -08001987 "src/math/exp-sse2-p5.c",
1988 "src/math/expminus-sse2-p5.c",
Marat Dukhan075088a2020-05-12 19:42:12 -07001989 "src/math/roundne-sse2-cvt.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07001990 "src/math/roundd-sse2-cvt.c",
1991 "src/math/roundu-sse2-cvt.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07001992 "src/math/roundz-sse2-cvt.c",
Marat Dukhanf8475d62020-09-17 15:01:43 -07001993 "src/math/sigmoid-sse2-rr2-p5-div.c",
1994 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07001995]
1996
1997SSSE3_UKERNELS = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07001998 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1999 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
2000 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002001 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
2002 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
2003 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002004 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
2005 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
2006 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002007 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
2008 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002009 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
2010 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002011 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
2012 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002013 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
2014 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002015 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002016 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
2017 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002018 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002019 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
2020 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
2021 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002022 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
2023 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002024 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
2025 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002026 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
2027 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002028 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002029 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
2030 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002031 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002032 "src/qs8-requantization/precise-ssse3.c",
2033 "src/qs8-requantization/q31-ssse3.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002034 "src/qu8-requantization/precise-ssse3.c",
2035 "src/qu8-requantization/q31-ssse3.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002036]
2037
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002038SSE41_UKERNELS = [
Marat Dukhan40a672f2019-11-25 03:08:22 -08002039 "src/f32-prelu/gen/sse41-2x4.c",
2040 "src/f32-prelu/gen/sse41-2x8.c",
Marat Dukhand243c1a2020-09-17 18:48:10 -07002041 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
2042 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
2043 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
2044 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
2045 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
2046 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002047 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002048 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002049 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
Marat Dukhan40a672f2019-11-25 03:08:22 -08002050 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002051 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
2052 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002053 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
2054 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002055 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
2056 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
2057 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
2058 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
2059 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
2060 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
2061 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
2062 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07002063 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
2064 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
2065 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
Marat Dukhan4ed53f42020-08-06 01:12:55 -07002066 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
2067 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
2068 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
Marat Dukhan159688f2020-08-06 10:34:29 -07002069 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
2070 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
2071 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002072 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
2073 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002074 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
2075 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002076 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
2077 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002078 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
2079 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002080 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002081 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
2082 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002083 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002084 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
2085 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
2086 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
Marat Dukhan14d3ce82020-07-31 16:19:15 -07002087 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
2088 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
Marat Dukhandee732b2020-07-31 17:32:04 -07002089 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
2090 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002091 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
2092 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002093 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
Marat Dukhan07bd2522020-07-31 19:12:39 -07002094 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
2095 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002096 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
Marat Dukhan2e23d2b2020-07-29 16:01:37 -07002097 "src/qs8-requantization/fp32-sse4.c",
2098 "src/qs8-requantization/precise-sse4.c",
2099 "src/qs8-requantization/q31-sse4.c",
Marat Dukhand9f3ad42020-08-10 12:30:58 -07002100 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
2101 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
2102 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
2103 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002104 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
2105 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
2106 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
2107 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan0270d9f2020-08-11 00:56:46 -07002108 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
2109 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
2110 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
2111 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002112 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
2113 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
2114 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
2115 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07002116 "src/qu8-requantization/precise-sse4.c",
2117 "src/qu8-requantization/q31-sse4.c",
Marat Dukhan8853b822020-05-07 12:19:01 -07002118 "src/math/roundne-sse41.c",
Marat Dukhanc9852ba2020-05-13 17:21:29 -07002119 "src/math/roundd-sse41.c",
2120 "src/math/roundu-sse41.c",
Marat Dukhan2dbb9442020-05-12 20:43:43 -07002121 "src/math/roundz-sse41.c",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08002122]
2123
Marat Dukhan08c4a432019-10-03 09:29:21 -07002124AVX_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002125 "src/f32-clamp/gen/avx-x8.c",
2126 "src/f32-clamp/gen/avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002127 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
2128 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
2129 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
2130 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
2131 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
2132 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
2133 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
2134 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
2135 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
2136 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
2137 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
2138 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
2139 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
2140 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
2141 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
2142 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
2143 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
2144 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
2145 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
2146 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
2147 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
2148 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
2149 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
2150 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
2151 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
2152 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
2153 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
2154 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
2155 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
2156 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002157 "src/f32-hswish/gen/hswish-avx-x8.c",
2158 "src/f32-hswish/gen/hswish-avx-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002159 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
2160 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
2161 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
2162 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
2163 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
2164 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
2165 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
2166 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
2167 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002168 "src/f32-prelu/gen/avx-2x8.c",
2169 "src/f32-prelu/gen/avx-2x16.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002170 "src/f32-relu/gen/avx-x16.c",
2171 "src/f32-relu/gen/avx-x8.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002172 "src/f32-rmax/avx.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002173 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
2174 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
2175 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
2176 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
2177 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
2178 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
2179 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
2180 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002181 "src/f32-vbinary/gen/vmax-avx-x8.c",
2182 "src/f32-vbinary/gen/vmax-avx-x16.c",
2183 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
2184 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
2185 "src/f32-vbinary/gen/vmin-avx-x8.c",
2186 "src/f32-vbinary/gen/vmin-avx-x16.c",
2187 "src/f32-vbinary/gen/vminc-avx-x8.c",
2188 "src/f32-vbinary/gen/vminc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002189 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
2190 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
2191 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
2192 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
2193 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
2194 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
2195 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
2196 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002197 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
2198 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
2199 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
2200 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002201 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
2202 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
2203 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
2204 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002205 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
2206 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002207 "src/f32-vscale/avx-unroll32.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002208 "src/f32-vrnd/gen/vrndne-avx-x8.c",
2209 "src/f32-vrnd/gen/vrndne-avx-x16.c",
2210 "src/f32-vrnd/gen/vrndz-avx-x8.c",
2211 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2212 "src/f32-vrnd/gen/vrndu-avx-x8.c",
2213 "src/f32-vrnd/gen/vrndu-avx-x16.c",
2214 "src/f32-vrnd/gen/vrndd-avx-x8.c",
2215 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002216 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2217 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002218 "src/f32-vunary/gen/vabs-avx-x8.c",
2219 "src/f32-vunary/gen/vabs-avx-x16.c",
2220 "src/f32-vunary/gen/vneg-avx-x8.c",
2221 "src/f32-vunary/gen/vneg-avx-x16.c",
2222 "src/f32-vunary/gen/vsqr-avx-x8.c",
2223 "src/f32-vunary/gen/vsqr-avx-x16.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002224]
2225
Marat Dukhan1566fee2020-08-02 21:55:41 -07002226XOP_UKERNELS = [
2227 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2228 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2229 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2230 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002231 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2232 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002233 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2234 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002235 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002236 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2237 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002238 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002239 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2240 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2241 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002242 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2243 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2244 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2245 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2246 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2247 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002248 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002249 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2250 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
Marat Dukhane7edc802020-08-03 16:43:57 -07002251 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
Marat Dukhanbb9225e2020-09-06 22:40:56 -07002252 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2253 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2254 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2255 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2256 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2257 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2258 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2259 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Marat Dukhan1566fee2020-08-02 21:55:41 -07002260]
2261
Marat Dukhanfda12b82019-11-21 12:27:59 -08002262FMA3_UKERNELS = [
Marat Dukhan1c587112020-04-08 20:04:28 -07002263 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2264 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2265 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2266 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
2267 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2268 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2269 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2270 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
2271 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2272 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
2273 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2274 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
2275 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2276 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2277 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2278 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2279 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2280 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
2281 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2282 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2283 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2284 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2285 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2286 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2287 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2288 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2289 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
2290 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
2291 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
2292 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2293 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2294 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
2295 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2296 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2297 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2298 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2299 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2300 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
2301 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
2302 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002303 "src/f32-hswish/gen/hswish-fma3-x8.c",
2304 "src/f32-hswish/gen/hswish-fma3-x16.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002305 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
2306 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
2307 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
2308 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2309 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2310 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
2311 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2312 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2313 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2314 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2315 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2316 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
2317 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
2318 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002319 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2320 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2321 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2322 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2323 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2324 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2325 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2326 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002327 "src/math/sqrt-fma3-nr1fma.c",
2328 "src/math/sqrt-fma3-nr2fma.c",
2329 "src/math/sqrt-fma3-nr1fma1adj.c",
Marat Dukhanfda12b82019-11-21 12:27:59 -08002330]
2331
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002332AVX2_UKERNELS = [
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002333 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
2334 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2335 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
2336 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
2337 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
2338 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
2339 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2340 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
2341 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
2342 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2343 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2344 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
2345 "src/f32-raddextexp/gen/avx2-p5-x64.c",
2346 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2347 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
2348 "src/f32-raddextexp/gen/avx2-p5-x72.c",
2349 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
2350 "src/f32-raddextexp/gen/avx2-p5-x80.c",
2351 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2352 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
2353 "src/f32-raddextexp/gen/avx2-p5-x96.c",
2354 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2355 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2356 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
2357 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
2358 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2359 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
2360 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
2361 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
2362 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
2363 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2364 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
2365 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
