Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1 | // 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 | |
| 15 | OPERATOR_SRCS = [ |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 16 | "src/operators/argmax-pooling-nhwc.c", |
| 17 | "src/operators/average-pooling-nhwc.c", |
| 18 | "src/operators/binary-elementwise-nd.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 19 | "src/operators/channel-shuffle-nc.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 20 | "src/operators/constant-pad-nd.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 21 | "src/operators/convolution-nchw.c", |
| 22 | "src/operators/convolution-nhwc.c", |
| 23 | "src/operators/deconvolution-nhwc.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 24 | "src/operators/depth-to-space-nchw2nhwc.c", |
| 25 | "src/operators/depth-to-space-nhwc.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 26 | "src/operators/fully-connected-nc.c", |
| 27 | "src/operators/global-average-pooling-ncw.c", |
| 28 | "src/operators/global-average-pooling-nwc.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 29 | "src/operators/leaky-relu-nc.c", |
| 30 | "src/operators/max-pooling-nhwc.c", |
| 31 | "src/operators/prelu-nc.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 32 | "src/operators/resize-bilinear-nchw.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 33 | "src/operators/resize-bilinear-nhwc.c", |
| 34 | "src/operators/sigmoid-nc.c", |
| 35 | "src/operators/softmax-nc.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 36 | "src/operators/unary-elementwise-nc.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 37 | "src/operators/unpooling-nhwc.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 38 | ] |
| 39 | |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 40 | LOGGING_SRCS = [ |
| 41 | "src/operator-strings.c", |
| 42 | "src/subgraph-strings.c", |
| 43 | ] |
| 44 | |
| 45 | SUBGRAPH_SRCS = [ |
| 46 | "src/subgraph/abs.c", |
| 47 | "src/subgraph/add2.c", |
| 48 | "src/subgraph/argmax-pooling-2d.c", |
| 49 | "src/subgraph/average-pooling-2d.c", |
| 50 | "src/subgraph/bankers-rounding.c", |
| 51 | "src/subgraph/ceiling.c", |
| 52 | "src/subgraph/clamp.c", |
| 53 | "src/subgraph/convolution-2d.c", |
| 54 | "src/subgraph/deconvolution-2d.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 55 | "src/subgraph/depth-to-space.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 56 | "src/subgraph/depthwise-convolution-2d.c", |
| 57 | "src/subgraph/divide.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 58 | "src/subgraph/floor.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 59 | "src/subgraph/fully-connected.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 60 | "src/subgraph/global-average-pooling-2d.c", |
| 61 | "src/subgraph/hardswish.c", |
| 62 | "src/subgraph/leaky-relu.c", |
| 63 | "src/subgraph/max-pooling-2d.c", |
| 64 | "src/subgraph/maximum2.c", |
| 65 | "src/subgraph/minimum2.c", |
| 66 | "src/subgraph/multiply2.c", |
| 67 | "src/subgraph/negate.c", |
| 68 | "src/subgraph/prelu.c", |
| 69 | "src/subgraph/sigmoid.c", |
| 70 | "src/subgraph/softmax.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 71 | "src/subgraph/square-root.c", |
| 72 | "src/subgraph/square.c", |
| 73 | "src/subgraph/squared-difference.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 74 | "src/subgraph/static-constant-pad.c", |
| 75 | "src/subgraph/static-reshape.c", |
| 76 | "src/subgraph/static-resize-bilinear-2d.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 77 | "src/subgraph/subtract.c", |
| 78 | "src/subgraph/unpooling-2d.c", |
| 79 | ] |
| 80 | |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 81 | TABLE_SRCS = [ |
| 82 | "src/tables/exp2-k-over-64.c", |
| 83 | "src/tables/exp2-k-over-2048.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 84 | "src/tables/exp2minus-k-over-4.c", |
| 85 | "src/tables/exp2minus-k-over-8.c", |
| 86 | "src/tables/exp2minus-k-over-16.c", |
| 87 | "src/tables/exp2minus-k-over-64.c", |
| 88 | "src/tables/exp2minus-k-over-2048.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 89 | ] |
| 90 | |
| 91 | SCALAR_UKERNELS = [ |
| 92 | "src/f32-argmaxpool/4x-scalar-c1.c", |
| 93 | "src/f32-argmaxpool/9p8x-scalar-c1.c", |
| 94 | "src/f32-argmaxpool/9x-scalar-c1.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 95 | "src/f32-avgpool/9p8x-minmax-scalar-c1.c", |
| 96 | "src/f32-avgpool/9x-minmax-scalar-c1.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 97 | "src/f32-clamp/gen/scalar-x1.c", |
| 98 | "src/f32-clamp/gen/scalar-x2.c", |
| 99 | "src/f32-clamp/gen/scalar-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 100 | "src/f32-conv-hwc/3x3s2p0p1c3x4-scalar-1x1.c", |
| 101 | "src/f32-conv-hwc/3x3s2p1c3x4-scalar-1x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 102 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-scalar-1x1.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 103 | "src/f32-dwconv/gen/up1x4-minmax-scalar-acc2.c", |
| 104 | "src/f32-dwconv/gen/up1x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 105 | "src/f32-dwconv/gen/up1x4-scalar-acc2.c", |
| 106 | "src/f32-dwconv/gen/up1x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 107 | "src/f32-dwconv/gen/up1x9-minmax-scalar-acc2.c", |
| 108 | "src/f32-dwconv/gen/up1x9-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 109 | "src/f32-dwconv/gen/up1x9-scalar-acc2.c", |
| 110 | "src/f32-dwconv/gen/up1x9-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 111 | "src/f32-dwconv/gen/up1x25-minmax-scalar-acc2.c", |
| 112 | "src/f32-dwconv/gen/up1x25-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 113 | "src/f32-dwconv/gen/up1x25-scalar-acc2.c", |
| 114 | "src/f32-dwconv/gen/up1x25-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 115 | "src/f32-dwconv/gen/up2x4-minmax-scalar-acc2.c", |
| 116 | "src/f32-dwconv/gen/up2x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 117 | "src/f32-dwconv/gen/up2x4-scalar-acc2.c", |
| 118 | "src/f32-dwconv/gen/up2x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 119 | "src/f32-dwconv/gen/up2x9-minmax-scalar-acc2.c", |
| 120 | "src/f32-dwconv/gen/up2x9-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 121 | "src/f32-dwconv/gen/up2x9-scalar-acc2.c", |
| 122 | "src/f32-dwconv/gen/up2x9-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 123 | "src/f32-dwconv/gen/up2x25-minmax-scalar-acc2.c", |
| 124 | "src/f32-dwconv/gen/up2x25-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 125 | "src/f32-dwconv/gen/up2x25-scalar-acc2.c", |
| 126 | "src/f32-dwconv/gen/up2x25-scalar.c", |
| 127 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc2.c", |
| 128 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc3.c", |
| 129 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1-acc4.c", |
| 130 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-1x1.c", |
| 131 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1-acc2.c", |
| 132 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-2x1.c", |
