blob: 16b757e16fcbbee85ec498bbba9ba84dbb38a1ba [file] [log] [blame]
Miao Wange9993472020-02-10 15:00:10 -08001// Copyright (C) 2020 The Android Open Source Project
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7// http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
Bob Badourc0116292021-02-12 20:34:13 -080015package {
16 default_applicable_licenses: ["external_XNNPACK_license"],
17}
18
19// Added automatically by a large-scale-change
20// See: http://go/android-license-faq
21license {
22 name: "external_XNNPACK_license",
23 visibility: [":__subpackages__"],
24 license_kinds: [
25 "SPDX-license-identifier-BSD",
26 ],
27 license_text: [
28 "LICENSE",
29 ],
30}
31
Miao Wange9993472020-02-10 15:00:10 -080032OPERATOR_SRCS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -070033 "src/operators/argmax-pooling-nhwc.c",
34 "src/operators/average-pooling-nhwc.c",
35 "src/operators/binary-elementwise-nd.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070036 "src/operators/channel-shuffle-nc.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070037 "src/operators/constant-pad-nd.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070038 "src/operators/convolution-nchw.c",
39 "src/operators/convolution-nhwc.c",
40 "src/operators/deconvolution-nhwc.c",
Miao Wang5eea8312020-12-07 09:12:40 -080041 "src/operators/depth-to-space-nchw2nhwc.c",
42 "src/operators/depth-to-space-nhwc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070043 "src/operators/fully-connected-nc.c",
44 "src/operators/global-average-pooling-ncw.c",
45 "src/operators/global-average-pooling-nwc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070046 "src/operators/leaky-relu-nc.c",
47 "src/operators/max-pooling-nhwc.c",
48 "src/operators/prelu-nc.c",
Miao Wang5eea8312020-12-07 09:12:40 -080049 "src/operators/resize-bilinear-nchw.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070050 "src/operators/resize-bilinear-nhwc.c",
51 "src/operators/sigmoid-nc.c",
52 "src/operators/softmax-nc.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070053 "src/operators/unary-elementwise-nc.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -070054 "src/operators/unpooling-nhwc.c",
Miao Wange9993472020-02-10 15:00:10 -080055]
56
Miao Wang86f5fbe2020-07-24 11:16:10 -070057LOGGING_SRCS = [
58 "src/operator-strings.c",
59 "src/subgraph-strings.c",
60]
61
62SUBGRAPH_SRCS = [
63 "src/subgraph/abs.c",
64 "src/subgraph/add2.c",
65 "src/subgraph/argmax-pooling-2d.c",
66 "src/subgraph/average-pooling-2d.c",
67 "src/subgraph/bankers-rounding.c",
68 "src/subgraph/ceiling.c",
69 "src/subgraph/clamp.c",
70 "src/subgraph/convolution-2d.c",
71 "src/subgraph/deconvolution-2d.c",
Miao Wang5eea8312020-12-07 09:12:40 -080072 "src/subgraph/depth-to-space.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070073 "src/subgraph/depthwise-convolution-2d.c",
74 "src/subgraph/divide.c",
Miao Wang55abe392021-02-03 14:54:41 -080075 "src/subgraph/elu.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070076 "src/subgraph/floor.c",
Miao Wang5eea8312020-12-07 09:12:40 -080077 "src/subgraph/fully-connected.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070078 "src/subgraph/global-average-pooling-2d.c",
79 "src/subgraph/hardswish.c",
80 "src/subgraph/leaky-relu.c",
81 "src/subgraph/max-pooling-2d.c",
82 "src/subgraph/maximum2.c",
83 "src/subgraph/minimum2.c",
84 "src/subgraph/multiply2.c",
85 "src/subgraph/negate.c",
86 "src/subgraph/prelu.c",
87 "src/subgraph/sigmoid.c",
88 "src/subgraph/softmax.c",
Miao Wang5eea8312020-12-07 09:12:40 -080089 "src/subgraph/square-root.c",
90 "src/subgraph/square.c",
91 "src/subgraph/squared-difference.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070092 "src/subgraph/static-constant-pad.c",
93 "src/subgraph/static-reshape.c",
94 "src/subgraph/static-resize-bilinear-2d.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -070095 "src/subgraph/subtract.c",
96 "src/subgraph/unpooling-2d.c",
97]
98
Miao Wange9993472020-02-10 15:00:10 -080099TABLE_SRCS = [
100 "src/tables/exp2-k-over-64.c",
101 "src/tables/exp2-k-over-2048.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800102 "src/tables/exp2minus-k-over-4.c",
103 "src/tables/exp2minus-k-over-8.c",
104 "src/tables/exp2minus-k-over-16.c",
105 "src/tables/exp2minus-k-over-64.c",
106 "src/tables/exp2minus-k-over-2048.c",
Miao Wange9993472020-02-10 15:00:10 -0800107]
108
109SCALAR_UKERNELS = [
110 "src/f32-argmaxpool/4x-scalar-c1.c",
111 "src/f32-argmaxpool/9p8x-scalar-c1.c",
112 "src/f32-argmaxpool/9x-scalar-c1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700113 "src/f32-avgpool/9p8x-minmax-scalar-c1.c",
114 "src/f32-avgpool/9x-minmax-scalar-c1.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700115 "src/f32-clamp/gen/scalar-x1.c",
116 "src/f32-clamp/gen/scalar-x2.c",
117 "src/f32-clamp/gen/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800118 "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c",
119 "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700120 "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700121 "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c",
122 "src/f32-dwconv/gen/up1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800123 "src/f32-dwconv/gen/up1x4-scalar-acc2.c",
124 "src/f32-dwconv/gen/up1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700125 "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c",
126 "src/f32-dwconv/gen/up1x9-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800127 "src/f32-dwconv/gen/up1x9-scalar-acc2.c",
128 "src/f32-dwconv/gen/up1x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700129 "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c",
130 "src/f32-dwconv/gen/up1x25-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800131 "src/f32-dwconv/gen/up1x25-scalar-acc2.c",
132 "src/f32-dwconv/gen/up1x25-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700133 "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c",
134 "src/f32-dwconv/gen/up2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800135 "src/f32-dwconv/gen/up2x4-scalar-acc2.c",
136 "src/f32-dwconv/gen/up2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700137 "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c",
138 "src/f32-dwconv/gen/up2x9-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800139 "src/f32-dwconv/gen/up2x9-scalar-acc2.c",
140 "src/f32-dwconv/gen/up2x9-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700141 "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c",
142 "src/f32-dwconv/gen/up2x25-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800143 "src/f32-dwconv/gen/up2x25-scalar-acc2.c",
144 "src/f32-dwconv/gen/up2x25-scalar.c",
145 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c",
146 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c",
147 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c",
148 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c",
149 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c",
150 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c",
151 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c",
152 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c",
153 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c",
154 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c",
155 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c",
156 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c",
157 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c",
158 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c",
159 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c",
160 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c",
161 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c",
162 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c",
163 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c",
164 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c",
165 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c",
166 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c",
167 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c",
168 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c",
169 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c",
170 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c",
171 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c",
172 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c",
173 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c",
174 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c",
175 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c",
176 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c",
177 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c",
178 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c",
179 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c",
180 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c",
181 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c",
182 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700183 "src/f32-gavgpool-cw/scalar-x1.c",
Miao Wang400e4042020-04-17 10:15:59 -0700184 "src/f32-gavgpool/7p7x-minmax-scalar-c1.c",
185 "src/f32-gavgpool/7x-minmax-scalar-c1.c",
186 "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c",
187 "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c",
188 "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700189 "src/f32-gemm/gen/1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800190 "src/f32-gemm/gen/1x4-relu-scalar.c",
191 "src/f32-gemm/gen/1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700192 "src/f32-gemm/gen/2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800193 "src/f32-gemm/gen/2x4-relu-scalar.c",
194 "src/f32-gemm/gen/2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700195 "src/f32-gemm/gen/4x2-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800196 "src/f32-gemm/gen/4x2-relu-scalar.c",
197 "src/f32-gemm/gen/4x2-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700198 "src/f32-gemm/gen/4x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800199 "src/f32-gemm/gen/4x4-relu-scalar.c",
200 "src/f32-gemm/gen/4x4-scalar.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700201 "src/f32-hswish/gen/hswish-scalar-x1.c",
202 "src/f32-hswish/gen/hswish-scalar-x2.c",
203 "src/f32-hswish/gen/hswish-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800204 "src/f32-ibilinear-chw/gen/scalar-p1.c",
205 "src/f32-ibilinear-chw/gen/scalar-p2.c",
206 "src/f32-ibilinear-chw/gen/scalar-p4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700207 "src/f32-ibilinear/gen/scalar-c1.c",
208 "src/f32-ibilinear/gen/scalar-c2.c",
209 "src/f32-ibilinear/gen/scalar-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700210 "src/f32-igemm/gen/1x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800211 "src/f32-igemm/gen/1x4-relu-scalar.c",
212 "src/f32-igemm/gen/1x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700213 "src/f32-igemm/gen/2x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800214 "src/f32-igemm/gen/2x4-relu-scalar.c",
215 "src/f32-igemm/gen/2x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700216 "src/f32-igemm/gen/4x2-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800217 "src/f32-igemm/gen/4x2-relu-scalar.c",
218 "src/f32-igemm/gen/4x2-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700219 "src/f32-igemm/gen/4x4-minmax-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800220 "src/f32-igemm/gen/4x4-relu-scalar.c",
221 "src/f32-igemm/gen/4x4-scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700222 "src/f32-maxpool/9p8x-minmax-scalar-c1.c",
223 "src/f32-pavgpool/9p8x-minmax-scalar-c1.c",
224 "src/f32-pavgpool/9x-minmax-scalar-c1.c",
225 "src/f32-ppmm/gen/2x4-minmax-scalar.c",
226 "src/f32-ppmm/gen/3x3-minmax-scalar.c",
227 "src/f32-ppmm/gen/4x2-minmax-scalar.c",
228 "src/f32-ppmm/gen/4x4-minmax-scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800229 "src/f32-prelu/gen/scalar-2x1.c",
230 "src/f32-prelu/gen/scalar-2x4.c",
231 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c",
Miao Wange9993472020-02-10 15:00:10 -0800232 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800233 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c",
Miao Wange9993472020-02-10 15:00:10 -0800234 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c",
235 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800236 "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800237 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c",
Miao Wange9993472020-02-10 15:00:10 -0800238 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800239 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c",
Miao Wange9993472020-02-10 15:00:10 -0800240 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c",
241 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800242 "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700243 "src/f32-relu/gen/scalar-x1.c",
244 "src/f32-relu/gen/scalar-x2.c",
245 "src/f32-relu/gen/scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800246 "src/f32-relu/gen/scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800247 "src/f32-rmax/scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800248 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c",
249 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c",
250 "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800251 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c",
252 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c",
253 "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800254 "src/f32-sigmoid/gen/scalar-p5-div-x1.c",
255 "src/f32-sigmoid/gen/scalar-p5-div-x2.c",
256 "src/f32-sigmoid/gen/scalar-p5-div-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700257 "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c",
258 "src/f32-spmm/gen/1x1-minmax-scalar.c",
259 "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c",
260 "src/f32-spmm/gen/2x1-minmax-scalar.c",
261 "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c",
262 "src/f32-spmm/gen/4x1-minmax-scalar.c",
263 "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c",
264 "src/f32-spmm/gen/8x1-minmax-scalar.c",
265 "src/f32-spmm/gen/8x2-minmax-scalar.c",
266 "src/f32-spmm/gen/8x4-minmax-scalar.c",
267 "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c",
268 "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c",
269 "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800270 "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c",
271 "src/f32-vbinary/gen/vadd-relu-scalar-x1.c",
272 "src/f32-vbinary/gen/vadd-relu-scalar-x2.c",
273 "src/f32-vbinary/gen/vadd-relu-scalar-x4.c",
274 "src/f32-vbinary/gen/vadd-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700275 "src/f32-vbinary/gen/vadd-scalar-x1.c",
276 "src/f32-vbinary/gen/vadd-scalar-x2.c",
277 "src/f32-vbinary/gen/vadd-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800278 "src/f32-vbinary/gen/vadd-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700279 "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c",
280 "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c",
281 "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800282 "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c",
283 "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c",
284 "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c",
285 "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c",
286 "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700287 "src/f32-vbinary/gen/vaddc-scalar-x1.c",
288 "src/f32-vbinary/gen/vaddc-scalar-x2.c",
289 "src/f32-vbinary/gen/vaddc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800290 "src/f32-vbinary/gen/vaddc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700291 "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c",
292 "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c",
293 "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800294 "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c",
295 "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c",
296 "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c",
297 "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c",
298 "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700299 "src/f32-vbinary/gen/vdiv-scalar-x1.c",
300 "src/f32-vbinary/gen/vdiv-scalar-x2.c",
301 "src/f32-vbinary/gen/vdiv-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800302 "src/f32-vbinary/gen/vdiv-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700303 "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c",
304 "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c",
305 "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800306 "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c",
307 "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c",
308 "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c",
309 "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c",
310 "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700311 "src/f32-vbinary/gen/vdivc-scalar-x1.c",
312 "src/f32-vbinary/gen/vdivc-scalar-x2.c",
313 "src/f32-vbinary/gen/vdivc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800314 "src/f32-vbinary/gen/vdivc-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800315 "src/f32-vbinary/gen/vmax-scalar-x1.c",
316 "src/f32-vbinary/gen/vmax-scalar-x2.c",
317 "src/f32-vbinary/gen/vmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800318 "src/f32-vbinary/gen/vmax-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800319 "src/f32-vbinary/gen/vmaxc-scalar-x1.c",
320 "src/f32-vbinary/gen/vmaxc-scalar-x2.c",
321 "src/f32-vbinary/gen/vmaxc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800322 "src/f32-vbinary/gen/vmaxc-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800323 "src/f32-vbinary/gen/vmin-scalar-x1.c",
324 "src/f32-vbinary/gen/vmin-scalar-x2.c",
325 "src/f32-vbinary/gen/vmin-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800326 "src/f32-vbinary/gen/vmin-scalar-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800327 "src/f32-vbinary/gen/vminc-scalar-x1.c",
328 "src/f32-vbinary/gen/vminc-scalar-x2.c",
329 "src/f32-vbinary/gen/vminc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800330 "src/f32-vbinary/gen/vminc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700331 "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c",
332 "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c",
333 "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800334 "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c",
335 "src/f32-vbinary/gen/vmul-relu-scalar-x1.c",
336 "src/f32-vbinary/gen/vmul-relu-scalar-x2.c",
337 "src/f32-vbinary/gen/vmul-relu-scalar-x4.c",
338 "src/f32-vbinary/gen/vmul-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700339 "src/f32-vbinary/gen/vmul-scalar-x1.c",
340 "src/f32-vbinary/gen/vmul-scalar-x2.c",
341 "src/f32-vbinary/gen/vmul-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800342 "src/f32-vbinary/gen/vmul-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700343 "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c",
344 "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c",
345 "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800346 "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c",
347 "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c",
348 "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c",
349 "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c",
350 "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700351 "src/f32-vbinary/gen/vmulc-scalar-x1.c",
352 "src/f32-vbinary/gen/vmulc-scalar-x2.c",
353 "src/f32-vbinary/gen/vmulc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800354 "src/f32-vbinary/gen/vmulc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700355 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c",
356 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c",
357 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800358 "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c",
359 "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c",
360 "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c",
361 "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c",
362 "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700363 "src/f32-vbinary/gen/vrdivc-scalar-x1.c",
364 "src/f32-vbinary/gen/vrdivc-scalar-x2.c",
365 "src/f32-vbinary/gen/vrdivc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800366 "src/f32-vbinary/gen/vrdivc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700367 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c",
368 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c",
369 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800370 "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c",
371 "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c",
372 "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c",
373 "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c",
374 "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700375 "src/f32-vbinary/gen/vrsubc-scalar-x1.c",
376 "src/f32-vbinary/gen/vrsubc-scalar-x2.c",
377 "src/f32-vbinary/gen/vrsubc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800378 "src/f32-vbinary/gen/vrsubc-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700379 "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c",
380 "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c",
381 "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800382 "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700383 "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c",
384 "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c",
385 "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800386 "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700387 "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c",
388 "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c",
389 "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800390 "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c",
391 "src/f32-vbinary/gen/vsub-relu-scalar-x1.c",
392 "src/f32-vbinary/gen/vsub-relu-scalar-x2.c",
393 "src/f32-vbinary/gen/vsub-relu-scalar-x4.c",
394 "src/f32-vbinary/gen/vsub-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700395 "src/f32-vbinary/gen/vsub-scalar-x1.c",
396 "src/f32-vbinary/gen/vsub-scalar-x2.c",
397 "src/f32-vbinary/gen/vsub-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800398 "src/f32-vbinary/gen/vsub-scalar-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700399 "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c",
400 "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c",
401 "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800402 "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c",
403 "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c",
404 "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c",
405 "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c",
406 "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700407 "src/f32-vbinary/gen/vsubc-scalar-x1.c",
408 "src/f32-vbinary/gen/vsubc-scalar-x2.c",
409 "src/f32-vbinary/gen/vsubc-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800410 "src/f32-vbinary/gen/vsubc-scalar-x8.c",
Miao Wang55abe392021-02-03 14:54:41 -0800411 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x1.c",
412 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x2.c",
413 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x3.c",
414 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x4.c",
415 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x5.c",
416 "src/f32-velu/gen/velu-scalar-rr2-lut16-p3-x6.c",