2366 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2367 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2368 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Marat Dukhanfa0a4322020-01-06 16:14:29 -08002369 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2370 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2371 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2372 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2373 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2374 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2375 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2376 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2377 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2378 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2379 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2380 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2381 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2382 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2383 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2384 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2385 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2386 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2387 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2388 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2389 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2390 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2391 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2392 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2393 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2394 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2395 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2396 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2397 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2398 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002399 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2400 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2401 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2402 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2403 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2404 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2405 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2406 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2407 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2408 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2409 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2410 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2411 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2412 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2413 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2414 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2415 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2416 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2417 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2418 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2419 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2420 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2421 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2422 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Marat Dukhan40135522020-08-07 01:21:00 -07002423 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2424 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
Marat Dukhand65a1522020-08-04 19:28:18 -07002425 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2426 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2427 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2428 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002429 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2430 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2431 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhan683fab32020-08-03 19:42:52 -07002432 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2433 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2434 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
Marat Dukhan12809522020-08-02 22:23:51 -07002435 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2436 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2437 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
Marat Dukhane6dc0b62020-09-08 23:57:14 -07002438 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2439 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2440 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2441 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2442 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2443 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2444 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2445 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002446 "src/math/exp-avx2-p5.c",
2447 "src/math/exp-avx2-perm-p3.c",
2448 "src/math/exp-avx2-perm-p4.c",
Marat Dukhan515c9772019-10-17 18:07:57 -07002449 "src/math/expminus-avx2-p5.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002450 "src/math/extexp-avx2-p5.c",
Marat Dukhan72416002020-01-05 21:53:19 -08002451 "src/math/sigmoid-avx2-rr2-p5-div.c",
2452 "src/math/sigmoid-avx2-rr1-p5-div.c",
2453 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2454 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2455 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2456 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002457]
2458
Marat Dukhan08c4a432019-10-03 09:29:21 -07002459AVX512F_UKERNELS = [
Marat Dukhan5c5fa962020-03-10 18:38:33 -07002460 "src/f32-clamp/gen/avx512f-x16.c",
2461 "src/f32-clamp/gen/avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002462 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2463 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2464 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2465 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
2466 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2467 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2468 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2469 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
2470 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2471 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
2472 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2473 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
2474 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2475 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2476 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2477 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2478 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2479 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
2480 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2481 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2482 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2483 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2484 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2485 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Marat Dukhanefaac272020-07-16 16:58:07 -07002486 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2487 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Marat Dukhan1c587112020-04-08 20:04:28 -07002488 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2489 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2490 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2491 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2492 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2493 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Marat Dukhan90eca0a2020-03-11 00:52:23 -07002494 "src/f32-prelu/gen/avx512f-2x16.c",
2495 "src/f32-prelu/gen/avx512f-2x32.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002496 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
2497 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2498 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2499 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
2500 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2501 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
2502 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2503 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2504 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
2505 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2506 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2507 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
2508 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
2509 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2510 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
2511 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
2512 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
2513 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
2514 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2515 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
2516 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
2517 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2518 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2519 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
2520 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
2521 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2522 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
2523 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
2524 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
2525 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
2526 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2527 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
2528 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
2529 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2530 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2531 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Frank Barchardfb158e22020-07-15 16:10:10 -07002532 "src/f32-relu/gen/avx512f-x16.c",
2533 "src/f32-relu/gen/avx512f-x32.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002534 "src/f32-rmax/avx512f.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002535 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2536 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2537 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2538 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2539 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2540 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2541 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2542 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Marat Dukhan9a88efe2019-12-10 15:54:24 -08002543 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2544 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2545 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2546 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2547 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2548 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2549 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2550 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002551 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2552 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2553 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2554 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2555 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2556 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2557 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2558 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Marat Dukhan13bafb02020-06-05 00:43:11 -07002559 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2560 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2561 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2562 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Marat Dukhan91cd2b72020-04-09 23:57:31 -07002563 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2564 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2565 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2566 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Marat Dukhan19dd91d2020-07-16 11:12:44 -07002567 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2568 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002569 "src/f32-vscale/avx512f-unroll64.c",
Marat Dukhan4c4eb002019-12-08 21:27:49 -08002570 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2571 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2572 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2573 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2574 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2575 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2576 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2577 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2578 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2579 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2580 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2581 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2582 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2583 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2584 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2585 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2586 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2587 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2588 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2589 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2590 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2591 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2592 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2593 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07002594 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2595 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2596 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2597 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2598 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2599 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2600 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2601 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
Marat Dukhanf4db2f32020-06-30 10:55:30 -07002602 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2603 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2604 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2605 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2606 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2607 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2608 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2609 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
Marat Dukhan5020b962020-06-08 13:30:10 -07002610 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2611 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2612 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2613 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2614 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2615 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002616 "src/math/exp-avx512f-p5-scalef.c",
2617 "src/math/exp-avx512f-p5.c",
2618 "src/math/exp-avx512f-perm-p3.c",
Marat Dukhanfeb49232019-10-28 11:03:31 -07002619 "src/math/exp-avx512f-perm2-p2.c",
Marat Dukhan98ba4412019-10-23 02:14:28 -07002620 "src/math/extexp-avx512f-p5.c",
Marat Dukhan84000762020-06-29 18:38:43 -07002621 "src/math/sqrt-avx512f-nr1fma.c",
2622 "src/math/sqrt-avx512f-nr2fma.c",
2623 "src/math/sqrt-avx512f-nr1fma1adj.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002624]
2625
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002626AVX512SKX_UKERNELS = [
Marat Dukhan2ffc5e62020-09-06 22:33:38 -07002627 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2628 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07002629 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2630 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2631 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2632 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2633 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2634 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2635 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2636 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
2637]
2638
Frank Barchardbcedc082020-08-17 18:00:51 -07002639WASM32_ASM_UKERNELS = [
2640 "src/f32-relu/wasm_shr_x1.S",
2641 "src/f32-relu/wasm_shr_x2.S",
2642 "src/f32-relu/wasm_shr_x4.S",
2643]
2644
Marat Dukhan08c4a432019-10-03 09:29:21 -07002645AARCH32_ASM_UKERNELS = [
Marat Dukhan32f93812020-05-17 20:31:21 -07002646 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
Marat Dukhan3b98f6b2020-05-17 10:09:22 -07002647 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002648 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2649 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002650 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002651 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Frank Barchard569561d2020-06-17 13:11:12 -07002652 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002653 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002654 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2655 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Frank Barchard490febe2020-07-16 18:42:17 -07002656 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2657 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2658 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2659 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002660]
2661
2662AARCH64_ASM_UKERNELS = [
Frank Barchard36b76b62020-04-10 12:39:17 -07002663 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002664 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2665 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard36b76b62020-04-10 12:39:17 -07002666 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchard683f5592020-04-10 00:48:26 -07002667 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
2668 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002669 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2670 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2671 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002672 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchardbddfbcd2020-04-15 12:32:41 -07002673 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2674 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
2675 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Frank Barchard3b8e5662020-04-20 12:12:53 -07002676 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002677 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2678 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002679 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002680 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2681 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2682 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2683 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2684 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2685 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2686 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2687 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2688 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2689 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2690 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2691 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2692 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2693 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2694 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2695 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2696 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2697 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002698 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
2699 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
Frank Barchard3cb54f92020-04-10 10:46:08 -07002700 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002701 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2702 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2703 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2704 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2705 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
2706 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2707 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2708 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2709 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2710 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2711 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
2712 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2713 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2714 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2715 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2716 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
2717 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2718 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan1c587112020-04-08 20:04:28 -07002719 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2720 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
2721 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2722 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2723 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2724 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2725 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2726 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2727 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2728 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2729 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2730 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2731 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2732 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2733 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2734 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2735 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2736 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002737]
2738
Marat Dukhan1b354632020-03-23 12:50:22 -07002739INTERNAL_MICROKERNEL_HDRS = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002740 "src/xnnpack/argmaxpool.h",
2741 "src/xnnpack/avgpool.h",