| 133 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-3x1.c", |
| 134 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-4x1.c", |
| 135 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-5x1.c", |
| 136 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-scalar-6x1.c", |
| 137 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc2.c", |
| 138 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc3.c", |
| 139 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1-acc4.c", |
| 140 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-1x1.c", |
| 141 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1-acc2.c", |
| 142 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-2x1.c", |
| 143 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-3x1.c", |
| 144 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-scalar-4x1.c", |
| 145 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc2.c", |
| 146 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc3.c", |
| 147 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc4.c", |
| 148 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1-acc5.c", |
| 149 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-1x1.c", |
| 150 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc2.c", |
| 151 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1-acc3.c", |
| 152 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-2x1.c", |
| 153 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1-acc2.c", |
| 154 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-scalar-3x1.c", |
| 155 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc2.c", |
| 156 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc3.c", |
| 157 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc4.c", |
| 158 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1-acc5.c", |
| 159 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-1x1.c", |
| 160 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc2.c", |
| 161 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1-acc3.c", |
| 162 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-2x1.c", |
| 163 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1-acc2.c", |
| 164 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-scalar-3x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 165 | "src/f32-gavgpool-cw/scalar-x1.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 166 | "src/f32-gavgpool/7p7x-minmax-scalar-c1.c", |
| 167 | "src/f32-gavgpool/7x-minmax-scalar-c1.c", |
| 168 | "src/f32-gemm/gen-inc/1x4inc-minmax-scalar.c", |
| 169 | "src/f32-gemm/gen-inc/2x4inc-minmax-scalar.c", |
| 170 | "src/f32-gemm/gen-inc/4x4inc-minmax-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 171 | "src/f32-gemm/gen/1x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 172 | "src/f32-gemm/gen/1x4-relu-scalar.c", |
| 173 | "src/f32-gemm/gen/1x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 174 | "src/f32-gemm/gen/2x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 175 | "src/f32-gemm/gen/2x4-relu-scalar.c", |
| 176 | "src/f32-gemm/gen/2x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 177 | "src/f32-gemm/gen/4x2-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 178 | "src/f32-gemm/gen/4x2-relu-scalar.c", |
| 179 | "src/f32-gemm/gen/4x2-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 180 | "src/f32-gemm/gen/4x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 181 | "src/f32-gemm/gen/4x4-relu-scalar.c", |
| 182 | "src/f32-gemm/gen/4x4-scalar.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 183 | "src/f32-hswish/gen/hswish-scalar-x1.c", |
| 184 | "src/f32-hswish/gen/hswish-scalar-x2.c", |
| 185 | "src/f32-hswish/gen/hswish-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 186 | "src/f32-ibilinear-chw/gen/scalar-p1.c", |
| 187 | "src/f32-ibilinear-chw/gen/scalar-p2.c", |
| 188 | "src/f32-ibilinear-chw/gen/scalar-p4.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 189 | "src/f32-ibilinear/gen/scalar-c1.c", |
| 190 | "src/f32-ibilinear/gen/scalar-c2.c", |
| 191 | "src/f32-ibilinear/gen/scalar-c4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 192 | "src/f32-igemm/gen/1x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 193 | "src/f32-igemm/gen/1x4-relu-scalar.c", |
| 194 | "src/f32-igemm/gen/1x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 195 | "src/f32-igemm/gen/2x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 196 | "src/f32-igemm/gen/2x4-relu-scalar.c", |
| 197 | "src/f32-igemm/gen/2x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 198 | "src/f32-igemm/gen/4x2-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 199 | "src/f32-igemm/gen/4x2-relu-scalar.c", |
| 200 | "src/f32-igemm/gen/4x2-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 201 | "src/f32-igemm/gen/4x4-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 202 | "src/f32-igemm/gen/4x4-relu-scalar.c", |
| 203 | "src/f32-igemm/gen/4x4-scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 204 | "src/f32-maxpool/9p8x-minmax-scalar-c1.c", |
| 205 | "src/f32-pavgpool/9p8x-minmax-scalar-c1.c", |
| 206 | "src/f32-pavgpool/9x-minmax-scalar-c1.c", |
| 207 | "src/f32-ppmm/gen/2x4-minmax-scalar.c", |
| 208 | "src/f32-ppmm/gen/3x3-minmax-scalar.c", |
| 209 | "src/f32-ppmm/gen/4x2-minmax-scalar.c", |
| 210 | "src/f32-ppmm/gen/4x4-minmax-scalar.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 211 | "src/f32-prelu/gen/scalar-2x1.c", |
| 212 | "src/f32-prelu/gen/scalar-2x4.c", |
| 213 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x1.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 214 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 215 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x2.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 216 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc2.c", |
| 217 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 218 | "src/f32-raddstoreexpminusmax/gen/scalar-lut64-p2-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 219 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x1.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 220 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 221 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x2.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 222 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc2.c", |
| 223 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 224 | "src/f32-raddstoreexpminusmax/gen/scalar-p5-x4.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 225 | "src/f32-relu/gen/scalar-x1.c", |
| 226 | "src/f32-relu/gen/scalar-x2.c", |
| 227 | "src/f32-relu/gen/scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 228 | "src/f32-relu/gen/scalar-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 229 | "src/f32-rmax/scalar.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 230 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x1.c", |
| 231 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x2.c", |