417 "src/f32-velu/gen/velu-scalar-rr2-p6-x1.c",
418 "src/f32-velu/gen/velu-scalar-rr2-p6-x2.c",
419 "src/f32-velu/gen/velu-scalar-rr2-p6-x3.c",
420 "src/f32-velu/gen/velu-scalar-rr2-p6-x4.c",
421 "src/f32-velu/gen/velu-scalar-rr2-p6-x5.c",
422 "src/f32-velu/gen/velu-scalar-rr2-p6-x6.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700423 "src/f32-vlrelu/gen/vlrelu-scalar-x1.c",
424 "src/f32-vlrelu/gen/vlrelu-scalar-x2.c",
425 "src/f32-vlrelu/gen/vlrelu-scalar-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700426 "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c",
427 "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c",
428 "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700429 "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c",
430 "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c",
431 "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800432 "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c",
433 "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c",
434 "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c",
435 "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c",
436 "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c",
437 "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c",
438 "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c",
439 "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c",
440 "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700441 "src/f32-vsqrt/gen/scalar-sqrt-x1.c",
442 "src/f32-vsqrt/gen/scalar-sqrt-x2.c",
443 "src/f32-vsqrt/gen/scalar-sqrt-x4.c",
444 "src/f32-vunary/gen/vabs-scalar-x1.c",
445 "src/f32-vunary/gen/vabs-scalar-x2.c",
446 "src/f32-vunary/gen/vabs-scalar-x4.c",
447 "src/f32-vunary/gen/vneg-scalar-x1.c",
448 "src/f32-vunary/gen/vneg-scalar-x2.c",
449 "src/f32-vunary/gen/vneg-scalar-x4.c",
450 "src/f32-vunary/gen/vsqr-scalar-x1.c",
451 "src/f32-vunary/gen/vsqr-scalar-x2.c",
452 "src/f32-vunary/gen/vsqr-scalar-x4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800453 "src/math/expm1minus-scalar-rr2-lut4-p4.c",
454 "src/math/expm1minus-scalar-rr2-lut8-p3.c",
455 "src/math/expm1minus-scalar-rr2-lut8-p4.c",
456 "src/math/expm1minus-scalar-rr2-lut16-p3.c",
457 "src/math/expm1minus-scalar-rr2-lut16-p4.c",
458 "src/math/expm1minus-scalar-rr2-p5.c",
459 "src/math/expm1minus-scalar-rr2-p6.c",
460 "src/math/expminus-scalar-rr2-lut64-p2.c",
461 "src/math/expminus-scalar-rr2-lut2048-p1.c",
462 "src/math/expminus-scalar-rr2-p5.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700463 "src/math/roundd-scalar-addsub.c",
464 "src/math/roundd-scalar-cvt.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700465 "src/math/roundd-scalar-floor.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800466 "src/math/roundne-scalar-addsub.c",
467 "src/math/roundne-scalar-nearbyint.c",
468 "src/math/roundne-scalar-rint.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700469 "src/math/roundu-scalar-addsub.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700470 "src/math/roundu-scalar-ceil.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700471 "src/math/roundu-scalar-cvt.c",
472 "src/math/roundz-scalar-addsub.c",
473 "src/math/roundz-scalar-cvt.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700474 "src/math/roundz-scalar-trunc.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800475 "src/math/sigmoid-scalar-rr2-lut64-p2-div.c",
476 "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c",
477 "src/math/sigmoid-scalar-rr2-p5-div.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800478 "src/qs8-requantization/fp32-scalar-lrintf.c",
479 "src/qs8-requantization/fp32-scalar-magic.c",
480 "src/qs8-requantization/precise-scalar-signed64.c",
481 "src/qs8-requantization/precise-scalar-unsigned32.c",
482 "src/qs8-requantization/precise-scalar-unsigned64.c",
483 "src/qs8-requantization/q31-scalar.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700484 "src/qu8-avgpool/9p8x-minmax-scalar-c1.c",
485 "src/qu8-avgpool/9x-minmax-scalar-c1.c",
486 "src/qu8-dwconv/up1x9-minmax-scalar.c",
487 "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c",
488 "src/qu8-gavgpool/7x-minmax-scalar-c1.c",
489 "src/qu8-gemm/2x2-minmax-scalar.c",
490 "src/qu8-igemm/2x2-minmax-scalar.c",
491 "src/qu8-requantization/fp32-scalar-lrintf.c",
492 "src/qu8-requantization/fp32-scalar-magic.c",
493 "src/qu8-requantization/precise-scalar-signed64.c",
494 "src/qu8-requantization/precise-scalar-unsigned32.c",
495 "src/qu8-requantization/precise-scalar-unsigned64.c",
496 "src/qu8-requantization/q31-scalar.c",
497 "src/qu8-vadd/minmax-scalar.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700498 "src/u8-clamp/scalar-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800499 "src/u8-lut32norm/scalar.c",
Miao Wang400e4042020-04-17 10:15:59 -0700500 "src/u8-maxpool/9p8x-minmax-scalar-c1.c",
Miao Wange9993472020-02-10 15:00:10 -0800501 "src/u8-rmax/scalar.c",
Miao Wange9993472020-02-10 15:00:10 -0800502 "src/x8-lut/scalar.c",
503 "src/x8-zip/x2-scalar.c",
504 "src/x8-zip/x3-scalar.c",
505 "src/x8-zip/x4-scalar.c",
506 "src/x8-zip/xm-scalar.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800507 "src/x32-depthtospace2d-chw2hwc/scalar.c",
508 "src/x32-fill/scalar-float.c",
509 "src/x32-fill/scalar-int.c",
510 "src/x32-packx/x2-scalar.c",
511 "src/x32-packx/x3-scalar.c",
512 "src/x32-packx/x4-scalar.c",
513 "src/x32-pad/scalar-float.c",
514 "src/x32-pad/scalar-int.c",
515 "src/x32-unpool/scalar.c",
516 "src/x32-zip/x2-scalar.c",
517 "src/x32-zip/x3-scalar.c",
518 "src/x32-zip/x4-scalar.c",
519 "src/x32-zip/xm-scalar.c",
520 "src/xx-copy/memcpy.c",
Miao Wange9993472020-02-10 15:00:10 -0800521]
522
Miao Wange9993472020-02-10 15:00:10 -0800523// ISA-specific micro-kernels
524NEON_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -0700525 "src/f32-argmaxpool/4x-neon-c4.c",
526 "src/f32-argmaxpool/9p8x-neon-c4.c",
527 "src/f32-argmaxpool/9x-neon-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700528 "src/f32-avgpool/9p8x-minmax-neon-c4.c",
529 "src/f32-avgpool/9x-minmax-neon-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700530 "src/f32-clamp/gen/neon-x4.c",
531 "src/f32-clamp/gen/neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700532 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700533 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800534 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700535 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c",
536 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700537 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800538 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700539 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c",
Miao Wang55abe392021-02-03 14:54:41 -0800540 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neon-2x2.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700541 "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800542 "src/f32-dwconv/gen/up4x4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700543 "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800544 "src/f32-dwconv/gen/up4x9-minmax-neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700545 "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800546 "src/f32-dwconv/gen/up4x25-minmax-neon.c",
547 "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c",
548 "src/f32-dwconv/gen/up8x4-minmax-neon.c",
549 "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c",
550 "src/f32-dwconv/gen/up8x9-minmax-neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700551 "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800552 "src/f32-dwconv/gen/up8x25-minmax-neon.c",
553 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c",
554 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c",
555 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c",
556 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c",
557 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c",
558 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c",
559 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c",
560 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c",
561 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c",
562 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c",
563 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c",
564 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c",
565 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c",
566 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c",
567 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c",
568 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c",
569 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c",
570 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c",
571 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c",
572 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c",
573 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c",
574 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c",
575 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c",
576 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c",
577 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c",
578 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c",
579 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c",
580 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c",
581 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c",
582 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c",
583 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c",
584 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c",
585 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c",
586 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c",
587 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c",
588 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c",
589 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c",
590 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c",
591 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c",
592 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c",
593 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700594 "src/f32-gavgpool-cw/neon-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700595 "src/f32-gavgpool/7p7x-minmax-neon-c4.c",
596 "src/f32-gavgpool/7x-minmax-neon-c4.c",
Miao Wang400e4042020-04-17 10:15:59 -0700597 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800598 "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c",
599 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700600 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800601 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c",
602 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c",
603 "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c",
604 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c",
605 "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700606 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c",
607 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800608 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c",
609 "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -0700610 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c",
611 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800612 "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c",
613 "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c",
614 "src/f32-gemm/gen/1x8s4-minmax-neon.c",
615 "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c",
616 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c",
617 "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c",
618 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c",
619 "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c",
620 "src/f32-gemm/gen/4x8s4-minmax-neon.c",
621 "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c",
622 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c",
623 "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c",
624 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c",
625 "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c",
626 "src/f32-gemm/gen/6x8s4-minmax-neon.c",
627 "src/f32-gemm/gen/8x8s4-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700628 "src/f32-hswish/gen/hswish-neon-x4.c",
629 "src/f32-hswish/gen/hswish-neon-x8.c",
630 "src/f32-hswish/gen/hswish-neon-x16.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700631 "src/f32-ibilinear-chw/gen/neon-p4.c",
632 "src/f32-ibilinear-chw/gen/neon-p8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700633 "src/f32-ibilinear/gen/neon-c4.c",
634 "src/f32-ibilinear/gen/neon-c8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800635 "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700636 "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800637 "src/f32-igemm/gen/1x8s4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700638 "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c",
639 "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700640 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800641 "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c",
642 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c",
643 "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c",
644 "src/f32-igemm/gen/4x8s4-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700645 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c",
646 "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800647 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c",
648 "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -0700649 "src/f32-igemm/gen/6x8s4-minmax-neon.c",
650 "src/f32-igemm/gen/8x8s4-minmax-neon.c",
651 "src/f32-maxpool/9p8x-minmax-neon-c4.c",
652 "src/f32-pavgpool/9p8x-minmax-neon-c4.c",
653 "src/f32-pavgpool/9x-minmax-neon-c4.c",
654 "src/f32-ppmm/gen/4x8-minmax-neon.c",
655 "src/f32-ppmm/gen/8x8-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700656 "src/f32-prelu/gen/neon-1x4.c",
657 "src/f32-prelu/gen/neon-1x8.c",
658 "src/f32-prelu/gen/neon-1x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800659 "src/f32-prelu/gen/neon-2x4.c",
660 "src/f32-prelu/gen/neon-2x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700661 "src/f32-prelu/gen/neon-2x16.c",
662 "src/f32-prelu/gen/neon-4x4.c",
663 "src/f32-prelu/gen/neon-4x8.c",
664 "src/f32-prelu/gen/neon-4x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800665 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -0800666 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800667 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800668 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c",
669 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800670 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c",
Miao Wange9993472020-02-10 15:00:10 -0800671 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c",
672 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800673 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c",
Miao Wange9993472020-02-10 15:00:10 -0800674 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c",
675 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800676 "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c",
677 "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c",
678 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c",
679 "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c",
680 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c",
681 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c",
682 "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c",
683 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c",
684 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c",
685 "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c",
686 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c",
687 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c",
688 "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700689 "src/f32-relu/gen/neon-x4.c",
690 "src/f32-relu/gen/neon-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800691 "src/f32-rmax/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800692 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c",
693 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c",
694 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c",
695 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c",
696 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c",
697 "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c",
698 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c",
699 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c",
700 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c",
701 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c",
702 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c",
703 "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800704 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c",
705 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c",
706 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c",
707 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c",
708 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c",
709 "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -0800710 "src/f32-spmm/gen/4x1-minmax-neon-pipelined.c",
711 "src/f32-spmm/gen/4x1-minmax-neon-x2.c",
712 "src/f32-spmm/gen/4x1-minmax-neon.c",
713 "src/f32-spmm/gen/8x1-minmax-neon-pipelined.c",
714 "src/f32-spmm/gen/8x1-minmax-neon-x2.c",
715 "src/f32-spmm/gen/8x1-minmax-neon.c",
716 "src/f32-spmm/gen/12x1-minmax-neon.c",
717 "src/f32-spmm/gen/16x1-minmax-neon-pipelined.c",
718 "src/f32-spmm/gen/16x1-minmax-neon-x2.c",
719 "src/f32-spmm/gen/16x1-minmax-neon.c",
720 "src/f32-spmm/gen/32x1-minmax-neon-pipelined.c",
721 "src/f32-spmm/gen/32x1-minmax-neon-x2.c",
722 "src/f32-spmm/gen/32x1-minmax-neon.c",
Miao Wang400e4042020-04-17 10:15:59 -0700723 "src/f32-vbinary/gen/vadd-minmax-neon-x4.c",
724 "src/f32-vbinary/gen/vadd-minmax-neon-x8.c",
725 "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c",
726 "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c",
Miao Wange9993472020-02-10 15:00:10 -0800727 "src/f32-vbinary/gen/vmax-neon-x4.c",
728 "src/f32-vbinary/gen/vmax-neon-x8.c",
729 "src/f32-vbinary/gen/vmaxc-neon-x4.c",
730 "src/f32-vbinary/gen/vmaxc-neon-x8.c",
731 "src/f32-vbinary/gen/vmin-neon-x4.c",
732 "src/f32-vbinary/gen/vmin-neon-x8.c",
733 "src/f32-vbinary/gen/vminc-neon-x4.c",
734 "src/f32-vbinary/gen/vminc-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700735 "src/f32-vbinary/gen/vmul-minmax-neon-x4.c",
736 "src/f32-vbinary/gen/vmul-minmax-neon-x8.c",
737 "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c",
738 "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c",
739 "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c",
740 "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700741 "src/f32-vbinary/gen/vsqrdiff-neon-x4.c",
742 "src/f32-vbinary/gen/vsqrdiff-neon-x8.c",
743 "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c",
744 "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700745 "src/f32-vbinary/gen/vsub-minmax-neon-x4.c",
746 "src/f32-vbinary/gen/vsub-minmax-neon-x8.c",
747 "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c",
748 "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c",
Miao Wang55abe392021-02-03 14:54:41 -0800749 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x4.c",
750 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x8.c",
751 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x12.c",
752 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x16.c",
753 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x20.c",
754 "src/f32-velu/gen/velu-neon-rr2-lut16-p3-x24.c",
755 "src/f32-velu/gen/velu-neon-rr2-p6-x4.c",
756 "src/f32-velu/gen/velu-neon-rr2-p6-x8.c",
757 "src/f32-velu/gen/velu-neon-rr2-p6-x12.c",
758 "src/f32-velu/gen/velu-neon-rr2-p6-x16.c",
759 "src/f32-velu/gen/velu-neon-rr2-p6-x20.c",
760 "src/f32-velu/gen/velu-neon-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700761 "src/f32-vlrelu/gen/vlrelu-neon-x4.c",
762 "src/f32-vlrelu/gen/vlrelu-neon-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700763 "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c",
764 "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700765 "src/f32-vrnd/gen/vrndd-neon-x4.c",
766 "src/f32-vrnd/gen/vrndd-neon-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800767 "src/f32-vrnd/gen/vrndne-neon-x4.c",
768 "src/f32-vrnd/gen/vrndne-neon-x8.c",
769 "src/f32-vrnd/gen/vrndu-neon-x4.c",
770 "src/f32-vrnd/gen/vrndu-neon-x8.c",
771 "src/f32-vrnd/gen/vrndz-neon-x4.c",
772 "src/f32-vrnd/gen/vrndz-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700773 "src/f32-vunary/gen/vabs-neon-x4.c",
774 "src/f32-vunary/gen/vabs-neon-x8.c",
775 "src/f32-vunary/gen/vneg-neon-x4.c",
776 "src/f32-vunary/gen/vneg-neon-x8.c",
777 "src/f32-vunary/gen/vsqr-neon-x4.c",
778 "src/f32-vunary/gen/vsqr-neon-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800779 "src/math/expm1minus-neon-rr2-lut16-p3.c",
780 "src/math/expm1minus-neon-rr2-p6.c",
781 "src/math/roundd-neon-addsub.c",
782 "src/math/roundd-neon-cvt.c",
783 "src/math/roundne-neon-addsub.c",
784 "src/math/roundu-neon-addsub.c",
785 "src/math/roundu-neon-cvt.c",
786 "src/math/roundz-neon-addsub.c",
787 "src/math/roundz-neon-cvt.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800788 "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c",
789 "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c",
790 "src/math/sigmoid-neon-rr2-p5-nr2recps.c",
791 "src/math/sqrt-neon-nr1rsqrts.c",
792 "src/math/sqrt-neon-nr2rsqrts.c",
793 "src/math/sqrt-neon-nr3rsqrts.c",
794 "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c",
795 "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c",
796 "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c",
797 "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c",
798 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c",
799 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c",
800 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c",
801 "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c",
802 "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c",
803 "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c",
804 "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c",
805 "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c",
806 "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700807 "src/qs8-gemm/gen/1x8-minmax-neon-mull-addw-dup.c",
808 "src/qs8-gemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
809 "src/qs8-gemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
810 "src/qs8-gemm/gen/1x8c8-minmax-neon-mlal-padal.c",
811 "src/qs8-gemm/gen/1x8c8-minmax-neon-mull-padal.c",
812 "src/qs8-gemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800813 "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700814 "src/qs8-gemm/gen/1x16-minmax-neon-mull-addw-dup.c",
815 "src/qs8-gemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
816 "src/qs8-gemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
817 "src/qs8-gemm/gen/1x16c8-minmax-neon-mlal-padal.c",
818 "src/qs8-gemm/gen/1x16c8-minmax-neon-mull-padal.c",
819 "src/qs8-gemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800820 "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700821 "src/qs8-gemm/gen/2x8-minmax-neon-mull-addw-dup.c",
822 "src/qs8-gemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
823 "src/qs8-gemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
824 "src/qs8-gemm/gen/2x8c8-minmax-neon-mlal-padal.c",