2742 "src/xnnpack/clamp.h",
2743 "src/xnnpack/common.h",
2744 "src/xnnpack/conv.h",
2745 "src/xnnpack/dwconv.h",
2746 "src/xnnpack/gavgpool.h",
2747 "src/xnnpack/gemm.h",
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07002748 "src/xnnpack/fill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002749 "src/xnnpack/hswish.h",
Marat Dukhan660fd192020-03-10 04:55:30 -07002750 "src/xnnpack/ibilinear.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002751 "src/xnnpack/igemm.h",
Marat Dukhancfb31342019-12-05 10:42:57 -08002752 "src/xnnpack/intrinsics-polyfill.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002753 "src/xnnpack/lut.h",
2754 "src/xnnpack/math.h",
2755 "src/xnnpack/maxpool.h",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002756 "src/xnnpack/memory.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002757 "src/xnnpack/packx.h",
2758 "src/xnnpack/pad.h",
2759 "src/xnnpack/params.h",
2760 "src/xnnpack/pavgpool.h",
2761 "src/xnnpack/ppmm.h",
2762 "src/xnnpack/prelu.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002763 "src/xnnpack/raddexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002764 "src/xnnpack/raddextexp.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002765 "src/xnnpack/raddstoreexpminusmax.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002766 "src/xnnpack/rmax.h",
2767 "src/xnnpack/scalar-utils.h",
2768 "src/xnnpack/spmm.h",
2769 "src/xnnpack/unpool.h",
2770 "src/xnnpack/vadd.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002771 "src/xnnpack/vbinary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002772 "src/xnnpack/vmulcaddc.h",
Marat Dukhan05ac8e32019-10-21 15:39:33 -07002773 "src/xnnpack/vscale.h",
Marat Dukhan97579532019-10-18 16:40:39 -07002774 "src/xnnpack/vscaleexpminusmax.h",
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07002775 "src/xnnpack/vscaleextexp.h",
Marat Dukhan1e782c42019-11-21 17:02:40 -08002776 "src/xnnpack/vunary.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002777 "src/xnnpack/zip.h",
Marat Dukhan1b354632020-03-23 12:50:22 -07002778]
2779
2780INTERNAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002781 "include/xnnpack.h",
2782 "src/xnnpack/allocator.h",
2783 "src/xnnpack/compute.h",
2784 "src/xnnpack/im2col.h",
2785 "src/xnnpack/indirection.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002786 "src/xnnpack/math-stubs.h",
Chao Mei6ddfc602020-05-13 22:29:36 -07002787 "src/xnnpack/memory-planner.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002788 "src/xnnpack/operator.h",
2789 "src/xnnpack/pack.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002790 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002791 "src/xnnpack/requantization-stubs.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002792 "src/xnnpack/requantization.h",
Marat Dukhan1d75a542020-02-03 12:23:01 -08002793 "src/xnnpack/subgraph.h",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002794]
2795
Marat Dukhan1b354632020-03-23 12:50:22 -07002796ACCURACY_EVAL_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07002797 "src/xnnpack/math-stubs.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002798]
2799
Marat Dukhan1b354632020-03-23 12:50:22 -07002800MICROKERNEL_BENCHMARK_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002801 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002802 "include/xnnpack.h",
2803]
2804
Marat Dukhan1b354632020-03-23 12:50:22 -07002805MICROKERNEL_TEST_HDRS = INTERNAL_MICROKERNEL_HDRS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07002806 "src/xnnpack/isa-checks.h",
Marat Dukhaneeaa7bd2019-10-25 17:31:25 -07002807 "src/xnnpack/params-init.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002808 "src/xnnpack/requantization.h",
2809 "include/xnnpack.h",
2810]
2811
2812OPERATOR_TEST_PARAMS_HDRS = [
2813 "src/xnnpack/params.h",
2814 "src/xnnpack/common.h",
2815]
2816
2817WEIGHTS_PACK_HDRS = [
2818 "src/xnnpack/pack.h",
2819 "src/xnnpack/operator.h",
2820 "src/xnnpack/compute.h",
2821]
2822
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002823LOGGING_COPTS = select({
2824 # No logging in optimized mode
2825 ":optimized_build": ["-DXNN_LOG_LEVEL=0"],
2826 # Full logging in debug mode
2827 ":debug_build": ["-DXNN_LOG_LEVEL=5"],
2828 # Error-only logging in default (fastbuild) mode
2829 "//conditions:default": ["-DXNN_LOG_LEVEL=2"],
2830})
2831
Marat Dukhan3b59de22020-06-03 20:15:19 -07002832LOGGING_SRCS = select({
2833 # No logging in optimized mode
2834 ":optimized_build": [],
2835 "//conditions:default": [
2836 "src/operator-strings.c",
2837 "src/subgraph-strings.c",
2838 ],
2839})
2840
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07002841LOGGING_HDRS = [
2842 "src/xnnpack/log.h",
2843]
2844
Marat Dukhan08c4a432019-10-03 09:29:21 -07002845xnnpack_cc_library(
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002846 name = "tables",
2847 srcs = TABLE_SRCS,
2848 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002849 gcc_copts = xnnpack_gcc_std_copts(),
2850 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002851)
2852
2853xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07002854 name = "scalar_ukernels",
2855 srcs = SCALAR_UKERNELS,
2856 hdrs = INTERNAL_HDRS,
2857 aarch32_copts = ["-marm"],
Marat Dukhan10a38082020-04-17 03:58:35 -07002858 gcc_copts = xnnpack_gcc_std_copts(),
2859 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002860 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002861 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002862 "@FP16",
2863 "@FXdiv",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002864 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002865 ],
2866)
2867
2868xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002869 name = "scalar_ukernels_test_mode",
2870 srcs = SCALAR_UKERNELS,
2871 hdrs = INTERNAL_HDRS,
2872 aarch32_copts = ["-marm"],
2873 copts = [
2874 "-UNDEBUG",
2875 "-DXNN_TEST_MODE=1",
2876 ],
2877 gcc_copts = xnnpack_gcc_std_copts(),
2878 msvc_copts = xnnpack_msvc_std_copts(),
2879 deps = [
2880 ":tables",
2881 "@FP16",
2882 "@FXdiv",
2883 "@pthreadpool",
2884 ],
2885)
2886
2887xnnpack_cc_library(
Marat Dukhan436ebe62019-12-04 15:10:12 -08002888 name = "wasm_ukernels",
2889 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07002890 gcc_copts = xnnpack_gcc_std_copts(),
2891 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan436ebe62019-12-04 15:10:12 -08002892 wasm_srcs = WASM_UKERNELS,
Marat Dukhan1c6e3892020-06-25 23:56:10 -07002893 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan436ebe62019-12-04 15:10:12 -08002894 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002895 ":tables",
Marat Dukhan436ebe62019-12-04 15:10:12 -08002896 "@FP16",
2897 "@FXdiv",
2898 "@pthreadpool",
2899 ],
2900)
2901
2902xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002903 name = "wasm_ukernels_test_mode",
2904 hdrs = INTERNAL_HDRS,
2905 copts = [
2906 "-UNDEBUG",
2907 "-DXNN_TEST_MODE=1",
2908 ],
2909 gcc_copts = xnnpack_gcc_std_copts(),
2910 msvc_copts = xnnpack_msvc_std_copts(),
2911 wasm_srcs = WASM_UKERNELS,
Marat Dukhan0e97d6f2020-06-26 19:35:09 -07002912 wasmsimd_srcs = WASM_UKERNELS + WASMSIMD_UKERNELS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07002913 deps = [
2914 ":tables",
2915 "@FP16",
2916 "@FXdiv",
2917 "@pthreadpool",
2918 ],
2919)
2920
2921xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002922 name = "psimd_fastmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002923 hdrs = INTERNAL_HDRS,
2924 aarch32_copts = [
2925 "-marm",
2926 "-mfpu=neon",
2927 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002928 gcc_copts = xnnpack_gcc_std_copts(),
2929 gcc_x86_copts = ["-msse2"],
2930 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07002931 optimized_copts = [
2932 "-O3",
2933 "-ffast-math",
2934 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002935 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07002936 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08002937 ":tables",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002938 "@FP16",
2939 "@psimd",
Marat Dukhan04f03be2019-11-19 12:36:47 -08002940 "@pthreadpool",
Marat Dukhan08c4a432019-10-03 09:29:21 -07002941 ],
2942)
2943
2944xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002945 name = "psimd_fastmath_ukernels_test_mode",
2946 hdrs = INTERNAL_HDRS,
2947 aarch32_copts = [
2948 "-marm",
2949 "-mfpu=neon",
2950 ],
2951 copts = [
2952 "-UNDEBUG",
2953 "-DXNN_TEST_MODE=1",
2954 ],
2955 gcc_copts = xnnpack_gcc_std_copts(),
2956 gcc_x86_copts = ["-msse2"],
2957 msvc_copts = xnnpack_msvc_std_copts(),
2958 optimized_copts = [
2959 "-O3",
2960 "-ffast-math",
2961 ],
2962 psimd_srcs = PSIMD_FASTMATH_UKERNELS,
2963 deps = [
2964 ":tables",
2965 "@FP16",
2966 "@psimd",
2967 "@pthreadpool",
2968 ],
2969)
2970
2971xnnpack_cc_library(
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002972 name = "psimd_accmath_ukernels",
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002973 hdrs = INTERNAL_HDRS,
2974 aarch32_copts = [
2975 "-marm",
2976 "-mfpu=neon",
2977 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07002978 gcc_copts = xnnpack_gcc_std_copts(),
2979 gcc_x86_copts = ["-msse2"],
2980 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002981 optimized_copts = [
2982 "-O3",
2983 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07002984 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08002985 deps = [
2986 ":tables",
2987 "@FP16",
2988 "@psimd",
2989 "@pthreadpool",
2990 ],
2991)
2992
2993xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07002994 name = "psimd_accmath_ukernels_test_mode",
2995 hdrs = INTERNAL_HDRS,
2996 aarch32_copts = [
2997 "-marm",
2998 "-mfpu=neon",
2999 ],
3000 copts = [
3001 "-UNDEBUG",
3002 "-DXNN_TEST_MODE=1",
3003 ],
3004 gcc_copts = xnnpack_gcc_std_copts(),
3005 gcc_x86_copts = ["-msse2"],
3006 msvc_copts = xnnpack_msvc_std_copts(),
3007 optimized_copts = [
3008 "-O3",
3009 ],
3010 psimd_srcs = PSIMD_ACCMATH_UKERNELS,
3011 deps = [
3012 ":tables",
3013 "@FP16",
3014 "@psimd",
3015 "@pthreadpool",
3016 ],
3017)
3018
3019xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003020 name = "neon_ukernels",
3021 hdrs = INTERNAL_HDRS,
3022 aarch32_copts = [
3023 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003024 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003025 "-mfpu=neon",
3026 ],
3027 aarch32_srcs = NEON_UKERNELS,
3028 aarch64_srcs = NEON_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003029 gcc_copts = xnnpack_gcc_std_copts(),
3030 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003031 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003032 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003033 "@FP16",
3034 "@pthreadpool",
3035 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003036)
3037
3038xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003039 name = "neon_ukernels_test_mode",
3040 hdrs = INTERNAL_HDRS,
3041 aarch32_copts = [
3042 "-marm",
3043 "-march=armv7-a",
3044 "-mfpu=neon",
3045 ],
3046 aarch32_srcs = NEON_UKERNELS,
3047 aarch64_srcs = NEON_UKERNELS,
3048 copts = [
3049 "-UNDEBUG",
3050 "-DXNN_TEST_MODE=1",
3051 ],
3052 gcc_copts = xnnpack_gcc_std_copts(),
3053 msvc_copts = xnnpack_msvc_std_copts(),
3054 deps = [
3055 ":tables",
3056 "@FP16",
3057 "@pthreadpool",
3058 ],
3059)
3060
3061xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003062 name = "neonfma_ukernels",
3063 hdrs = INTERNAL_HDRS,
3064 aarch32_copts = [
3065 "-marm",
Marat Dukhan8853b822020-05-07 12:19:01 -07003066 "-march=armv7-a",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003067 "-mfpu=neon-vfpv4",
3068 ],
3069 aarch32_srcs = NEONFMA_UKERNELS,
3070 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003071 apple_aarch32_copts = [
3072 "-mcpu=swift",
3073 "-mtune=generic",
3074 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003075 gcc_copts = xnnpack_gcc_std_copts(),
3076 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003077 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003078 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003079 "@FP16",
3080 "@pthreadpool",
3081 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003082)
3083
3084xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003085 name = "neonfma_ukernels_test_mode",
3086 hdrs = INTERNAL_HDRS,
3087 aarch32_copts = [
3088 "-marm",
3089 "-march=armv7-a",
3090 "-mfpu=neon-vfpv4",
3091 ],
3092 aarch32_srcs = NEONFMA_UKERNELS,
3093 aarch64_srcs = NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003094 apple_aarch32_copts = [
3095 "-mcpu=swift",
3096 "-mtune=generic",
3097 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003098 copts = [
3099 "-UNDEBUG",
3100 "-DXNN_TEST_MODE=1",
3101 ],
3102 gcc_copts = xnnpack_gcc_std_copts(),
3103 msvc_copts = xnnpack_msvc_std_copts(),
3104 deps = [
3105 ":tables",
3106 "@FP16",
3107 "@pthreadpool",
3108 ],
3109)
3110
3111xnnpack_cc_library(
Marat Dukhan8853b822020-05-07 12:19:01 -07003112 name = "neonv8_ukernels",
3113 hdrs = INTERNAL_HDRS,
3114 aarch32_copts = [
3115 "-marm",
3116 "-march=armv8-a",
3117 "-mfpu=neon-fp-armv8",
3118 ],
3119 aarch32_srcs = NEONV8_UKERNELS,
3120 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003121 apple_aarch32_copts = [
3122 "-mcpu=cyclone",
3123 "-mtune=generic",
3124 ],
Marat Dukhan8853b822020-05-07 12:19:01 -07003125 gcc_copts = xnnpack_gcc_std_copts(),
3126 msvc_copts = xnnpack_msvc_std_copts(),
3127 deps = [
3128 ":tables",
3129 "@FP16",
3130 "@pthreadpool",
3131 ],
3132)
3133
3134xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003135 name = "neonv8_ukernels_test_mode",
3136 hdrs = INTERNAL_HDRS,
3137 aarch32_copts = [
3138 "-marm",
3139 "-march=armv8-a",
3140 "-mfpu=neon-fp-armv8",
3141 ],
3142 aarch32_srcs = NEONV8_UKERNELS,
3143 aarch64_srcs = NEONV8_UKERNELS,
Marat Dukhanbc69ed62020-06-09 21:34:56 -07003144 apple_aarch32_copts = [
3145 "-mcpu=cyclone",
3146 "-mtune=generic",
3147 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003148 copts = [
3149 "-UNDEBUG",
3150 "-DXNN_TEST_MODE=1",
3151 ],
3152 gcc_copts = xnnpack_gcc_std_copts(),
3153 msvc_copts = xnnpack_msvc_std_copts(),
3154 deps = [
3155 ":tables",
3156 "@FP16",
3157 "@pthreadpool",
3158 ],
3159)
3160
3161xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003162 name = "neonfp16arith_ukernels",
3163 hdrs = INTERNAL_HDRS,
3164 aarch64_copts = ["-march=armv8.2-a+fp16"],
3165 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003166 gcc_copts = xnnpack_gcc_std_copts(),
3167 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan04f03be2019-11-19 12:36:47 -08003168 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003169 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003170 "@FP16",
3171 "@pthreadpool",
3172 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003173)
3174
3175xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003176 name = "neonfp16arith_ukernels_test_mode",
3177 hdrs = INTERNAL_HDRS,
3178 aarch64_copts = ["-march=armv8.2-a+fp16"],
3179 aarch64_srcs = AARCH64_NEONFP16ARITH_UKERNELS,
3180 copts = [
3181 "-UNDEBUG",
3182 "-DXNN_TEST_MODE=1",
3183 ],
3184 gcc_copts = xnnpack_gcc_std_copts(),
3185 msvc_copts = xnnpack_msvc_std_copts(),
3186 deps = [
3187 ":tables",
3188 "@FP16",
3189 "@pthreadpool",
3190 ],
3191)
3192
3193xnnpack_cc_library(
Benoit Jacoba9644732020-08-13 12:48:55 -07003194 name = "neondot_ukernels",
3195 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003196 aarch32_copts = [
3197 "-marm",
3198 "-march=armv8.2-a+dotprod",
3199 "-mfpu=neon-fp-armv8",
3200 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003201 aarch32_srcs = NEONDOT_UKERNELS,
3202 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3203 aarch64_srcs = NEONDOT_UKERNELS,
3204 gcc_copts = xnnpack_gcc_std_copts(),
3205 msvc_copts = xnnpack_msvc_std_copts(),
3206 deps = [
3207 ":tables",
3208 "@FP16",
3209 "@pthreadpool",
3210 ],
3211)
3212
3213xnnpack_cc_library(
3214 name = "neondot_ukernels_test_mode",
3215 hdrs = INTERNAL_HDRS,
Marat Dukhan799ac752020-08-13 16:08:03 -07003216 aarch32_copts = [
3217 "-marm",
3218 "-march=armv8.2-a+dotprod",
3219 "-mfpu=neon-fp-armv8",
3220 ],
Benoit Jacoba9644732020-08-13 12:48:55 -07003221 aarch32_srcs = NEONDOT_UKERNELS,
3222 aarch64_copts = ["-march=armv8.2-a+dotprod"],
3223 aarch64_srcs = NEONDOT_UKERNELS,
3224 copts = [
3225 "-UNDEBUG",
3226 "-DXNN_TEST_MODE=1",
3227 ],
3228 gcc_copts = xnnpack_gcc_std_copts(),
3229 msvc_copts = xnnpack_msvc_std_copts(),
3230 deps = [
3231 ":tables",
3232 "@FP16",
3233 "@pthreadpool",
3234 ],
3235)
3236
3237xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003238 name = "sse2_ukernels",
3239 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003240 gcc_copts = xnnpack_gcc_std_copts(),
3241 gcc_x86_copts = ["-msse2"],
3242 msvc_copts = xnnpack_msvc_std_copts(),
3243 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003244 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003245 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003246 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003247 "@FP16",
3248 "@pthreadpool",
3249 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003250)
3251
3252xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003253 name = "sse2_ukernels_test_mode",
3254 hdrs = INTERNAL_HDRS,
3255 copts = [
3256 "-UNDEBUG",
3257 "-DXNN_TEST_MODE=1",
3258 ],
3259 gcc_copts = xnnpack_gcc_std_copts(),
3260 gcc_x86_copts = ["-msse2"],
3261 msvc_copts = xnnpack_msvc_std_copts(),
3262 msvc_x86_32_copts = ["/arch:SSE2"],
3263 x86_srcs = SSE_UKERNELS + SSE2_UKERNELS,
3264 deps = [
3265 ":tables",
3266 "@FP16",
3267 "@pthreadpool",
3268 ],
3269)
3270
3271xnnpack_cc_library(
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003272 name = "ssse3_ukernels",
3273 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003274 gcc_copts = xnnpack_gcc_std_copts(),
3275 gcc_x86_copts = ["-mssse3"],
3276 msvc_copts = xnnpack_msvc_std_copts(),
3277 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003278 x86_srcs = SSSE3_UKERNELS,
3279 deps = [
3280 ":tables",
3281 "@FP16",
3282 "@pthreadpool",
3283 ],
3284)
3285
3286xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003287 name = "ssse3_ukernels_test_mode",
3288 hdrs = INTERNAL_HDRS,
3289 copts = [
3290 "-UNDEBUG",
3291 "-DXNN_TEST_MODE=1",
3292 ],
3293 gcc_copts = xnnpack_gcc_std_copts(),
3294 gcc_x86_copts = ["-mssse3"],
3295 msvc_copts = xnnpack_msvc_std_copts(),
3296 msvc_x86_32_copts = ["/arch:SSE2"],
3297 x86_srcs = SSSE3_UKERNELS,
3298 deps = [
3299 ":tables",
3300 "@FP16",
3301 "@pthreadpool",
3302 ],
3303)
3304
3305xnnpack_cc_library(
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003306 name = "sse41_ukernels",
3307 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003308 gcc_copts = xnnpack_gcc_std_copts(),
3309 gcc_x86_copts = ["-msse4.1"],
3310 msvc_copts = xnnpack_msvc_std_copts(),
3311 msvc_x86_32_copts = ["/arch:SSE2"],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003312 x86_srcs = SSE41_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003313 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003314 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003315 "@FP16",
3316 "@pthreadpool",
3317 ],
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003318)
3319
3320xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003321 name = "sse41_ukernels_test_mode",
3322 hdrs = INTERNAL_HDRS,
3323 copts = [
3324 "-UNDEBUG",
3325 "-DXNN_TEST_MODE=1",
3326 ],
3327 gcc_copts = xnnpack_gcc_std_copts(),
3328 gcc_x86_copts = ["-msse4.1"],
3329 msvc_copts = xnnpack_msvc_std_copts(),
3330 msvc_x86_32_copts = ["/arch:SSE2"],
3331 x86_srcs = SSE41_UKERNELS,
3332 deps = [
3333 ":tables",
3334 "@FP16",
3335 "@pthreadpool",
3336 ],
3337)
3338
3339xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003340 name = "avx_ukernels",
3341 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003342 gcc_copts = xnnpack_gcc_std_copts(),
3343 gcc_x86_copts = ["-mavx"],
3344 msvc_copts = xnnpack_msvc_std_copts(),
3345 msvc_x86_32_copts = ["/arch:AVX"],
3346 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003347 x86_srcs = AVX_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003348 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003349 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003350 "@FP16",
3351 "@pthreadpool",
3352 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003353)
3354
3355xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003356 name = "avx_ukernels_test_mode",
3357 hdrs = INTERNAL_HDRS,
3358 copts = [
3359 "-UNDEBUG",
3360 "-DXNN_TEST_MODE=1",
3361 ],
3362 gcc_copts = xnnpack_gcc_std_copts(),
3363 gcc_x86_copts = ["-mavx"],
3364 msvc_copts = xnnpack_msvc_std_copts(),
3365 msvc_x86_32_copts = ["/arch:AVX"],
3366 msvc_x86_64_copts = ["/arch:AVX"],
3367 x86_srcs = AVX_UKERNELS,
3368 deps = [
3369 ":tables",
3370 "@FP16",
3371 "@pthreadpool",
3372 ],
3373)
3374
3375xnnpack_cc_library(
Marat Dukhan1566fee2020-08-02 21:55:41 -07003376 name = "xop_ukernels",
3377 hdrs = INTERNAL_HDRS,
3378 gcc_copts = xnnpack_gcc_std_copts(),
3379 gcc_x86_copts = ["-mxop"],
3380 msvc_copts = xnnpack_msvc_std_copts(),
3381 msvc_x86_32_copts = ["/arch:AVX"],
3382 msvc_x86_64_copts = ["/arch:AVX"],
3383 x86_srcs = XOP_UKERNELS,
3384 deps = [
3385 ":tables",
3386 "@FP16",
3387 "@pthreadpool",
3388 ],
3389)
3390
3391xnnpack_cc_library(
3392 name = "xop_ukernels_test_mode",
3393 hdrs = INTERNAL_HDRS,
3394 copts = [
3395 "-UNDEBUG",
3396 "-DXNN_TEST_MODE=1",
3397 ],
3398 gcc_copts = xnnpack_gcc_std_copts(),
3399 gcc_x86_copts = ["-mxop"],
3400 msvc_copts = xnnpack_msvc_std_copts(),
3401 msvc_x86_32_copts = ["/arch:AVX"],
3402 msvc_x86_64_copts = ["/arch:AVX"],
3403 x86_srcs = XOP_UKERNELS,
3404 deps = [
3405 ":tables",
3406 "@FP16",
3407 "@pthreadpool",
3408 ],
3409)
3410
3411xnnpack_cc_library(
Marat Dukhanfda12b82019-11-21 12:27:59 -08003412 name = "fma3_ukernels",
3413 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003414 gcc_copts = xnnpack_gcc_std_copts(),
3415 gcc_x86_copts = ["-mfma"],
3416 msvc_copts = xnnpack_msvc_std_copts(),
3417 msvc_x86_32_copts = ["/arch:AVX"],
3418 msvc_x86_64_copts = ["/arch:AVX"],
Marat Dukhanfda12b82019-11-21 12:27:59 -08003419 x86_srcs = FMA3_UKERNELS,
3420 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003421 ":tables",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003422 "@FP16",
3423 "@pthreadpool",
3424 ],
3425)
3426