| 232 | "src/f32-sigmoid/gen/scalar-lut64-p2-div-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 233 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x1.c", |
| 234 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x2.c", |
| 235 | "src/f32-sigmoid/gen/scalar-lut2048-p1-div-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 236 | "src/f32-sigmoid/gen/scalar-p5-div-x1.c", |
| 237 | "src/f32-sigmoid/gen/scalar-p5-div-x2.c", |
| 238 | "src/f32-sigmoid/gen/scalar-p5-div-x4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 239 | "src/f32-spmm/gen/1x1-minmax-scalar-pipelined.c", |
| 240 | "src/f32-spmm/gen/1x1-minmax-scalar.c", |
| 241 | "src/f32-spmm/gen/2x1-minmax-scalar-pipelined.c", |
| 242 | "src/f32-spmm/gen/2x1-minmax-scalar.c", |
| 243 | "src/f32-spmm/gen/4x1-minmax-scalar-pipelined.c", |
| 244 | "src/f32-spmm/gen/4x1-minmax-scalar.c", |
| 245 | "src/f32-spmm/gen/8x1-minmax-scalar-pipelined.c", |
| 246 | "src/f32-spmm/gen/8x1-minmax-scalar.c", |
| 247 | "src/f32-spmm/gen/8x2-minmax-scalar.c", |
| 248 | "src/f32-spmm/gen/8x4-minmax-scalar.c", |
| 249 | "src/f32-vbinary/gen/vadd-minmax-scalar-x1.c", |
| 250 | "src/f32-vbinary/gen/vadd-minmax-scalar-x2.c", |
| 251 | "src/f32-vbinary/gen/vadd-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 252 | "src/f32-vbinary/gen/vadd-minmax-scalar-x8.c", |
| 253 | "src/f32-vbinary/gen/vadd-relu-scalar-x1.c", |
| 254 | "src/f32-vbinary/gen/vadd-relu-scalar-x2.c", |
| 255 | "src/f32-vbinary/gen/vadd-relu-scalar-x4.c", |
| 256 | "src/f32-vbinary/gen/vadd-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 257 | "src/f32-vbinary/gen/vadd-scalar-x1.c", |
| 258 | "src/f32-vbinary/gen/vadd-scalar-x2.c", |
| 259 | "src/f32-vbinary/gen/vadd-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 260 | "src/f32-vbinary/gen/vadd-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 261 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x1.c", |
| 262 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x2.c", |
| 263 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 264 | "src/f32-vbinary/gen/vaddc-minmax-scalar-x8.c", |
| 265 | "src/f32-vbinary/gen/vaddc-relu-scalar-x1.c", |
| 266 | "src/f32-vbinary/gen/vaddc-relu-scalar-x2.c", |
| 267 | "src/f32-vbinary/gen/vaddc-relu-scalar-x4.c", |
| 268 | "src/f32-vbinary/gen/vaddc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 269 | "src/f32-vbinary/gen/vaddc-scalar-x1.c", |
| 270 | "src/f32-vbinary/gen/vaddc-scalar-x2.c", |
| 271 | "src/f32-vbinary/gen/vaddc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 272 | "src/f32-vbinary/gen/vaddc-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 273 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x1.c", |
| 274 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x2.c", |
| 275 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 276 | "src/f32-vbinary/gen/vdiv-minmax-scalar-x8.c", |
| 277 | "src/f32-vbinary/gen/vdiv-relu-scalar-x1.c", |
| 278 | "src/f32-vbinary/gen/vdiv-relu-scalar-x2.c", |
| 279 | "src/f32-vbinary/gen/vdiv-relu-scalar-x4.c", |
| 280 | "src/f32-vbinary/gen/vdiv-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 281 | "src/f32-vbinary/gen/vdiv-scalar-x1.c", |
| 282 | "src/f32-vbinary/gen/vdiv-scalar-x2.c", |
| 283 | "src/f32-vbinary/gen/vdiv-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 284 | "src/f32-vbinary/gen/vdiv-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 285 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x1.c", |
| 286 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x2.c", |
| 287 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 288 | "src/f32-vbinary/gen/vdivc-minmax-scalar-x8.c", |
| 289 | "src/f32-vbinary/gen/vdivc-relu-scalar-x1.c", |
| 290 | "src/f32-vbinary/gen/vdivc-relu-scalar-x2.c", |
| 291 | "src/f32-vbinary/gen/vdivc-relu-scalar-x4.c", |
| 292 | "src/f32-vbinary/gen/vdivc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 293 | "src/f32-vbinary/gen/vdivc-scalar-x1.c", |
| 294 | "src/f32-vbinary/gen/vdivc-scalar-x2.c", |
| 295 | "src/f32-vbinary/gen/vdivc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 296 | "src/f32-vbinary/gen/vdivc-scalar-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 297 | "src/f32-vbinary/gen/vmax-scalar-x1.c", |
| 298 | "src/f32-vbinary/gen/vmax-scalar-x2.c", |
| 299 | "src/f32-vbinary/gen/vmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 300 | "src/f32-vbinary/gen/vmax-scalar-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 301 | "src/f32-vbinary/gen/vmaxc-scalar-x1.c", |
| 302 | "src/f32-vbinary/gen/vmaxc-scalar-x2.c", |
| 303 | "src/f32-vbinary/gen/vmaxc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 304 | "src/f32-vbinary/gen/vmaxc-scalar-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 305 | "src/f32-vbinary/gen/vmin-scalar-x1.c", |
| 306 | "src/f32-vbinary/gen/vmin-scalar-x2.c", |
| 307 | "src/f32-vbinary/gen/vmin-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 308 | "src/f32-vbinary/gen/vmin-scalar-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 309 | "src/f32-vbinary/gen/vminc-scalar-x1.c", |
| 310 | "src/f32-vbinary/gen/vminc-scalar-x2.c", |
| 311 | "src/f32-vbinary/gen/vminc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 312 | "src/f32-vbinary/gen/vminc-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 313 | "src/f32-vbinary/gen/vmul-minmax-scalar-x1.c", |
| 314 | "src/f32-vbinary/gen/vmul-minmax-scalar-x2.c", |
| 315 | "src/f32-vbinary/gen/vmul-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 316 | "src/f32-vbinary/gen/vmul-minmax-scalar-x8.c", |
| 317 | "src/f32-vbinary/gen/vmul-relu-scalar-x1.c", |
| 318 | "src/f32-vbinary/gen/vmul-relu-scalar-x2.c", |
| 319 | "src/f32-vbinary/gen/vmul-relu-scalar-x4.c", |
| 320 | "src/f32-vbinary/gen/vmul-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 321 | "src/f32-vbinary/gen/vmul-scalar-x1.c", |
| 322 | "src/f32-vbinary/gen/vmul-scalar-x2.c", |
| 323 | "src/f32-vbinary/gen/vmul-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 324 | "src/f32-vbinary/gen/vmul-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 325 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x1.c", |
| 326 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x2.c", |
| 327 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 328 | "src/f32-vbinary/gen/vmulc-minmax-scalar-x8.c", |
| 329 | "src/f32-vbinary/gen/vmulc-relu-scalar-x1.c", |
| 330 | "src/f32-vbinary/gen/vmulc-relu-scalar-x2.c", |
| 331 | "src/f32-vbinary/gen/vmulc-relu-scalar-x4.c", |
| 332 | "src/f32-vbinary/gen/vmulc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 333 | "src/f32-vbinary/gen/vmulc-scalar-x1.c", |
| 334 | "src/f32-vbinary/gen/vmulc-scalar-x2.c", |
| 335 | "src/f32-vbinary/gen/vmulc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 336 | "src/f32-vbinary/gen/vmulc-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 337 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x1.c", |
| 338 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x2.c", |