825 "src/qs8-gemm/gen/2x8c8-minmax-neon-mull-padal.c",
826 "src/qs8-gemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800827 "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700828 "src/qs8-gemm/gen/2x16-minmax-neon-mull-addw-dup.c",
829 "src/qs8-gemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
830 "src/qs8-gemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
831 "src/qs8-gemm/gen/2x16c8-minmax-neon-mlal-padal.c",
832 "src/qs8-gemm/gen/2x16c8-minmax-neon-mull-padal.c",
833 "src/qs8-gemm/gen/2x16c16-minmax-neon-mlal-padal.c",
834 "src/qs8-gemm/gen/3x8-minmax-neon-mlal-lane.c",
835 "src/qs8-gemm/gen/3x8-minmax-neon-mull-addw-dup.c",
836 "src/qs8-gemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
837 "src/qs8-gemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
838 "src/qs8-gemm/gen/3x8c8-minmax-neon-mlal-padal.c",
839 "src/qs8-gemm/gen/3x8c8-minmax-neon-mull-padal.c",
840 "src/qs8-gemm/gen/3x8c16-minmax-neon-mlal-padal.c",
841 "src/qs8-gemm/gen/3x16-minmax-neon-mlal-lane.c",
842 "src/qs8-gemm/gen/3x16-minmax-neon-mull-addw-dup.c",
843 "src/qs8-gemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
844 "src/qs8-gemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
845 "src/qs8-gemm/gen/3x16c8-minmax-neon-mlal-padal.c",
846 "src/qs8-gemm/gen/3x16c8-minmax-neon-mull-padal.c",
847 "src/qs8-gemm/gen/3x16c16-minmax-neon-mlal-padal.c",
848 "src/qs8-gemm/gen/4x8-minmax-neon-mlal-lane.c",
849 "src/qs8-gemm/gen/4x8-minmax-neon-mull-addw-dup.c",
850 "src/qs8-gemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
851 "src/qs8-gemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
852 "src/qs8-gemm/gen/4x8c8-minmax-neon-mlal-padal.c",
853 "src/qs8-gemm/gen/4x8c8-minmax-neon-mull-padal.c",
854 "src/qs8-gemm/gen/4x8c16-minmax-neon-mlal-padal.c",
855 "src/qs8-gemm/gen/4x16-minmax-neon-mlal-lane.c",
856 "src/qs8-gemm/gen/4x16-minmax-neon-mull-addw-dup.c",
857 "src/qs8-gemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
858 "src/qs8-gemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
859 "src/qs8-gemm/gen/4x16c8-minmax-neon-mlal-padal.c",
860 "src/qs8-gemm/gen/4x16c8-minmax-neon-mull-padal.c",
861 "src/qs8-gemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800862 "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700863 "src/qs8-igemm/gen/1x8-minmax-neon-mull-addw-dup.c",
864 "src/qs8-igemm/gen/1x8c2-minmax-neon-mlal-padal-dup.c",
865 "src/qs8-igemm/gen/1x8c2-minmax-neon-mull-padal-dup.c",
866 "src/qs8-igemm/gen/1x8c8-minmax-neon-mlal-padal.c",
867 "src/qs8-igemm/gen/1x8c8-minmax-neon-mull-padal.c",
868 "src/qs8-igemm/gen/1x8c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800869 "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700870 "src/qs8-igemm/gen/1x16-minmax-neon-mull-addw-dup.c",
871 "src/qs8-igemm/gen/1x16c2-minmax-neon-mlal-padal-dup.c",
872 "src/qs8-igemm/gen/1x16c2-minmax-neon-mull-padal-dup.c",
873 "src/qs8-igemm/gen/1x16c8-minmax-neon-mlal-padal.c",
874 "src/qs8-igemm/gen/1x16c8-minmax-neon-mull-padal.c",
875 "src/qs8-igemm/gen/1x16c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800876 "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700877 "src/qs8-igemm/gen/2x8-minmax-neon-mull-addw-dup.c",
878 "src/qs8-igemm/gen/2x8c2-minmax-neon-mlal-padal-dup.c",
879 "src/qs8-igemm/gen/2x8c2-minmax-neon-mull-padal-dup.c",
880 "src/qs8-igemm/gen/2x8c8-minmax-neon-mlal-padal.c",
881 "src/qs8-igemm/gen/2x8c8-minmax-neon-mull-padal.c",
882 "src/qs8-igemm/gen/2x8c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800883 "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700884 "src/qs8-igemm/gen/2x16-minmax-neon-mull-addw-dup.c",
885 "src/qs8-igemm/gen/2x16c2-minmax-neon-mlal-padal-dup.c",
886 "src/qs8-igemm/gen/2x16c2-minmax-neon-mull-padal-dup.c",
887 "src/qs8-igemm/gen/2x16c8-minmax-neon-mlal-padal.c",
888 "src/qs8-igemm/gen/2x16c8-minmax-neon-mull-padal.c",
889 "src/qs8-igemm/gen/2x16c16-minmax-neon-mlal-padal.c",
890 "src/qs8-igemm/gen/3x8-minmax-neon-mlal-lane.c",
891 "src/qs8-igemm/gen/3x8-minmax-neon-mull-addw-dup.c",
892 "src/qs8-igemm/gen/3x8c2-minmax-neon-mlal-padal-dup.c",
893 "src/qs8-igemm/gen/3x8c2-minmax-neon-mull-padal-dup.c",
894 "src/qs8-igemm/gen/3x8c8-minmax-neon-mlal-padal.c",
895 "src/qs8-igemm/gen/3x8c8-minmax-neon-mull-padal.c",
896 "src/qs8-igemm/gen/3x8c16-minmax-neon-mlal-padal.c",
897 "src/qs8-igemm/gen/3x16-minmax-neon-mlal-lane.c",
898 "src/qs8-igemm/gen/3x16-minmax-neon-mull-addw-dup.c",
899 "src/qs8-igemm/gen/3x16c2-minmax-neon-mlal-padal-dup.c",
900 "src/qs8-igemm/gen/3x16c2-minmax-neon-mull-padal-dup.c",
901 "src/qs8-igemm/gen/3x16c8-minmax-neon-mlal-padal.c",
902 "src/qs8-igemm/gen/3x16c8-minmax-neon-mull-padal.c",
903 "src/qs8-igemm/gen/3x16c16-minmax-neon-mlal-padal.c",
904 "src/qs8-igemm/gen/4x8-minmax-neon-mlal-lane.c",
905 "src/qs8-igemm/gen/4x8-minmax-neon-mull-addw-dup.c",
906 "src/qs8-igemm/gen/4x8c2-minmax-neon-mlal-padal-dup.c",
907 "src/qs8-igemm/gen/4x8c2-minmax-neon-mull-padal-dup.c",
908 "src/qs8-igemm/gen/4x8c8-minmax-neon-mlal-padal.c",
909 "src/qs8-igemm/gen/4x8c8-minmax-neon-mull-padal.c",
910 "src/qs8-igemm/gen/4x8c16-minmax-neon-mlal-padal.c",
911 "src/qs8-igemm/gen/4x16-minmax-neon-mlal-lane.c",
912 "src/qs8-igemm/gen/4x16-minmax-neon-mull-addw-dup.c",
913 "src/qs8-igemm/gen/4x16c2-minmax-neon-mlal-padal-dup.c",
914 "src/qs8-igemm/gen/4x16c2-minmax-neon-mull-padal-dup.c",
915 "src/qs8-igemm/gen/4x16c8-minmax-neon-mlal-padal.c",
916 "src/qs8-igemm/gen/4x16c8-minmax-neon-mull-padal.c",
917 "src/qs8-igemm/gen/4x16c16-minmax-neon-mlal-padal.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800918 "src/qs8-requantization/fp32-neon.c",
919 "src/qs8-requantization/precise-neon.c",
920 "src/qs8-requantization/q31-neon.c",
921 "src/qs8-vadd/gen/minmax-neon-ld64-x8.c",
922 "src/qs8-vadd/gen/minmax-neon-ld64-x16.c",
923 "src/qs8-vadd/gen/minmax-neon-ld64-x24.c",
924 "src/qs8-vadd/gen/minmax-neon-ld64-x32.c",
925 "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c",
926 "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c",
927 "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c",
928 "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700929 "src/qu8-avgpool/9p8x-minmax-neon-c8.c",
930 "src/qu8-avgpool/9x-minmax-neon-c8.c",
931 "src/qu8-dwconv/up8x9-minmax-neon.c",
932 "src/qu8-gavgpool/7p7x-minmax-neon-c8.c",
933 "src/qu8-gavgpool/7x-minmax-neon-c8.c",
934 "src/qu8-gemm/4x8-minmax-neon.c",
935 "src/qu8-gemm/8x8-minmax-neon.c",
936 "src/qu8-igemm/4x8-minmax-neon.c",
937 "src/qu8-igemm/8x8-minmax-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700938 "src/qu8-requantization/fp32-neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800939 "src/qu8-requantization/precise-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700940 "src/qu8-requantization/q31-neon.c",
941 "src/qu8-vadd/minmax-neon.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700942 "src/u8-clamp/neon-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -0700943 "src/u8-maxpool/9p8x-minmax-neon-c16.c",
Miao Wange9993472020-02-10 15:00:10 -0800944 "src/u8-rmax/neon.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800945 "src/x8-zip/x2-neon.c",
946 "src/x8-zip/x3-neon.c",
947 "src/x8-zip/x4-neon.c",
948 "src/x8-zip/xm-neon.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700949 "src/x32-fill/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800950 "src/x32-packx/x4-neon-st4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -0700951 "src/x32-pad/neon.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -0700952 "src/x32-unpool/neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800953 "src/x32-zip/x2-neon.c",
954 "src/x32-zip/x3-neon.c",
955 "src/x32-zip/x4-neon.c",
956 "src/x32-zip/xm-neon.c",
Miao Wange9993472020-02-10 15:00:10 -0800957]
958
959NEONFMA_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -0800960 "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c",
961 "src/f32-dwconv/gen/up4x4-minmax-neonfma.c",
962 "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c",
963 "src/f32-dwconv/gen/up4x9-minmax-neonfma.c",
964 "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c",
965 "src/f32-dwconv/gen/up4x25-minmax-neonfma.c",
966 "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c",
967 "src/f32-dwconv/gen/up8x4-minmax-neonfma.c",
968 "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c",
969 "src/f32-dwconv/gen/up8x9-minmax-neonfma.c",
970 "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c",
971 "src/f32-dwconv/gen/up8x25-minmax-neonfma.c",
972 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c",
973 "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c",
974 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c",
975 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c",
976 "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c",
977 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c",
978 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c",
979 "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c",
980 "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c",
981 "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c",
982 "src/f32-gemm/gen/1x8s4-minmax-neonfma.c",
983 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c",
984 "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c",
985 "src/f32-gemm/gen/4x8s4-minmax-neonfma.c",
986 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c",
987 "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c",
988 "src/f32-gemm/gen/6x8s4-minmax-neonfma.c",
989 "src/f32-gemm/gen/8x8s4-minmax-neonfma.c",
Miao Wang338e80e2021-03-23 08:56:09 -0700990 "src/f32-ibilinear-chw/gen/neonfma-p4.c",
991 "src/f32-ibilinear-chw/gen/neonfma-p8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -0700992 "src/f32-ibilinear/gen/neonfma-c4.c",
993 "src/f32-ibilinear/gen/neonfma-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -0700994 "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800995 "src/f32-igemm/gen/1x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -0700996 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -0800997 "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c",
998 "src/f32-igemm/gen/4x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -0700999 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c",
1000 "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001001 "src/f32-igemm/gen/6x8s4-minmax-neonfma.c",
1002 "src/f32-igemm/gen/8x8s4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -07001003 "src/f32-ppmm/gen/4x8-minmax-neonfma.c",
1004 "src/f32-ppmm/gen/8x8-minmax-neonfma.c",
Miao Wange9993472020-02-10 15:00:10 -08001005 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c",
Miao Wange9993472020-02-10 15:00:10 -08001006 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001007 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001008 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c",
1009 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001010 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c",
Miao Wange9993472020-02-10 15:00:10 -08001011 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c",
1012 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001013 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001014 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c",
1015 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001016 "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c",
1017 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c",
1018 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c",
1019 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c",
1020 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c",
1021 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c",
1022 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c",
1023 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c",
1024 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c",
1025 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c",
1026 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c",
1027 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c",
1028 "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c",
Miao Wange9993472020-02-10 15:00:10 -08001029 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c",
1030 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c",
1031 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c",
1032 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c",
1033 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c",
1034 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001035 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c",
1036 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c",
1037 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c",
1038 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c",
1039 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c",
1040 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001041 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c",
1042 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c",
1043 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c",
1044 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c",
1045 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c",
1046 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001047 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c",
1048 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c",
1049 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c",
1050 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c",
1051 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c",
1052 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001053 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c",
1054 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c",
1055 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c",
1056 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c",
1057 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c",
1058 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001059 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c",
1060 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c",
1061 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c",
1062 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c",
1063 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c",
1064 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001065 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c",
1066 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c",
1067 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c",
1068 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c",
1069 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c",
1070 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c",
1071 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c",
1072 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c",
1073 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c",
1074 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c",
1075 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c",
1076 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c",
1077 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c",
1078 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c",
1079 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c",
1080 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c",
1081 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c",
1082 "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -08001083 "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c",
1084 "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c",
1085 "src/f32-spmm/gen/4x1-minmax-neonfma.c",
1086 "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c",
1087 "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c",
1088 "src/f32-spmm/gen/8x1-minmax-neonfma.c",
1089 "src/f32-spmm/gen/12x1-minmax-neonfma.c",
1090 "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c",
1091 "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c",
1092 "src/f32-spmm/gen/16x1-minmax-neonfma.c",
1093 "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c",
1094 "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c",
1095 "src/f32-spmm/gen/32x1-minmax-neonfma.c",
1096 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x4.c",
1097 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x8.c",
1098 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x12.c",
1099 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x16.c",
1100 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x20.c",
1101 "src/f32-velu/gen/velu-neonfma-rr1-lut16-p3-x24.c",
1102 "src/f32-velu/gen/velu-neonfma-rr1-p6-x4.c",
1103 "src/f32-velu/gen/velu-neonfma-rr1-p6-x8.c",
1104 "src/f32-velu/gen/velu-neonfma-rr1-p6-x12.c",
1105 "src/f32-velu/gen/velu-neonfma-rr1-p6-x16.c",
1106 "src/f32-velu/gen/velu-neonfma-rr1-p6-x20.c",
1107 "src/f32-velu/gen/velu-neonfma-rr1-p6-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -07001108 "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c",
1109 "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001110 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c",
1111 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c",
1112 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c",
1113 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c",
1114 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c",
1115 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c",
1116 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c",
1117 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c",
1118 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c",
1119 "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c",
1120 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c",
1121 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c",
1122 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c",
1123 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c",
1124 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c",
1125 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c",
1126 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c",
1127 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c",
1128 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c",
1129 "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001130 "src/math/exp-neonfma-rr2-lut64-p2.c",
1131 "src/math/exp-neonfma-rr2-p5.c",
Miao Wang55abe392021-02-03 14:54:41 -08001132 "src/math/expm1minus-neonfma-rr1-lut16-p3.c",
1133 "src/math/expm1minus-neonfma-rr1-p6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001134 "src/math/expminus-neonfma-rr2-lut64-p2.c",
1135 "src/math/expminus-neonfma-rr2-lut2048-p1.c",
1136 "src/math/expminus-neonfma-rr2-p5.c",
Miao Wange9993472020-02-10 15:00:10 -08001137 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c",
1138 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c",
1139 "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001140 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c",
1141 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c",
1142 "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001143 "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c",
1144 "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c",
1145 "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001146 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c",
1147 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c",
1148 "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001149 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c",
1150 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c",
1151 "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c",
Miao Wange9993472020-02-10 15:00:10 -08001152 "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c",
1153 "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c",
1154 "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001155 "src/math/sqrt-neonfma-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001156 "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001157 "src/math/sqrt-neonfma-nr2fma.c",
1158 "src/math/sqrt-neonfma-nr2fma1adj.c",
1159 "src/math/sqrt-neonfma-nr3fma.c",
Miao Wange9993472020-02-10 15:00:10 -08001160]
1161
1162AARCH64_NEONFMA_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -08001163 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c",
1164 "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c",
1165 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c",
1166 "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c",
1167 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c",
1168 "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c",
1169 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c",
1170 "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c",
1171 "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c",
1172 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c",
1173 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c",
1174 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c",
1175 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c",
1176 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c",
1177 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c",
1178 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c",
1179 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c",
1180 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c",
1181 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c",
1182 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c",
1183 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c",
1184 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c",
1185 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c",
1186 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c",
1187 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c",
1188 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c",
1189 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c",
1190 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c",
1191 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c",
1192 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c",
1193 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c",
1194 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c",
1195 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c",
1196 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c",
1197 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c",
1198 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c",
1199 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c",
1200 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c",
1201 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c",
1202 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c",
1203 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c",
1204 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c",
1205 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c",
1206 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c",
1207 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c",
1208 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c",
1209 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c",
1210 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c",
1211 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c",
1212 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c",
Miao Wang400e4042020-04-17 10:15:59 -07001213 "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001214 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001215 "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001216 "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c",