3427xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003428 name = "fma3_ukernels_test_mode",
3429 hdrs = INTERNAL_HDRS,
3430 copts = [
3431 "-UNDEBUG",
3432 "-DXNN_TEST_MODE=1",
3433 ],
3434 gcc_copts = xnnpack_gcc_std_copts(),
3435 gcc_x86_copts = ["-mfma"],
3436 msvc_copts = xnnpack_msvc_std_copts(),
3437 msvc_x86_32_copts = ["/arch:AVX"],
3438 msvc_x86_64_copts = ["/arch:AVX"],
3439 x86_srcs = FMA3_UKERNELS,
3440 deps = [
3441 ":tables",
3442 "@FP16",
3443 "@pthreadpool",
3444 ],
3445)
3446
3447xnnpack_cc_library(
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003448 name = "avx2_ukernels",
3449 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003450 gcc_copts = xnnpack_gcc_std_copts(),
3451 gcc_x86_copts = [
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003452 "-mfma",
3453 "-mavx2",
3454 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003455 msvc_copts = xnnpack_msvc_std_copts(),
3456 msvc_x86_32_copts = ["/arch:AVX2"],
3457 msvc_x86_64_copts = ["/arch:AVX2"],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003458 x86_srcs = AVX2_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003459 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003460 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003461 "@FP16",
3462 "@pthreadpool",
3463 ],
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003464)
3465
3466xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003467 name = "avx2_ukernels_test_mode",
3468 hdrs = INTERNAL_HDRS,
3469 copts = [
3470 "-UNDEBUG",
3471 "-DXNN_TEST_MODE=1",
3472 ],
3473 gcc_copts = xnnpack_gcc_std_copts(),
3474 gcc_x86_copts = [
3475 "-mfma",
3476 "-mavx2",
3477 ],
3478 msvc_copts = xnnpack_msvc_std_copts(),
3479 msvc_x86_32_copts = ["/arch:AVX2"],
3480 msvc_x86_64_copts = ["/arch:AVX2"],
3481 x86_srcs = AVX2_UKERNELS,
3482 deps = [
3483 ":tables",
3484 "@FP16",
3485 "@pthreadpool",
3486 ],
3487)
3488
3489xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003490 name = "avx512f_ukernels",
3491 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003492 gcc_copts = xnnpack_gcc_std_copts(),
3493 gcc_x86_copts = ["-mavx512f"],
3494 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3495 msvc_copts = xnnpack_msvc_std_copts(),
3496 msvc_x86_32_copts = ["/arch:AVX512"],
3497 msvc_x86_64_copts = ["/arch:AVX512"],
3498 msys_copts = ["-fno-asynchronous-unwind-tables"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003499 x86_srcs = AVX512F_UKERNELS,
Marat Dukhan04f03be2019-11-19 12:36:47 -08003500 deps = [
Marat Dukhan3a77ea72019-12-23 12:10:24 -08003501 ":tables",
Marat Dukhan04f03be2019-11-19 12:36:47 -08003502 "@FP16",
3503 "@pthreadpool",
3504 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003505)
3506
3507xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003508 name = "avx512f_ukernels_test_mode",
3509 hdrs = INTERNAL_HDRS,
3510 copts = [
3511 "-UNDEBUG",
3512 "-DXNN_TEST_MODE=1",
3513 ],
3514 gcc_copts = xnnpack_gcc_std_copts(),
3515 gcc_x86_copts = ["-mavx512f"],
3516 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3517 msvc_copts = xnnpack_msvc_std_copts(),
3518 msvc_x86_32_copts = ["/arch:AVX512"],
3519 msvc_x86_64_copts = ["/arch:AVX512"],
3520 msys_copts = ["-fno-asynchronous-unwind-tables"],
3521 x86_srcs = AVX512F_UKERNELS,
3522 deps = [
3523 ":tables",
3524 "@FP16",
3525 "@pthreadpool",
3526 ],
3527)
3528
3529xnnpack_cc_library(
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003530 name = "avx512skx_ukernels",
3531 hdrs = INTERNAL_HDRS,
3532 gcc_copts = xnnpack_gcc_std_copts(),
3533 gcc_x86_copts = [
3534 "-mavx512f",
3535 "-mavx512cd",
3536 "-mavx512bw",
3537 "-mavx512dq",
3538 "-mavx512vl",
3539 ],
3540 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3541 msvc_copts = xnnpack_msvc_std_copts(),
3542 msvc_x86_32_copts = ["/arch:AVX512"],
3543 msvc_x86_64_copts = ["/arch:AVX512"],
3544 msys_copts = ["-fno-asynchronous-unwind-tables"],
3545 x86_srcs = AVX512SKX_UKERNELS,
3546 deps = [
3547 ":tables",
3548 "@FP16",
3549 "@pthreadpool",
3550 ],
3551)
3552
3553xnnpack_cc_library(
3554 name = "avx512skx_ukernels_test_mode",
3555 hdrs = INTERNAL_HDRS,
3556 copts = [
3557 "-UNDEBUG",
3558 "-DXNN_TEST_MODE=1",
3559 ],
3560 gcc_copts = xnnpack_gcc_std_copts(),
3561 gcc_x86_copts = [
3562 "-mavx512f",
3563 "-mavx512cd",
3564 "-mavx512bw",
3565 "-mavx512dq",
3566 "-mavx512vl",
3567 ],
3568 mingw_copts = ["-fno-asynchronous-unwind-tables"],
3569 msvc_copts = xnnpack_msvc_std_copts(),
3570 msvc_x86_32_copts = ["/arch:AVX512"],
3571 msvc_x86_64_copts = ["/arch:AVX512"],
3572 msys_copts = ["-fno-asynchronous-unwind-tables"],
3573 x86_srcs = AVX512SKX_UKERNELS,
3574 deps = [
3575 ":tables",
3576 "@FP16",
3577 "@pthreadpool",
3578 ],
3579)
3580
3581xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003582 name = "asm_ukernels",
3583 hdrs = ["src/xnnpack/assembly.h"],
3584 aarch32_srcs = AARCH32_ASM_UKERNELS,
Frank Barchard683f5592020-04-10 00:48:26 -07003585 aarch64_copts = ["-march=armv8.2-a+fp16"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003586 aarch64_srcs = AARCH64_ASM_UKERNELS,
Frank Barchardbcedc082020-08-17 18:00:51 -07003587 wasm_srcs = WASM32_ASM_UKERNELS,
3588 wasmsimd_srcs = WASM32_ASM_UKERNELS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07003589)
3590
Marat Dukhan3b59de22020-06-03 20:15:19 -07003591xnnpack_cc_library(
3592 name = "logging_utils",
3593 srcs = LOGGING_SRCS,
3594 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3595 copts = LOGGING_COPTS + [
3596 "-Isrc",
3597 "-Iinclude",
3598 ] + select({
3599 ":debug_build": [],
3600 "//conditions:default": xnnpack_min_size_copts(),
3601 }),
3602 gcc_copts = xnnpack_gcc_std_copts(),
3603 msvc_copts = xnnpack_msvc_std_copts(),
3604 visibility = xnnpack_visibility(),
3605 deps = [
3606 "@FP16",
3607 "@clog",
3608 "@pthreadpool",
3609 ],
3610)
3611
Marat Dukhan08c4a432019-10-03 09:29:21 -07003612xnnpack_aggregate_library(
3613 name = "ukernels",
3614 aarch32_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003615 ":neon_ukernels",
3616 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003617 ":neonv8_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003618 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003619 ":asm_ukernels",
3620 ],
3621 aarch64_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003622 ":neon_ukernels",
3623 ":neonfma_ukernels",
Marat Dukhan8853b822020-05-07 12:19:01 -07003624 ":neonv8_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003625 ":neonfp16arith_ukernels",
Benoit Jacoba9644732020-08-13 12:48:55 -07003626 ":neondot_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003627 ":asm_ukernels",
3628 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003629 generic_deps = [
3630 ":scalar_ukernels",
3631 ],
Marat Dukhan500b8892020-04-15 17:09:50 -07003632 psimd_deps = [
Marat Dukhan8d3c07e2020-01-02 01:20:59 -08003633 ":psimd_fastmath_ukernels",
3634 ":psimd_accmath_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003635 ],
Marat Dukhan33fcf782020-05-24 14:27:15 -07003636 wasm_deps = [
3637 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003638 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003639 ],
3640 wasmsimd_deps = [
3641 ":wasm_ukernels",
Frank Barchardbcedc082020-08-17 18:00:51 -07003642 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003643 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003644 x86_deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003645 ":sse2_ukernels",
Marat Dukhanfe7acb62020-03-09 19:30:05 -07003646 ":ssse3_ukernels",
Marat Dukhan69c3f2c2019-11-06 12:30:01 -08003647 ":sse41_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003648 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003649 ":xop_ukernels",
Marat Dukhanfda12b82019-11-21 12:27:59 -08003650 ":fma3_ukernels",
Marat Dukhan6adff4e2019-10-14 18:32:07 -07003651 ":avx2_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003652 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003653 ":avx512skx_ukernels",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003654 ],
3655)
3656
Marat Dukhan33fcf782020-05-24 14:27:15 -07003657xnnpack_aggregate_library(
3658 name = "ukernels_test_mode",
3659 aarch32_deps = [
3660 ":neon_ukernels_test_mode",
3661 ":neonfma_ukernels_test_mode",
3662 ":neonv8_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003663 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003664 ":asm_ukernels",
3665 ],
3666 aarch64_deps = [
3667 ":neon_ukernels_test_mode",
3668 ":neonfma_ukernels_test_mode",
3669 ":neonv8_ukernels_test_mode",
3670 ":neonfp16arith_ukernels_test_mode",
Benoit Jacoba9644732020-08-13 12:48:55 -07003671 ":neondot_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003672 ":asm_ukernels",
3673 ],
3674 generic_deps = [
3675 ":scalar_ukernels_test_mode",
3676 ],
3677 psimd_deps = [
3678 ":psimd_fastmath_ukernels_test_mode",
3679 ":psimd_accmath_ukernels_test_mode",
3680 ],
3681 wasm_deps = [
3682 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003683 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003684 ],
3685 wasmsimd_deps = [
3686 ":wasm_ukernels_test_mode",
Frank Barchardbcedc082020-08-17 18:00:51 -07003687 ":asm_ukernels",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003688 ],
3689 x86_deps = [
3690 ":sse2_ukernels_test_mode",
3691 ":ssse3_ukernels_test_mode",
3692 ":sse41_ukernels_test_mode",
3693 ":avx_ukernels_test_mode",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003694 ":xop_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003695 ":fma3_ukernels_test_mode",
3696 ":avx2_ukernels_test_mode",
3697 ":avx512f_ukernels_test_mode",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003698 ":avx512skx_ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003699 ],
3700)
3701
Marat Dukhan86df3842020-06-17 14:03:31 -07003702# This target must be compatible with Android NDK r17 and compile with Clang 6.
3703# It is similar to :ukernels target, but excludes microkernels which require
3704# NEON FP16 arithmetics, NEON Dot Product, and newer intrinsics.
Marat Dukhanae046f52020-06-15 13:16:14 -07003705xnnpack_aggregate_library(
Marat Dukhan86df3842020-06-17 14:03:31 -07003706 name = "ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07003707 aarch32_deps = [
3708 ":neon_ukernels",
3709 ":neonfma_ukernels",
3710 ":neonv8_ukernels",
3711 ":asm_ukernels",
3712 ],
3713 aarch64_deps = [
3714 ":neon_ukernels",
3715 ":neonfma_ukernels",
3716 ":neonv8_ukernels",
3717 ":asm_ukernels",
3718 ],
3719 generic_deps = [
3720 ":scalar_ukernels",
3721 ],
3722 psimd_deps = [
3723 ":psimd_fastmath_ukernels",
3724 ":psimd_accmath_ukernels",
3725 ],
3726 wasm_deps = [
3727 ":wasm_ukernels",
3728 ],
3729 wasmsimd_deps = [
3730 ":wasm_ukernels",
3731 ],
3732 x86_deps = [
3733 ":sse2_ukernels",
3734 ":ssse3_ukernels",
3735 ":sse41_ukernels",
3736 ":avx_ukernels",
Marat Dukhan1566fee2020-08-02 21:55:41 -07003737 ":xop_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003738 ":fma3_ukernels",
3739 ":avx2_ukernels",
3740 ":avx512f_ukernels",
Marat Dukhanbb00b1d2020-08-10 11:37:23 -07003741 ":avx512skx_ukernels",
Marat Dukhanae046f52020-06-15 13:16:14 -07003742 ],
3743)
3744
Marat Dukhan08c4a432019-10-03 09:29:21 -07003745xnnpack_cc_library(
3746 name = "im2col",
3747 srcs = ["src/im2col.c"],
3748 hdrs = [
3749 "src/xnnpack/common.h",
3750 "src/xnnpack/im2col.h",
3751 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07003752 gcc_copts = xnnpack_gcc_std_copts(),
3753 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003754)
3755
3756xnnpack_cc_library(
3757 name = "indirection",
3758 srcs = ["src/indirection.c"],
3759 hdrs = INTERNAL_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003760 gcc_copts = xnnpack_gcc_std_copts(),
3761 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003762 deps = [
3763 "@FP16",
3764 "@FXdiv",
3765 "@pthreadpool",
3766 ],
3767)
3768
3769xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003770 name = "indirection_test_mode",
3771 srcs = ["src/indirection.c"],
3772 hdrs = INTERNAL_HDRS,
3773 copts = [
3774 "-UNDEBUG",
3775 "-DXNN_TEST_MODE=1",
3776 ],
3777 gcc_copts = xnnpack_gcc_std_copts(),
3778 msvc_copts = xnnpack_msvc_std_copts(),
3779 deps = [
3780 "@FP16",
3781 "@FXdiv",
3782 "@pthreadpool",
3783 ],
3784)
3785
3786xnnpack_cc_library(
Marat Dukhanab582382020-07-06 13:32:08 -07003787 name = "packing",
3788 srcs = ["src/packing.c"],
3789 hdrs = INTERNAL_HDRS,
3790 gcc_copts = xnnpack_gcc_std_copts(),
3791 msvc_copts = xnnpack_msvc_std_copts(),
3792 deps = [
3793 "@FP16",
3794 "@FXdiv",
3795 "@pthreadpool",
3796 ],
3797)
3798
3799xnnpack_cc_library(
3800 name = "packing_test_mode",
3801 srcs = ["src/packing.c"],
3802 hdrs = INTERNAL_HDRS,
3803 copts = [
3804 "-UNDEBUG",
3805 "-DXNN_TEST_MODE=1",
3806 ],
3807 gcc_copts = xnnpack_gcc_std_copts(),
3808 msvc_copts = xnnpack_msvc_std_copts(),
3809 deps = [
3810 "@FP16",
3811 "@FXdiv",
3812 "@pthreadpool",
3813 ],
3814)
3815
3816xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003817 name = "operator_run",
3818 srcs = ["src/operator-run.c"],
Marat Dukhanc8e00eb2019-10-04 14:55:26 -07003819 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003820 copts = LOGGING_COPTS + select({
Marat Dukhan05702cf2020-03-26 15:41:33 -07003821 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3822 "//conditions:default": [],
3823 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003824 gcc_copts = xnnpack_gcc_std_copts(),
3825 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003826 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003827 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003828 "@FP16",
3829 "@FXdiv",
3830 "@clog",
3831 "@pthreadpool",
3832 ],
3833)
3834
Chao Mei6ddfc602020-05-13 22:29:36 -07003835xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003836 name = "operator_run_test_mode",
3837 srcs = ["src/operator-run.c"],
3838 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3839 copts = LOGGING_COPTS + [
3840 "-UNDEBUG",
3841 "-DXNN_TEST_MODE=1",
3842 ] + select({
3843 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3844 "//conditions:default": [],
3845 }),
3846 gcc_copts = xnnpack_gcc_std_copts(),
3847 msvc_copts = xnnpack_msvc_std_copts(),
3848 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003849 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003850 "@FP16",
3851 "@FXdiv",
3852 "@clog",
3853 "@pthreadpool",
3854 ],
3855)
3856
3857xnnpack_cc_library(
Chao Mei6ddfc602020-05-13 22:29:36 -07003858 name = "memory_planner",
3859 srcs = ["src/memory-planner.c"],
3860 hdrs = INTERNAL_HDRS,
3861 defines = select({
3862 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3863 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3864 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3865 }),
3866 gcc_copts = xnnpack_gcc_std_copts(),
3867 msvc_copts = xnnpack_msvc_std_copts(),
3868 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003869 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07003870 "@pthreadpool",
3871 ],
3872)
3873
Marat Dukhan33fcf782020-05-24 14:27:15 -07003874xnnpack_cc_library(
3875 name = "memory_planner_test_mode",
3876 srcs = ["src/memory-planner.c"],
3877 hdrs = INTERNAL_HDRS,
3878 copts = [
3879 "-UNDEBUG",
3880 "-DXNN_TEST_MODE=1",
3881 ],
3882 defines = select({
3883 ":xnn_enable_memopt_explicit_true": ["XNN_ENABLE_MEMOPT=1"],
3884 ":xnn_enable_memopt_explicit_false": ["XNN_ENABLE_MEMOPT=0"],
3885 "//conditions:default": ["XNN_ENABLE_MEMOPT=1"],
3886 }),
3887 gcc_copts = xnnpack_gcc_std_copts(),
3888 msvc_copts = xnnpack_msvc_std_copts(),
3889 deps = [
Marat Dukhan3b59de22020-06-03 20:15:19 -07003890 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003891 "@pthreadpool",
3892 ],
3893)
3894
Marat Dukhan08c4a432019-10-03 09:29:21 -07003895cc_library(
3896 name = "enable_assembly",
3897 defines = select({
3898 ":xnn_enable_assembly_explicit_true": ["XNN_ENABLE_ASSEMBLY=1"],
3899 ":xnn_enable_assembly_explicit_false": ["XNN_ENABLE_ASSEMBLY=0"],
Frank Barchard810171d2019-10-10 10:34:51 -07003900 "//conditions:default": ["XNN_ENABLE_ASSEMBLY=1"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003901 }),
3902)
3903
Marat Dukhan9de90e02020-06-18 16:04:12 -07003904cc_library(
3905 name = "enable_sparse",
3906 defines = select({
3907 ":xnn_enable_sparse_explicit_true": ["XNN_ENABLE_SPARSE=1"],
3908 ":xnn_enable_sparse_explicit_false": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan8b283aa2020-06-18 16:48:57 -07003909 "//conditions:default": ["XNN_ENABLE_SPARSE=0"],
Marat Dukhan9de90e02020-06-18 16:04:12 -07003910 }),
3911)
3912
Marat Dukhancf056b22019-10-07 10:26:29 -07003913xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07003914 name = "operators",
3915 srcs = OPERATOR_SRCS + [
Marat Dukhan04f03be2019-11-19 12:36:47 -08003916 "src/memory.c",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003917 "src/operator-delete.c",
Marat Dukhancf056b22019-10-07 10:26:29 -07003918 ],
3919 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003920 copts = LOGGING_COPTS + [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003921 "-Isrc",
3922 "-Iinclude",
3923 ] + select({
3924 ":debug_build": [],
3925 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003926 }) + select({
3927 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3928 "//conditions:default": [],
Marat Dukhan08c4a432019-10-03 09:29:21 -07003929 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003930 gcc_copts = xnnpack_gcc_std_copts(),
3931 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07003932 deps = [
Marat Dukhan08c4a432019-10-03 09:29:21 -07003933 ":indirection",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003934 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003935 ":packing",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003936 "@FP16",
3937 "@FXdiv",
3938 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07003939 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003940 ],
3941)
3942
Marat Dukhan10a38082020-04-17 03:58:35 -07003943xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07003944 name = "operators_test_mode",
3945 srcs = OPERATOR_SRCS + [
3946 "src/memory.c",
3947 "src/operator-delete.c",
3948 ],
3949 hdrs = INTERNAL_HDRS + LOGGING_HDRS,
3950 copts = LOGGING_COPTS + [
3951 "-Isrc",
3952 "-Iinclude",
3953 "-UNDEBUG",
3954 "-DXNN_TEST_MODE=1",
3955 ] + select({
3956 ":debug_build": [],
3957 "//conditions:default": xnnpack_min_size_copts(),
3958 }) + select({
3959 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3960 "//conditions:default": [],
3961 }),
3962 gcc_copts = xnnpack_gcc_std_copts(),
3963 msvc_copts = xnnpack_msvc_std_copts(),
3964 deps = [
3965 ":indirection_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07003966 ":logging_utils",
Marat Dukhanab582382020-07-06 13:32:08 -07003967 ":packing_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07003968 "@FP16",
3969 "@FXdiv",
3970 "@clog",
3971 "@pthreadpool",
3972 ],
3973)
3974
3975xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003976 name = "XNNPACK",
3977 srcs = [
3978 "src/init.c",
Marat Dukhanccfdbd12020-02-03 14:27:45 -08003979 "src/runtime.c",
3980 "src/subgraph.c",
3981 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07003982 ] + SUBGRAPH_SRCS,
Marat Dukhan10a38082020-04-17 03:58:35 -07003983 hdrs = ["include/xnnpack.h"],
3984 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003985 "-Isrc",
3986 "-Iinclude",
3987 ] + select({
3988 ":debug_build": [],
3989 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07003990 }) + select({
3991 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
3992 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003993 }),
Marat Dukhan10a38082020-04-17 03:58:35 -07003994 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003995 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07003996 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan8fe54e42019-10-10 14:12:59 -07003997 visibility = xnnpack_visibility(),
3998 deps = [
3999 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004000 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004001 ":logging_utils",
Chao Mei6ddfc602020-05-13 22:29:36 -07004002 ":memory_planner",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004003 ":operator_run",
4004 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004005 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004006 "@clog",
Marat Dukhanab2946c2020-05-21 20:04:13 -07004007 "@FP16",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004008 "@pthreadpool",
Marat Dukhand343c222019-10-07 09:22:14 -07004009 ] + select({
4010 ":emscripten": [],
4011 "//conditions:default": ["@cpuinfo"],
4012 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004013)
4014
Marat Dukhan10a38082020-04-17 03:58:35 -07004015xnnpack_cc_library(
Marat Dukhan33fcf782020-05-24 14:27:15 -07004016 name = "XNNPACK_test_mode",
4017 srcs = [
4018 "src/init.c",
4019 "src/runtime.c",
4020 "src/subgraph.c",
4021 "src/tensor.c",
Marat Dukhanf03da0d2020-06-10 16:00:20 -07004022 ] + SUBGRAPH_SRCS,
Marat Dukhan33fcf782020-05-24 14:27:15 -07004023 hdrs = ["include/xnnpack.h"],
4024 copts = LOGGING_COPTS + [
4025 "-Isrc",
4026 "-Iinclude",
4027 "-UNDEBUG",
4028 "-DXNN_TEST_MODE=1",
4029 ] + select({
4030 ":debug_build": [],
4031 "//conditions:default": xnnpack_min_size_copts(),
4032 }) + select({
4033 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4034 "//conditions:default": [],
4035 }),
4036 gcc_copts = xnnpack_gcc_std_copts(),
4037 includes = ["include"],
4038 msvc_copts = xnnpack_msvc_std_copts(),
4039 visibility = xnnpack_visibility(),
4040 deps = [
4041 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004042 ":enable_sparse",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004043 ":logging_utils",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004044 ":memory_planner_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004045 ":operator_run_test_mode",
4046 ":operators_test_mode",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004047 ":ukernels_test_mode",
Marat Dukhan33fcf782020-05-24 14:27:15 -07004048 "@clog",
4049 "@FP16",
4050 "@pthreadpool",
4051 ] + select({
4052 ":emscripten": [],
4053 "//conditions:default": ["@cpuinfo"],
4054 }),
4055)
4056
4057xnnpack_cc_library(
Marat Dukhanae046f52020-06-15 13:16:14 -07004058 name = "xnnpack_f32",
4059 srcs = [
4060 "src/init.c",
4061 "src/runtime.c",
4062 "src/subgraph.c",
4063 "src/tensor.c",
4064 ] + SUBGRAPH_SRCS,
4065 hdrs = ["include/xnnpack.h"],