| 339 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 340 | "src/f32-vbinary/gen/vrdivc-minmax-scalar-x8.c", |
| 341 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x1.c", |
| 342 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x2.c", |
| 343 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x4.c", |
| 344 | "src/f32-vbinary/gen/vrdivc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 345 | "src/f32-vbinary/gen/vrdivc-scalar-x1.c", |
| 346 | "src/f32-vbinary/gen/vrdivc-scalar-x2.c", |
| 347 | "src/f32-vbinary/gen/vrdivc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 348 | "src/f32-vbinary/gen/vrdivc-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 349 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x1.c", |
| 350 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x2.c", |
| 351 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 352 | "src/f32-vbinary/gen/vrsubc-minmax-scalar-x8.c", |
| 353 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x1.c", |
| 354 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x2.c", |
| 355 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x4.c", |
| 356 | "src/f32-vbinary/gen/vrsubc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 357 | "src/f32-vbinary/gen/vrsubc-scalar-x1.c", |
| 358 | "src/f32-vbinary/gen/vrsubc-scalar-x2.c", |
| 359 | "src/f32-vbinary/gen/vrsubc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 360 | "src/f32-vbinary/gen/vrsubc-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 361 | "src/f32-vbinary/gen/vsqrdiff-scalar-x1.c", |
| 362 | "src/f32-vbinary/gen/vsqrdiff-scalar-x2.c", |
| 363 | "src/f32-vbinary/gen/vsqrdiff-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 364 | "src/f32-vbinary/gen/vsqrdiff-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 365 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x1.c", |
| 366 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x2.c", |
| 367 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 368 | "src/f32-vbinary/gen/vsqrdiffc-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 369 | "src/f32-vbinary/gen/vsub-minmax-scalar-x1.c", |
| 370 | "src/f32-vbinary/gen/vsub-minmax-scalar-x2.c", |
| 371 | "src/f32-vbinary/gen/vsub-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 372 | "src/f32-vbinary/gen/vsub-minmax-scalar-x8.c", |
| 373 | "src/f32-vbinary/gen/vsub-relu-scalar-x1.c", |
| 374 | "src/f32-vbinary/gen/vsub-relu-scalar-x2.c", |
| 375 | "src/f32-vbinary/gen/vsub-relu-scalar-x4.c", |
| 376 | "src/f32-vbinary/gen/vsub-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 377 | "src/f32-vbinary/gen/vsub-scalar-x1.c", |
| 378 | "src/f32-vbinary/gen/vsub-scalar-x2.c", |
| 379 | "src/f32-vbinary/gen/vsub-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 380 | "src/f32-vbinary/gen/vsub-scalar-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 381 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x1.c", |
| 382 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x2.c", |
| 383 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 384 | "src/f32-vbinary/gen/vsubc-minmax-scalar-x8.c", |
| 385 | "src/f32-vbinary/gen/vsubc-relu-scalar-x1.c", |
| 386 | "src/f32-vbinary/gen/vsubc-relu-scalar-x2.c", |
| 387 | "src/f32-vbinary/gen/vsubc-relu-scalar-x4.c", |
| 388 | "src/f32-vbinary/gen/vsubc-relu-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 389 | "src/f32-vbinary/gen/vsubc-scalar-x1.c", |
| 390 | "src/f32-vbinary/gen/vsubc-scalar-x2.c", |
| 391 | "src/f32-vbinary/gen/vsubc-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 392 | "src/f32-vbinary/gen/vsubc-scalar-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 393 | "src/f32-vlrelu/gen/vlrelu-scalar-x1.c", |
| 394 | "src/f32-vlrelu/gen/vlrelu-scalar-x2.c", |
| 395 | "src/f32-vlrelu/gen/vlrelu-scalar-x4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 396 | "src/f32-vmulcaddc/gen/c1-minmax-scalar-2x.c", |
| 397 | "src/f32-vmulcaddc/gen/c2-minmax-scalar-2x.c", |
| 398 | "src/f32-vmulcaddc/gen/c4-minmax-scalar-2x.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 399 | "src/f32-vrnd/gen/vrndd-scalar-libm-x1.c", |
| 400 | "src/f32-vrnd/gen/vrndd-scalar-libm-x2.c", |
| 401 | "src/f32-vrnd/gen/vrndd-scalar-libm-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 402 | "src/f32-vrnd/gen/vrndne-scalar-libm-x1.c", |
| 403 | "src/f32-vrnd/gen/vrndne-scalar-libm-x2.c", |
| 404 | "src/f32-vrnd/gen/vrndne-scalar-libm-x4.c", |
| 405 | "src/f32-vrnd/gen/vrndu-scalar-libm-x1.c", |
| 406 | "src/f32-vrnd/gen/vrndu-scalar-libm-x2.c", |
| 407 | "src/f32-vrnd/gen/vrndu-scalar-libm-x4.c", |
| 408 | "src/f32-vrnd/gen/vrndz-scalar-libm-x1.c", |
| 409 | "src/f32-vrnd/gen/vrndz-scalar-libm-x2.c", |
| 410 | "src/f32-vrnd/gen/vrndz-scalar-libm-x4.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 411 | "src/f32-vsqrt/gen/scalar-sqrt-x1.c", |
| 412 | "src/f32-vsqrt/gen/scalar-sqrt-x2.c", |
| 413 | "src/f32-vsqrt/gen/scalar-sqrt-x4.c", |
| 414 | "src/f32-vunary/gen/vabs-scalar-x1.c", |
| 415 | "src/f32-vunary/gen/vabs-scalar-x2.c", |
| 416 | "src/f32-vunary/gen/vabs-scalar-x4.c", |
| 417 | "src/f32-vunary/gen/vneg-scalar-x1.c", |
| 418 | "src/f32-vunary/gen/vneg-scalar-x2.c", |
| 419 | "src/f32-vunary/gen/vneg-scalar-x4.c", |
| 420 | "src/f32-vunary/gen/vsqr-scalar-x1.c", |
| 421 | "src/f32-vunary/gen/vsqr-scalar-x2.c", |
| 422 | "src/f32-vunary/gen/vsqr-scalar-x4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 423 | "src/math/expm1minus-scalar-rr2-lut4-p4.c", |
| 424 | "src/math/expm1minus-scalar-rr2-lut8-p3.c", |
| 425 | "src/math/expm1minus-scalar-rr2-lut8-p4.c", |
| 426 | "src/math/expm1minus-scalar-rr2-lut16-p3.c", |
| 427 | "src/math/expm1minus-scalar-rr2-lut16-p4.c", |
| 428 | "src/math/expm1minus-scalar-rr2-p5.c", |
| 429 | "src/math/expm1minus-scalar-rr2-p6.c", |
| 430 | "src/math/expminus-scalar-rr2-lut64-p2.c", |
| 431 | "src/math/expminus-scalar-rr2-lut2048-p1.c", |
| 432 | "src/math/expminus-scalar-rr2-p5.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 433 | "src/math/roundd-scalar-addsub.c", |
| 434 | "src/math/roundd-scalar-cvt.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 435 | "src/math/roundd-scalar-floor.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 436 | "src/math/roundne-scalar-addsub.c", |
| 437 | "src/math/roundne-scalar-nearbyint.c", |
| 438 | "src/math/roundne-scalar-rint.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 439 | "src/math/roundu-scalar-addsub.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 440 | "src/math/roundu-scalar-ceil.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 441 | "src/math/roundu-scalar-cvt.c", |
| 442 | "src/math/roundz-scalar-addsub.c", |
| 443 | "src/math/roundz-scalar-cvt.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 444 | "src/math/roundz-scalar-trunc.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 445 | "src/math/sigmoid-scalar-rr2-lut64-p2-div.c", |
| 446 | "src/math/sigmoid-scalar-rr2-lut2048-p1-div.c", |
| 447 | "src/math/sigmoid-scalar-rr2-p5-div.c", |