1217 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c",
1218 "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001219 "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1220 "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1221 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c",
1222 "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c",
1223 "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c",
1224 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1225 "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001226 "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c",
1227 "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c",
1228 "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001229 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001230 "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c",
Miao Wang400e4042020-04-17 10:15:59 -07001231 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c",
1232 "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c",
Miao Wange9993472020-02-10 15:00:10 -08001233 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c",
1234 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c",
1235 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c",
1236 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c",
1237 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c",
1238 "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c",
1239 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c",
1240 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c",
1241 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c",
1242 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c",
1243 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c",
1244 "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001245 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c",
1246 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c",
1247 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c",
1248 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c",
1249 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c",
1250 "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c",
Miao Wang400e4042020-04-17 10:15:59 -07001251 "src/f32-spmm/gen/4x2-minmax-neonfma.c",
1252 "src/f32-spmm/gen/4x4-minmax-neonfma.c",
Miao Wang400e4042020-04-17 10:15:59 -07001253 "src/f32-spmm/gen/8x2-minmax-neonfma.c",
1254 "src/f32-spmm/gen/8x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001255 "src/f32-spmm/gen/12x2-minmax-neonfma.c",
1256 "src/f32-spmm/gen/12x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001257 "src/f32-spmm/gen/16x2-minmax-neonfma.c",
1258 "src/f32-spmm/gen/16x4-minmax-neonfma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001259 "src/f32-spmm/gen/32x2-minmax-neonfma.c",
1260 "src/f32-spmm/gen/32x4-minmax-neonfma.c",
1261 "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c",
1262 "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c",
1263 "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c",
1264 "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c",
1265 "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c",
1266 "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001267 "src/f32-vsqrt/gen/neon-sqrt-x4.c",
1268 "src/f32-vsqrt/gen/neon-sqrt-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001269 "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001270 "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001271 "src/math/sigmoid-neonfma-rr1-p5-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001272 "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001273 "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c",
Miao Wange9993472020-02-10 15:00:10 -08001274 "src/math/sigmoid-neonfma-rr2-p5-div.c",
1275]
1276
Miao Wangc0aa11a2020-06-10 13:41:26 -07001277NEONV8_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -07001278 "src/f32-vrnd/gen/vrndd-neonv8-x4.c",
1279 "src/f32-vrnd/gen/vrndd-neonv8-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001280 "src/f32-vrnd/gen/vrndne-neonv8-x4.c",
1281 "src/f32-vrnd/gen/vrndne-neonv8-x8.c",
1282 "src/f32-vrnd/gen/vrndu-neonv8-x4.c",
1283 "src/f32-vrnd/gen/vrndu-neonv8-x8.c",
1284 "src/f32-vrnd/gen/vrndz-neonv8-x4.c",
1285 "src/f32-vrnd/gen/vrndz-neonv8-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001286 "src/math/roundd-neonv8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001287 "src/math/roundne-neonv8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001288 "src/math/roundu-neonv8.c",
1289 "src/math/roundz-neonv8.c",
1290]
1291
Miao Wange9993472020-02-10 15:00:10 -08001292AARCH64_NEONFP16ARITH_UKERNELS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -07001293 "src/f16-clamp/gen/neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001294 "src/f16-clamp/gen/neonfp16arith-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001295 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c",
1296 "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c",
1297 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c",
1298 "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001299 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c",
1300 "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c",
1301 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c",
1302 "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c",
1303 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c",
1304 "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c",
1305 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c",
1306 "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001307 "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c",
1308 "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001309 "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c",
1310 "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c",
1311 "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c",
1312 "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c",
1313 "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c",
1314 "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c",
1315 "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c",
1316 "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c",
1317 "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1318 "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1319 "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1320 "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1321 "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1322 "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1323 "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1324 "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001325 "src/f16-hswish/gen/hswish-neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001326 "src/f16-hswish/gen/hswish-neonfp16arith-x16.c",
1327 "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c",
1328 "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c",
1329 "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c",
1330 "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c",
1331 "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c",
1332 "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c",
1333 "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c",
1334 "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001335 "src/f16-prelu/gen/neonfp16arith-2x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001336 "src/f16-prelu/gen/neonfp16arith-2x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001337 "src/f16-relu/gen/neonfp16arith-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001338 "src/f16-relu/gen/neonfp16arith-x16.c",
1339 "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001340 "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001341 "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001342 "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001343 "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001344 "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001345 "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001346 "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001347 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c",
1348 "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c",
1349 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c",
1350 "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c",
1351 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c",
1352 "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c",
1353 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c",
1354 "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c",
1355 "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c",
1356 "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c",
1357 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c",
1358 "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c",
1359 "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c",
1360 "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c",
1361 "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c",
1362 "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c",
1363 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c",
1364 "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c",
1365 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c",
1366 "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c",
1367 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c",
1368 "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c",
1369 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c",
1370 "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c",
1371 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c",
1372 "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c",
1373 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c",
1374 "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c",
1375 "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c",
1376 "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c",
1377]
1378
1379NEONDOT_UKERNELS = [
1380 "src/qs8-gemm/gen/1x8c4-minmax-neondot.c",
1381 "src/qs8-gemm/gen/1x16c4-minmax-neondot.c",
1382 "src/qs8-gemm/gen/4x8c4-minmax-neondot.c",
1383 "src/qs8-gemm/gen/4x16c4-minmax-neondot.c",
1384 "src/qs8-gemm/gen/6x8c4-minmax-neondot.c",
1385 "src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
1386 "src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
1387 "src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001388 "src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
1389 "src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
1390 "src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
1391 "src/qs8-igemm/gen/4x16c4-minmax-neondot.c",
1392 "src/qs8-igemm/gen/6x8c4-minmax-neondot.c",
1393 "src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
1394 "src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
1395 "src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
Miao Wange9993472020-02-10 15:00:10 -08001396]
1397
1398SSE_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07001399 "src/f32-avgpool/9p8x-minmax-sse-c4.c",
1400 "src/f32-avgpool/9x-minmax-sse-c4.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001401 "src/f32-clamp/gen/sse-x4.c",
1402 "src/f32-clamp/gen/sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001403 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c",
1404 "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c",
Miao Wang400e4042020-04-17 10:15:59 -07001405 "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c",
1406 "src/f32-dwconv/gen/up4x4-minmax-sse.c",
1407 "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c",
1408 "src/f32-dwconv/gen/up4x9-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001409 "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c",
1410 "src/f32-dwconv/gen/up4x25-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001411 "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c",
1412 "src/f32-dwconv/gen/up8x4-minmax-sse.c",
1413 "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c",
1414 "src/f32-dwconv/gen/up8x9-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001415 "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c",
1416 "src/f32-dwconv/gen/up8x25-minmax-sse.c",
1417 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c",
1418 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c",
1419 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c",
1420 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c",
1421 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c",
1422 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c",
1423 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c",
1424 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c",
1425 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c",
1426 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c",
1427 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c",
1428 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c",
1429 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c",
1430 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c",
1431 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c",
1432 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c",
1433 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c",
1434 "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c",
1435 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c",
1436 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c",
1437 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c",
1438 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c",
1439 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c",
1440 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c",
1441 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c",
1442 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c",
1443 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c",
1444 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c",
1445 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c",
1446 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c",
1447 "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c",
1448 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c",
1449 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c",
1450 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c",
1451 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c",
1452 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c",
1453 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c",
1454 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c",
1455 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c",
1456 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c",
1457 "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001458 "src/f32-gavgpool-cw/sse-x4.c",
Miao Wang400e4042020-04-17 10:15:59 -07001459 "src/f32-gavgpool/7p7x-minmax-sse-c4.c",
1460 "src/f32-gavgpool/7x-minmax-sse-c4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001461 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c",
1462 "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c",
1463 "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001464 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-dup.c",
1465 "src/f32-gemm/gen-inc/3x8inc-minmax-sse-load1.c",
1466 "src/f32-gemm/gen-inc/3x8s4inc-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001467 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c",
1468 "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c",
1469 "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001470 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-dup.c",
1471 "src/f32-gemm/gen-inc/5x8inc-minmax-sse-load1.c",
1472 "src/f32-gemm/gen-inc/5x8s4inc-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001473 "src/f32-gemm/gen/1x8-minmax-sse-dup.c",
1474 "src/f32-gemm/gen/1x8-minmax-sse-load1.c",
1475 "src/f32-gemm/gen/1x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001476 "src/f32-gemm/gen/3x8-minmax-sse-dup.c",
1477 "src/f32-gemm/gen/3x8-minmax-sse-load1.c",
1478 "src/f32-gemm/gen/3x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001479 "src/f32-gemm/gen/4x2c4-minmax-sse.c",
1480 "src/f32-gemm/gen/4x8-minmax-sse-dup.c",
1481 "src/f32-gemm/gen/4x8-minmax-sse-load1.c",
1482 "src/f32-gemm/gen/4x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001483 "src/f32-gemm/gen/5x8-minmax-sse-dup.c",
1484 "src/f32-gemm/gen/5x8-minmax-sse-load1.c",
1485 "src/f32-gemm/gen/5x8s4-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001486 "src/f32-hswish/gen/hswish-sse-x4.c",
1487 "src/f32-hswish/gen/hswish-sse-x8.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001488 "src/f32-ibilinear/gen/sse-c4.c",
1489 "src/f32-ibilinear/gen/sse-c8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001490 "src/f32-igemm/gen/1x8-minmax-sse-dup.c",
1491 "src/f32-igemm/gen/1x8-minmax-sse-load1.c",
1492 "src/f32-igemm/gen/1x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001493 "src/f32-igemm/gen/3x8-minmax-sse-dup.c",
1494 "src/f32-igemm/gen/3x8-minmax-sse-load1.c",
1495 "src/f32-igemm/gen/3x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001496 "src/f32-igemm/gen/4x2c4-minmax-sse.c",
1497 "src/f32-igemm/gen/4x8-minmax-sse-dup.c",
1498 "src/f32-igemm/gen/4x8-minmax-sse-load1.c",
1499 "src/f32-igemm/gen/4x8s4-minmax-sse.c",
Miao Wang55abe392021-02-03 14:54:41 -08001500 "src/f32-igemm/gen/5x8-minmax-sse-dup.c",
1501 "src/f32-igemm/gen/5x8-minmax-sse-load1.c",
1502 "src/f32-igemm/gen/5x8s4-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001503 "src/f32-maxpool/9p8x-minmax-sse-c4.c",
1504 "src/f32-pavgpool/9p8x-minmax-sse-c4.c",
1505 "src/f32-pavgpool/9x-minmax-sse-c4.c",
1506 "src/f32-ppmm/gen/4x8-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001507 "src/f32-prelu/gen/sse-2x4.c",
1508 "src/f32-prelu/gen/sse-2x8.c",
1509 "src/f32-relu/gen/sse-x4.c",
1510 "src/f32-relu/gen/sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001511 "src/f32-rmax/sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001512 "src/f32-spmm/gen/4x1-minmax-sse.c",
1513 "src/f32-spmm/gen/8x1-minmax-sse.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001514 "src/f32-spmm/gen/16x1-minmax-sse.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001515 "src/f32-spmm/gen/32x1-minmax-sse.c",
Miao Wang400e4042020-04-17 10:15:59 -07001516 "src/f32-vbinary/gen/vadd-minmax-sse-x4.c",
1517 "src/f32-vbinary/gen/vadd-minmax-sse-x8.c",
1518 "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c",
1519 "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c",
1520 "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c",
1521 "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c",
1522 "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c",
1523 "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001524 "src/f32-vbinary/gen/vmax-sse-x4.c",
1525 "src/f32-vbinary/gen/vmax-sse-x8.c",
1526 "src/f32-vbinary/gen/vmaxc-sse-x4.c",
1527 "src/f32-vbinary/gen/vmaxc-sse-x8.c",
1528 "src/f32-vbinary/gen/vmin-sse-x4.c",
1529 "src/f32-vbinary/gen/vmin-sse-x8.c",
1530 "src/f32-vbinary/gen/vminc-sse-x4.c",
1531 "src/f32-vbinary/gen/vminc-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001532 "src/f32-vbinary/gen/vmul-minmax-sse-x4.c",
1533 "src/f32-vbinary/gen/vmul-minmax-sse-x8.c",
1534 "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c",
1535 "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c",
1536 "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c",
1537 "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c",
1538 "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c",
1539 "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001540 "src/f32-vbinary/gen/vsqrdiff-sse-x4.c",
1541 "src/f32-vbinary/gen/vsqrdiff-sse-x8.c",
1542 "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c",
1543 "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001544 "src/f32-vbinary/gen/vsub-minmax-sse-x4.c",
1545 "src/f32-vbinary/gen/vsub-minmax-sse-x8.c",
1546 "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c",
1547 "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001548 "src/f32-vlrelu/gen/vlrelu-sse-x4.c",
1549 "src/f32-vlrelu/gen/vlrelu-sse-x8.c",
Miao Wang400e4042020-04-17 10:15:59 -07001550 "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c",
1551 "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001552 "src/f32-vsqrt/gen/sse-sqrt-x4.c",
1553 "src/f32-vsqrt/gen/sse-sqrt-x8.c",
1554 "src/f32-vunary/gen/vabs-sse-x4.c",
1555 "src/f32-vunary/gen/vabs-sse-x8.c",
1556 "src/f32-vunary/gen/vneg-sse-x4.c",
1557 "src/f32-vunary/gen/vneg-sse-x8.c",
1558 "src/f32-vunary/gen/vsqr-sse-x4.c",
1559 "src/f32-vunary/gen/vsqr-sse-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001560 "src/math/roundd-sse-addsub.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001561 "src/math/roundne-sse-addsub.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001562 "src/math/roundu-sse-addsub.c",
1563 "src/math/roundz-sse-addsub.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001564 "src/math/sqrt-sse-hh1mac.c",
1565 "src/math/sqrt-sse-nr1mac.c",
1566 "src/math/sqrt-sse-nr2mac.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001567 "src/x32-fill/sse.c",
1568 "src/x32-packx/x4-sse.c",
1569 "src/x32-pad/sse.c",
Miao Wange9993472020-02-10 15:00:10 -08001570]
1571
1572SSE2_UKERNELS = [
Miao Wange9993472020-02-10 15:00:10 -08001573 "src/f32-argmaxpool/4x-sse2-c4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001574 "src/f32-argmaxpool/9p8x-sse2-c4.c",
Miao Wange9993472020-02-10 15:00:10 -08001575 "src/f32-argmaxpool/9x-sse2-c4.c",
Miao Wang55abe392021-02-03 14:54:41 -08001576 "src/f32-gemm/gen-inc/1x8inc-minmax-sse2-dup.c",
1577 "src/f32-gemm/gen-inc/3x8inc-minmax-sse2-dup.c",
1578 "src/f32-gemm/gen-inc/4x8inc-minmax-sse2-dup.c",
1579 "src/f32-gemm/gen-inc/5x8inc-minmax-sse2-dup.c",
1580 "src/f32-gemm/gen/1x8-minmax-sse2-dup.c",
1581 "src/f32-gemm/gen/3x8-minmax-sse2-dup.c",
1582 "src/f32-gemm/gen/4x8-minmax-sse2-dup.c",
1583 "src/f32-gemm/gen/5x8-minmax-sse2-dup.c",
1584 "src/f32-igemm/gen/1x8-minmax-sse2-dup.c",
1585 "src/f32-igemm/gen/3x8-minmax-sse2-dup.c",
1586 "src/f32-igemm/gen/4x8-minmax-sse2-dup.c",
1587 "src/f32-igemm/gen/5x8-minmax-sse2-dup.c",
Miao Wange9993472020-02-10 15:00:10 -08001588 "src/f32-prelu/gen/sse2-2x4.c",
1589 "src/f32-prelu/gen/sse2-2x8.c",
1590 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c",
Miao Wange9993472020-02-10 15:00:10 -08001591 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001592 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c",
Miao Wange9993472020-02-10 15:00:10 -08001593 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c",
1594 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001595 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c",
Miao Wange9993472020-02-10 15:00:10 -08001596 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c",
1597 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001598 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001599 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c",
1600 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001601 "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c",
1602 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c",
1603 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c",
1604 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c",
1605 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c",