4066 copts = LOGGING_COPTS + [
4067 "-Isrc",
4068 "-Iinclude",
4069 ] + select({
4070 ":debug_build": [],
4071 "//conditions:default": xnnpack_min_size_copts(),
4072 }) + select({
4073 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4074 "//conditions:default": [],
4075 }),
4076 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004077 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004078 "XNN_NO_QU8_OPERATORS",
Marat Dukhanae046f52020-06-15 13:16:14 -07004079 "XNN_NO_U8_OPERATORS",
4080 "XNN_NO_X8_OPERATORS",
4081 "XNN_NO_F16_OPERATORS",
4082 "XNN_NO_X16_OPERATORS",
4083 ],
4084 gcc_copts = xnnpack_gcc_std_copts(),
4085 includes = ["include"],
4086 msvc_copts = xnnpack_msvc_std_copts(),
4087 visibility = xnnpack_visibility(),
4088 deps = [
4089 ":enable_assembly",
Marat Dukhan9de90e02020-06-18 16:04:12 -07004090 ":enable_sparse",
Marat Dukhanae046f52020-06-15 13:16:14 -07004091 ":logging_utils",
4092 ":memory_planner",
4093 ":operator_run",
4094 ":operators",
Marat Dukhan86df3842020-06-17 14:03:31 -07004095 ":ukernels_max_compatibility",
Marat Dukhanae046f52020-06-15 13:16:14 -07004096 "@clog",
4097 "@FP16",
4098 "@pthreadpool",
4099 ] + select({
4100 ":emscripten": [],
4101 "//conditions:default": ["@cpuinfo"],
4102 }),
4103)
4104
4105xnnpack_cc_library(
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004106 name = "xnnpack_operators_nhwc_f32",
4107 srcs = [
4108 "src/init.c",
4109 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004110 hdrs = ["include/xnnpack.h"],
4111 copts = LOGGING_COPTS + [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004112 "-Isrc",
4113 "-Iinclude",
4114 ] + select({
4115 ":debug_build": [],
4116 "//conditions:default": xnnpack_min_size_copts(),
Marat Dukhan05702cf2020-03-26 15:41:33 -07004117 }) + select({
4118 ":xnn_enable_hmp_explicit_false": ["-DXNN_MAX_UARCH_TYPES=1"],
4119 "//conditions:default": [],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004120 }),
4121 defines = [
Marat Dukhan16f1e1a2020-08-04 16:38:22 -07004122 "XNN_NO_QS8_OPERATORS",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004123 "XNN_NO_QU8_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004124 "XNN_NO_U8_OPERATORS",
4125 "XNN_NO_X8_OPERATORS",
Marat Dukhanefc47b82019-11-18 09:25:38 -08004126 "XNN_NO_NCHW_OPERATORS",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004127 ],
Marat Dukhan10a38082020-04-17 03:58:35 -07004128 gcc_copts = xnnpack_gcc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004129 includes = ["include"],
Marat Dukhan10a38082020-04-17 03:58:35 -07004130 msvc_copts = xnnpack_msvc_std_copts(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004131 visibility = xnnpack_visibility(),
4132 deps = [
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004133 ":enable_assembly",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004134 ":logging_utils",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004135 ":operator_run",
4136 ":operators",
Marat Dukhan3b59de22020-06-03 20:15:19 -07004137 ":ukernels",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004138 "@clog",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004139 "@pthreadpool",
Marat Dukhan8fe54e42019-10-10 14:12:59 -07004140 ] + select({
4141 ":emscripten": [],
4142 "//conditions:default": ["@cpuinfo"],
4143 }),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004144)
4145
Marat Dukhancf056b22019-10-07 10:26:29 -07004146xnnpack_cc_library(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004147 name = "bench_utils",
4148 srcs = ["bench/utils.cc"],
4149 hdrs = ["bench/utils.h"],
Marat Dukhanbad48fe2019-11-04 10:35:22 -08004150 deps = [
4151 "@com_google_benchmark//:benchmark",
4152 "@cpuinfo",
4153 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004154)
4155
Frank Barchard7e955972019-10-11 10:34:25 -07004156######################### Benchmarks for micro-kernels #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07004157
4158xnnpack_benchmark(
Marat Dukhan595e1702020-07-31 10:12:52 -07004159 name = "qs8_gemm_bench",
4160 srcs = [
4161 "bench/gemm.h",
4162 "bench/qs8-gemm.cc",
4163 "src/xnnpack/AlignedAllocator.h",
4164 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
4165 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"],
4166)
4167
4168xnnpack_benchmark(
Marat Dukhan56bdd4a2020-08-03 19:47:04 -07004169 name = "qs8_requantization_bench",
4170 srcs = [
4171 "bench/qs8-requantization.cc",
4172 "src/xnnpack/requantization-stubs.h",
4173 "src/xnnpack/AlignedAllocator.h",
4174 ] + MICROKERNEL_BENCHMARK_HDRS,
4175 deps = MICROKERNEL_BENCHMARK_DEPS,
4176)
4177
4178xnnpack_benchmark(
Marat Dukhan08b7a972020-07-14 18:17:29 -07004179 name = "qu8_gemm_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004180 srcs = [
4181 "bench/gemm.h",
Marat Dukhan08b7a972020-07-14 18:17:29 -07004182 "bench/qu8-gemm.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004183 "src/xnnpack/AlignedAllocator.h",
4184 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004185 copts = xnnpack_optional_ruy_copts() + xnnpack_optional_gemmlowp_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004186 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps() + xnnpack_optional_gemmlowp_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004187)
4188
4189xnnpack_benchmark(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07004190 name = "qu8_requantization_bench",
4191 srcs = [
4192 "bench/qu8-requantization.cc",
4193 "src/xnnpack/requantization-stubs.h",
4194 "src/xnnpack/AlignedAllocator.h",
4195 ] + MICROKERNEL_BENCHMARK_HDRS,
4196 deps = MICROKERNEL_BENCHMARK_DEPS,
4197)
4198
4199xnnpack_benchmark(
Frank Barchard40d20fe2020-05-05 00:37:45 -07004200 name = "f16_igemm_bench",
4201 srcs = [
4202 "bench/f16-igemm.cc",
4203 "bench/conv.h",
4204 "bench/google/conv.h",
4205 "src/xnnpack/AlignedAllocator.h",
4206 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004207 deps = MICROKERNEL_BENCHMARK_DEPS + [
4208 ":indirection",
4209 ":packing",
4210 ],
Frank Barchard40d20fe2020-05-05 00:37:45 -07004211)
4212
4213xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004214 name = "f16_gemm_bench",
4215 srcs = [
4216 "bench/f16-gemm.cc",
4217 "bench/gemm.h",
4218 "src/xnnpack/AlignedAllocator.h",
4219 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004220 deps = MICROKERNEL_BENCHMARK_DEPS + [
4221 ":packing",
4222 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004223)
4224
4225xnnpack_benchmark(
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004226 name = "f16_spmm_bench",
4227 srcs = [
4228 "bench/f16-spmm.cc",
4229 "bench/gemm.h",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004230 "src/xnnpack/AlignedAllocator.h",
4231 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004232 deps = MICROKERNEL_BENCHMARK_DEPS,
4233)
4234
4235xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004236 name = "f32_igemm_bench",
4237 srcs = [
4238 "bench/f32-igemm.cc",
4239 "bench/conv.h",
4240 "src/xnnpack/AlignedAllocator.h",
4241 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004242 deps = MICROKERNEL_BENCHMARK_DEPS + [
4243 ":indirection",
4244 ":packing",
4245 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004246)
4247
4248xnnpack_benchmark(
Frank Barchardc2010062020-07-19 10:04:12 -07004249 name = "f16_relu_bench",
4250 srcs = [
4251 "bench/f16-relu.cc",
4252 "src/xnnpack/AlignedAllocator.h",
4253 ] + MICROKERNEL_BENCHMARK_HDRS,
4254 deps = MICROKERNEL_BENCHMARK_DEPS,
4255)
4256
4257xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004258 name = "f32_conv_hwc_bench",
4259 srcs = [
4260 "bench/f32-conv-hwc.cc",
4261 "bench/dconv.h",
4262 "src/xnnpack/AlignedAllocator.h",
4263 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004264 deps = MICROKERNEL_BENCHMARK_DEPS + [
4265 ":packing",
4266 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004267)
4268
4269xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004270 name = "f32_conv_hwc2chw_bench",
Erich Elsen563df5f2019-10-23 08:02:21 -07004271 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004272 "bench/f32-conv-hwc2chw.cc",
Erich Elsen563df5f2019-10-23 08:02:21 -07004273 "bench/dconv.h",
4274 "src/xnnpack/AlignedAllocator.h",
4275 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004276 deps = MICROKERNEL_BENCHMARK_DEPS + [
4277 ":packing",
4278 ],
Erich Elsen563df5f2019-10-23 08:02:21 -07004279)
4280
4281xnnpack_benchmark(
Frank Barchard5a599a62020-06-04 20:12:44 -07004282 name = "f16_dwconv_bench",
4283 srcs = [
4284 "bench/f16-dwconv.cc",
4285 "bench/dwconv.h",
4286 "bench/google/dwconv.h",
4287 "src/xnnpack/AlignedAllocator.h",
4288 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004289 deps = MICROKERNEL_BENCHMARK_DEPS + [
4290 ":indirection",
4291 ":packing",
4292 ],
Frank Barchard5a599a62020-06-04 20:12:44 -07004293)
4294
4295xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004296 name = "f32_dwconv_bench",
4297 srcs = [
4298 "bench/f32-dwconv.cc",
4299 "bench/dwconv.h",
4300 "src/xnnpack/AlignedAllocator.h",
4301 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004302 deps = MICROKERNEL_BENCHMARK_DEPS + [
4303 ":indirection",
4304 ":packing",
4305 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004306)
4307
4308xnnpack_benchmark(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004309 name = "f32_dwconv_chw_bench",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004310 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004311 "bench/f32-dwconv-chw.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004312 "bench/dwconv.h",
4313 "src/xnnpack/AlignedAllocator.h",
4314 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004315 deps = MICROKERNEL_BENCHMARK_DEPS + [
4316 ":indirection",
4317 ":packing",
4318 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004319)
4320
4321xnnpack_benchmark(
4322 name = "f32_gemm_bench",
4323 srcs = [
4324 "bench/f32-gemm.cc",
4325 "bench/gemm.h",
4326 "src/xnnpack/AlignedAllocator.h",
4327 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004328 copts = xnnpack_optional_ruy_copts(),
Marat Dukhanab582382020-07-06 13:32:08 -07004329 deps = MICROKERNEL_BENCHMARK_DEPS + [":packing"] + xnnpack_optional_ruy_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004330)
4331
4332xnnpack_benchmark(
Marat Dukhana11ca342020-06-25 23:45:07 -07004333 name = "f32_hswish_bench",
4334 srcs = [
4335 "bench/f32-hswish.cc",
4336 "src/xnnpack/AlignedAllocator.h",
4337 ] + MICROKERNEL_BENCHMARK_HDRS,
4338 deps = MICROKERNEL_BENCHMARK_DEPS,
4339)
4340
4341xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004342 name = "f32_raddexpminusmax_bench",
4343 srcs = [
4344 "bench/f32-raddexpminusmax.cc",
4345 "src/xnnpack/AlignedAllocator.h",
4346 ] + MICROKERNEL_BENCHMARK_HDRS,
4347 deps = MICROKERNEL_BENCHMARK_DEPS,
4348)
4349
4350xnnpack_benchmark(
4351 name = "f32_raddextexp_bench",
4352 srcs = [
4353 "bench/f32-raddextexp.cc",
4354 "src/xnnpack/AlignedAllocator.h",
4355 ] + MICROKERNEL_BENCHMARK_HDRS,
4356 deps = MICROKERNEL_BENCHMARK_DEPS,
4357)
4358
4359xnnpack_benchmark(
4360 name = "f32_raddstoreexpminusmax_bench",
4361 srcs = [
4362 "bench/f32-raddstoreexpminusmax.cc",
4363 "src/xnnpack/AlignedAllocator.h",
4364 ] + MICROKERNEL_BENCHMARK_HDRS,
4365 deps = MICROKERNEL_BENCHMARK_DEPS,
4366)
4367
4368xnnpack_benchmark(
Frank Barchard3c4a9522020-07-18 16:11:22 -07004369 name = "f32_relu_bench",
4370 srcs = [
4371 "bench/f32-relu.cc",
4372 "src/xnnpack/AlignedAllocator.h",
4373 ] + MICROKERNEL_BENCHMARK_HDRS,
4374 deps = MICROKERNEL_BENCHMARK_DEPS,
4375)
4376
4377xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004378 name = "f32_rmax_bench",
4379 srcs = [
4380 "bench/f32-rmax.cc",
4381 "src/xnnpack/AlignedAllocator.h",
4382 ] + MICROKERNEL_BENCHMARK_HDRS,
4383 deps = MICROKERNEL_BENCHMARK_DEPS,
4384)
4385
4386xnnpack_benchmark(
Marat Dukhan14bec502019-11-18 11:35:31 -08004387 name = "f32_sigmoid_bench",
4388 srcs = [
4389 "bench/f32-sigmoid.cc",
4390 "src/xnnpack/AlignedAllocator.h",
4391 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan14bec502019-11-18 11:35:31 -08004392 deps = MICROKERNEL_BENCHMARK_DEPS,
4393)
4394
4395xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004396 name = "f32_spmm_bench",
4397 srcs = [
4398 "bench/f32-spmm.cc",
4399 "bench/gemm.h",
4400 "src/xnnpack/AlignedAllocator.h",
4401 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004402 deps = MICROKERNEL_BENCHMARK_DEPS,
4403)
4404
4405xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004406 name = "f32_softmax_bench",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004407 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004408 "bench/f32-softmax.cc",
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004409 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan10a38082020-04-17 03:58:35 -07004410 copts = xnnpack_optional_dnnl_copts(),
Marat Dukhan8d3c6932020-03-06 20:27:27 -08004411 deps = MICROKERNEL_BENCHMARK_DEPS + xnnpack_optional_dnnl_deps(),
Marat Dukhan4a4a7fa2019-10-21 13:46:14 -07004412)
4413
4414xnnpack_benchmark(
Marat Dukhan4c4eb002019-12-08 21:27:49 -08004415 name = "f32_vscaleexpminusmax_bench",
4416 srcs = [
4417 "bench/f32-vscaleexpminusmax.cc",
4418 "src/xnnpack/AlignedAllocator.h",
4419 ] + MICROKERNEL_BENCHMARK_HDRS,
4420 deps = MICROKERNEL_BENCHMARK_DEPS,
4421)
4422
4423xnnpack_benchmark(
4424 name = "f32_vscaleextexp_bench",
4425 srcs = [
4426 "bench/f32-vscaleextexp.cc",
4427 "src/xnnpack/AlignedAllocator.h",
4428 ] + MICROKERNEL_BENCHMARK_HDRS,
4429 deps = MICROKERNEL_BENCHMARK_DEPS,
4430)
4431
4432xnnpack_benchmark(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07004433 name = "f32_vsqrt_bench",
4434 srcs = [
4435 "bench/f32-vsqrt.cc",
4436 "src/xnnpack/AlignedAllocator.h",
4437 ] + MICROKERNEL_BENCHMARK_HDRS,
4438 deps = MICROKERNEL_BENCHMARK_DEPS,
4439)
4440
4441xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004442 name = "f32_im2col_gemm_bench",
4443 srcs = [
4444 "bench/f32-im2col-gemm.cc",
4445 "bench/conv.h",
4446 "src/xnnpack/AlignedAllocator.h",
4447 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004448 deps = MICROKERNEL_BENCHMARK_DEPS + [
4449 ":im2col",
4450 ":packing",
4451 ],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004452)
4453
Marat Dukhanfe7acb62020-03-09 19:30:05 -07004454xnnpack_benchmark(
Marat Dukhanffbf96a2020-05-14 02:59:08 -07004455 name = "rounding_bench",
4456 srcs = [
4457 "bench/rounding.cc",
4458 "src/xnnpack/math-stubs.h",
4459 "src/xnnpack/AlignedAllocator.h",
4460 ] + MICROKERNEL_BENCHMARK_HDRS,
4461 deps = MICROKERNEL_BENCHMARK_DEPS,
4462)
4463
Marat Dukhan08c4a432019-10-03 09:29:21 -07004464########################### Benchmarks for operators ###########################
4465
4466xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004467 name = "average_pooling_bench",
4468 srcs = ["bench/average-pooling.cc"],
Marat Dukhan7a16d8b2020-03-11 04:22:44 -07004469 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004470 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004471 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004472)
4473
4474xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004475 name = "bankers_rounding_bench",
4476 srcs = ["bench/bankers-rounding.cc"],
4477 copts = xnnpack_optional_tflite_copts(),
4478 tags = ["nowin32"],
4479 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4480)
4481
4482xnnpack_benchmark(
4483 name = "ceiling_bench",
4484 srcs = ["bench/ceiling.cc"],
4485 copts = xnnpack_optional_tflite_copts(),
4486 tags = ["nowin32"],
4487 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4488)
4489
4490xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004491 name = "channel_shuffle_bench",
4492 srcs = ["bench/channel-shuffle.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004493 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004494)
4495
4496xnnpack_benchmark(
4497 name = "convolution_bench",
4498 srcs = ["bench/convolution.cc"],
4499 copts = xnnpack_optional_tflite_copts() + xnnpack_optional_armcl_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004500 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004501 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps() + xnnpack_optional_armcl_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004502)
4503
4504xnnpack_benchmark(
4505 name = "deconvolution_bench",
4506 srcs = ["bench/deconvolution.cc"],
4507 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004508 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004509 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004510)
4511
4512xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004513 name = "floor_bench",
4514 srcs = ["bench/floor.cc"],
4515 copts = xnnpack_optional_tflite_copts(),
4516 tags = ["nowin32"],
4517 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4518)
4519
4520xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004521 name = "global_average_pooling_bench",
4522 srcs = ["bench/global-average-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004523 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004524)
4525
4526xnnpack_benchmark(
Marat Dukhanad352602020-06-25 21:50:54 -07004527 name = "hardswish_bench",
4528 srcs = ["bench/hardswish.cc"],
4529 copts = xnnpack_optional_tflite_copts(),
4530 tags = ["nowin32"],
4531 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4532)
4533
4534xnnpack_benchmark(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004535 name = "max_pooling_bench",
4536 srcs = ["bench/max-pooling.cc"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004537 deps = OPERATOR_BENCHMARK_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07004538)
4539
4540xnnpack_benchmark(
4541 name = "sigmoid_bench",
4542 srcs = ["bench/sigmoid.cc"],
Marat Dukhanc3b9e862019-11-17 13:18:54 -08004543 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004544 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004545 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004546)
4547
4548xnnpack_benchmark(
Marat Dukhan95b22432019-10-30 16:30:14 -07004549 name = "prelu_bench",
4550 srcs = ["bench/prelu.cc"],
4551 copts = xnnpack_optional_tflite_copts(),
Marat Dukhan8ea0b072020-04-23 16:12:18 -07004552 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004553 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan95b22432019-10-30 16:30:14 -07004554)
4555
4556xnnpack_benchmark(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08004557 name = "softmax_bench",
4558 srcs = ["bench/softmax.cc"],
Marat Dukhan9c0db962020-01-28 12:30:14 -08004559 copts = xnnpack_optional_tflite_copts(),
Marat Dukhanca2ba702020-04-24 01:31:47 -07004560 tags = ["nowin32"],
Marat Dukhan1b354632020-03-23 12:50:22 -07004561 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
Marat Dukhan08c4a432019-10-03 09:29:21 -07004562)
4563
Marat Dukhan87727142020-06-24 15:24:10 -07004564xnnpack_benchmark(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07004565 name = "square_root_bench",
4566 srcs = ["bench/square-root.cc"],
4567 copts = xnnpack_optional_tflite_copts(),
4568 tags = ["nowin32"],
4569 deps = OPERATOR_BENCHMARK_DEPS + xnnpack_optional_tflite_deps(),
4570)
4571
4572xnnpack_benchmark(
Marat Dukhan87727142020-06-24 15:24:10 -07004573 name = "truncation_bench",
4574 srcs = ["bench/truncation.cc"],
4575 deps = OPERATOR_BENCHMARK_DEPS,
4576)
4577
Marat Dukhanc068bb62019-10-04 13:24:39 -07004578############################# End-to-end benchmarks ############################
4579
4580cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004581 name = "fp32_mobilenet_v1",
4582 srcs = ["models/fp32-mobilenet-v1.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004583 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004584 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004585 linkstatic = True,
4586 deps = [
4587 ":XNNPACK",
4588 "@pthreadpool",
4589 ],
4590)
4591
4592cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004593 name = "fp16_mobilenet_v1",
4594 srcs = ["models/fp16-mobilenet-v1.cc"],
4595 hdrs = ["models/models.h"],
4596 copts = xnnpack_std_cxxopts(),
4597 linkstatic = True,
4598 deps = [
4599 ":XNNPACK",
4600 "@FP16",
4601 "@pthreadpool",
4602 ],
4603)
4604
4605cc_library(
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004606 name = "qs8_mobilenet_v1",
4607 srcs = ["models/qs8-mobilenet-v1.cc"],
4608 hdrs = ["models/models.h"],
4609 copts = xnnpack_std_cxxopts(),
4610 linkstatic = True,
4611 deps = [
4612 ":XNNPACK",
4613 "@pthreadpool",
4614 ],
4615)
4616
4617cc_library(
Marat Dukhan70a96182020-09-03 17:13:58 -07004618 name = "qs8_mobilenet_v2",
4619 srcs = ["models/qs8-mobilenet-v2.cc"],
4620 hdrs = ["models/models.h"],
4621 copts = xnnpack_std_cxxopts(),
4622 linkstatic = True,
4623 deps = [
4624 ":XNNPACK",
4625 "@pthreadpool",
4626 ],
4627)
4628
4629cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004630 name = "fp32_mobilenet_v2",
4631 srcs = ["models/fp32-mobilenet-v2.cc"],
Marat Dukhanc068bb62019-10-04 13:24:39 -07004632 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004633 copts = xnnpack_std_cxxopts(),
Marat Dukhanc068bb62019-10-04 13:24:39 -07004634 linkstatic = True,
4635 deps = [
4636 ":XNNPACK",
4637 "@pthreadpool",
4638 ],
4639)
4640
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004641cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004642 name = "fp16_mobilenet_v2",
4643 srcs = ["models/fp16-mobilenet-v2.cc"],
4644 hdrs = ["models/models.h"],
4645 copts = xnnpack_std_cxxopts(),
4646 linkstatic = True,
4647 deps = [
4648 ":XNNPACK",
4649 "@FP16",
4650 "@pthreadpool",
4651 ],
4652)
4653
4654cc_library(
4655 name = "fp32_mobilenet_v3_large",
4656 srcs = ["models/fp32-mobilenet-v3-large.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004657 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004658 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004659 linkstatic = True,
4660 deps = [
4661 ":XNNPACK",
4662 "@pthreadpool",
4663 ],
4664)
4665
4666cc_library(
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004667 name = "fp16_mobilenet_v3_large",
4668 srcs = ["models/fp16-mobilenet-v3-large.cc"],
4669 hdrs = ["models/models.h"],
4670 copts = xnnpack_std_cxxopts(),
4671 linkstatic = True,
4672 deps = [
4673 ":XNNPACK",
4674 "@FP16",
4675 "@pthreadpool",
4676 ],
4677)
4678
4679cc_library(
Marat Dukhan270a2c42020-06-26 16:45:52 -07004680 name = "fp32_mobilenet_v3_small",
4681 srcs = ["models/fp32-mobilenet-v3-small.cc"],