| 448 | "src/qs8-gemm/gen/8x8c4-minmax-scalar.c", |
| 449 | "src/qs8-gemm/gen/12x4c4-minmax-scalar.c", |
| 450 | "src/qs8-requantization/fp32-scalar-lrintf.c", |
| 451 | "src/qs8-requantization/fp32-scalar-magic.c", |
| 452 | "src/qs8-requantization/precise-scalar-signed64.c", |
| 453 | "src/qs8-requantization/precise-scalar-unsigned32.c", |
| 454 | "src/qs8-requantization/precise-scalar-unsigned64.c", |
| 455 | "src/qs8-requantization/q31-scalar.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 456 | "src/qu8-avgpool/9p8x-minmax-scalar-c1.c", |
| 457 | "src/qu8-avgpool/9x-minmax-scalar-c1.c", |
| 458 | "src/qu8-dwconv/up1x9-minmax-scalar.c", |
| 459 | "src/qu8-gavgpool/7p7x-minmax-scalar-c1.c", |
| 460 | "src/qu8-gavgpool/7x-minmax-scalar-c1.c", |
| 461 | "src/qu8-gemm/2x2-minmax-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 462 | "src/qu8-gemm/gen/8x8c4-minmax-scalar.c", |
| 463 | "src/qu8-gemm/gen/12x4c4-minmax-scalar.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 464 | "src/qu8-igemm/2x2-minmax-scalar.c", |
| 465 | "src/qu8-requantization/fp32-scalar-lrintf.c", |
| 466 | "src/qu8-requantization/fp32-scalar-magic.c", |
| 467 | "src/qu8-requantization/precise-scalar-signed64.c", |
| 468 | "src/qu8-requantization/precise-scalar-unsigned32.c", |
| 469 | "src/qu8-requantization/precise-scalar-unsigned64.c", |
| 470 | "src/qu8-requantization/q31-scalar.c", |
| 471 | "src/qu8-vadd/minmax-scalar.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 472 | "src/u8-clamp/scalar-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 473 | "src/u8-lut32norm/scalar.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 474 | "src/u8-maxpool/9p8x-minmax-scalar-c1.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 475 | "src/u8-rmax/scalar.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 476 | "src/x8-lut/scalar.c", |
| 477 | "src/x8-zip/x2-scalar.c", |
| 478 | "src/x8-zip/x3-scalar.c", |
| 479 | "src/x8-zip/x4-scalar.c", |
| 480 | "src/x8-zip/xm-scalar.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 481 | "src/x32-depthtospace2d-chw2hwc/scalar.c", |
| 482 | "src/x32-fill/scalar-float.c", |
| 483 | "src/x32-fill/scalar-int.c", |
| 484 | "src/x32-packx/x2-scalar.c", |
| 485 | "src/x32-packx/x3-scalar.c", |
| 486 | "src/x32-packx/x4-scalar.c", |
| 487 | "src/x32-pad/scalar-float.c", |
| 488 | "src/x32-pad/scalar-int.c", |
| 489 | "src/x32-unpool/scalar.c", |
| 490 | "src/x32-zip/x2-scalar.c", |
| 491 | "src/x32-zip/x3-scalar.c", |
| 492 | "src/x32-zip/x4-scalar.c", |
| 493 | "src/x32-zip/xm-scalar.c", |
| 494 | "src/xx-copy/memcpy.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 495 | ] |
| 496 | |
| 497 | PSIMD_ACCMATH_UKERNELS = [ |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 498 | "src/qs8-requantization/fp32-psimd.c", |
| 499 | "src/qs8-requantization/precise-psimd.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 500 | "src/qu8-requantization/fp32-psimd.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 501 | "src/qu8-requantization/precise-psimd.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 502 | ] |
| 503 | |
| 504 | // ISA-specific micro-kernels |
| 505 | NEON_UKERNELS = [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 506 | "src/f32-argmaxpool/4x-neon-c4.c", |
| 507 | "src/f32-argmaxpool/9p8x-neon-c4.c", |
| 508 | "src/f32-argmaxpool/9x-neon-c4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 509 | "src/f32-avgpool/9p8x-minmax-neon-c4.c", |
| 510 | "src/f32-avgpool/9x-minmax-neon-c4.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 511 | "src/f32-clamp/gen/neon-x4.c", |
| 512 | "src/f32-clamp/gen/neon-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 513 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 514 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neon-2x2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 515 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 516 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neon-2x2.c", |
| 517 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 518 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neon-2x2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 519 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x1.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 520 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neon-2x2.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 521 | "src/f32-dwconv/gen/up4x4-minmax-neon-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 522 | "src/f32-dwconv/gen/up4x4-minmax-neon.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 523 | "src/f32-dwconv/gen/up4x9-minmax-neon-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 524 | "src/f32-dwconv/gen/up4x9-minmax-neon.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 525 | "src/f32-dwconv/gen/up4x25-minmax-neon-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 526 | "src/f32-dwconv/gen/up4x25-minmax-neon.c", |
| 527 | "src/f32-dwconv/gen/up8x4-minmax-neon-acc2.c", |
| 528 | "src/f32-dwconv/gen/up8x4-minmax-neon.c", |
| 529 | "src/f32-dwconv/gen/up8x9-minmax-neon-acc2.c", |
| 530 | "src/f32-dwconv/gen/up8x9-minmax-neon.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 531 | "src/f32-dwconv/gen/up8x25-minmax-neon-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 532 | "src/f32-dwconv/gen/up8x25-minmax-neon.c", |
| 533 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc2.c", |
| 534 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc3.c", |
| 535 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4-acc4.c", |
| 536 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-1x4.c", |
| 537 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4-acc2.c", |
| 538 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-2x4.c", |
| 539 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-3x4.c", |
| 540 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-4x4.c", |
| 541 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-5x4.c", |
| 542 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neon-6x4.c", |
| 543 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc2.c", |
| 544 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc3.c", |
| 545 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4-acc4.c", |
| 546 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-1x4.c", |
| 547 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4-acc2.c", |
| 548 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-2x4.c", |
| 549 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-3x4.c", |
| 550 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neon-4x4.c", |
| 551 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc2.c", |
| 552 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc3.c", |
| 553 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc4.c", |
| 554 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4-acc5.c", |
| 555 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-1x4.c", |
| 556 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc2.c", |