1606 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c",
1607 "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001608 "src/f32-sigmoid/gen/sse2-p5-div-x4.c",
1609 "src/f32-sigmoid/gen/sse2-p5-div-x8.c",
1610 "src/f32-sigmoid/gen/sse2-p5-div-x12.c",
1611 "src/f32-sigmoid/gen/sse2-p5-div-x16.c",
1612 "src/f32-sigmoid/gen/sse2-p5-div-x20.c",
1613 "src/f32-sigmoid/gen/sse2-p5-div-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -08001614 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x4.c",
1615 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x8.c",
1616 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x12.c",
1617 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x16.c",
1618 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x20.c",
1619 "src/f32-velu/gen/velu-sse2-rr2-lut16-p3-x24.c",
1620 "src/f32-velu/gen/velu-sse2-rr2-p6-x4.c",
1621 "src/f32-velu/gen/velu-sse2-rr2-p6-x8.c",
1622 "src/f32-velu/gen/velu-sse2-rr2-p6-x12.c",
1623 "src/f32-velu/gen/velu-sse2-rr2-p6-x16.c",
1624 "src/f32-velu/gen/velu-sse2-rr2-p6-x20.c",
1625 "src/f32-velu/gen/velu-sse2-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001626 "src/f32-vlrelu/gen/vlrelu-sse2-x4.c",
1627 "src/f32-vlrelu/gen/vlrelu-sse2-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001628 "src/f32-vrnd/gen/vrndd-sse2-x4.c",
1629 "src/f32-vrnd/gen/vrndd-sse2-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001630 "src/f32-vrnd/gen/vrndne-sse2-x4.c",
1631 "src/f32-vrnd/gen/vrndne-sse2-x8.c",
1632 "src/f32-vrnd/gen/vrndu-sse2-x4.c",
1633 "src/f32-vrnd/gen/vrndu-sse2-x8.c",
1634 "src/f32-vrnd/gen/vrndz-sse2-x4.c",
1635 "src/f32-vrnd/gen/vrndz-sse2-x8.c",
1636 "src/math/exp-sse2-rr2-lut64-p2.c",
1637 "src/math/exp-sse2-rr2-p5.c",
1638 "src/math/expm1minus-sse2-rr2-lut16-p3.c",
1639 "src/math/expm1minus-sse2-rr2-p6.c",
1640 "src/math/expminus-sse2-rr2-p5.c",
1641 "src/math/roundd-sse2-cvt.c",
1642 "src/math/roundne-sse2-cvt.c",
1643 "src/math/roundu-sse2-cvt.c",
1644 "src/math/roundz-sse2-cvt.c",
1645 "src/math/sigmoid-sse2-rr2-lut64-p2-div.c",
1646 "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c",
1647 "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c",
1648 "src/math/sigmoid-sse2-rr2-p5-div.c",
1649 "src/math/sigmoid-sse2-rr2-p5-nr1.c",
1650 "src/math/sigmoid-sse2-rr2-p5-nr2.c",
1651 "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c",
1652 "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c",
1653 "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c",
1654 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c",
1655 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c",
1656 "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c",
1657 "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c",
1658 "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c",
1659 "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c",
1660 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c",
1661 "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c",
1662 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c",
1663 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c",
1664 "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c",
1665 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c",
1666 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c",
1667 "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c",
1668 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c",
1669 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c",
1670 "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c",
1671 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c",
1672 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c",
1673 "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c",
1674 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c",
1675 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c",
1676 "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c",
1677 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c",
1678 "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c",
1679 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c",
1680 "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c",
1681 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c",
1682 "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c",
1683 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c",
1684 "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c",
1685 "src/qs8-requantization/fp32-sse2.c",
1686 "src/qs8-requantization/precise-sse2.c",
1687 "src/qs8-requantization/q31-sse2.c",
1688 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c",
1689 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c",
1690 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c",
1691 "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c",
1692 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
1693 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c",
1694 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c",
1695 "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001696 "src/qu8-avgpool/9p8x-minmax-sse2-c8.c",
1697 "src/qu8-avgpool/9x-minmax-sse2-c8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001698 "src/qu8-dwconv/up8x9-minmax-sse2.c",
1699 "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c",
1700 "src/qu8-gavgpool/7x-minmax-sse2-c8.c",
1701 "src/qu8-gemm/2x4c8-minmax-sse2.c",
1702 "src/qu8-gemm/4x4c2-minmax-sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001703 "src/qu8-igemm/4x4c2-minmax-sse2.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001704 "src/qu8-requantization/fp32-sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001705 "src/qu8-requantization/precise-sse2.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001706 "src/qu8-requantization/q31-sse2.c",
1707 "src/qu8-vadd/minmax-sse2.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001708 "src/u8-clamp/sse2-x64.c",
Miao Wang400e4042020-04-17 10:15:59 -07001709 "src/u8-maxpool/9p8x-minmax-sse2-c16.c",
Miao Wange9993472020-02-10 15:00:10 -08001710 "src/u8-rmax/sse2.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001711 "src/x8-zip/x2-sse2.c",
1712 "src/x8-zip/x3-sse2.c",
1713 "src/x8-zip/x4-sse2.c",
1714 "src/x8-zip/xm-sse2.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001715 "src/x32-unpool/sse2.c",
Miao Wange9993472020-02-10 15:00:10 -08001716 "src/x32-zip/x2-sse2.c",
1717 "src/x32-zip/x3-sse2.c",
1718 "src/x32-zip/x4-sse2.c",
1719 "src/x32-zip/xm-sse2.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001720]
1721
1722SSSE3_UKERNELS = [
Miao Wang5eea8312020-12-07 09:12:40 -08001723 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c",
1724 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c",
1725 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c",
1726 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c",
1727 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c",
1728 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c",
1729 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c",
1730 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c",
1731 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c",
1732 "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c",
1733 "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c",
1734 "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c",
1735 "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c",
1736 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c",
1737 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c",
1738 "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c",
1739 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c",
1740 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c",
1741 "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c",
1742 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c",
1743 "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c",
1744 "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c",
1745 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c",
1746 "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c",
1747 "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c",
1748 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c",
1749 "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c",
1750 "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c",
1751 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c",
1752 "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c",
1753 "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c",
1754 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c",
1755 "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c",
1756 "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c",
1757 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c",
1758 "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c",
1759 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c",
1760 "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c",
1761 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c",
1762 "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c",
1763 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c",
1764 "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c",
1765 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c",
1766 "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c",
1767 "src/qs8-requantization/precise-ssse3.c",
1768 "src/qs8-requantization/q31-ssse3.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001769 "src/qu8-requantization/precise-ssse3.c",
1770 "src/qu8-requantization/q31-ssse3.c",
Miao Wange9993472020-02-10 15:00:10 -08001771]
1772
1773SSE41_UKERNELS = [
1774 "src/f32-prelu/gen/sse41-2x4.c",
1775 "src/f32-prelu/gen/sse41-2x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001776 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c",
1777 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c",
1778 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c",
1779 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c",
1780 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c",
1781 "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c",
Miao Wange9993472020-02-10 15:00:10 -08001782 "src/f32-sigmoid/gen/sse41-p5-div-x4.c",
1783 "src/f32-sigmoid/gen/sse41-p5-div-x8.c",
1784 "src/f32-sigmoid/gen/sse41-p5-div-x12.c",
1785 "src/f32-sigmoid/gen/sse41-p5-div-x16.c",
1786 "src/f32-sigmoid/gen/sse41-p5-div-x20.c",
1787 "src/f32-sigmoid/gen/sse41-p5-div-x24.c",
Miao Wang55abe392021-02-03 14:54:41 -08001788 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x4.c",
1789 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x8.c",
1790 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x12.c",
1791 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x16.c",
1792 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x20.c",
1793 "src/f32-velu/gen/velu-sse41-rr2-lut16-p3-x24.c",
1794 "src/f32-velu/gen/velu-sse41-rr2-p6-x4.c",
1795 "src/f32-velu/gen/velu-sse41-rr2-p6-x8.c",
1796 "src/f32-velu/gen/velu-sse41-rr2-p6-x12.c",
1797 "src/f32-velu/gen/velu-sse41-rr2-p6-x16.c",
1798 "src/f32-velu/gen/velu-sse41-rr2-p6-x20.c",
1799 "src/f32-velu/gen/velu-sse41-rr2-p6-x24.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001800 "src/f32-vlrelu/gen/vlrelu-sse41-x4.c",
1801 "src/f32-vlrelu/gen/vlrelu-sse41-x8.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001802 "src/f32-vrnd/gen/vrndd-sse41-x4.c",
1803 "src/f32-vrnd/gen/vrndd-sse41-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001804 "src/f32-vrnd/gen/vrndne-sse41-x4.c",
1805 "src/f32-vrnd/gen/vrndne-sse41-x8.c",
1806 "src/f32-vrnd/gen/vrndu-sse41-x4.c",
1807 "src/f32-vrnd/gen/vrndu-sse41-x8.c",
1808 "src/f32-vrnd/gen/vrndz-sse41-x4.c",
1809 "src/f32-vrnd/gen/vrndz-sse41-x8.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001810 "src/math/roundd-sse41.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001811 "src/math/roundne-sse41.c",
Miao Wangc0aa11a2020-06-10 13:41:26 -07001812 "src/math/roundu-sse41.c",
1813 "src/math/roundz-sse41.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001814 "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c",
1815 "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c",
1816 "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c",
1817 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c",
1818 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c",
1819 "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c",
1820 "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c",
1821 "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c",
1822 "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c",
1823 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c",
1824 "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c",
1825 "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c",
1826 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c",
1827 "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c",
1828 "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c",
1829 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c",
1830 "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c",
1831 "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c",
1832 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c",
1833 "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c",
1834 "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c",
1835 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c",
1836 "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c",
1837 "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c",
1838 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c",
1839 "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c",
1840 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c",
1841 "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c",
1842 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c",
1843 "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c",
1844 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c",
1845 "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c",
1846 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c",
1847 "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c",
1848 "src/qs8-requantization/fp32-sse4.c",
1849 "src/qs8-requantization/precise-sse4.c",
1850 "src/qs8-requantization/q31-sse4.c",
1851 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c",
1852 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c",
1853 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c",
1854 "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c",
1855 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c",
1856 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c",
1857 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c",
1858 "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c",
1859 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
1860 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c",
1861 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c",
1862 "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c",
1863 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c",
1864 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c",
1865 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c",
1866 "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c",
1867 "src/qu8-requantization/precise-sse4.c",
1868 "src/qu8-requantization/q31-sse4.c",
Miao Wange9993472020-02-10 15:00:10 -08001869]
1870
1871AVX_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -07001872 "src/f32-clamp/gen/avx-x8.c",
1873 "src/f32-clamp/gen/avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001874 "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c",
1875 "src/f32-dwconv/gen/up8x4-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001876 "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c",
1877 "src/f32-dwconv/gen/up8x9-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001878 "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c",
1879 "src/f32-dwconv/gen/up8x25-minmax-avx.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001880 "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c",
1881 "src/f32-dwconv/gen/up16x4-minmax-avx.c",
1882 "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c",
1883 "src/f32-dwconv/gen/up16x9-minmax-avx.c",
1884 "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c",
1885 "src/f32-dwconv/gen/up16x25-minmax-avx.c",
Miao Wang400e4042020-04-17 10:15:59 -07001886 "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001887 "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c",
1888 "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001889 "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001890 "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001891 "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001892 "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001893 "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c",
1894 "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c",
1895 "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c",
1896 "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c",
1897 "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c",
1898 "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c",
1899 "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c",
1900 "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c",
1901 "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c",
1902 "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c",
1903 "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001904 "src/f32-hswish/gen/hswish-avx-x8.c",
1905 "src/f32-hswish/gen/hswish-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001906 "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001907 "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c",
1908 "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001909 "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001910 "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001911 "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07001912 "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001913 "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c",
1914 "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07001915 "src/f32-prelu/gen/avx-2x8.c",
1916 "src/f32-prelu/gen/avx-2x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001917 "src/f32-relu/gen/avx-x8.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001918 "src/f32-relu/gen/avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001919 "src/f32-rmax/avx.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001920 "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c",
1921 "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c",
1922 "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c",
1923 "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c",
1924 "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c",
1925 "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c",
1926 "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c",
1927 "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c",
1928 "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c",
1929 "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c",
1930 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c",
1931 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c",
1932 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c",
1933 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c",
1934 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c",
1935 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c",
1936 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c",
1937 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c",
1938 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c",
1939 "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c",
Miao Wang400e4042020-04-17 10:15:59 -07001940 "src/f32-vbinary/gen/vadd-minmax-avx-x8.c",
1941 "src/f32-vbinary/gen/vadd-minmax-avx-x16.c",
1942 "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c",
1943 "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c",
1944 "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c",
1945 "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c",
1946 "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c",
1947 "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c",
Miao Wange9993472020-02-10 15:00:10 -08001948 "src/f32-vbinary/gen/vmax-avx-x8.c",
1949 "src/f32-vbinary/gen/vmax-avx-x16.c",
1950 "src/f32-vbinary/gen/vmaxc-avx-x8.c",
1951 "src/f32-vbinary/gen/vmaxc-avx-x16.c",
1952 "src/f32-vbinary/gen/vmin-avx-x8.c",
1953 "src/f32-vbinary/gen/vmin-avx-x16.c",
1954 "src/f32-vbinary/gen/vminc-avx-x8.c",
1955 "src/f32-vbinary/gen/vminc-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001956 "src/f32-vbinary/gen/vmul-minmax-avx-x8.c",
1957 "src/f32-vbinary/gen/vmul-minmax-avx-x16.c",
1958 "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c",
1959 "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c",
1960 "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c",
1961 "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c",
1962 "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c",
1963 "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001964 "src/f32-vbinary/gen/vsqrdiff-avx-x8.c",
1965 "src/f32-vbinary/gen/vsqrdiff-avx-x16.c",
1966 "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c",
1967 "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07001968 "src/f32-vbinary/gen/vsub-minmax-avx-x8.c",
1969 "src/f32-vbinary/gen/vsub-minmax-avx-x16.c",
1970 "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c",
1971 "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c",
Miao Wang55abe392021-02-03 14:54:41 -08001972 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x8.c",
1973 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x16.c",
1974 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x24.c",
1975 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x32.c",
1976 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x40.c",
1977 "src/f32-velu/gen/velu-avx-rr2-lut4-p4-perm-x48.c",
1978 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x8.c",
1979 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x16.c",
1980 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x24.c",
1981 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x32.c",
1982 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x40.c",
1983 "src/f32-velu/gen/velu-avx-rr2-lut16-p3-x48.c",
1984 "src/f32-velu/gen/velu-avx-rr2-p6-x8.c",
1985 "src/f32-velu/gen/velu-avx-rr2-p6-x16.c",
1986 "src/f32-velu/gen/velu-avx-rr2-p6-x24.c",
1987 "src/f32-velu/gen/velu-avx-rr2-p6-x32.c",
1988 "src/f32-velu/gen/velu-avx-rr2-p6-x40.c",
1989 "src/f32-velu/gen/velu-avx-rr2-p6-x48.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001990 "src/f32-vlrelu/gen/vlrelu-avx-x8.c",
1991 "src/f32-vlrelu/gen/vlrelu-avx-x16.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07001992 "src/f32-vrnd/gen/vrndd-avx-x8.c",
1993 "src/f32-vrnd/gen/vrndd-avx-x16.c",
Miao Wang5eea8312020-12-07 09:12:40 -08001994 "src/f32-vrnd/gen/vrndne-avx-x8.c",
1995 "src/f32-vrnd/gen/vrndne-avx-x16.c",
1996 "src/f32-vrnd/gen/vrndu-avx-x8.c",
1997 "src/f32-vrnd/gen/vrndu-avx-x16.c",
1998 "src/f32-vrnd/gen/vrndz-avx-x8.c",
1999 "src/f32-vrnd/gen/vrndz-avx-x16.c",
2000 "src/f32-vscale/avx-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002001 "src/f32-vsqrt/gen/avx-sqrt-x8.c",
2002 "src/f32-vsqrt/gen/avx-sqrt-x16.c",
2003 "src/f32-vunary/gen/vabs-avx-x8.c",
2004 "src/f32-vunary/gen/vabs-avx-x16.c",