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004682 hdrs = ["models/models.h"],
Marat Dukhana84e40b2019-12-11 15:38:03 -08004683 copts = xnnpack_std_cxxopts(),
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004684 linkstatic = True,
4685 deps = [
4686 ":XNNPACK",
4687 "@pthreadpool",
4688 ],
4689)
4690
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004691cc_library(
4692 name = "fp16_mobilenet_v3_small",
4693 srcs = ["models/fp16-mobilenet-v3-small.cc"],
4694 hdrs = ["models/models.h"],
4695 copts = xnnpack_std_cxxopts(),
4696 linkstatic = True,
4697 deps = [
4698 ":XNNPACK",
4699 "@FP16",
4700 "@pthreadpool",
4701 ],
4702)
4703
Marat Dukhanc068bb62019-10-04 13:24:39 -07004704xnnpack_benchmark(
Marat Dukhanef4416e2019-10-31 13:44:40 -07004705 name = "f32_dwconv_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004706 srcs = [
4707 "bench/f32-dwconv-e2e.cc",
4708 "bench/end2end.h",
4709 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhanef4416e2019-10-31 13:44:40 -07004710 deps = MICROKERNEL_BENCHMARK_DEPS + [
4711 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004712 ":fp32_mobilenet_v1",
4713 ":fp32_mobilenet_v2",
4714 ":fp32_mobilenet_v3_large",
4715 ":fp32_mobilenet_v3_small",
Marat Dukhanef4416e2019-10-31 13:44:40 -07004716 ],
4717)
4718
4719xnnpack_benchmark(
Marat Dukhan5f18d262019-10-31 10:24:14 -07004720 name = "f32_gemm_e2e_bench",
Marat Dukhanc08cdf52019-12-09 09:17:51 -08004721 srcs = [
4722 "bench/f32-gemm-e2e.cc",
4723 "bench/end2end.h",
4724 ] + MICROKERNEL_BENCHMARK_HDRS,
Marat Dukhan5f18d262019-10-31 10:24:14 -07004725 deps = MICROKERNEL_BENCHMARK_DEPS + [
4726 ":XNNPACK",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004727 ":fp32_mobilenet_v1",
4728 ":fp32_mobilenet_v2",
4729 ":fp32_mobilenet_v3_large",
4730 ":fp32_mobilenet_v3_small",
Marat Dukhan5f18d262019-10-31 10:24:14 -07004731 ],
4732)
4733
4734xnnpack_benchmark(
Marat Dukhanc068bb62019-10-04 13:24:39 -07004735 name = "end2end_bench",
4736 srcs = ["bench/end2end.cc"],
4737 deps = [
4738 ":XNNPACK",
Frank Barchardc712fa42019-10-31 14:00:21 -07004739 ":bench_utils",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004740 ":fp16_mobilenet_v1",
4741 ":fp16_mobilenet_v2",
Marat Dukhan66f3ccd2020-09-14 16:09:38 -07004742 ":fp16_mobilenet_v3_large",
4743 ":fp16_mobilenet_v3_small",
Marat Dukhan270a2c42020-06-26 16:45:52 -07004744 ":fp32_mobilenet_v1",
4745 ":fp32_mobilenet_v2",
4746 ":fp32_mobilenet_v3_large",
4747 ":fp32_mobilenet_v3_small",
Marat Dukhan0743cdf2020-08-04 18:52:07 -07004748 ":qs8_mobilenet_v1",
Marat Dukhan70a96182020-09-03 17:13:58 -07004749 ":qs8_mobilenet_v2",
Marat Dukhanc068bb62019-10-04 13:24:39 -07004750 "@pthreadpool",
4751 ],
4752)
4753
Marat Dukhan6adff4e2019-10-14 18:32:07 -07004754#################### Accuracy evaluation for math functions ####################
4755
4756xnnpack_benchmark(
4757 name = "f32_exp_eval",
4758 srcs = [
4759 "eval/f32-exp.cc",
4760 "src/xnnpack/AlignedAllocator.h",
4761 ] + ACCURACY_EVAL_HDRS,
4762 deps = ACCURACY_EVAL_DEPS,
4763)
4764
Marat Dukhan515c9772019-10-17 18:07:57 -07004765xnnpack_benchmark(
4766 name = "f32_expminus_eval",
4767 srcs = [
4768 "eval/f32-expminus.cc",
4769 "src/xnnpack/AlignedAllocator.h",
4770 ] + ACCURACY_EVAL_HDRS,
4771 deps = ACCURACY_EVAL_DEPS,
4772)
4773
Marat Dukhan98ba4412019-10-23 02:14:28 -07004774xnnpack_benchmark(
4775 name = "f32_extexp_eval",
4776 srcs = [
4777 "eval/f32-extexp.cc",
4778 "src/xnnpack/AlignedAllocator.h",
4779 ] + ACCURACY_EVAL_HDRS,
4780 deps = ACCURACY_EVAL_DEPS,
4781)
4782
Marat Dukhan8853b822020-05-07 12:19:01 -07004783xnnpack_unit_test(
4784 name = "f32_roundne_eval",
4785 srcs = [
4786 "eval/f32-roundne.cc",
4787 "src/xnnpack/AlignedAllocator.h",
4788 "src/xnnpack/math-stubs.h",
4789 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhan22eed3d2020-05-11 20:13:37 -07004790 automatic = False,
Marat Dukhan8853b822020-05-07 12:19:01 -07004791 deps = MICROKERNEL_TEST_DEPS,
4792)
4793
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004794xnnpack_unit_test(
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004795 name = "f32_roundd_eval",
4796 srcs = [
4797 "eval/f32-roundd.cc",
4798 "src/xnnpack/AlignedAllocator.h",
4799 "src/xnnpack/math-stubs.h",
4800 ] + MICROKERNEL_TEST_HDRS,
4801 automatic = False,
4802 deps = MICROKERNEL_TEST_DEPS,
4803)
4804
4805xnnpack_unit_test(
4806 name = "f32_roundu_eval",
4807 srcs = [
4808 "eval/f32-roundu.cc",
4809 "src/xnnpack/AlignedAllocator.h",
4810 "src/xnnpack/math-stubs.h",
4811 ] + MICROKERNEL_TEST_HDRS,
4812 automatic = False,
4813 deps = MICROKERNEL_TEST_DEPS,
4814)
4815
4816xnnpack_unit_test(
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004817 name = "f32_roundz_eval",
4818 srcs = [
4819 "eval/f32-roundz.cc",
4820 "src/xnnpack/AlignedAllocator.h",
4821 "src/xnnpack/math-stubs.h",
4822 ] + MICROKERNEL_TEST_HDRS,
Marat Dukhanc9852ba2020-05-13 17:21:29 -07004823 automatic = False,
Marat Dukhan2dbb9442020-05-12 20:43:43 -07004824 deps = MICROKERNEL_TEST_DEPS,
4825)
4826
Marat Dukhan346a9e52019-11-15 09:06:30 -08004827xnnpack_benchmark(
4828 name = "f32_sigmoid_eval",
4829 srcs = [
4830 "eval/f32-sigmoid.cc",
4831 "src/xnnpack/AlignedAllocator.h",
4832 ] + ACCURACY_EVAL_HDRS,
4833 deps = ACCURACY_EVAL_DEPS,
4834)
4835
Marat Dukhan84000762020-06-29 18:38:43 -07004836xnnpack_benchmark(
4837 name = "f32_sqrt_eval",
4838 srcs = [
4839 "eval/f32-sqrt.cc",
4840 "src/xnnpack/AlignedAllocator.h",
4841 ] + ACCURACY_EVAL_HDRS,
4842 deps = ACCURACY_EVAL_DEPS + [
4843 ":bench_utils",
4844 ],
4845)
4846
Marat Dukhan08c4a432019-10-03 09:29:21 -07004847######################### Unit tests for micro-kernels #########################
4848
4849xnnpack_unit_test(
Marat Dukhande06f492020-04-09 00:19:31 -07004850 name = "f16_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004851 srcs = [
Marat Dukhande06f492020-04-09 00:19:31 -07004852 "test/f16-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004853 "test/gemm-microkernel-tester.h",
4854 "src/xnnpack/AlignedAllocator.h",
4855 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004856 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004857)
4858
4859xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07004860 name = "f16_spmm_minmax_test",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004861 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07004862 "test/f16-spmm-minmax.cc",
Marat Dukhanbdb56f52020-02-05 21:42:49 -08004863 "test/spmm-microkernel-tester.h",
4864 "src/xnnpack/AlignedAllocator.h",
4865 ] + MICROKERNEL_TEST_HDRS,
4866 deps = MICROKERNEL_TEST_DEPS,
4867)
4868
4869xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004870 name = "f32_argmaxpool_test",
4871 srcs = [
4872 "test/f32-argmaxpool.cc",
4873 "test/argmaxpool-microkernel-tester.h",
4874 "src/xnnpack/AlignedAllocator.h",
4875 ] + MICROKERNEL_TEST_HDRS,
4876 deps = MICROKERNEL_TEST_DEPS,
4877)
4878
4879xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07004880 name = "f32_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004881 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07004882 "test/f32-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004883 "test/avgpool-microkernel-tester.h",
4884 "src/xnnpack/AlignedAllocator.h",
4885 ] + MICROKERNEL_TEST_HDRS,
4886 deps = MICROKERNEL_TEST_DEPS,
4887)
4888
4889xnnpack_unit_test(
Marat Dukhan660fd192020-03-10 04:55:30 -07004890 name = "f32_ibilinear_test",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004891 srcs = [
Marat Dukhan660fd192020-03-10 04:55:30 -07004892 "test/f32-ibilinear.cc",
4893 "test/ibilinear-microkernel-tester.h",
Marat Dukhan35dacfb2019-11-07 19:18:16 -08004894 "src/xnnpack/AlignedAllocator.h",
4895 ] + MICROKERNEL_TEST_HDRS,
4896 deps = MICROKERNEL_TEST_DEPS,
4897)
4898
4899xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07004900 name = "f16_clamp_test",
4901 srcs = [
4902 "test/f16-clamp.cc",
4903 "test/clamp-microkernel-tester.h",
4904 ] + MICROKERNEL_TEST_HDRS,
4905 deps = MICROKERNEL_TEST_DEPS,
4906)
4907
4908xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004909 name = "f32_clamp_test",
4910 srcs = [
4911 "test/f32-clamp.cc",
4912 "test/clamp-microkernel-tester.h",
4913 ] + MICROKERNEL_TEST_HDRS,
4914 deps = MICROKERNEL_TEST_DEPS,
4915)
4916
4917xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004918 name = "f32_igemm_test",
4919 srcs = [
4920 "test/f32-igemm.cc",
4921 "test/gemm-microkernel-tester.h",
4922 "src/xnnpack/AlignedAllocator.h",
4923 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004924 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004925)
4926
4927xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07004928 name = "f32_igemm_relu_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004929 srcs = [
Marat Dukhan467f6362020-05-22 23:21:55 -07004930 "test/f32-igemm-relu.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004931 "test/gemm-microkernel-tester.h",
4932 "src/xnnpack/AlignedAllocator.h",
4933 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004934 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004935)
4936
4937xnnpack_unit_test(
Marat Dukhane207b7b2020-05-28 16:27:42 -07004938 name = "f32_igemm_minmax_test",
4939 srcs = [
4940 "test/f32-igemm-minmax.cc",
4941 "test/gemm-microkernel-tester.h",
4942 "src/xnnpack/AlignedAllocator.h",
4943 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004944 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhane207b7b2020-05-28 16:27:42 -07004945)
4946
4947xnnpack_unit_test(
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004948 name = "f16_igemm_minmax_test",
4949 srcs = [
4950 "test/f16-igemm-minmax.cc",
4951 "test/gemm-microkernel-tester.h",
4952 "src/xnnpack/AlignedAllocator.h",
4953 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004954 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchardb0e4fae2020-05-04 15:27:51 -07004955)
4956
4957xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07004958 name = "f32_conv_hwc_test",
4959 srcs = [
4960 "test/f32-conv-hwc.cc",
4961 "test/conv-hwc-microkernel-tester.h",
4962 "src/xnnpack/AlignedAllocator.h",
4963 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004964 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004965)
4966
4967xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07004968 name = "f32_conv_hwc2chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004969 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07004970 "test/f32-conv-hwc2chw.cc",
4971 "test/conv-hwc2chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004972 "src/xnnpack/AlignedAllocator.h",
4973 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004974 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07004975)
4976
4977xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07004978 name = "f32_dwconv_test",
4979 srcs = [
4980 "test/f32-dwconv.cc",
4981 "test/dwconv-microkernel-tester.h",
4982 "src/xnnpack/AlignedAllocator.h",
4983 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004984 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07004985)
4986
4987xnnpack_unit_test(
Frank Barchard5a599a62020-06-04 20:12:44 -07004988 name = "f16_dwconv_minmax_test",
4989 srcs = [
4990 "test/f16-dwconv-minmax.cc",
4991 "test/dwconv-microkernel-tester.h",
4992 "src/xnnpack/AlignedAllocator.h",
4993 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07004994 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard5a599a62020-06-04 20:12:44 -07004995)
4996
4997xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07004998 name = "f32_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07004999 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005000 "test/f32-dwconv-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005001 "test/dwconv-microkernel-tester.h",
5002 "src/xnnpack/AlignedAllocator.h",
5003 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005004 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005005)
5006
5007xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005008 name = "f32_dwconv_chw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005009 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005010 "test/f32-dwconv-chw.cc",
5011 "test/dwconv-chw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005012 "src/xnnpack/AlignedAllocator.h",
5013 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005014 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005015)
5016
5017xnnpack_unit_test(
Frank Barchard0bb49a72020-06-04 11:35:11 -07005018 name = "f16_gavgpool_minmax_test",
5019 srcs = [
5020 "test/f16-gavgpool-minmax.cc",
5021 "test/gavgpool-microkernel-tester.h",
5022 "src/xnnpack/AlignedAllocator.h",
5023 ] + MICROKERNEL_TEST_HDRS,
5024 deps = MICROKERNEL_TEST_DEPS,
5025)
5026
5027xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005028 name = "f32_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005029 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005030 "test/f32-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005031 "test/gavgpool-microkernel-tester.h",
5032 "src/xnnpack/AlignedAllocator.h",
5033 ] + MICROKERNEL_TEST_HDRS,
5034 deps = MICROKERNEL_TEST_DEPS,
5035)
5036
5037xnnpack_unit_test(
Marat Dukhan1f29b802020-05-15 23:46:39 -07005038 name = "f32_gavgpool_cw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005039 srcs = [
Marat Dukhan1f29b802020-05-15 23:46:39 -07005040 "test/f32-gavgpool-cw.cc",
5041 "test/gavgpool-cw-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005042 "src/xnnpack/AlignedAllocator.h",
5043 ] + MICROKERNEL_TEST_HDRS,
5044 deps = MICROKERNEL_TEST_DEPS,
5045)
5046
5047xnnpack_unit_test(
Marat Dukhan163a7e62020-04-09 04:19:26 -07005048 name = "f32_gemm_test",
5049 srcs = [
5050 "test/f32-gemm.cc",
5051 "test/gemm-microkernel-tester.h",
5052 "src/xnnpack/AlignedAllocator.h",
5053 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005054 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan163a7e62020-04-09 04:19:26 -07005055)
5056
5057xnnpack_unit_test(
Marat Dukhan467f6362020-05-22 23:21:55 -07005058 name = "f32_gemm_relu_test",
5059 srcs = [
5060 "test/f32-gemm-relu.cc",
5061 "test/gemm-microkernel-tester.h",
5062 "src/xnnpack/AlignedAllocator.h",
5063 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005064 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan467f6362020-05-22 23:21:55 -07005065)
5066
5067xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005068 name = "f32_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005069 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005070 "test/f32-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005071 "test/gemm-microkernel-tester.h",
5072 "src/xnnpack/AlignedAllocator.h",
5073 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005074 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005075)
5076
5077xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005078 name = "f32_gemminc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005079 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005080 "test/f32-gemminc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005081 "test/gemm-microkernel-tester.h",
5082 "src/xnnpack/AlignedAllocator.h",
5083 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005084 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005085)
5086
5087xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005088 name = "f16_hswish_test",
5089 srcs = [
5090 "test/f16-hswish.cc",
5091 "test/hswish-microkernel-tester.h",
5092 ] + MICROKERNEL_TEST_HDRS,
5093 deps = MICROKERNEL_TEST_DEPS,
5094)
5095
5096xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005097 name = "f32_hswish_test",
5098 srcs = [
5099 "test/f32-hswish.cc",
5100 "test/hswish-microkernel-tester.h",
5101 ] + MICROKERNEL_TEST_HDRS,
5102 deps = MICROKERNEL_TEST_DEPS,
5103)
5104
5105xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005106 name = "f32_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005107 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005108 "test/f32-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005109 "test/maxpool-microkernel-tester.h",
5110 ] + MICROKERNEL_TEST_HDRS,
5111 deps = MICROKERNEL_TEST_DEPS,
5112)
5113
5114xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005115 name = "f32_pavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005116 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005117 "test/f32-pavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005118 "test/avgpool-microkernel-tester.h",
5119 "src/xnnpack/AlignedAllocator.h",
5120 ] + MICROKERNEL_TEST_HDRS,
5121 deps = MICROKERNEL_TEST_DEPS,
5122)
5123
5124xnnpack_unit_test(
Marat Dukhan1c587112020-04-08 20:04:28 -07005125 name = "f32_ppmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005126 srcs = [
Marat Dukhan1c587112020-04-08 20:04:28 -07005127 "test/f32-ppmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005128 "test/gemm-microkernel-tester.h",
5129 "src/xnnpack/AlignedAllocator.h",
5130 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005131 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005132)
5133
5134xnnpack_unit_test(
Frank Barchardb1966592020-05-12 13:47:06 -07005135 name = "f16_prelu_test",
5136 srcs = [
5137 "test/f16-prelu.cc",
5138 "test/prelu-microkernel-tester.h",
5139 "src/xnnpack/AlignedAllocator.h",
5140 ] + MICROKERNEL_TEST_HDRS,
5141 deps = MICROKERNEL_TEST_DEPS,
5142)
5143
5144xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005145 name = "f32_prelu_test",
5146 srcs = [
5147 "test/f32-prelu.cc",
5148 "test/prelu-microkernel-tester.h",
5149 "src/xnnpack/AlignedAllocator.h",
5150 ] + MICROKERNEL_TEST_HDRS,
5151 deps = MICROKERNEL_TEST_DEPS,
5152)
5153
5154xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005155 name = "f32_raddexpminusmax_test",
5156 srcs = [
5157 "test/f32-raddexpminusmax.cc",
5158 "test/raddexpminusmax-microkernel-tester.h",
5159 ] + MICROKERNEL_TEST_HDRS,
5160 deps = MICROKERNEL_TEST_DEPS,
5161)
5162
5163xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005164 name = "f32_raddextexp_test",
5165 srcs = [
5166 "test/f32-raddextexp.cc",
5167 "test/raddextexp-microkernel-tester.h",
5168 ] + MICROKERNEL_TEST_HDRS,
5169 deps = MICROKERNEL_TEST_DEPS,
5170)
5171
5172xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005173 name = "f32_raddstoreexpminusmax_test",
5174 srcs = [
5175 "test/f32-raddstoreexpminusmax.cc",
5176 "test/raddstoreexpminusmax-microkernel-tester.h",
5177 ] + MICROKERNEL_TEST_HDRS,
5178 deps = MICROKERNEL_TEST_DEPS,
5179)
5180
5181xnnpack_unit_test(
Frank Barchardfb158e22020-07-15 16:10:10 -07005182 name = "f32_relu_test",
5183 srcs = [
5184 "test/f32-relu.cc",
5185 "test/vunary-microkernel-tester.h",
5186 ] + MICROKERNEL_TEST_HDRS,
5187 deps = MICROKERNEL_TEST_DEPS,
5188)
5189
5190xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005191 name = "f32_rmax_test",
5192 srcs = [
5193 "test/f32-rmax.cc",
5194 "test/rmax-microkernel-tester.h",
5195 ] + MICROKERNEL_TEST_HDRS,
5196 deps = MICROKERNEL_TEST_DEPS,
5197)
5198
5199xnnpack_unit_test(
Marat Dukhan346a9e52019-11-15 09:06:30 -08005200 name = "f32_sigmoid_test",
5201 srcs = [
5202 "test/f32-sigmoid.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005203 "test/vunary-microkernel-tester.h",
Marat Dukhan346a9e52019-11-15 09:06:30 -08005204 ] + MICROKERNEL_TEST_HDRS,
5205 deps = MICROKERNEL_TEST_DEPS,
5206)
5207
5208xnnpack_unit_test(
Marat Dukhan355ab432020-04-09 19:01:52 -07005209 name = "f32_spmm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005210 srcs = [
Marat Dukhan355ab432020-04-09 19:01:52 -07005211 "test/f32-spmm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005212 "test/spmm-microkernel-tester.h",
5213 "src/xnnpack/AlignedAllocator.h",
5214 ] + MICROKERNEL_TEST_HDRS,
5215 deps = MICROKERNEL_TEST_DEPS,
5216)
5217
5218xnnpack_unit_test(
Frank Barchardd793f6c2020-05-08 13:37:43 -07005219 name = "f16_vadd_minmax_test",
5220 srcs = [
5221 "test/f16-vadd-minmax.cc",
5222 "test/vbinary-microkernel-tester.h",
5223 ] + MICROKERNEL_TEST_HDRS,
5224 deps = MICROKERNEL_TEST_DEPS,
5225)
5226
5227xnnpack_unit_test(
5228 name = "f16_vaddc_minmax_test",
5229 srcs = [
5230 "test/f16-vaddc-minmax.cc",
5231 "test/vbinaryc-microkernel-tester.h",
5232 ] + MICROKERNEL_TEST_HDRS,
5233 deps = MICROKERNEL_TEST_DEPS,
5234)
5235
5236xnnpack_unit_test(
5237 name = "f16_vdiv_minmax_test",
5238 srcs = [
5239 "test/f16-vdiv-minmax.cc",
5240 "test/vbinary-microkernel-tester.h",
5241 ] + MICROKERNEL_TEST_HDRS,
5242 deps = MICROKERNEL_TEST_DEPS,
5243)
5244
5245xnnpack_unit_test(
5246 name = "f16_vdivc_minmax_test",
5247 srcs = [
5248 "test/f16-vdivc-minmax.cc",
5249 "test/vbinaryc-microkernel-tester.h",
5250 ] + MICROKERNEL_TEST_HDRS,
5251 deps = MICROKERNEL_TEST_DEPS,
5252)
5253
5254xnnpack_unit_test(
5255 name = "f16_vrdivc_minmax_test",
5256 srcs = [
5257 "test/f16-vrdivc-minmax.cc",
5258 "test/vbinaryc-microkernel-tester.h",
5259 ] + MICROKERNEL_TEST_HDRS,
5260 deps = MICROKERNEL_TEST_DEPS,
5261)
5262
5263xnnpack_unit_test(
5264 name = "f16_vmax_test",
5265 srcs = [
5266 "test/f16-vmax.cc",
5267 "test/vbinary-microkernel-tester.h",
5268 ] + MICROKERNEL_TEST_HDRS,
5269 deps = MICROKERNEL_TEST_DEPS,
5270)
5271
5272xnnpack_unit_test(
5273 name = "f16_vmaxc_test",
5274 srcs = [
5275 "test/f16-vmaxc.cc",
5276 "test/vbinaryc-microkernel-tester.h",
5277 ] + MICROKERNEL_TEST_HDRS,
5278 deps = MICROKERNEL_TEST_DEPS,
5279)
5280
5281xnnpack_unit_test(
5282 name = "f16_vmin_test",
5283 srcs = [
5284 "test/f16-vmin.cc",
5285 "test/vbinary-microkernel-tester.h",
5286 ] + MICROKERNEL_TEST_HDRS,
5287 deps = MICROKERNEL_TEST_DEPS,
5288)
5289
5290xnnpack_unit_test(
5291 name = "f16_vminc_test",
5292 srcs = [
5293 "test/f16-vminc.cc",
5294 "test/vbinaryc-microkernel-tester.h",
5295 ] + MICROKERNEL_TEST_HDRS,
5296 deps = MICROKERNEL_TEST_DEPS,
5297)
5298
5299xnnpack_unit_test(
5300 name = "f16_vmul_minmax_test",
5301 srcs = [
5302 "test/f16-vmul-minmax.cc",
5303 "test/vbinary-microkernel-tester.h",
5304 ] + MICROKERNEL_TEST_HDRS,
5305 deps = MICROKERNEL_TEST_DEPS,
5306)
5307
5308xnnpack_unit_test(
5309 name = "f16_vmulc_minmax_test",
5310 srcs = [
5311 "test/f16-vmulc-minmax.cc",
5312 "test/vbinaryc-microkernel-tester.h",
5313 ] + MICROKERNEL_TEST_HDRS,
5314 deps = MICROKERNEL_TEST_DEPS,
5315)
5316
5317xnnpack_unit_test(
5318 name = "f16_vsub_minmax_test",
5319 srcs = [
5320 "test/f16-vsub-minmax.cc",
5321 "test/vbinary-microkernel-tester.h",
5322 ] + MICROKERNEL_TEST_HDRS,