| 557 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4-acc3.c", |
| 558 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-2x4.c", |
| 559 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4-acc2.c", |
| 560 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-3x4.c", |
| 561 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4-acc2.c", |
| 562 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-4x4.c", |
| 563 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neon-5x4.c", |
| 564 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc2.c", |
| 565 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc3.c", |
| 566 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc4.c", |
| 567 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4-acc5.c", |
| 568 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-1x4.c", |
| 569 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc2.c", |
| 570 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4-acc3.c", |
| 571 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-2x4.c", |
| 572 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4-acc2.c", |
| 573 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neon-3x4.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 574 | "src/f32-gavgpool-cw/neon-x4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 575 | "src/f32-gavgpool/7p7x-minmax-neon-c4.c", |
| 576 | "src/f32-gavgpool/7x-minmax-neon-c4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 577 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-dup-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 578 | "src/f32-gemm/gen-inc/1x8inc-minmax-neon-lane-ld64.c", |
| 579 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neon.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 580 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 581 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-dup-ld128.c", |
| 582 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld64.c", |
| 583 | "src/f32-gemm/gen-inc/4x8inc-minmax-neon-lane-ld128.c", |
| 584 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neon.c", |
| 585 | "src/f32-gemm/gen-inc/5x8inc-minmax-neon-lane-ld64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 586 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld64.c", |
| 587 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-dup-ld128.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 588 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld64.c", |
| 589 | "src/f32-gemm/gen-inc/6x8inc-minmax-neon-lane-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 590 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neon.c", |
| 591 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neon.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 592 | "src/f32-gemm/gen/1x8-minmax-neon-dup-ld64.c", |
| 593 | "src/f32-gemm/gen/1x8-minmax-neon-lane-ld64.c", |
| 594 | "src/f32-gemm/gen/1x8s4-minmax-neon.c", |
| 595 | "src/f32-gemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 596 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld64.c", |
| 597 | "src/f32-gemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 598 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 599 | "src/f32-gemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 600 | "src/f32-gemm/gen/4x8s4-minmax-neon.c", |
| 601 | "src/f32-gemm/gen/5x8-minmax-neon-lane-ld64.c", |
| 602 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 603 | "src/f32-gemm/gen/6x8-minmax-neon-dup-ld128.c", |
| 604 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 605 | "src/f32-gemm/gen/6x8-minmax-neon-lane-ld128.c", |
| 606 | "src/f32-gemm/gen/6x8s4-minmax-neon.c", |
| 607 | "src/f32-gemm/gen/8x8s4-minmax-neon.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 608 | "src/f32-hswish/gen/hswish-neon-x4.c", |
| 609 | "src/f32-hswish/gen/hswish-neon-x8.c", |
| 610 | "src/f32-hswish/gen/hswish-neon-x16.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 611 | "src/f32-ibilinear/gen/neon-c4.c", |
| 612 | "src/f32-ibilinear/gen/neon-c8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 613 | "src/f32-igemm/gen/1x8-minmax-neon-dup-ld64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 614 | "src/f32-igemm/gen/1x8-minmax-neon-lane-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 615 | "src/f32-igemm/gen/1x8s4-minmax-neon.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 616 | "src/f32-igemm/gen/4x2-minmax-neon-lane-ld64.c", |
| 617 | "src/f32-igemm/gen/4x4-minmax-neon-lane-ld64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 618 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 619 | "src/f32-igemm/gen/4x8-minmax-neon-dup-ld128.c", |
| 620 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld64.c", |
| 621 | "src/f32-igemm/gen/4x8-minmax-neon-lane-ld128.c", |
| 622 | "src/f32-igemm/gen/4x8s4-minmax-neon.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 623 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld64.c", |
| 624 | "src/f32-igemm/gen/6x8-minmax-neon-dup-ld128.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 625 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld64.c", |
| 626 | "src/f32-igemm/gen/6x8-minmax-neon-lane-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 627 | "src/f32-igemm/gen/6x8s4-minmax-neon.c", |
| 628 | "src/f32-igemm/gen/8x8s4-minmax-neon.c", |
| 629 | "src/f32-maxpool/9p8x-minmax-neon-c4.c", |
| 630 | "src/f32-pavgpool/9p8x-minmax-neon-c4.c", |
| 631 | "src/f32-pavgpool/9x-minmax-neon-c4.c", |
| 632 | "src/f32-ppmm/gen/4x8-minmax-neon.c", |
| 633 | "src/f32-ppmm/gen/8x8-minmax-neon.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 634 | "src/f32-prelu/gen/neon-1x4.c", |
| 635 | "src/f32-prelu/gen/neon-1x8.c", |
| 636 | "src/f32-prelu/gen/neon-1x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 637 | "src/f32-prelu/gen/neon-2x4.c", |
| 638 | "src/f32-prelu/gen/neon-2x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 639 | "src/f32-prelu/gen/neon-2x16.c", |
| 640 | "src/f32-prelu/gen/neon-4x4.c", |
| 641 | "src/f32-prelu/gen/neon-4x8.c", |
| 642 | "src/f32-prelu/gen/neon-4x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 643 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 644 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 645 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 646 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc2.c", |
| 647 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 648 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x12.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 649 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc2.c", |
| 650 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 651 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 652 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc2.c", |
| 653 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 654 | "src/f32-raddstoreexpminusmax/gen/neon-lut64-p2-x20.c", |
| 655 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x4.c", |