2005 "src/f32-vunary/gen/vneg-avx-x8.c",
2006 "src/f32-vunary/gen/vneg-avx-x16.c",
2007 "src/f32-vunary/gen/vsqr-avx-x8.c",
2008 "src/f32-vunary/gen/vsqr-avx-x16.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002009 "src/math/exp-avx-rr2-p5.c",
2010 "src/math/expm1minus-avx-rr2-lut4-p4-perm.c",
2011 "src/math/expm1minus-avx-rr2-lut16-p3.c",
2012 "src/math/expm1minus-avx-rr2-p6.c",
2013 "src/math/sigmoid-avx-rr2-lut64-p2-div.c",
2014 "src/math/sigmoid-avx-rr2-p5-div.c",
2015 "src/math/sigmoid-avx-rr2-p5-nr1.c",
2016 "src/math/sigmoid-avx-rr2-p5-nr2.c",
2017]
2018
2019XOP_UKERNELS = [
2020 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c",
2021 "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c",
2022 "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c",
2023 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c",
2024 "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c",
2025 "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c",
2026 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c",
2027 "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c",
2028 "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c",
2029 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c",
2030 "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c",
2031 "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c",
2032 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c",
2033 "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c",
2034 "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c",
2035 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c",
2036 "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c",
2037 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c",
2038 "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c",
2039 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c",
2040 "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c",
2041 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c",
2042 "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c",
2043 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c",
2044 "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c",
2045 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c",
2046 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c",
2047 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c",
2048 "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c",
2049 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c",
2050 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c",
2051 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c",
2052 "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002053]
2054
2055FMA3_UKERNELS = [
Miao Wang400e4042020-04-17 10:15:59 -07002056 "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c",
2057 "src/f32-dwconv/gen/up8x4-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07002058 "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c",
2059 "src/f32-dwconv/gen/up8x9-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07002060 "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c",
2061 "src/f32-dwconv/gen/up8x25-minmax-fma3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002062 "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c",
2063 "src/f32-dwconv/gen/up16x4-minmax-fma3.c",
2064 "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c",
2065 "src/f32-dwconv/gen/up16x9-minmax-fma3.c",
2066 "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c",
2067 "src/f32-dwconv/gen/up16x25-minmax-fma3.c",
Miao Wang400e4042020-04-17 10:15:59 -07002068 "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002069 "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c",
2070 "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c",
2071 "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c",
2072 "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002073 "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002074 "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c",
2075 "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002076 "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002077 "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c",
2078 "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002079 "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c",
2080 "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c",
2081 "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002082 "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c",
2083 "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c",
2084 "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c",
2085 "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c",
2086 "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c",
2087 "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c",
2088 "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c",
2089 "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c",
2090 "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c",
2091 "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c",
2092 "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c",
2093 "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c",
2094 "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c",
2095 "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002096 "src/f32-hswish/gen/hswish-fma3-x8.c",
2097 "src/f32-hswish/gen/hswish-fma3-x16.c",
Miao Wang400e4042020-04-17 10:15:59 -07002098 "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002099 "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c",
2100 "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c",
2101 "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c",
2102 "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002103 "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002104 "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c",
2105 "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002106 "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002107 "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c",
2108 "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c",
Miao Wang400e4042020-04-17 10:15:59 -07002109 "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c",
2110 "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c",
2111 "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002112 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c",
2113 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c",
2114 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c",
2115 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c",
2116 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c",
2117 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c",
2118 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c",
2119 "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c",
2120 "src/math/sqrt-fma3-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002121 "src/math/sqrt-fma3-nr1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002122 "src/math/sqrt-fma3-nr2fma.c",
Miao Wange9993472020-02-10 15:00:10 -08002123]
2124
2125AVX2_UKERNELS = [
Miao Wange9993472020-02-10 15:00:10 -08002126 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c",
2127 "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002128 "src/f32-raddexpminusmax/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002129 "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002130 "src/f32-raddexpminusmax/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002131 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c",
2132 "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002133 "src/f32-raddexpminusmax/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002134 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c",
2135 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c",
2136 "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002137 "src/f32-raddexpminusmax/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002138 "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c",
2139 "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002140 "src/f32-raddextexp/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002141 "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002142 "src/f32-raddextexp/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002143 "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c",
2144 "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002145 "src/f32-raddextexp/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002146 "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c",
2147 "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c",
2148 "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002149 "src/f32-raddextexp/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002150 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c",
2151 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002152 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002153 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002154 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c",
Miao Wange9993472020-02-10 15:00:10 -08002155 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c",
2156 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002157 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002158 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c",
2159 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c",
2160 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002161 "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c",
Miao Wange9993472020-02-10 15:00:10 -08002162 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c",
2163 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c",
2164 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c",
2165 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c",
2166 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c",
2167 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c",
2168 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c",
2169 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c",
2170 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c",
2171 "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c",
2172 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c",
2173 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c",
2174 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c",
2175 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c",
2176 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c",
2177 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c",
2178 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c",
2179 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c",
2180 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c",
2181 "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c",
2182 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c",
2183 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c",
2184 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c",
2185 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c",
2186 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c",
2187 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c",
2188 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c",
2189 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c",
2190 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c",
2191 "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c",
Miao Wang55abe392021-02-03 14:54:41 -08002192 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x8.c",
2193 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x16.c",
2194 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x24.c",
2195 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x32.c",
2196 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x40.c",
2197 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x48.c",
2198 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x56.c",
2199 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x64.c",
2200 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x72.c",
2201 "src/f32-velu/gen/velu-avx2-rr1-lut4-p4-perm-x80.c",
2202 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x8.c",
2203 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x16.c",
2204 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x24.c",
2205 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x32.c",
2206 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x40.c",
2207 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x48.c",
2208 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x56.c",
2209 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x64.c",
2210 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x72.c",
2211 "src/f32-velu/gen/velu-avx2-rr1-lut8-p4-perm-x80.c",
2212 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x8.c",
2213 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x16.c",
2214 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x24.c",
2215 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x32.c",
2216 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x40.c",
2217 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x48.c",
2218 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x56.c",
2219 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x64.c",
2220 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x72.c",
2221 "src/f32-velu/gen/velu-avx2-rr1-lut16-p3-gather-x80.c",
2222 "src/f32-velu/gen/velu-avx2-rr1-p6-x8.c",
2223 "src/f32-velu/gen/velu-avx2-rr1-p6-x16.c",
2224 "src/f32-velu/gen/velu-avx2-rr1-p6-x24.c",
2225 "src/f32-velu/gen/velu-avx2-rr1-p6-x32.c",
2226 "src/f32-velu/gen/velu-avx2-rr1-p6-x40.c",
2227 "src/f32-velu/gen/velu-avx2-rr1-p6-x48.c",
2228 "src/f32-velu/gen/velu-avx2-rr1-p6-x56.c",
2229 "src/f32-velu/gen/velu-avx2-rr1-p6-x64.c",
2230 "src/f32-velu/gen/velu-avx2-rr1-p6-x72.c",
2231 "src/f32-velu/gen/velu-avx2-rr1-p6-x80.c",
Miao Wange9993472020-02-10 15:00:10 -08002232 "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c",
2233 "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c",
2234 "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c",
2235 "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c",
2236 "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c",
2237 "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c",
2238 "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c",
2239 "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c",
2240 "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c",
2241 "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c",
2242 "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c",
2243 "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c",
2244 "src/f32-vscaleextexp/gen/avx2-p5-x8.c",
2245 "src/f32-vscaleextexp/gen/avx2-p5-x16.c",
2246 "src/f32-vscaleextexp/gen/avx2-p5-x24.c",
2247 "src/f32-vscaleextexp/gen/avx2-p5-x32.c",
2248 "src/f32-vscaleextexp/gen/avx2-p5-x40.c",
2249 "src/f32-vscaleextexp/gen/avx2-p5-x48.c",
2250 "src/f32-vscaleextexp/gen/avx2-p5-x56.c",
2251 "src/f32-vscaleextexp/gen/avx2-p5-x64.c",
2252 "src/f32-vscaleextexp/gen/avx2-p5-x72.c",
2253 "src/f32-vscaleextexp/gen/avx2-p5-x80.c",
2254 "src/f32-vscaleextexp/gen/avx2-p5-x88.c",
2255 "src/f32-vscaleextexp/gen/avx2-p5-x96.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002256 "src/math/exp-avx2-rr2-lut8-p3-perm.c",
2257 "src/math/exp-avx2-rr2-lut8-p4-perm.c",
2258 "src/math/exp-avx2-rr2-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002259 "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c",
2260 "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c",
2261 "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c",
2262 "src/math/expm1minus-avx2-rr1-p6.c",
Miao Wang55abe392021-02-03 14:54:41 -08002263 "src/math/expminus-avx2-rr2-p5.c",
Miao Wange9993472020-02-10 15:00:10 -08002264 "src/math/extexp-avx2-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002265 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c",
2266 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c",
2267 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c",
2268 "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c",
Miao Wange9993472020-02-10 15:00:10 -08002269 "src/math/sigmoid-avx2-rr1-p5-div.c",
Miao Wange9993472020-02-10 15:00:10 -08002270 "src/math/sigmoid-avx2-rr1-p5-nr1fma.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002271 "src/math/sigmoid-avx2-rr1-p5-nr2fma.c",
2272 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c",
2273 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c",
2274 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c",
2275 "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c",
2276 "src/math/sigmoid-avx2-rr2-p5-div.c",
2277 "src/math/sigmoid-avx2-rr2-p5-nr1fma.c",
2278 "src/math/sigmoid-avx2-rr2-p5-nr2fma.c",
2279 "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c",
2280 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c",
2281 "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c",
2282 "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c",
2283 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c",
2284 "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c",
2285 "src/qs8-gemm/gen/1x8c8-minmax-avx2.c",
2286 "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c",
2287 "src/qs8-gemm/gen/2x8c8-minmax-avx2.c",
2288 "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c",
2289 "src/qs8-gemm/gen/3x8c8-minmax-avx2.c",
2290 "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c",
2291 "src/qs8-igemm/gen/1x8c8-minmax-avx2.c",
2292 "src/qs8-igemm/gen/2x8c8-minmax-avx2.c",
2293 "src/qs8-igemm/gen/3x8c8-minmax-avx2.c",
2294 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c",
2295 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c",
2296 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c",
2297 "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c",
2298 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c",
2299 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c",
2300 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c",
2301 "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002302]
2303
2304AVX512F_UKERNELS = [
Miao Wang2534c2f2020-03-16 11:58:04 -07002305 "src/f32-clamp/gen/avx512f-x16.c",
2306 "src/f32-clamp/gen/avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002307 "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c",
2308 "src/f32-dwconv/gen/up16x4-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002309 "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c",
2310 "src/f32-dwconv/gen/up16x9-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002311 "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c",
2312 "src/f32-dwconv/gen/up16x25-minmax-avx512f.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002313 "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c",
2314 "src/f32-dwconv/gen/up32x4-minmax-avx512f.c",
2315 "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c",
2316 "src/f32-dwconv/gen/up32x9-minmax-avx512f.c",
2317 "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c",
2318 "src/f32-dwconv/gen/up32x25-minmax-avx512f.c",
Miao Wang400e4042020-04-17 10:15:59 -07002319 "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c",
2320 "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c",
2321 "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c",
2322 "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c",
2323 "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c",
2324 "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002325 "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c",
2326 "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c",
2327 "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c",
2328 "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c",
2329 "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c",
2330 "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002331 "src/f32-hswish/gen/hswish-avx512f-x16.c",
2332 "src/f32-hswish/gen/hswish-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002333 "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c",
2334 "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c",
2335 "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c",
2336 "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c",
2337 "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c",
2338 "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c",
Miao Wang2534c2f2020-03-16 11:58:04 -07002339 "src/f32-prelu/gen/avx512f-2x16.c",
2340 "src/f32-prelu/gen/avx512f-2x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002341 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2342 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002343 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002344 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002345 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002346 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2347 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002348 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002349 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2350 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2351 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002352 "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c",
Miao Wange9993472020-02-10 15:00:10 -08002353 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c",
2354 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002355 "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002356 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002357 "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002358 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c",
2359 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002360 "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002361 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c",
2362 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c",
2363 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002364 "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c",
Miao Wange9993472020-02-10 15:00:10 -08002365 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c",
2366 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002367 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c",
Miao Wange9993472020-02-10 15:00:10 -08002368 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002369 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c",
Miao Wange9993472020-02-10 15:00:10 -08002370 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c",
2371 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002372 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c",
Miao Wange9993472020-02-10 15:00:10 -08002373 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c",
2374 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c",
2375 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002376 "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002377 "src/f32-relu/gen/avx512f-x16.c",
2378 "src/f32-relu/gen/avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002379 "src/f32-rmax/avx512f.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002380 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c",
2381 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c",
2382 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c",
2383 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c",
2384 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c",
2385 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c",