5323 deps = MICROKERNEL_TEST_DEPS,
5324)
5325
5326xnnpack_unit_test(
5327 name = "f16_vsubc_minmax_test",
5328 srcs = [
5329 "test/f16-vsubc-minmax.cc",
5330 "test/vbinaryc-microkernel-tester.h",
5331 ] + MICROKERNEL_TEST_HDRS,
5332 deps = MICROKERNEL_TEST_DEPS,
5333)
5334
5335xnnpack_unit_test(
5336 name = "f16_vrsubc_minmax_test",
5337 srcs = [
5338 "test/f16-vrsubc-minmax.cc",
5339 "test/vbinaryc-microkernel-tester.h",
5340 ] + MICROKERNEL_TEST_HDRS,
5341 deps = MICROKERNEL_TEST_DEPS,
5342)
5343
5344xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005345 name = "f32_vabs_test",
5346 srcs = [
5347 "test/f32-vabs.cc",
5348 "test/vunary-microkernel-tester.h",
5349 ] + MICROKERNEL_TEST_HDRS,
5350 deps = MICROKERNEL_TEST_DEPS,
5351)
5352
5353xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005354 name = "f32_vadd_test",
5355 srcs = [
5356 "test/f32-vadd.cc",
5357 "test/vbinary-microkernel-tester.h",
5358 ] + MICROKERNEL_TEST_HDRS,
5359 deps = MICROKERNEL_TEST_DEPS,
5360)
5361
5362xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005363 name = "f32_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005364 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005365 "test/f32-vadd-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005366 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005367 ] + MICROKERNEL_TEST_HDRS,
5368 deps = MICROKERNEL_TEST_DEPS,
5369)
5370
5371xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005372 name = "f32_vadd_relu_test",
5373 srcs = [
5374 "test/f32-vadd-relu.cc",
5375 "test/vbinary-microkernel-tester.h",
5376 ] + MICROKERNEL_TEST_HDRS,
5377 deps = MICROKERNEL_TEST_DEPS,
5378)
5379
5380xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005381 name = "f32_vaddc_test",
5382 srcs = [
5383 "test/f32-vaddc.cc",
5384 "test/vbinaryc-microkernel-tester.h",
5385 ] + MICROKERNEL_TEST_HDRS,
5386 deps = MICROKERNEL_TEST_DEPS,
5387)
5388
5389xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005390 name = "f32_vaddc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005391 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005392 "test/f32-vaddc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005393 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005394 ] + MICROKERNEL_TEST_HDRS,
5395 deps = MICROKERNEL_TEST_DEPS,
5396)
5397
5398xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005399 name = "f32_vaddc_relu_test",
5400 srcs = [
5401 "test/f32-vaddc-relu.cc",
5402 "test/vbinaryc-microkernel-tester.h",
5403 ] + MICROKERNEL_TEST_HDRS,
5404 deps = MICROKERNEL_TEST_DEPS,
5405)
5406
5407xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005408 name = "f32_vdiv_test",
5409 srcs = [
5410 "test/f32-vdiv.cc",
5411 "test/vbinary-microkernel-tester.h",
5412 ] + MICROKERNEL_TEST_HDRS,
5413 deps = MICROKERNEL_TEST_DEPS,
5414)
5415
5416xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005417 name = "f32_vdiv_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005418 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005419 "test/f32-vdiv-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005420 "test/vbinary-microkernel-tester.h",
5421 ] + MICROKERNEL_TEST_HDRS,
5422 deps = MICROKERNEL_TEST_DEPS,
5423)
5424
5425xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005426 name = "f32_vdiv_relu_test",
5427 srcs = [
5428 "test/f32-vdiv-relu.cc",
5429 "test/vbinary-microkernel-tester.h",
5430 ] + MICROKERNEL_TEST_HDRS,
5431 deps = MICROKERNEL_TEST_DEPS,
5432)
5433
5434xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005435 name = "f32_vdivc_test",
5436 srcs = [
5437 "test/f32-vdivc.cc",
5438 "test/vbinaryc-microkernel-tester.h",
5439 ] + MICROKERNEL_TEST_HDRS,
5440 deps = MICROKERNEL_TEST_DEPS,
5441)
5442
5443xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005444 name = "f32_vdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005445 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005446 "test/f32-vdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005447 "test/vbinaryc-microkernel-tester.h",
5448 ] + MICROKERNEL_TEST_HDRS,
5449 deps = MICROKERNEL_TEST_DEPS,
5450)
5451
5452xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005453 name = "f32_vdivc_relu_test",
5454 srcs = [
5455 "test/f32-vdivc-relu.cc",
5456 "test/vbinaryc-microkernel-tester.h",
5457 ] + MICROKERNEL_TEST_HDRS,
5458 deps = MICROKERNEL_TEST_DEPS,
5459)
5460
5461xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005462 name = "f32_vrdivc_test",
5463 srcs = [
5464 "test/f32-vrdivc.cc",
5465 "test/vbinaryc-microkernel-tester.h",
5466 ] + MICROKERNEL_TEST_HDRS,
5467 deps = MICROKERNEL_TEST_DEPS,
5468)
5469
5470xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005471 name = "f32_vrdivc_minmax_test",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005472 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005473 "test/f32-vrdivc-minmax.cc",
Marat Dukhan77ca6302019-12-06 12:48:15 -08005474 "test/vbinaryc-microkernel-tester.h",
5475 ] + MICROKERNEL_TEST_HDRS,
5476 deps = MICROKERNEL_TEST_DEPS,
5477)
5478
5479xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005480 name = "f32_vrdivc_relu_test",
5481 srcs = [
5482 "test/f32-vrdivc-relu.cc",
5483 "test/vbinaryc-microkernel-tester.h",
5484 ] + MICROKERNEL_TEST_HDRS,
5485 deps = MICROKERNEL_TEST_DEPS,
5486)
5487
5488xnnpack_unit_test(
Marat Dukhan403b7d42019-12-05 12:49:11 -08005489 name = "f32_vmax_test",
5490 srcs = [
5491 "test/f32-vmax.cc",
5492 "test/vbinary-microkernel-tester.h",
5493 ] + MICROKERNEL_TEST_HDRS,
5494 deps = MICROKERNEL_TEST_DEPS,
5495)
5496
5497xnnpack_unit_test(
5498 name = "f32_vmaxc_test",
5499 srcs = [
5500 "test/f32-vmaxc.cc",
5501 "test/vbinaryc-microkernel-tester.h",
5502 ] + MICROKERNEL_TEST_HDRS,
5503 deps = MICROKERNEL_TEST_DEPS,
5504)
5505
5506xnnpack_unit_test(
5507 name = "f32_vmin_test",
5508 srcs = [
5509 "test/f32-vmin.cc",
5510 "test/vbinary-microkernel-tester.h",
5511 ] + MICROKERNEL_TEST_HDRS,
5512 deps = MICROKERNEL_TEST_DEPS,
5513)
5514
5515xnnpack_unit_test(
5516 name = "f32_vminc_test",
5517 srcs = [
5518 "test/f32-vminc.cc",
5519 "test/vbinaryc-microkernel-tester.h",
5520 ] + MICROKERNEL_TEST_HDRS,
5521 deps = MICROKERNEL_TEST_DEPS,
5522)
5523
5524xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005525 name = "f32_vmul_test",
5526 srcs = [
5527 "test/f32-vmul.cc",
5528 "test/vbinary-microkernel-tester.h",
5529 ] + MICROKERNEL_TEST_HDRS,
5530 deps = MICROKERNEL_TEST_DEPS,
5531)
5532
5533xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005534 name = "f32_vmul_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005535 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005536 "test/f32-vmul-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005537 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005538 ] + MICROKERNEL_TEST_HDRS,
5539 deps = MICROKERNEL_TEST_DEPS,
5540)
5541
5542xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005543 name = "f32_vmul_relu_test",
5544 srcs = [
5545 "test/f32-vmul-relu.cc",
5546 "test/vbinary-microkernel-tester.h",
5547 ] + MICROKERNEL_TEST_HDRS,
5548 deps = MICROKERNEL_TEST_DEPS,
5549)
5550
5551xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005552 name = "f32_vmulc_test",
5553 srcs = [
5554 "test/f32-vmulc.cc",
5555 "test/vbinaryc-microkernel-tester.h",
5556 ] + MICROKERNEL_TEST_HDRS,
5557 deps = MICROKERNEL_TEST_DEPS,
5558)
5559
5560xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005561 name = "f32_vmulc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005562 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005563 "test/f32-vmulc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005564 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005565 ] + MICROKERNEL_TEST_HDRS,
5566 deps = MICROKERNEL_TEST_DEPS,
5567)
5568
5569xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005570 name = "f32_vmulc_relu_test",
5571 srcs = [
5572 "test/f32-vmulc-relu.cc",
5573 "test/vbinaryc-microkernel-tester.h",
5574 ] + MICROKERNEL_TEST_HDRS,
5575 deps = MICROKERNEL_TEST_DEPS,
5576)
5577
5578xnnpack_unit_test(
Frank Barchard2a1049c2020-06-03 02:31:27 -07005579 name = "f16_vmulcaddc_minmax_test",
5580 srcs = [
5581 "test/f16-vmulcaddc-minmax.cc",
5582 "test/vmulcaddc-microkernel-tester.h",
5583 "src/xnnpack/AlignedAllocator.h",
5584 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005585 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Frank Barchard2a1049c2020-06-03 02:31:27 -07005586)
5587
5588xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005589 name = "f32_vmulcaddc_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005590 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005591 "test/f32-vmulcaddc-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005592 "test/vmulcaddc-microkernel-tester.h",
5593 "src/xnnpack/AlignedAllocator.h",
5594 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005595 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005596)
5597
5598xnnpack_unit_test(
Marat Dukhan8cc7efe2020-06-10 16:24:27 -07005599 name = "f32_vlrelu_test",
5600 srcs = [
5601 "test/f32-vlrelu.cc",
5602 "test/vunary-microkernel-tester.h",
5603 ] + MICROKERNEL_TEST_HDRS,
5604 deps = MICROKERNEL_TEST_DEPS,
5605)
5606
5607xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005608 name = "f32_vneg_test",
5609 srcs = [
5610 "test/f32-vneg.cc",
5611 "test/vunary-microkernel-tester.h",
5612 ] + MICROKERNEL_TEST_HDRS,
5613 deps = MICROKERNEL_TEST_DEPS,
5614)
5615
5616xnnpack_unit_test(
Marat Dukhaneecf8fd2020-06-09 08:59:37 -07005617 name = "f32_vrndne_test",
5618 srcs = [
5619 "test/f32-vrndne.cc",
5620 "test/vunary-microkernel-tester.h",
5621 ] + MICROKERNEL_TEST_HDRS,
5622 deps = MICROKERNEL_TEST_DEPS,
5623)
5624
5625xnnpack_unit_test(
5626 name = "f32_vrndz_test",
5627 srcs = [
5628 "test/f32-vrndz.cc",
5629 "test/vunary-microkernel-tester.h",
5630 ] + MICROKERNEL_TEST_HDRS,
5631 deps = MICROKERNEL_TEST_DEPS,
5632)
5633
5634xnnpack_unit_test(
5635 name = "f32_vrndu_test",
5636 srcs = [
5637 "test/f32-vrndu.cc",
5638 "test/vunary-microkernel-tester.h",
5639 ] + MICROKERNEL_TEST_HDRS,
5640 deps = MICROKERNEL_TEST_DEPS,
5641)
5642
5643xnnpack_unit_test(
5644 name = "f32_vrndd_test",
5645 srcs = [
5646 "test/f32-vrndd.cc",
5647 "test/vunary-microkernel-tester.h",
5648 ] + MICROKERNEL_TEST_HDRS,
5649 deps = MICROKERNEL_TEST_DEPS,
5650)
5651
5652xnnpack_unit_test(
Marat Dukhan05ac8e32019-10-21 15:39:33 -07005653 name = "f32_vscale_test",
5654 srcs = [
5655 "test/f32-vscale.cc",
5656 "test/vscale-microkernel-tester.h",
5657 ] + MICROKERNEL_TEST_HDRS,
5658 deps = MICROKERNEL_TEST_DEPS,
5659)
5660
5661xnnpack_unit_test(
Marat Dukhan97579532019-10-18 16:40:39 -07005662 name = "f32_vscaleexpminusmax_test",
5663 srcs = [
5664 "test/f32-vscaleexpminusmax.cc",
5665 "test/vscaleexpminusmax-microkernel-tester.h",
5666 ] + MICROKERNEL_TEST_HDRS,
5667 deps = MICROKERNEL_TEST_DEPS,
5668)
5669
5670xnnpack_unit_test(
Marat Dukhan6f8d4d32019-10-25 17:07:09 -07005671 name = "f32_vscaleextexp_test",
5672 srcs = [
5673 "test/f32-vscaleextexp.cc",
5674 "test/vscaleextexp-microkernel-tester.h",
5675 ] + MICROKERNEL_TEST_HDRS,
5676 deps = MICROKERNEL_TEST_DEPS,
5677)
5678
5679xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07005680 name = "f32_vsqr_test",
5681 srcs = [
5682 "test/f32-vsqr.cc",
5683 "test/vunary-microkernel-tester.h",
5684 ] + MICROKERNEL_TEST_HDRS,
5685 deps = MICROKERNEL_TEST_DEPS,
5686)
5687
5688xnnpack_unit_test(
Marat Dukhan13bafb02020-06-05 00:43:11 -07005689 name = "f32_vsqrdiff_test",
5690 srcs = [
5691 "test/f32-vsqrdiff.cc",
5692 "test/vbinary-microkernel-tester.h",
5693 ] + MICROKERNEL_TEST_HDRS,
5694 deps = MICROKERNEL_TEST_DEPS,
5695)
5696
5697xnnpack_unit_test(
5698 name = "f32_vsqrdiffc_test",
5699 srcs = [
5700 "test/f32-vsqrdiffc.cc",
5701 "test/vbinaryc-microkernel-tester.h",
5702 ] + MICROKERNEL_TEST_HDRS,
5703 deps = MICROKERNEL_TEST_DEPS,
5704)
5705
5706xnnpack_unit_test(
Marat Dukhanf4db2f32020-06-30 10:55:30 -07005707 name = "f32_vsqrt_test",
5708 srcs = [
5709 "test/f32-vsqrt.cc",
5710 "test/vunary-microkernel-tester.h",
5711 ] + MICROKERNEL_TEST_HDRS,
5712 deps = MICROKERNEL_TEST_DEPS,
5713)
5714
5715xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005716 name = "f32_vsub_test",
5717 srcs = [
5718 "test/f32-vsub.cc",
5719 "test/vbinary-microkernel-tester.h",
5720 ] + MICROKERNEL_TEST_HDRS,
5721 deps = MICROKERNEL_TEST_DEPS,
5722)
5723
5724xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005725 name = "f32_vsub_minmax_test",
Marat Dukhan97579532019-10-18 16:40:39 -07005726 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005727 "test/f32-vsub-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005728 "test/vbinary-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005729 ] + MICROKERNEL_TEST_HDRS,
5730 deps = MICROKERNEL_TEST_DEPS,
5731)
5732
5733xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005734 name = "f32_vsub_relu_test",
5735 srcs = [
5736 "test/f32-vsub-relu.cc",
5737 "test/vbinary-microkernel-tester.h",
5738 ] + MICROKERNEL_TEST_HDRS,
5739 deps = MICROKERNEL_TEST_DEPS,
5740)
5741
5742xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005743 name = "f32_vsubc_test",
5744 srcs = [
5745 "test/f32-vsubc.cc",
5746 "test/vbinaryc-microkernel-tester.h",
5747 ] + MICROKERNEL_TEST_HDRS,
5748 deps = MICROKERNEL_TEST_DEPS,
5749)
5750
5751xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005752 name = "f32_vsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005753 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005754 "test/f32-vsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005755 "test/vbinaryc-microkernel-tester.h",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005756 ] + MICROKERNEL_TEST_HDRS,
5757 deps = MICROKERNEL_TEST_DEPS,
5758)
5759
5760xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005761 name = "f32_vsubc_relu_test",
5762 srcs = [
5763 "test/f32-vsubc-relu.cc",
5764 "test/vbinaryc-microkernel-tester.h",
5765 ] + MICROKERNEL_TEST_HDRS,
5766 deps = MICROKERNEL_TEST_DEPS,
5767)
5768
5769xnnpack_unit_test(
Frank Barchardd5b9f1c2020-07-01 15:00:19 -07005770 name = "f32_vrsubc_test",
5771 srcs = [
5772 "test/f32-vrsubc.cc",
5773 "test/vbinaryc-microkernel-tester.h",
5774 ] + MICROKERNEL_TEST_HDRS,
5775 deps = MICROKERNEL_TEST_DEPS,
5776)
5777
5778xnnpack_unit_test(
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005779 name = "f32_vrsubc_minmax_test",
Marat Dukhanc07cb7f2019-11-14 15:32:05 -08005780 srcs = [
Marat Dukhan91cd2b72020-04-09 23:57:31 -07005781 "test/f32-vrsubc-minmax.cc",
Marat Dukhan1e782c42019-11-21 17:02:40 -08005782 "test/vbinaryc-microkernel-tester.h",
Marat Dukhan97579532019-10-18 16:40:39 -07005783 ] + MICROKERNEL_TEST_HDRS,
5784 deps = MICROKERNEL_TEST_DEPS,
5785)
5786
5787xnnpack_unit_test(
Frank Barchard674778d2020-08-08 10:17:25 -07005788 name = "f32_vrsubc_relu_test",
5789 srcs = [
5790 "test/f32-vrsubc-relu.cc",
5791 "test/vbinaryc-microkernel-tester.h",
5792 ] + MICROKERNEL_TEST_HDRS,
5793 deps = MICROKERNEL_TEST_DEPS,
5794)
5795
5796xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005797 name = "qs8_dwconv_minmax_test",
5798 srcs = [
5799 "test/qs8-dwconv-minmax.cc",
5800 "test/dwconv-microkernel-tester.h",
5801 "src/xnnpack/AlignedAllocator.h",
5802 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5803 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5804)
5805
5806xnnpack_unit_test(
Marat Dukhan4ed53f42020-08-06 01:12:55 -07005807 name = "qs8_gavgpool_minmax_test",
5808 srcs = [
5809 "test/qs8-gavgpool-minmax.cc",
5810 "test/gavgpool-microkernel-tester.h",
5811 "src/xnnpack/AlignedAllocator.h",
5812 ] + MICROKERNEL_TEST_HDRS,
5813 deps = MICROKERNEL_TEST_DEPS,
5814)
5815
5816xnnpack_unit_test(
Marat Dukhan595e1702020-07-31 10:12:52 -07005817 name = "qs8_gemm_minmax_test",
5818 srcs = [
5819 "test/qs8-gemm-minmax.cc",
5820 "test/gemm-microkernel-tester.h",
5821 "src/xnnpack/AlignedAllocator.h",
5822 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5823 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5824)
5825
5826xnnpack_unit_test(
Marat Dukhanf9480682020-07-31 14:50:24 -07005827 name = "qs8_igemm_minmax_test",
5828 srcs = [
5829 "test/qs8-igemm-minmax.cc",
5830 "test/gemm-microkernel-tester.h",
5831 "src/xnnpack/AlignedAllocator.h",
5832 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
5833 deps = MICROKERNEL_TEST_DEPS + [":packing"],
5834)
5835
5836xnnpack_unit_test(
5837 name = "qs8_requantization_test",
5838 srcs = [
5839 "src/xnnpack/requantization-stubs.h",
5840 "test/qs8-requantization.cc",
5841 "test/requantization-tester.h",
5842 ] + MICROKERNEL_TEST_HDRS,
5843 deps = MICROKERNEL_TEST_DEPS,
5844)
5845
5846xnnpack_unit_test(
Marat Dukhand9f3ad42020-08-10 12:30:58 -07005847 name = "qs8_vadd_minmax_test",
5848 srcs = [
5849 "test/qs8-vadd-minmax.cc",
5850 "test/vadd-microkernel-tester.h",
5851 ] + MICROKERNEL_TEST_HDRS,
5852 deps = MICROKERNEL_TEST_DEPS,
5853)
5854
5855xnnpack_unit_test(
Marat Dukhan0270d9f2020-08-11 00:56:46 -07005856 name = "qs8_vaddc_minmax_test",
5857 srcs = [
5858 "test/qs8-vaddc-minmax.cc",
5859 "test/vaddc-microkernel-tester.h",
5860 ] + MICROKERNEL_TEST_HDRS,
5861 deps = MICROKERNEL_TEST_DEPS,
5862)
5863
5864xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005865 name = "qu8_avgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005866 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005867 "test/qu8-avgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005868 "test/avgpool-microkernel-tester.h",
5869 "src/xnnpack/AlignedAllocator.h",
5870 ] + MICROKERNEL_TEST_HDRS,
5871 deps = MICROKERNEL_TEST_DEPS,
5872)
5873
5874xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005875 name = "qu8_dwconv_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005876 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005877 "test/qu8-dwconv-minmax.cc",
5878 "test/dwconv-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005879 "src/xnnpack/AlignedAllocator.h",
5880 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005881 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005882)
5883
5884xnnpack_unit_test(
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005885 name = "qu8_igemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005886 srcs = [
Marat Dukhanf62bbdc2020-08-04 13:59:04 -07005887 "test/qu8-igemm-minmax.cc",
5888 "test/gemm-microkernel-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005889 "src/xnnpack/AlignedAllocator.h",
5890 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005891 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005892)
5893
5894xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005895 name = "qu8_gavgpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005896 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005897 "test/qu8-gavgpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005898 "test/gavgpool-microkernel-tester.h",
5899 "src/xnnpack/AlignedAllocator.h",
5900 ] + MICROKERNEL_TEST_HDRS,
5901 deps = MICROKERNEL_TEST_DEPS,
5902)
5903
5904xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005905 name = "qu8_gemm_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005906 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005907 "test/qu8-gemm-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005908 "test/gemm-microkernel-tester.h",
5909 "src/xnnpack/AlignedAllocator.h",
5910 ] + WEIGHTS_PACK_HDRS + MICROKERNEL_TEST_HDRS,
Marat Dukhanab582382020-07-06 13:32:08 -07005911 deps = MICROKERNEL_TEST_DEPS + [":packing"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07005912)
5913
5914xnnpack_unit_test(
Marat Dukhan5b69f8b2020-07-24 15:26:48 -07005915 name = "qu8_requantization_test",
5916 srcs = [
5917 "src/xnnpack/requantization-stubs.h",
5918 "test/qu8-requantization.cc",
5919 "test/requantization-tester.h",
5920 ] + MICROKERNEL_TEST_HDRS,
5921 deps = MICROKERNEL_TEST_DEPS,
5922)
5923
5924xnnpack_unit_test(
Marat Dukhan08b7a972020-07-14 18:17:29 -07005925 name = "qu8_vadd_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005926 srcs = [
Marat Dukhan08b7a972020-07-14 18:17:29 -07005927 "test/qu8-vadd-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005928 "test/vadd-microkernel-tester.h",
5929 ] + MICROKERNEL_TEST_HDRS,
5930 deps = MICROKERNEL_TEST_DEPS,
5931)
5932
5933xnnpack_unit_test(
5934 name = "u8_clamp_test",
5935 srcs = [
5936 "test/u8-clamp.cc",
5937 "test/clamp-microkernel-tester.h",
5938 ] + MICROKERNEL_TEST_HDRS,
5939 deps = MICROKERNEL_TEST_DEPS,
5940)
5941
5942xnnpack_unit_test(
5943 name = "u8_lut32norm_test",
5944 srcs = [
5945 "test/u8-lut32norm.cc",
5946 "test/lut-norm-microkernel-tester.h",
5947 ] + MICROKERNEL_TEST_HDRS,
5948 deps = MICROKERNEL_TEST_DEPS,
5949)
5950
5951xnnpack_unit_test(
Marat Dukhan99936602020-04-11 16:47:01 -07005952 name = "u8_maxpool_minmax_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005953 srcs = [
Marat Dukhan99936602020-04-11 16:47:01 -07005954 "test/u8-maxpool-minmax.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07005955 "test/maxpool-microkernel-tester.h",
5956 ] + MICROKERNEL_TEST_HDRS,
5957 deps = MICROKERNEL_TEST_DEPS,
5958)
5959
5960xnnpack_unit_test(
5961 name = "u8_rmax_test",
5962 srcs = [
5963 "test/u8-rmax.cc",
5964 "test/rmax-microkernel-tester.h",
5965 ] + MICROKERNEL_TEST_HDRS,
5966 deps = MICROKERNEL_TEST_DEPS,
5967)
5968
5969xnnpack_unit_test(
Marat Dukhan3bb3bfc2020-05-19 17:42:46 -07005970 name = "x32_fill_test",
5971 srcs = [
5972 "test/x32-fill.cc",
5973 "test/fill-microkernel-tester.h",
5974 ] + MICROKERNEL_TEST_HDRS,
5975 deps = MICROKERNEL_TEST_DEPS,
5976)
5977
5978xnnpack_unit_test(
Marat Dukhan08c4a432019-10-03 09:29:21 -07005979 name = "x32_packx_test",
5980 srcs = [
5981 "test/x32-packx.cc",
5982 "test/pack-microkernel-tester.h",
5983 "src/xnnpack/AlignedAllocator.h",
5984 ] + MICROKERNEL_TEST_HDRS,
5985 deps = MICROKERNEL_TEST_DEPS,
5986)
5987
5988xnnpack_unit_test(
5989 name = "x32_pad_test",
5990 srcs = [
5991 "test/x32-pad.cc",
5992 "test/pad-microkernel-tester.h",
5993 ] + MICROKERNEL_TEST_HDRS,
5994 deps = MICROKERNEL_TEST_DEPS,
5995)
5996
5997xnnpack_unit_test(
5998 name = "x32_unpool_test",
5999 srcs = [
6000 "test/x32-unpool.cc",
6001 "test/unpool-microkernel-tester.h",
6002 ] + MICROKERNEL_TEST_HDRS,
6003 deps = MICROKERNEL_TEST_DEPS,
6004)
6005
6006xnnpack_unit_test(
6007 name = "x32_zip_test",
6008 srcs = [
6009 "test/x32-zip.cc",
6010 "test/zip-microkernel-tester.h",