| 656 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8-acc2.c", |
| 657 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x8.c", |
| 658 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc2.c", |
| 659 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12-acc3.c", |
| 660 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x12.c", |
| 661 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc2.c", |
| 662 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16-acc4.c", |
| 663 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x16.c", |
| 664 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc2.c", |
| 665 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20-acc5.c", |
| 666 | "src/f32-raddstoreexpminusmax/gen/neon-p5-x20.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 667 | "src/f32-relu/gen/neon-x4.c", |
| 668 | "src/f32-relu/gen/neon-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 669 | "src/f32-rmax/neon.c", |
| 670 | "src/f32-sigmoid/gen/neon-frac-p9-p10-nr1recps-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 671 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x4.c", |
| 672 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x8.c", |
| 673 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x12.c", |
| 674 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x16.c", |
| 675 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x20.c", |
| 676 | "src/f32-sigmoid/gen/neon-rr2-lut64-p2-nr2recps-x24.c", |
| 677 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x4.c", |
| 678 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x8.c", |
| 679 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x12.c", |
| 680 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x16.c", |
| 681 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x20.c", |
| 682 | "src/f32-sigmoid/gen/neon-rr2-lut2048-p1-nr2recps-x24.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 683 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x4.c", |
| 684 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x8.c", |
| 685 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x12.c", |
| 686 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x16.c", |
| 687 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x20.c", |
| 688 | "src/f32-sigmoid/gen/neon-rr2-p5-nr2recps-x24.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 689 | "src/f32-vbinary/gen/vadd-minmax-neon-x4.c", |
| 690 | "src/f32-vbinary/gen/vadd-minmax-neon-x8.c", |
| 691 | "src/f32-vbinary/gen/vaddc-minmax-neon-x4.c", |
| 692 | "src/f32-vbinary/gen/vaddc-minmax-neon-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 693 | "src/f32-vbinary/gen/vmax-neon-x4.c", |
| 694 | "src/f32-vbinary/gen/vmax-neon-x8.c", |
| 695 | "src/f32-vbinary/gen/vmaxc-neon-x4.c", |
| 696 | "src/f32-vbinary/gen/vmaxc-neon-x8.c", |
| 697 | "src/f32-vbinary/gen/vmin-neon-x4.c", |
| 698 | "src/f32-vbinary/gen/vmin-neon-x8.c", |
| 699 | "src/f32-vbinary/gen/vminc-neon-x4.c", |
| 700 | "src/f32-vbinary/gen/vminc-neon-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 701 | "src/f32-vbinary/gen/vmul-minmax-neon-x4.c", |
| 702 | "src/f32-vbinary/gen/vmul-minmax-neon-x8.c", |
| 703 | "src/f32-vbinary/gen/vmulc-minmax-neon-x4.c", |
| 704 | "src/f32-vbinary/gen/vmulc-minmax-neon-x8.c", |
| 705 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x4.c", |
| 706 | "src/f32-vbinary/gen/vrsubc-minmax-neon-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 707 | "src/f32-vbinary/gen/vsqrdiff-neon-x4.c", |
| 708 | "src/f32-vbinary/gen/vsqrdiff-neon-x8.c", |
| 709 | "src/f32-vbinary/gen/vsqrdiffc-neon-x4.c", |
| 710 | "src/f32-vbinary/gen/vsqrdiffc-neon-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 711 | "src/f32-vbinary/gen/vsub-minmax-neon-x4.c", |
| 712 | "src/f32-vbinary/gen/vsub-minmax-neon-x8.c", |
| 713 | "src/f32-vbinary/gen/vsubc-minmax-neon-x4.c", |
| 714 | "src/f32-vbinary/gen/vsubc-minmax-neon-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 715 | "src/f32-vlrelu/gen/vlrelu-neon-x4.c", |
| 716 | "src/f32-vlrelu/gen/vlrelu-neon-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 717 | "src/f32-vmulcaddc/gen/c4-minmax-neon-2x.c", |
| 718 | "src/f32-vmulcaddc/gen/c8-minmax-neon-2x.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 719 | "src/f32-vrnd/gen/vrndd-neon-x4.c", |
| 720 | "src/f32-vrnd/gen/vrndd-neon-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 721 | "src/f32-vrnd/gen/vrndne-neon-x4.c", |
| 722 | "src/f32-vrnd/gen/vrndne-neon-x8.c", |
| 723 | "src/f32-vrnd/gen/vrndu-neon-x4.c", |
| 724 | "src/f32-vrnd/gen/vrndu-neon-x8.c", |
| 725 | "src/f32-vrnd/gen/vrndz-neon-x4.c", |
| 726 | "src/f32-vrnd/gen/vrndz-neon-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 727 | "src/f32-vunary/gen/vabs-neon-x4.c", |
| 728 | "src/f32-vunary/gen/vabs-neon-x8.c", |
| 729 | "src/f32-vunary/gen/vneg-neon-x4.c", |
| 730 | "src/f32-vunary/gen/vneg-neon-x8.c", |
| 731 | "src/f32-vunary/gen/vsqr-neon-x4.c", |
| 732 | "src/f32-vunary/gen/vsqr-neon-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 733 | "src/math/expm1minus-neon-rr2-lut16-p3.c", |
| 734 | "src/math/expm1minus-neon-rr2-p6.c", |
| 735 | "src/math/roundd-neon-addsub.c", |
| 736 | "src/math/roundd-neon-cvt.c", |
| 737 | "src/math/roundne-neon-addsub.c", |
| 738 | "src/math/roundu-neon-addsub.c", |
| 739 | "src/math/roundu-neon-cvt.c", |
| 740 | "src/math/roundz-neon-addsub.c", |
| 741 | "src/math/roundz-neon-cvt.c", |
| 742 | "src/math/sigmoid-neon-frac-p9-p10-nr1recps.c", |
| 743 | "src/math/sigmoid-neon-rr1-lut64-p2-nr2recps.c", |
| 744 | "src/math/sigmoid-neon-rr1-lut2048-p1-nr2recps.c", |
| 745 | "src/math/sigmoid-neon-rr1-p5-nr2recps.c", |
| 746 | "src/math/sigmoid-neon-rr2-lut64-p2-nr2recps.c", |
| 747 | "src/math/sigmoid-neon-rr2-lut2048-p1-nr2recps.c", |
| 748 | "src/math/sigmoid-neon-rr2-p5-nr2recps.c", |
| 749 | "src/math/sqrt-neon-nr1rsqrts.c", |
| 750 | "src/math/sqrt-neon-nr2rsqrts.c", |
| 751 | "src/math/sqrt-neon-nr3rsqrts.c", |
| 752 | "src/qs8-dwconv/gen/up8x9-minmax-neon-mul16.c", |
| 753 | "src/qs8-dwconv/gen/up16x9-minmax-neon-mul16.c", |
| 754 | "src/qs8-dwconv/gen/up24x9-minmax-neon-mul16.c", |
| 755 | "src/qs8-dwconv/gen/up32x9-minmax-neon-mul16.c", |
| 756 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c8-acc2.c", |
| 757 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c16-acc2.c", |
| 758 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c24-acc2.c", |
| 759 | "src/qs8-gavgpool/gen/7p7x-minmax-neon-c32-acc2.c", |
| 760 | "src/qs8-gavgpool/gen/7x-minmax-neon-c8-acc2.c", |
| 761 | "src/qs8-gavgpool/gen/7x-minmax-neon-c16-acc2.c", |
| 762 | "src/qs8-gavgpool/gen/7x-minmax-neon-c24-acc2.c", |
| 763 | "src/qs8-gavgpool/gen/7x-minmax-neon-c32-acc2.c", |
| 764 | "src/qs8-gemm/gen/1x8-minmax-neon-mlal-lane.c", |
| 765 | "src/qs8-gemm/gen/1x16-minmax-neon-mlal-lane.c", |
| 766 | "src/qs8-gemm/gen/2x8-minmax-neon-mlal-lane.c", |
| 767 | "src/qs8-gemm/gen/2x16-minmax-neon-mlal-lane.c", |