2386 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c",
2387 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c",
2388 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c",
2389 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c",
2390 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c",
2391 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c",
2392 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c",
2393 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c",
2394 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c",
2395 "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c",
2396 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c",
2397 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c",
2398 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c",
2399 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c",
2400 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c",
2401 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c",
2402 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c",
2403 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c",
2404 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c",
2405 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c",
2406 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c",
2407 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c",
2408 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c",
2409 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c",
2410 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c",
2411 "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c",
2412 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c",
2413 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c",
2414 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c",
2415 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c",
2416 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c",
2417 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c",
2418 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c",
2419 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c",
2420 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c",
2421 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c",
2422 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c",
2423 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c",
2424 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c",
2425 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c",
2426 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c",
2427 "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c",
Miao Wang400e4042020-04-17 10:15:59 -07002428 "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c",
2429 "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c",
2430 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c",
2431 "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c",
2432 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c",
2433 "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c",
2434 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c",
2435 "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c",
Miao Wange9993472020-02-10 15:00:10 -08002436 "src/f32-vbinary/gen/vmax-avx512f-x16.c",
2437 "src/f32-vbinary/gen/vmax-avx512f-x32.c",
2438 "src/f32-vbinary/gen/vmaxc-avx512f-x16.c",
2439 "src/f32-vbinary/gen/vmaxc-avx512f-x32.c",
2440 "src/f32-vbinary/gen/vmin-avx512f-x16.c",
2441 "src/f32-vbinary/gen/vmin-avx512f-x32.c",
2442 "src/f32-vbinary/gen/vminc-avx512f-x16.c",
2443 "src/f32-vbinary/gen/vminc-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002444 "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c",
2445 "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c",
2446 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c",
2447 "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c",
2448 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c",
2449 "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c",
2450 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c",
2451 "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002452 "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c",
2453 "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c",
2454 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c",
2455 "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c",
Miao Wang400e4042020-04-17 10:15:59 -07002456 "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c",
2457 "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c",
2458 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c",
2459 "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c",
Miao Wang55abe392021-02-03 14:54:41 -08002460 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x16.c",
2461 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x32.c",
2462 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x48.c",
2463 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x64.c",
2464 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x80.c",
2465 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x96.c",
2466 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x112.c",
2467 "src/f32-velu/gen/velu-avx512f-rr1-lut16-p3-perm-x128.c",
2468 "src/f32-velu/gen/velu-avx512f-rr1-p6-x16.c",
2469 "src/f32-velu/gen/velu-avx512f-rr1-p6-x32.c",
2470 "src/f32-velu/gen/velu-avx512f-rr1-p6-x48.c",
2471 "src/f32-velu/gen/velu-avx512f-rr1-p6-x64.c",
2472 "src/f32-velu/gen/velu-avx512f-rr1-p6-x80.c",
2473 "src/f32-velu/gen/velu-avx512f-rr1-p6-x96.c",
2474 "src/f32-velu/gen/velu-avx512f-rr1-p6-x112.c",
2475 "src/f32-velu/gen/velu-avx512f-rr1-p6-x128.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002476 "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c",
2477 "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002478 "src/f32-vrnd/gen/vrndd-avx512f-x16.c",
2479 "src/f32-vrnd/gen/vrndd-avx512f-x32.c",
2480 "src/f32-vrnd/gen/vrndne-avx512f-x16.c",
2481 "src/f32-vrnd/gen/vrndne-avx512f-x32.c",
2482 "src/f32-vrnd/gen/vrndu-avx512f-x16.c",
2483 "src/f32-vrnd/gen/vrndu-avx512f-x32.c",
2484 "src/f32-vrnd/gen/vrndz-avx512f-x16.c",
2485 "src/f32-vrnd/gen/vrndz-avx512f-x32.c",
2486 "src/f32-vscale/avx512f-x64.c",
Miao Wange9993472020-02-10 15:00:10 -08002487 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c",
2488 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c",
2489 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c",
2490 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c",
2491 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c",
2492 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c",
2493 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c",
2494 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c",
2495 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c",
2496 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c",
2497 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c",
2498 "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c",
2499 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c",
2500 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c",
2501 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c",
2502 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c",
2503 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c",
2504 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c",
2505 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c",
2506 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c",
2507 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c",
2508 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c",
2509 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c",
2510 "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002511 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c",
2512 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c",
2513 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c",
2514 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c",
2515 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c",
2516 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c",
2517 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c",
2518 "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c",
2519 "src/f32-vunary/gen/vabs-avx512f-x16.c",
2520 "src/f32-vunary/gen/vabs-avx512f-x32.c",
2521 "src/f32-vunary/gen/vneg-avx512f-x16.c",
2522 "src/f32-vunary/gen/vneg-avx512f-x32.c",
2523 "src/f32-vunary/gen/vsqr-avx512f-x16.c",
2524 "src/f32-vunary/gen/vsqr-avx512f-x32.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002525 "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c",
2526 "src/math/exp-avx512f-rr2-lut16-p3-perm.c",
2527 "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c",
2528 "src/math/exp-avx512f-rr2-lut32-p2-perm2.c",
2529 "src/math/exp-avx512f-rr2-p5-scalef.c",
2530 "src/math/exp-avx512f-rr2-p5.c",
2531 "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c",
2532 "src/math/expm1minus-avx512f-rr1-p6.c",
Miao Wange9993472020-02-10 15:00:10 -08002533 "src/math/extexp-avx512f-p5.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002534 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c",
2535 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c",
2536 "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c",
2537 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c",
2538 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c",
2539 "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c",
2540 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c",
2541 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c",
2542 "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c",
2543 "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c",
2544 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c",
2545 "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c",
2546 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c",
2547 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c",
2548 "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c",
2549 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c",
2550 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c",
2551 "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c",
2552 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c",
2553 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c",
2554 "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c",
2555 "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c",
2556 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c",
2557 "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002558 "src/math/sqrt-avx512f-nr1fma.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002559 "src/math/sqrt-avx512f-nr1fma1adj.c",
Miao Wang5eea8312020-12-07 09:12:40 -08002560 "src/math/sqrt-avx512f-nr2fma.c",
2561]
2562
2563AVX512SKX_UKERNELS = [
2564 "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c",
2565 "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c",
2566 "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c",
2567 "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c",
2568 "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c",
2569 "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c",
2570 "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c",
2571 "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c",
2572 "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c",
2573 "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c",
Miao Wange9993472020-02-10 15:00:10 -08002574]
2575
2576AARCH32_ASM_UKERNELS = [
Miao Wang86f5fbe2020-07-24 11:16:10 -07002577 "src/f32-gemm/4x4-aarch32-vfp-ld64.S",
2578 "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002579 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2580 "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002581 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
Miao Wang400e4042020-04-17 10:15:59 -07002582 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002583 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002584 "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002585 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S",
2586 "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002587 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S",
2588 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S",
2589 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S",
2590 "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S",
Miao Wange9993472020-02-10 15:00:10 -08002591]
2592
2593AARCH64_ASM_UKERNELS = [
Miao Wangc0aa11a2020-06-10 13:41:26 -07002594 "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002595 "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002596 "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002597 "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002598 "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002599 "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002600 "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002601 "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S",
2602 "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S",
2603 "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S",
2604 "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S",
2605 "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S",
2606 "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S",
2607 "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S",
Miao Wang400e4042020-04-17 10:15:59 -07002608 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S",
2609 "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S",
Miao Wang400e4042020-04-17 10:15:59 -07002610 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2611 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2612 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002613 "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S",
2614 "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002615 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2616 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S",
2617 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2618 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002619 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002620 "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S",
2621 "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002622 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S",
2623 "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S",
2624 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S",
2625 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S",
Miao Wang400e4042020-04-17 10:15:59 -07002626 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002627 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S",
Miao Wang400e4042020-04-17 10:15:59 -07002628 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002629 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002630 "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S",
2631 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S",
2632 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2633 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
2634 "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S",
2635 "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S",
2636 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2637 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S",
2638 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2639 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2640 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S",
2641 "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S",
2642 "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2643 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2644 "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
2645 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2646 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2647 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2648 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2649 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
2650 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S",
2651 "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S",
Miao Wang400e4042020-04-17 10:15:59 -07002652 "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002653 "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S",
Miao Wang400e4042020-04-17 10:15:59 -07002654 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S",
2655 "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002656 "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S",
2657 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S",
2658 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S",
2659 "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S",
2660 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S",
2661 "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002662 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S",
2663 "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S",
2664 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S",
2665 "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang400e4042020-04-17 10:15:59 -07002666 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S",
2667 "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S",
Miao Wang338e80e2021-03-23 08:56:09 -07002668 "src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002669 "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S",
Miao Wang338e80e2021-03-23 08:56:09 -07002670 "src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S",
2671 "src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S",
2672 "src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002673 "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S",
Miao Wang338e80e2021-03-23 08:56:09 -07002674 "src/qs8-gemm/4x16c4-aarch64-neondot-ld32.S",
Miao Wang5eea8312020-12-07 09:12:40 -08002675 "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S",
Miao Wang338e80e2021-03-23 08:56:09 -07002676 "src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S",
2677 "src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S",
2678 "src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S",
2679 "src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S",
Miao Wange9993472020-02-10 15:00:10 -08002680]
2681
2682cc_defaults {
2683 name: "xnnpack_internal_default",
2684 vendor_available: true,
2685 sdk_version: "current",
2686 local_include_dirs: [
2687 "include",
2688 "src",
2689 ],
2690 cflags: [
2691 "-std=c99",
2692 "-DXNN_LOG_LEVEL=2",
2693 "-Wno-unused-parameter",
2694 "-Wno-missing-field-initializers",
2695 "-Wno-pointer-arith",
Miao Wangef75bc92020-07-25 09:15:12 -07002696 "-Wno-implicit-function-declaration",
Miao Wange9993472020-02-10 15:00:10 -08002697 ],
2698 stl: "libc++_static",
2699}
2700
2701cc_library_static {
2702 name: "xnnpack_tables",
2703 defaults: ["xnnpack_internal_default"],
2704 srcs: TABLE_SRCS,
2705}
2706
2707cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07002708 name: "xnnpack_logging_utils",
2709 defaults: ["xnnpack_internal_default"],
2710 srcs: LOGGING_SRCS,
2711 header_libs: [
2712 "fp16_headers",
2713 ],
2714 static_libs: [
2715 "libclog",
2716 "libpthreadpool",
2717 ],
2718}
2719
2720cc_library_static {
2721 name: "xnnpack_memory_planner",
2722 defaults: ["xnnpack_internal_default"],
2723 srcs: [
2724 "src/memory-planner.c",
2725 ],
2726 cflags: [
2727 "-DXNN_ENABLE_MEMOPT=1",
2728 ],
2729 header_libs: [
2730 "fp16_headers",
2731 ],
2732 static_libs: [
2733 "libpthreadpool",
2734 "xnnpack_logging_utils",
2735 ],
2736}
2737
2738cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002739 name: "xnnpack_im2col",
2740 defaults: ["xnnpack_internal_default"],
2741 srcs: [
2742 "src/im2col.c",
2743 ],
2744}
2745
2746cc_library_static {
2747 name: "xnnpack_indirection",
2748 defaults: ["xnnpack_internal_default"],
2749 srcs: [
2750 "src/indirection.c",
2751 ],
2752 header_libs: [
2753 "fp16_headers",
2754 "fxdiv_headers",
2755 ],
2756 static_libs: [
2757 "libpthreadpool",
2758 ],
2759}
2760
2761cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07002762 name: "xnnpack_packing",
2763 defaults: ["xnnpack_internal_default"],
2764 srcs: [
2765 "src/packing.c",
2766 ],
2767 header_libs: [
2768 "fp16_headers",
2769 "fxdiv_headers",
2770 ],
2771 static_libs: [
2772 "libpthreadpool",
2773 ],
2774}
2775
2776cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002777 name: "xnnpack_operator_run",
2778 defaults: ["xnnpack_internal_default"],
2779 srcs: [
2780 "src/operator-run.c",
2781 ],
2782 cflags: [
2783 "-Wno-vla",
2784 ],
2785 header_libs: [
2786 "fp16_headers",
2787 "fxdiv_headers",
2788 ],
2789 static_libs: [
2790 "libclog",
2791 "libpthreadpool",
2792 ],
2793}
2794
2795cc_library_static {
2796 name: "xnnpack_operators",
2797 defaults: ["xnnpack_internal_default"],
2798 srcs: OPERATOR_SRCS + [
2799 "src/memory.c",
2800 "src/operator-delete.c",
2801 ],
2802 header_libs: [
2803 "fp16_headers",
2804 "fxdiv_headers",
2805 ],
2806 static_libs: [
2807 "libclog",
2808 "libpthreadpool",
Miao Wang86f5fbe2020-07-24 11:16:10 -07002809 "xnnpack_logging_utils",
2810 "xnnpack_packing",
Miao Wange9993472020-02-10 15:00:10 -08002811 ],
2812 whole_static_libs: [
2813 "xnnpack_indirection",
2814 ],
2815}
2816
2817cc_library_static {
2818 name: "xnnpack_scalar_ukernels",
2819 defaults: ["xnnpack_internal_default"],
2820 srcs: SCALAR_UKERNELS,
2821 header_libs: [
2822 "fp16_headers",
2823 "fxdiv_headers",
2824 ],
2825 static_libs: [
2826 "libpthreadpool",
2827 "xnnpack_tables",
2828 ],
2829}
2830
2831cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002832 name: "xnnpack_neon_ukernels",
2833 defaults: ["xnnpack_internal_default"],
2834 arch: {
2835 arm: {
2836 srcs: NEON_UKERNELS,
2837 cflags: [
2838 "-marm",
2839 "-mfpu=neon",
2840 ],
2841 },
2842 arm64: {
2843 srcs: NEON_UKERNELS,
2844 },
2845 x86: { enabled: false, },
2846 x86_64: { enabled: false, },
2847 },
2848 header_libs: [
2849 "fp16_headers",
2850 ],
2851 static_libs: [
2852 "libpthreadpool",
2853 "xnnpack_tables",
2854 ],
2855}
2856
2857cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08002858 name: "xnnpack_neondot_ukernels",
2859 defaults: ["xnnpack_internal_default"],
2860 arch: {
2861 arm: {
2862 srcs: NEONDOT_UKERNELS,
2863 cflags: [
2864 "-marm",
2865 "-march=armv8.2-a+dotprod",
2866 "-mfpu=neon-fp-armv8",
2867 ],
2868 },
2869 arm64: {
2870 srcs: NEONDOT_UKERNELS,
2871 cflags: [
2872 "-march=armv8.2-a+dotprod",
2873 ],
2874 },
2875 x86: { enabled: false, },
2876 x86_64: { enabled: false, },
2877 },
2878 header_libs: [
2879 "fp16_headers",
2880 ],
2881 static_libs: [
2882 "libpthreadpool",
2883 "xnnpack_tables",
2884 ],
2885}
2886
2887cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002888 name: "xnnpack_neonfma_ukernels",
2889 defaults: ["xnnpack_internal_default"],
2890 arch: {
2891 arm: {
2892 srcs: NEONFMA_UKERNELS,
2893 cflags: [
2894 "-marm",
Miao Wangc0aa11a2020-06-10 13:41:26 -07002895 "-march=armv7-a",
Miao Wange9993472020-02-10 15:00:10 -08002896 "-mfpu=neon-vfpv4",
2897 ],
2898 },
2899 arm64: {
2900 srcs: NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS,
2901 },
2902 x86: { enabled: false, },
2903 x86_64: { enabled: false, },
2904 },
2905 header_libs: [
2906 "fp16_headers",
2907 ],
2908 static_libs: [
2909 "libpthreadpool",
2910 "xnnpack_tables",
2911 ],
2912}
2913
2914cc_library_static {
Miao Wangc0aa11a2020-06-10 13:41:26 -07002915 name: "xnnpack_neonv8_ukernels",
2916 defaults: ["xnnpack_internal_default"],
2917 arch: {
2918 arm: {
2919 srcs: NEONV8_UKERNELS,
2920 cflags: [
2921 "-marm",