6011 ] + MICROKERNEL_TEST_HDRS,
6012 deps = MICROKERNEL_TEST_DEPS,
6013)
6014
6015xnnpack_unit_test(
6016 name = "x8_lut_test",
6017 srcs = [
6018 "test/x8-lut.cc",
6019 "test/lut-microkernel-tester.h",
6020 ] + MICROKERNEL_TEST_HDRS,
6021 deps = MICROKERNEL_TEST_DEPS,
6022)
6023
6024xnnpack_unit_test(
6025 name = "x8_zip_test",
6026 srcs = [
6027 "test/x8-zip.cc",
6028 "test/zip-microkernel-tester.h",
6029 ] + MICROKERNEL_TEST_HDRS,
6030 deps = MICROKERNEL_TEST_DEPS,
6031)
6032
Marat Dukhan20c3b922020-03-10 03:45:06 -07006033########################## Size tests for the library #########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006034
6035xnnpack_binary(
Marat Dukhan20c3b922020-03-10 03:45:06 -07006036 name = "operator_size_test",
6037 srcs = ["test/operator-size.c"],
Marat Dukhan8fe54e42019-10-10 14:12:59 -07006038 deps = [":xnnpack_operators_nhwc_f32"],
Marat Dukhan08c4a432019-10-03 09:29:21 -07006039)
6040
Marat Dukhan20c3b922020-03-10 03:45:06 -07006041xnnpack_binary(
6042 name = "subgraph_size_test",
6043 srcs = ["test/subgraph-size.c"],
6044 deps = [":XNNPACK"],
6045)
6046
6047########################### Unit tests for operators ##########################
Marat Dukhan08c4a432019-10-03 09:29:21 -07006048
6049xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006050 name = "abs_nc_test",
6051 srcs = [
6052 "test/abs-nc.cc",
6053 "test/abs-operator-tester.h",
6054 ],
6055 deps = OPERATOR_TEST_DEPS,
6056)
6057
6058xnnpack_unit_test(
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006059 name = "add_nd_test",
6060 srcs = [
6061 "test/add-nd.cc",
6062 "test/binary-elementwise-operator-tester.h",
6063 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006064 deps = OPERATOR_TEST_DEPS,
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006065)
6066
6067xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006068 name = "argmax_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006069 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006070 "test/argmax-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006071 "test/argmax-pooling-operator-tester.h",
6072 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006073 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006074)
6075
6076xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006077 name = "average_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006078 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006079 "test/average-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006080 "test/average-pooling-operator-tester.h",
6081 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006082 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006083)
6084
6085xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006086 name = "bankers_rounding_nc_test",
6087 srcs = [
6088 "test/bankers-rounding-nc.cc",
6089 "test/bankers-rounding-operator-tester.h",
6090 ],
6091 deps = OPERATOR_TEST_DEPS,
6092)
6093
6094xnnpack_unit_test(
6095 name = "ceiling_nc_test",
6096 srcs = [
6097 "test/ceiling-nc.cc",
6098 "test/ceiling-operator-tester.h",
6099 ],
6100 deps = OPERATOR_TEST_DEPS,
6101)
6102
6103xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006104 name = "channel_shuffle_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006105 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006106 "test/channel-shuffle-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006107 "test/channel-shuffle-operator-tester.h",
6108 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006109 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006110)
6111
6112xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006113 name = "clamp_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006114 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006115 "test/clamp-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006116 "test/clamp-operator-tester.h",
6117 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006118 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006119)
6120
6121xnnpack_unit_test(
Marat Dukhan065b11e2020-05-22 09:49:41 -07006122 name = "constant_pad_nd_test",
6123 srcs = [
6124 "test/constant-pad-nd.cc",
6125 "test/constant-pad-operator-tester.h",
6126 ],
6127 deps = OPERATOR_TEST_DEPS,
6128)
6129
6130xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006131 name = "convolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006132 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006133 "test/convolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006134 "test/convolution-operator-tester.h",
6135 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006136 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006137)
6138
6139xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006140 name = "convolution_nchw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006141 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006142 "test/convolution-nchw.cc",
6143 "test/convolution-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006144 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006145 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006146)
6147
6148xnnpack_unit_test(
Marat Dukhan4e21b272020-06-04 18:45:01 -07006149 name = "copy_nc_test",
6150 srcs = [
6151 "test/copy-nc.cc",
6152 "test/copy-operator-tester.h",
6153 ],
6154 deps = OPERATOR_TEST_DEPS,
6155)
6156
6157xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006158 name = "deconvolution_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006159 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006160 "test/deconvolution-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006161 "test/deconvolution-operator-tester.h",
6162 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006163 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006164)
6165
6166xnnpack_unit_test(
Marat Dukhan69180502019-12-06 15:00:31 -08006167 name = "divide_nd_test",
6168 srcs = [
6169 "test/binary-elementwise-operator-tester.h",
6170 "test/divide-nd.cc",
6171 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006172 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69180502019-12-06 15:00:31 -08006173)
6174
6175xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006176 name = "fully_connected_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006177 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006178 "test/fully-connected-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006179 "test/fully-connected-operator-tester.h",
6180 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006181 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006182)
6183
6184xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006185 name = "floor_nc_test",
6186 srcs = [
6187 "test/floor-nc.cc",
6188 "test/floor-operator-tester.h",
6189 ],
6190 deps = OPERATOR_TEST_DEPS,
6191)
6192
6193xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006194 name = "global_average_pooling_nwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006195 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006196 "test/global-average-pooling-nwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006197 "test/global-average-pooling-operator-tester.h",
Marat Dukhanef61d022020-06-19 13:54:49 -07006198 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006199 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006200)
6201
6202xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006203 name = "global_average_pooling_ncw_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006204 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006205 "test/global-average-pooling-ncw.cc",
6206 "test/global-average-pooling-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006207 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006208 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006209)
6210
6211xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006212 name = "hardswish_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006213 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006214 "test/hardswish-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006215 "test/hardswish-operator-tester.h",
6216 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006217 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006218)
6219
6220xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006221 name = "leaky_relu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006222 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006223 "test/leaky-relu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006224 "test/leaky-relu-operator-tester.h",
6225 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006226 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006227)
6228
6229xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006230 name = "max_pooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006231 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006232 "test/max-pooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006233 "test/max-pooling-operator-tester.h",
6234 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006235 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006236)
6237
6238xnnpack_unit_test(
Marat Dukhan79e7f842019-12-05 14:35:50 -08006239 name = "maximum_nd_test",
6240 srcs = [
6241 "test/binary-elementwise-operator-tester.h",
6242 "test/maximum-nd.cc",
6243 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006244 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006245)
6246
6247xnnpack_unit_test(
6248 name = "minimum_nd_test",
6249 srcs = [
6250 "test/binary-elementwise-operator-tester.h",
6251 "test/minimum-nd.cc",
6252 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006253 deps = OPERATOR_TEST_DEPS,
Marat Dukhan79e7f842019-12-05 14:35:50 -08006254)
6255
6256xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006257 name = "multiply_nd_test",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006258 srcs = [
Marat Dukhanb1a0fc32019-12-02 19:32:02 -08006259 "test/binary-elementwise-operator-tester.h",
Marat Dukhanefc47b82019-11-18 09:25:38 -08006260 "test/multiply-nd.cc",
Marat Dukhanca2733c2019-11-15 23:21:17 -08006261 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006262 deps = OPERATOR_TEST_DEPS,
Marat Dukhanca2733c2019-11-15 23:21:17 -08006263)
6264
6265xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006266 name = "negate_nc_test",
6267 srcs = [
6268 "test/negate-nc.cc",
6269 "test/negate-operator-tester.h",
6270 ],
6271 deps = OPERATOR_TEST_DEPS,
6272)
6273
6274xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006275 name = "prelu_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006276 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006277 "test/prelu-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006278 "test/prelu-operator-tester.h",
6279 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006280 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006281)
6282
6283xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006284 name = "resize_bilinear_nhwc_test",
Marat Dukhan69722492019-11-11 19:55:50 -08006285 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006286 "test/resize-bilinear-nhwc.cc",
Marat Dukhan69722492019-11-11 19:55:50 -08006287 "test/resize-bilinear-operator-tester.h",
6288 ] + OPERATOR_TEST_PARAMS_HDRS,
Marat Dukhan1b354632020-03-23 12:50:22 -07006289 deps = OPERATOR_TEST_DEPS,
Marat Dukhan69722492019-11-11 19:55:50 -08006290)
6291
6292xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006293 name = "sigmoid_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006294 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006295 "test/sigmoid-nc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006296 "test/sigmoid-operator-tester.h",
6297 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006298 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006299)
6300
6301xnnpack_unit_test(
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006302 name = "softmax_nc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006303 srcs = [
Marat Dukhanfd8e6892020-01-27 15:25:25 -08006304 "test/softmax-nc.cc",
6305 "test/softmax-operator-tester.h",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006306 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006307 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006308)
6309
6310xnnpack_unit_test(
Marat Dukhan5020b962020-06-08 13:30:10 -07006311 name = "square_nc_test",
6312 srcs = [
6313 "test/square-nc.cc",
6314 "test/square-operator-tester.h",
6315 ],
6316 deps = OPERATOR_TEST_DEPS,
6317)
6318
6319xnnpack_unit_test(
Marat Dukhan6804bbd2020-06-30 19:26:11 -07006320 name = "square_root_nc_test",
6321 srcs = [
6322 "test/square-root-nc.cc",
6323 "test/square-root-operator-tester.h",
6324 ],
6325 deps = OPERATOR_TEST_DEPS,
6326)
6327
6328xnnpack_unit_test(
Marat Dukhanf7399262020-06-05 10:58:44 -07006329 name = "squared_difference_nd_test",
6330 srcs = [
6331 "test/binary-elementwise-operator-tester.h",
6332 "test/squared-difference-nd.cc",
6333 ],
6334 deps = OPERATOR_TEST_DEPS,
6335)
6336
6337xnnpack_unit_test(
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006338 name = "subtract_nd_test",
6339 srcs = [
6340 "test/binary-elementwise-operator-tester.h",
6341 "test/subtract-nd.cc",
6342 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006343 deps = OPERATOR_TEST_DEPS,
Marat Dukhan05f3f6d2019-12-03 15:13:53 -08006344)
6345
6346xnnpack_unit_test(
Marat Dukhan64e52512020-06-09 13:41:16 -07006347 name = "truncation_nc_test",
6348 srcs = [
6349 "test/truncation-nc.cc",
6350 "test/truncation-operator-tester.h",
6351 ],
6352 deps = OPERATOR_TEST_DEPS,
6353)
6354
6355xnnpack_unit_test(
Marat Dukhanefc47b82019-11-18 09:25:38 -08006356 name = "unpooling_nhwc_test",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006357 srcs = [
Marat Dukhanefc47b82019-11-18 09:25:38 -08006358 "test/unpooling-nhwc.cc",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006359 "test/unpooling-operator-tester.h",
6360 ],
Marat Dukhan1b354632020-03-23 12:50:22 -07006361 deps = OPERATOR_TEST_DEPS,
Marat Dukhan08c4a432019-10-03 09:29:21 -07006362)
6363
Chao Mei6ddfc602020-05-13 22:29:36 -07006364############################### Misc unit tests ###############################
6365
6366xnnpack_unit_test(
6367 name = "memory_planner_test",
6368 srcs = [
6369 "test/memory-planner-test.cc",
6370 ],
6371 deps = [
6372 ":XNNPACK",
6373 ":memory_planner",
6374 ],
6375)
6376
Marat Dukhan08c4a432019-10-03 09:29:21 -07006377############################# Build configurations #############################
6378
Marat Dukhanb8642352019-10-30 15:43:02 -07006379# Enables usage of assembly kernels.
Marat Dukhan08c4a432019-10-03 09:29:21 -07006380config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006381 name = "xnn_enable_assembly_explicit_true",
6382 define_values = {"xnn_enable_assembly": "true"},
6383)
6384
6385# Disables usage of assembly kernels.
6386config_setting(
6387 name = "xnn_enable_assembly_explicit_false",
6388 define_values = {"xnn_enable_assembly": "false"},
6389)
6390
Marat Dukhan9de90e02020-06-18 16:04:12 -07006391# Enables usage of sparse inference.
6392config_setting(
6393 name = "xnn_enable_sparse_explicit_true",
6394 define_values = {"xnn_enable_sparse": "true"},
6395)
6396
6397# Disables usage of sparse inference.
6398config_setting(
6399 name = "xnn_enable_sparse_explicit_false",
6400 define_values = {"xnn_enable_sparse": "false"},
6401)
6402
Marat Dukhan05702cf2020-03-26 15:41:33 -07006403# Disables usage of HMP-aware optimizations.
6404config_setting(
6405 name = "xnn_enable_hmp_explicit_false",
6406 define_values = {"xnn_enable_hmp": "false"},
6407)
6408
Chao Mei6ddfc602020-05-13 22:29:36 -07006409# Enable usage of optimized memory allocation
6410config_setting(
6411 name = "xnn_enable_memopt_explicit_true",
6412 define_values = {"xnnpack_enable_memopt": "true"},
6413)
6414
6415# Disable usage of optimized memory allocation
6416config_setting(
6417 name = "xnn_enable_memopt_explicit_false",
6418 define_values = {"xnnpack_enable_memopt": "false"},
6419)
6420
Marat Dukhanb8642352019-10-30 15:43:02 -07006421# Builds with -c dbg
6422config_setting(
6423 name = "debug_build",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006424 values = {
Marat Dukhanb8642352019-10-30 15:43:02 -07006425 "compilation_mode": "dbg",
6426 },
6427)
6428
6429# Builds with -c opt
6430config_setting(
6431 name = "optimized_build",
6432 values = {
6433 "compilation_mode": "opt",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006434 },
6435)
6436
6437config_setting(
Marat Dukhanb8642352019-10-30 15:43:02 -07006438 name = "linux_k8",
6439 values = {"cpu": "k8"},
6440)
6441
6442config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006443 name = "linux_arm",
6444 values = {"cpu": "arm"},
Marat Dukhan4e45e662019-10-03 15:40:24 -07006445)
6446
6447config_setting(
Marat Dukhanf0bd4de2020-06-15 15:53:19 -07006448 name = "linux_armeabi",
6449 values = {"cpu": "armeabi"},
6450)
6451
6452config_setting(
Terry Heo68eef3f2020-04-13 22:53:52 -07006453 name = "linux_armhf",
6454 values = {"cpu": "armhf"},
6455)
6456
6457config_setting(
Marat Dukhana720e932020-06-10 13:01:11 -07006458 name = "linux_armv7a",
6459 values = {"cpu": "armv7a"},
6460)
6461
6462config_setting(
Marat Dukhan582094e2020-04-30 17:21:25 -07006463 name = "linux_aarch64",
6464 values = {"cpu": "aarch64"},
6465)
6466
6467config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006468 name = "android",
6469 values = {"crosstool_top": "//external:android/crosstool"},
6470)
6471
6472config_setting(
6473 name = "android_armv7",
6474 values = {
6475 "crosstool_top": "//external:android/crosstool",
6476 "cpu": "armeabi-v7a",
6477 },
6478)
6479
6480config_setting(
6481 name = "android_arm64",
6482 values = {
6483 "crosstool_top": "//external:android/crosstool",
6484 "cpu": "arm64-v8a",
6485 },
6486)
6487
6488config_setting(
6489 name = "android_x86",
6490 values = {
6491 "crosstool_top": "//external:android/crosstool",
6492 "cpu": "x86",
6493 },
6494)
6495
6496config_setting(
6497 name = "android_x86_64",
6498 values = {
6499 "crosstool_top": "//external:android/crosstool",
6500 "cpu": "x86_64",
6501 },
6502)
6503
6504config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006505 name = "windows_x86_64",
6506 values = {"cpu": "x64_windows"},
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006507)
6508
6509config_setting(
Marat Dukhan10a38082020-04-17 03:58:35 -07006510 name = "windows_x86_64_clang",
6511 values = {
6512 "compiler": "clang-cl",
6513 "cpu": "x64_windows",
6514 },
6515)
6516
6517config_setting(
6518 name = "windows_x86_64_mingw",
6519 values = {
6520 "compiler": "mingw-gcc",
6521 "cpu": "x64_windows",
6522 },
6523)
6524
6525config_setting(
6526 name = "windows_x86_64_msys",
6527 values = {
6528 "compiler": "msys-gcc",
6529 "cpu": "x64_windows",
6530 },
Marat Dukhan9fe932e2020-04-11 17:14:15 -07006531)
6532
6533config_setting(
Marat Dukhan885ca242019-10-07 09:17:32 -07006534 name = "macos_x86_64",
6535 values = {
6536 "apple_platform_type": "macos",
6537 "cpu": "darwin",
6538 },
6539)
6540
6541config_setting(
Marat Dukhan08c4a432019-10-03 09:29:21 -07006542 name = "emscripten",
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006543 values = {"crosstool_top": "//toolchain:emscripten"},
Marat Dukhan08c4a432019-10-03 09:29:21 -07006544)
6545
6546config_setting(
6547 name = "emscripten_wasm",
6548 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006549 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006550 "cpu": "wasm",
6551 },
6552)
6553
6554config_setting(
6555 name = "emscripten_wasmsimd",
6556 values = {
Marat Dukhan1a729ec2019-10-07 09:31:44 -07006557 "crosstool_top": "//toolchain:emscripten",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006558 "cpu": "wasm",
Marat Dukhan81c62602020-05-29 13:22:49 -07006559 "copt": "-msimd128",
Marat Dukhan08c4a432019-10-03 09:29:21 -07006560 },
6561)
6562
6563config_setting(
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006564 name = "ios_armv7",
6565 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006566 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006567 "cpu": "ios_armv7",
6568 },
6569)
6570
6571config_setting(
6572 name = "ios_arm64",
6573 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006574 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006575 "cpu": "ios_arm64",
6576 },
6577)
6578
6579config_setting(
6580 name = "ios_arm64e",
6581 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006582 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006583 "cpu": "ios_arm64e",
6584 },
6585)
6586
6587config_setting(
6588 name = "ios_x86",
6589 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006590 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006591 "cpu": "ios_i386",
6592 },
6593)
6594
6595config_setting(
6596 name = "ios_x86_64",
6597 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006598 "apple_platform_type": "ios",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006599 "cpu": "ios_x86_64",
6600 },
6601)
6602
6603config_setting(
6604 name = "watchos_armv7k",
6605 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006606 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006607 "cpu": "watchos_armv7k",
6608 },
6609)
6610
6611config_setting(
6612 name = "watchos_arm64_32",
6613 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006614 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006615 "cpu": "watchos_arm64_32",
6616 },
6617)
6618
6619config_setting(
6620 name = "watchos_x86",
6621 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006622 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006623 "cpu": "watchos_i386",
6624 },
6625)
6626
6627config_setting(
6628 name = "watchos_x86_64",
6629 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006630 "apple_platform_type": "watchos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006631 "cpu": "watchos_x86_64",
6632 },
6633)
6634
6635config_setting(
6636 name = "tvos_arm64",
6637 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006638 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006639 "cpu": "tvos_arm64",
6640 },
6641)
6642
6643config_setting(
6644 name = "tvos_x86_64",
6645 values = {
Marat Dukhanf85fc332020-02-13 00:05:20 -08006646 "apple_platform_type": "tvos",
Marat Dukhan1498d1d2020-02-11 20:00:05 -08006647 "cpu": "tvos_x86_64",
6648 },
6649)