2922 "-march=armv8-a",
2923 "-mfpu=neon-fp-armv8",
2924 ],
2925 },
2926 arm64: {
2927 srcs: NEONV8_UKERNELS,
2928 },
2929 x86: { enabled: false, },
2930 x86_64: { enabled: false, },
2931 },
2932 header_libs: [
2933 "fp16_headers",
2934 ],
2935 static_libs: [
2936 "libpthreadpool",
2937 "xnnpack_tables",
2938 ],
2939}
2940
2941cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08002942 name: "xnnpack_neonfp16arith_ukernels",
2943 defaults: ["xnnpack_internal_default"],
2944 arch: {
2945 arm: { enabled: false, },
2946 arm64: {
2947 srcs: AARCH64_NEONFP16ARITH_UKERNELS,
2948 cflags: [
2949 "-march=armv8.2-a+fp16",
2950 ],
2951 },
2952 x86: { enabled: false, },
2953 x86_64: { enabled: false, },
2954 },
2955 header_libs: [
2956 "fp16_headers",
2957 ],
2958 static_libs: [
2959 "libpthreadpool",
2960 "xnnpack_tables",
2961 ],
2962}
2963
2964cc_library_static {
2965 name: "xnnpack_asm_ukernels",
2966 defaults: ["xnnpack_internal_default"],
2967 arch: {
2968 arm: {
2969 srcs: AARCH32_ASM_UKERNELS,
2970 },
2971 arm64: {
2972 srcs: AARCH64_ASM_UKERNELS,
Miao Wang400e4042020-04-17 10:15:59 -07002973 clang_asflags: [
Miao Wang5eea8312020-12-07 09:12:40 -08002974 "-march=armv8.2-a+fp16+dotprod",
Miao Wang400e4042020-04-17 10:15:59 -07002975 ],
Miao Wange9993472020-02-10 15:00:10 -08002976 },
2977 x86: { enabled: false, },
2978 x86_64: { enabled: false, },
2979 },
2980}
2981
2982cc_library_static {
2983 name: "xnnpack_sse2_ukernels",
2984 defaults: ["xnnpack_internal_default"],
2985 arch: {
2986 arm: { enabled: false, },
2987 arm64: { enabled: false, },
2988 x86: {
2989 srcs: SSE_UKERNELS + SSE2_UKERNELS,
2990 cflags: [
2991 "-msse2",
2992 ],
2993 },
2994 x86_64: {
2995 srcs: SSE_UKERNELS + SSE2_UKERNELS,
2996 cflags: [
2997 "-msse2",
2998 ],
2999 },
3000 },
3001 header_libs: [
3002 "fp16_headers",
3003 ],
3004 static_libs: [
3005 "libpthreadpool",
3006 "xnnpack_tables",
3007 ],
3008}
3009
3010cc_library_static {
Miao Wang2534c2f2020-03-16 11:58:04 -07003011 name: "xnnpack_ssse3_ukernels",
3012 defaults: ["xnnpack_internal_default"],
3013 arch: {
3014 arm: { enabled: false, },
3015 arm64: { enabled: false, },
3016 x86: {
3017 srcs: SSSE3_UKERNELS,
3018 cflags: [
3019 "-mssse3",
3020 ],
3021 },
3022 x86_64: {
3023 srcs: SSSE3_UKERNELS,
3024 cflags: [
3025 "-mssse3",
3026 ],
3027 },
3028 },
3029 header_libs: [
3030 "fp16_headers",
3031 ],
3032 static_libs: [
3033 "libpthreadpool",
3034 "xnnpack_tables",
3035 ],
3036}
3037
3038cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08003039 name: "xnnpack_sse41_ukernels",
3040 defaults: ["xnnpack_internal_default"],
3041 arch: {
3042 arm: { enabled: false, },
3043 arm64: { enabled: false, },
3044 x86: {
3045 srcs: SSE41_UKERNELS,
3046 cflags: [
3047 "-msse4.1",
3048 ],
3049 },
3050 x86_64: {
3051 srcs: SSE41_UKERNELS,
3052 cflags: [
3053 "-msse4.1",
3054 ],
3055 },
3056 },
3057 header_libs: [
3058 "fp16_headers",
3059 ],
3060 static_libs: [
3061 "libpthreadpool",
3062 "xnnpack_tables",
3063 ],
3064}
3065
3066cc_library_static {
3067 name: "xnnpack_avx_ukernels",
3068 defaults: ["xnnpack_internal_default"],
3069 arch: {
3070 arm: { enabled: false, },
3071 arm64: { enabled: false, },
3072 x86: {
3073 srcs: AVX_UKERNELS,
3074 cflags: [
3075 "-mavx",
3076 ],
3077 },
3078 x86_64: {
3079 srcs: AVX_UKERNELS,
3080 cflags: [
3081 "-mavx",
3082 ],
3083 },
3084 },
3085 header_libs: [
3086 "fp16_headers",
3087 ],
3088 static_libs: [
3089 "libpthreadpool",
3090 "xnnpack_tables",
3091 ],
3092}
3093
3094cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003095 name: "xnnpack_xop_ukernels",
3096 defaults: ["xnnpack_internal_default"],
3097 arch: {
3098 arm: { enabled: false, },
3099 arm64: { enabled: false, },
3100 x86: {
3101 srcs: XOP_UKERNELS,
3102 cflags: [
3103 "-mxop",
3104 ],
3105 },
3106 x86_64: {
3107 srcs: XOP_UKERNELS,
3108 cflags: [
3109 "-mxop",
3110 ],
3111 },
3112 },
3113 header_libs: [
3114 "fp16_headers",
3115 ],
3116 static_libs: [
3117 "libpthreadpool",
3118 "xnnpack_tables",
3119 ],
3120}
3121
3122cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08003123 name: "xnnpack_fma3_ukernels",
3124 defaults: ["xnnpack_internal_default"],
3125 arch: {
3126 arm: { enabled: false, },
3127 arm64: { enabled: false, },
3128 x86: {
3129 srcs: FMA3_UKERNELS,
3130 cflags: [
3131 "-mfma",
3132 ],
3133 },
3134 x86_64: {
3135 srcs: FMA3_UKERNELS,
3136 cflags: [
3137 "-mfma",
3138 ],
3139 },
3140 },
3141 header_libs: [
3142 "fp16_headers",
3143 ],
3144 static_libs: [
3145 "libpthreadpool",
3146 "xnnpack_tables",
3147 ],
3148}
3149
3150cc_library_static {
3151 name: "xnnpack_avx2_ukernels",
3152 defaults: ["xnnpack_internal_default"],
3153 arch: {
3154 arm: { enabled: false, },
3155 arm64: { enabled: false, },
3156 x86: {
3157 srcs: AVX2_UKERNELS,
3158 cflags: [
3159 "-mfma",
3160 "-mavx2",
3161 ],
3162 },
3163 x86_64: {
3164 srcs: AVX2_UKERNELS,
3165 cflags: [
3166 "-mfma",
3167 "-mavx2",
3168 ],
3169 },
3170 },
3171 header_libs: [
3172 "fp16_headers",
3173 ],
3174 static_libs: [
3175 "libpthreadpool",
3176 "xnnpack_tables",
3177 ],
3178}
3179
3180cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003181 name: "xnnpack_avx512skx_ukernels",
3182 defaults: ["xnnpack_internal_default"],
3183 arch: {
3184 arm: { enabled: false, },
3185 arm64: { enabled: false, },
3186 x86: {
3187 srcs: AVX512SKX_UKERNELS,
3188 cflags: [
3189 "-mavx512f",
3190 "-mavx512cd",
3191 "-mavx512bw",
3192 "-mavx512dq",
3193 "-mavx512vl",
3194 ],
3195 },
3196 x86_64: {
3197 srcs: AVX512SKX_UKERNELS,
3198 cflags: [
3199 "-mavx512f",
3200 "-mavx512cd",
3201 "-mavx512bw",
3202 "-mavx512dq",
3203 "-mavx512vl",
3204 ],
3205 },
3206 },
3207 header_libs: [
3208 "fp16_headers",
3209 ],
3210 static_libs: [
3211 "libpthreadpool",
3212 "xnnpack_tables",
3213 ],
3214}
3215
3216cc_library_static {
Miao Wange9993472020-02-10 15:00:10 -08003217 name: "xnnpack_avx512f_ukernels",
3218 defaults: ["xnnpack_internal_default"],
3219 arch: {
3220 arm: { enabled: false, },
3221 arm64: { enabled: false, },
3222 x86: {
3223 srcs: AVX512F_UKERNELS,
3224 cflags: [
3225 "-mavx512f",
3226 ],
3227 },
3228 x86_64: {
3229 srcs: AVX512F_UKERNELS,
3230 cflags: [
3231 "-mavx512f",
3232 ],
3233 },
3234 },
3235 header_libs: [
3236 "fp16_headers",
3237 ],
3238 static_libs: [
3239 "libpthreadpool",
3240 "xnnpack_tables",
3241 ],
3242}
3243
3244cc_library_static {
3245 name: "xnnpack_ukernels",
3246 defaults: ["xnnpack_internal_default"],
3247 arch: {
3248 arm: {
3249 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003250 "xnnpack_neon_ukernels",
3251 "xnnpack_neonfma_ukernels",
Miao Wangc0aa11a2020-06-10 13:41:26 -07003252 "xnnpack_neonv8_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003253 "xnnpack_neondot_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003254 "xnnpack_asm_ukernels",
3255 ],
3256 },
3257 arm64: {
3258 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003259 "xnnpack_neon_ukernels",
3260 "xnnpack_neonfma_ukernels",
Miao Wangc0aa11a2020-06-10 13:41:26 -07003261 "xnnpack_neonv8_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003262 "xnnpack_neonfp16arith_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003263 "xnnpack_neondot_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003264 "xnnpack_asm_ukernels",
3265 ],
3266 },
3267 x86: {
3268 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003269 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07003270 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003271 "xnnpack_sse41_ukernels",
3272 "xnnpack_avx_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003273 "xnnpack_xop_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003274 "xnnpack_fma3_ukernels",
3275 "xnnpack_avx2_ukernels",
3276 "xnnpack_avx512f_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003277 "xnnpack_avx512skx_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003278 ],
3279 },
3280 x86_64: {
3281 whole_static_libs: [
Miao Wange9993472020-02-10 15:00:10 -08003282 "xnnpack_sse2_ukernels",
Miao Wang2534c2f2020-03-16 11:58:04 -07003283 "xnnpack_ssse3_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003284 "xnnpack_sse41_ukernels",
3285 "xnnpack_avx_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003286 "xnnpack_xop_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003287 "xnnpack_fma3_ukernels",
3288 "xnnpack_avx2_ukernels",
3289 "xnnpack_avx512f_ukernels",
Miao Wang5eea8312020-12-07 09:12:40 -08003290 "xnnpack_avx512skx_ukernels",
Miao Wange9993472020-02-10 15:00:10 -08003291 ],
3292 },
3293 },
3294 whole_static_libs: [
3295 "xnnpack_scalar_ukernels",
3296 "xnnpack_tables",
3297 ],
3298}
3299
3300cc_library_static {
3301 name: "libXNNPACK",
3302 defaults: ["xnnpack_internal_default"],
3303 export_include_dirs: ["include"],
3304 srcs: [
3305 "src/init.c",
3306 "src/runtime.c",
3307 "src/subgraph.c",
3308 "src/tensor.c",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003309 ] + SUBGRAPH_SRCS,
3310 header_libs: [
3311 "fp16_headers",
Miao Wange9993472020-02-10 15:00:10 -08003312 ],
3313 whole_static_libs: [
3314 "libclog",
3315 "libcpuinfo",
3316 "libpthreadpool",
3317 "xnnpack_ukernels",
3318 "xnnpack_operator_run",
3319 "xnnpack_operators",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003320 "xnnpack_logging_utils",
3321 "xnnpack_memory_planner",
3322 "xnnpack_packing",
Miao Wange9993472020-02-10 15:00:10 -08003323 ],
3324}
3325
3326// Tests and benchmarks
3327cc_defaults {
3328 name: "xnnpack_tests_default",
3329 vendor_available: true,
3330 stl: "libc++_static",
3331 local_include_dirs: [
3332 "bench",
3333 "models",
3334 "test",
3335 "src",
3336 ],
3337 cflags: [
Miao Wang400e4042020-04-17 10:15:59 -07003338 "-Wno-unused-function",
3339 "-Wno-unused-parameter",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003340 "-Wno-unused-private-field",
Miao Wange9993472020-02-10 15:00:10 -08003341 ],
3342 header_libs: [
3343 "fp16_headers",
3344 ],
3345 static_libs: [
3346 "libXNNPACK",
3347 "libpthreadpool",
3348 "libgmock",
3349 ],
3350 shared_libs: [
3351 "liblog",
3352 ],
3353}
3354
3355cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003356 name: "xnnpack_mobilenet_v1_fp32",
Miao Wange9993472020-02-10 15:00:10 -08003357 defaults: ["xnnpack_tests_default"],
3358 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003359 "models/fp32-mobilenet-v1.cc",
Miao Wange9993472020-02-10 15:00:10 -08003360 ],
3361}
3362
3363cc_library_static {
Miao Wang338e80e2021-03-23 08:56:09 -07003364 name: "xnnpack_mobilenet_v1_fp32_sparse",
3365 defaults: ["xnnpack_tests_default"],
3366 srcs: [
3367 "models/fp32-sparse-mobilenet-v1.cc",
3368 ],
3369}
3370
3371cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003372 name: "xnnpack_qs8_mobilenet_v1",
3373 defaults: ["xnnpack_tests_default"],
3374 srcs: [
3375 "models/qs8-mobilenet-v1.cc",
3376 ],
3377}
3378
3379cc_library_static {
Miao Wang338e80e2021-03-23 08:56:09 -07003380 name: "xnnpack_qu8_mobilenet_v1",
3381 defaults: ["xnnpack_tests_default"],
3382 srcs: [
3383 "models/qu8-mobilenet-v1.cc",
3384 ],
3385}
3386
3387cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003388 name: "xnnpack_mobilenet_v1_fp16",
Miao Wange9993472020-02-10 15:00:10 -08003389 defaults: ["xnnpack_tests_default"],
3390 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003391 "models/fp16-mobilenet-v1.cc",
Miao Wange9993472020-02-10 15:00:10 -08003392 ],
3393}
3394
3395cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003396 name: "xnnpack_qs8_mobilenet_v2",
3397 defaults: ["xnnpack_tests_default"],
3398 srcs: [
3399 "models/qs8-mobilenet-v2.cc",
3400 ],
3401}
3402
3403cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003404 name: "xnnpack_mobilenet_v2_fp32",
Miao Wange9993472020-02-10 15:00:10 -08003405 defaults: ["xnnpack_tests_default"],
3406 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003407 "models/fp32-mobilenet-v2.cc",
Miao Wange9993472020-02-10 15:00:10 -08003408 ],
3409}
3410
3411cc_library_static {
Miao Wang338e80e2021-03-23 08:56:09 -07003412 name: "xnnpack_mobilenet_v2_fp32_sparse",
3413 defaults: ["xnnpack_tests_default"],
3414 srcs: [
3415 "models/fp32-sparse-mobilenet-v2.cc",
3416 ],
3417}
3418
3419cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003420 name: "xnnpack_mobilenet_v2_fp16",
Miao Wange9993472020-02-10 15:00:10 -08003421 defaults: ["xnnpack_tests_default"],
3422 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003423 "models/fp16-mobilenet-v2.cc",
3424 ],
3425}
3426
3427
3428cc_library_static {
3429 name: "xnnpack_mobilenet_v3_large_fp32",
3430 defaults: ["xnnpack_tests_default"],
3431 srcs: [
3432 "models/fp32-mobilenet-v3-large.cc",
3433 ],
3434}
3435
3436cc_library_static {
Miao Wang338e80e2021-03-23 08:56:09 -07003437 name: "xnnpack_mobilenet_v3_large_fp32_sparse",
3438 defaults: ["xnnpack_tests_default"],
3439 srcs: [
3440 "models/fp32-sparse-mobilenet-v3-large.cc",
3441 ],
3442}
3443
3444cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003445 name: "xnnpack_mobilenet_v3_large_fp16",
3446 defaults: ["xnnpack_tests_default"],
3447 srcs: [
3448 "models/fp16-mobilenet-v3-large.cc",
3449 ],
3450}
3451
3452cc_library_static {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003453 name: "xnnpack_mobilenet_v3_small_fp32",
3454 defaults: ["xnnpack_tests_default"],
3455 srcs: [
3456 "models/fp32-mobilenet-v3-small.cc",
Miao Wange9993472020-02-10 15:00:10 -08003457 ],
3458}
3459
Miao Wang5eea8312020-12-07 09:12:40 -08003460cc_library_static {
Miao Wang338e80e2021-03-23 08:56:09 -07003461 name: "xnnpack_mobilenet_v3_small_fp32_sparse",
3462 defaults: ["xnnpack_tests_default"],
3463 srcs: [
3464 "models/fp32-sparse-mobilenet-v3-small.cc",
3465 ],
3466}
3467
3468cc_library_static {
Miao Wang5eea8312020-12-07 09:12:40 -08003469 name: "xnnpack_mobilenet_v3_small_fp16",
3470 defaults: ["xnnpack_tests_default"],
3471 srcs: [
3472 "models/fp16-mobilenet-v3-small.cc",
3473 ],
3474}
3475
Miao Wange9993472020-02-10 15:00:10 -08003476cc_benchmark {
3477 name: "xnnpack_end2end_bench",
3478 defaults: ["xnnpack_tests_default"],
3479 srcs: [
3480 "bench/end2end.cc",
3481 "bench/utils.cc",
3482 ],
3483 cflags: [
3484 "-Wno-unused-result"
3485 ],
3486 static_libs: [
3487 "libcpuinfo",
3488 "libgoogle-benchmark",
Miao Wang5eea8312020-12-07 09:12:40 -08003489 "xnnpack_qs8_mobilenet_v1",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003490 "xnnpack_mobilenet_v1_fp32",
Miao Wang338e80e2021-03-23 08:56:09 -07003491 "xnnpack_mobilenet_v1_fp32_sparse",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003492 "xnnpack_mobilenet_v1_fp16",
Miao Wang5eea8312020-12-07 09:12:40 -08003493 "xnnpack_qs8_mobilenet_v2",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003494 "xnnpack_mobilenet_v2_fp32",
Miao Wang338e80e2021-03-23 08:56:09 -07003495 "xnnpack_mobilenet_v2_fp32_sparse",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003496 "xnnpack_mobilenet_v2_fp16",
3497 "xnnpack_mobilenet_v3_large_fp32",
Miao Wang338e80e2021-03-23 08:56:09 -07003498 "xnnpack_mobilenet_v3_large_fp32_sparse",
Miao Wang5eea8312020-12-07 09:12:40 -08003499 "xnnpack_mobilenet_v3_large_fp16",
Miao Wang86f5fbe2020-07-24 11:16:10 -07003500 "xnnpack_mobilenet_v3_small_fp32",
Miao Wang338e80e2021-03-23 08:56:09 -07003501 "xnnpack_mobilenet_v3_small_fp32_sparse",
Miao Wang5eea8312020-12-07 09:12:40 -08003502 "xnnpack_mobilenet_v3_small_fp16",
Miao Wang338e80e2021-03-23 08:56:09 -07003503 "xnnpack_qu8_mobilenet_v1",
Miao Wange9993472020-02-10 15:00:10 -08003504 ],
3505}
3506
3507cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003508 name: "xnnpack_abs_nc_test",
Miao Wange9993472020-02-10 15:00:10 -08003509 defaults: ["xnnpack_tests_default"],
3510 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003511 "test/abs-nc.cc",
Miao Wange9993472020-02-10 15:00:10 -08003512 ],
3513 test_suites: [
3514 "general-tests",
3515 ],
3516}
3517
3518cc_test {
3519 name: "xnnpack_add_nd_test",
3520 defaults: ["xnnpack_tests_default"],
3521 srcs: [
3522 "test/add-nd.cc",
3523 ],
3524 test_suites: [
3525 "general-tests",
3526 ],
3527}
3528
3529cc_test {
3530 name: "xnnpack_argmax_pooling_nhwc_test",
3531 defaults: ["xnnpack_tests_default"],
3532 srcs: [
3533 "test/argmax-pooling-nhwc.cc",
3534 ],
3535 test_suites: [
3536 "general-tests",
3537 ],
3538}
3539
3540cc_test {
3541 name: "xnnpack_average_pooling_nhwc_test",
3542 defaults: ["xnnpack_tests_default"],
3543 srcs: [
3544 "test/average-pooling-nhwc.cc",
3545 ],
3546 test_suites: [
3547 "general-tests",
3548 ],
3549}
3550
3551cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003552 name: "xnnpack_bankers_rounding_nc_test",
Miao Wange9993472020-02-10 15:00:10 -08003553 defaults: ["xnnpack_tests_default"],
3554 srcs: [
Miao Wang86f5fbe2020-07-24 11:16:10 -07003555 "test/bankers-rounding-nc.cc",
3556 ],
3557 test_suites: [
3558 "general-tests",
3559 ],
3560}
3561
3562cc_test {
3563 name: "xnnpack_ceiling_nc_test",
3564 defaults: ["xnnpack_tests_default"],
3565 srcs: [
3566 "test/ceiling-nc.cc",
Miao Wange9993472020-02-10 15:00:10 -08003567 ],
3568 test_suites: [
3569 "general-tests",
3570 ],
3571}
3572
3573cc_test {
3574 name: "xnnpack_channel_shuffle_nc_test",
3575 defaults: ["xnnpack_tests_default"],
3576 srcs: [
3577 "test/channel-shuffle-nc.cc",
3578 ],
3579 test_suites: [
3580 "general-tests",
3581 ],
3582}
3583
3584cc_test {
3585 name: "xnnpack_clamp_nc_test",
3586 defaults: ["xnnpack_tests_default"],
3587 srcs: [
3588 "test/clamp-nc.cc",
3589 ],
3590 test_suites: [
3591 "general-tests",
3592 ],
3593}
3594
3595cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003596 name: "xnnpack_constant_pad_nd_test",
3597 defaults: ["xnnpack_tests_default"],
3598 srcs: [
3599 "test/constant-pad-nd.cc",
3600 ],
3601 test_suites: [
3602 "general-tests",
3603 ],
3604}
3605
3606cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003607 name: "xnnpack_convolution_nhwc_test",
3608 defaults: ["xnnpack_tests_default"],
3609 srcs: [
3610 "test/convolution-nhwc.cc",
3611 ],
3612 test_suites: [
3613 "general-tests",
3614 ],
3615}
3616
3617cc_test {
3618 name: "xnnpack_convolution_nchw_test",
3619 defaults: ["xnnpack_tests_default"],
3620 srcs: [
3621 "test/convolution-nchw.cc",
3622 ],
3623 test_suites: [
3624 "general-tests",
3625 ],
3626}
3627
3628cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003629 name: "xnnpack_copy_nc_test",
3630 defaults: ["xnnpack_tests_default"],
3631 srcs: [
3632 "test/copy-nc.cc",
3633 ],
3634 test_suites: [
3635 "general-tests",
3636 ],
3637}
3638
3639cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003640 name: "xnnpack_deconvolution_nhwc_test",
3641 defaults: ["xnnpack_tests_default"],
3642 srcs: [
3643 "test/deconvolution-nhwc.cc",
3644 ],
3645 test_suites: [
3646 "general-tests",
3647 ],
3648}
3649
3650cc_test {
3651 name: "xnnpack_divide_nd_test",
3652 defaults: ["xnnpack_tests_default"],
3653 srcs: [
3654 "test/divide-nd.cc",
3655 ],
3656 test_suites: [
3657 "general-tests",
3658 ],
3659}
3660
3661cc_test {
3662 name: "xnnpack_fully_connected_nc_test",
3663 defaults: ["xnnpack_tests_default"],
3664 srcs: [
3665 "test/fully-connected-nc.cc",
3666 ],
3667 test_suites: [
3668 "general-tests",
3669 ],
3670}
3671
3672cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003673 name: "xnnpack_floor_nc_test",
3674 defaults: ["xnnpack_tests_default"],
3675 srcs: [
3676 "test/floor-nc.cc",
3677 ],
3678 test_suites: [
3679 "general-tests",
3680 ],
3681}
3682
3683cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003684 name: "xnnpack_global_average_pooling_nwc_test",
3685 defaults: ["xnnpack_tests_default"],
3686 srcs: [
3687 "test/global-average-pooling-nwc.cc",
3688 ],
3689 test_suites: [
3690 "general-tests",
3691 ],
3692}
3693
3694cc_test {
3695 name: "xnnpack_global_average_pooling_ncw_test",
3696 defaults: ["xnnpack_tests_default"],
3697 srcs: [
3698 "test/global-average-pooling-ncw.cc",
3699 ],
3700 test_suites: [
3701 "general-tests",
3702 ],
3703}
3704
3705cc_test {
3706 name: "xnnpack_hardswish_nc_test",
3707 defaults: ["xnnpack_tests_default"],
3708 srcs: [
3709 "test/hardswish-nc.cc",
3710 ],
3711 test_suites: [
3712 "general-tests",
3713 ],
3714}
3715
3716cc_test {
3717 name: "xnnpack_leaky_relu_nc_test",
3718 defaults: ["xnnpack_tests_default"],
3719 srcs: [
3720 "test/leaky-relu-nc.cc",
3721 ],
3722 test_suites: [
3723 "general-tests",
3724 ],
3725}
3726
3727cc_test {
3728 name: "xnnpack_max_pooling_nhwc_test",
3729 defaults: ["xnnpack_tests_default"],
3730 srcs: [
3731 "test/max-pooling-nhwc.cc",
3732 ],
3733 test_suites: [
3734 "general-tests",
3735 ],
3736}
3737
3738cc_test {
3739 name: "xnnpack_maximum_nd_test",
3740 defaults: ["xnnpack_tests_default"],
3741 srcs: [
3742 "test/maximum-nd.cc",
3743 ],
3744 test_suites: [
3745 "general-tests",
3746 ],
3747}
3748
3749cc_test {
3750 name: "xnnpack_minimum_nd_test",
3751 defaults: ["xnnpack_tests_default"],
3752 srcs: [
3753 "test/minimum-nd.cc",
3754 ],
3755 test_suites: [
3756 "general-tests",
3757 ],
3758}
3759
3760cc_test {
3761 name: "xnnpack_multiply_nd_test",
3762 defaults: ["xnnpack_tests_default"],
3763 srcs: [
3764 "test/multiply-nd.cc",
3765 ],
3766 test_suites: [
3767 "general-tests",
3768 ],
3769}
3770
3771cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003772 name: "xnnpack_negate_nc_test",
3773 defaults: ["xnnpack_tests_default"],
3774 srcs: [
3775 "test/negate-nc.cc",
3776 ],
3777 test_suites: [
3778 "general-tests",
3779 ],
3780}
3781
3782cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003783 name: "xnnpack_prelu_nc_test",
3784 defaults: ["xnnpack_tests_default"],
3785 srcs: [
3786 "test/prelu-nc.cc",
3787 ],
3788 test_suites: [
3789 "general-tests",
3790 ],
3791}
3792
3793cc_test {
3794 name: "xnnpack_resize_bilinear_nhwc_test",
3795 defaults: ["xnnpack_tests_default"],
3796 srcs: [
3797 "test/resize-bilinear-nhwc.cc",
3798 ],
3799 test_suites: [
3800 "general-tests",
3801 ],
3802}
3803
3804cc_test {
3805 name: "xnnpack_sigmoid_nc_test",
3806 defaults: ["xnnpack_tests_default"],
3807 srcs: [
3808 "test/sigmoid-nc.cc",
3809 ],
3810 test_suites: [
3811 "general-tests",
3812 ],
3813}
3814
3815cc_test {
3816 name: "xnnpack_softmax_nc_test",
3817 defaults: ["xnnpack_tests_default"],
3818 srcs: [
3819 "test/softmax-nc.cc",
3820 ],
3821 test_suites: [
3822 "general-tests",
3823 ],
3824}
3825
3826cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003827 name: "xnnpack_square_nc_test",
3828 defaults: ["xnnpack_tests_default"],
3829 srcs: [
3830 "test/square-nc.cc",
3831 ],
3832 test_suites: [
3833 "general-tests",
3834 ],
3835}
3836
3837cc_test {
3838 name: "xnnpack_square_root_nc_test",
3839 defaults: ["xnnpack_tests_default"],
3840 srcs: [
3841 "test/square-root-nc.cc",
3842 ],
3843 test_suites: [
3844 "general-tests",
3845 ],
3846}
3847
3848cc_test {
3849 name: "xnnpack_square_difference_nd_test",
3850 defaults: ["xnnpack_tests_default"],
3851 srcs: [
3852 "test/squared-difference-nd.cc",
3853 ],
3854 test_suites: [
3855 "general-tests",
3856 ],
3857}
3858
3859cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003860 name: "xnnpack_subtract_nd_test",
3861 defaults: ["xnnpack_tests_default"],
3862 srcs: [
3863 "test/subtract-nd.cc",
3864 ],
3865 test_suites: [
3866 "general-tests",
3867 ],
3868}
3869
3870cc_test {
Miao Wang86f5fbe2020-07-24 11:16:10 -07003871 name: "xnnpack_truncation_nc_test",
3872 defaults: ["xnnpack_tests_default"],
3873 srcs: [
3874 "test/truncation-nc.cc",
3875 ],
3876 test_suites: [
3877 "general-tests",
3878 ],
3879}
3880
3881cc_test {
Miao Wange9993472020-02-10 15:00:10 -08003882 name: "xnnpack_unpooling_nhwc_test",
3883 defaults: ["xnnpack_tests_default"],
3884 srcs: [
3885 "test/unpooling-nhwc.cc",
3886 ],
3887 test_suites: [
3888 "general-tests",
3889 ],
3890}