| 768 | "src/qs8-igemm/gen/1x8-minmax-neon-mlal-lane.c", |
| 769 | "src/qs8-igemm/gen/1x16-minmax-neon-mlal-lane.c", |
| 770 | "src/qs8-igemm/gen/2x8-minmax-neon-mlal-lane.c", |
| 771 | "src/qs8-igemm/gen/2x16-minmax-neon-mlal-lane.c", |
| 772 | "src/qs8-requantization/fp32-neon.c", |
| 773 | "src/qs8-requantization/precise-neon.c", |
| 774 | "src/qs8-requantization/q31-neon.c", |
| 775 | "src/qs8-vadd/gen/minmax-neon-ld64-x8.c", |
| 776 | "src/qs8-vadd/gen/minmax-neon-ld64-x16.c", |
| 777 | "src/qs8-vadd/gen/minmax-neon-ld64-x24.c", |
| 778 | "src/qs8-vadd/gen/minmax-neon-ld64-x32.c", |
| 779 | "src/qs8-vaddc/gen/minmax-neon-ld64-x8.c", |
| 780 | "src/qs8-vaddc/gen/minmax-neon-ld64-x16.c", |
| 781 | "src/qs8-vaddc/gen/minmax-neon-ld64-x24.c", |
| 782 | "src/qs8-vaddc/gen/minmax-neon-ld64-x32.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 783 | "src/qu8-avgpool/9p8x-minmax-neon-c8.c", |
| 784 | "src/qu8-avgpool/9x-minmax-neon-c8.c", |
| 785 | "src/qu8-dwconv/up8x9-minmax-neon.c", |
| 786 | "src/qu8-gavgpool/7p7x-minmax-neon-c8.c", |
| 787 | "src/qu8-gavgpool/7x-minmax-neon-c8.c", |
| 788 | "src/qu8-gemm/4x8-minmax-neon.c", |
| 789 | "src/qu8-gemm/8x8-minmax-neon.c", |
| 790 | "src/qu8-igemm/4x8-minmax-neon.c", |
| 791 | "src/qu8-igemm/8x8-minmax-neon.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 792 | "src/qu8-requantization/fp32-neon.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 793 | "src/qu8-requantization/precise-neon.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 794 | "src/qu8-requantization/q31-neon.c", |
| 795 | "src/qu8-vadd/minmax-neon.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 796 | "src/u8-clamp/neon-x64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 797 | "src/u8-maxpool/9p8x-minmax-neon-c16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 798 | "src/u8-rmax/neon.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 799 | "src/x8-zip/x2-neon.c", |
| 800 | "src/x8-zip/x3-neon.c", |
| 801 | "src/x8-zip/x4-neon.c", |
| 802 | "src/x8-zip/xm-neon.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 803 | "src/x32-fill/neon.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 804 | "src/x32-packx/x4-neon-st4.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 805 | "src/x32-pad/neon.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 806 | "src/x32-unpool/neon.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 807 | "src/x32-zip/x2-neon.c", |
| 808 | "src/x32-zip/x3-neon.c", |
| 809 | "src/x32-zip/x4-neon.c", |
| 810 | "src/x32-zip/xm-neon.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 811 | ] |
| 812 | |
| 813 | NEONFMA_UKERNELS = [ |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 814 | "src/f32-dwconv/gen/up4x4-minmax-neonfma-acc2.c", |
| 815 | "src/f32-dwconv/gen/up4x4-minmax-neonfma.c", |
| 816 | "src/f32-dwconv/gen/up4x9-minmax-neonfma-acc2.c", |
| 817 | "src/f32-dwconv/gen/up4x9-minmax-neonfma.c", |
| 818 | "src/f32-dwconv/gen/up4x25-minmax-neonfma-acc2.c", |
| 819 | "src/f32-dwconv/gen/up4x25-minmax-neonfma.c", |
| 820 | "src/f32-dwconv/gen/up8x4-minmax-neonfma-acc2.c", |
| 821 | "src/f32-dwconv/gen/up8x4-minmax-neonfma.c", |
| 822 | "src/f32-dwconv/gen/up8x9-minmax-neonfma-acc2.c", |
| 823 | "src/f32-dwconv/gen/up8x9-minmax-neonfma.c", |
| 824 | "src/f32-dwconv/gen/up8x25-minmax-neonfma-acc2.c", |
| 825 | "src/f32-dwconv/gen/up8x25-minmax-neonfma.c", |
| 826 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-dup-ld64.c", |
| 827 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-neonfma.c", |
| 828 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld64.c", |
| 829 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-dup-ld128.c", |
| 830 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-neonfma.c", |
| 831 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld64.c", |
| 832 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-dup-ld128.c", |
| 833 | "src/f32-gemm/gen-inc/6x8s4inc-minmax-neonfma.c", |
| 834 | "src/f32-gemm/gen-inc/8x8s4inc-minmax-neonfma.c", |
| 835 | "src/f32-gemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
| 836 | "src/f32-gemm/gen/1x8s4-minmax-neonfma.c", |
| 837 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
| 838 | "src/f32-gemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 839 | "src/f32-gemm/gen/4x8s4-minmax-neonfma.c", |
| 840 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 841 | "src/f32-gemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
| 842 | "src/f32-gemm/gen/6x8s4-minmax-neonfma.c", |
| 843 | "src/f32-gemm/gen/8x8s4-minmax-neonfma.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 844 | "src/f32-ibilinear/gen/neonfma-c4.c", |
| 845 | "src/f32-ibilinear/gen/neonfma-c8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 846 | "src/f32-igemm/gen/1x8-minmax-neonfma-dup-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 847 | "src/f32-igemm/gen/1x8s4-minmax-neonfma.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 848 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 849 | "src/f32-igemm/gen/4x8-minmax-neonfma-dup-ld128.c", |
| 850 | "src/f32-igemm/gen/4x8s4-minmax-neonfma.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 851 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld64.c", |
| 852 | "src/f32-igemm/gen/6x8-minmax-neonfma-dup-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 853 | "src/f32-igemm/gen/6x8s4-minmax-neonfma.c", |
| 854 | "src/f32-igemm/gen/8x8s4-minmax-neonfma.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 855 | "src/f32-ppmm/gen/4x8-minmax-neonfma.c", |
| 856 | "src/f32-ppmm/gen/8x8-minmax-neonfma.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 857 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 858 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 859 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 860 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc2.c", |
| 861 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 862 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x12.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 863 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc2.c", |
| 864 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 865 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 866 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc2.c", |
| 867 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 868 | "src/f32-raddstoreexpminusmax/gen/neonfma-lut64-p2-x20.c", |
| 869 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x4.c", |
| 870 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8-acc2.c", |
| 871 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x8.c", |
| 872 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc2.c", |
| 873 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12-acc3.c", |
| 874 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x12.c", |
| 875 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc2.c", |
| 876 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16-acc4.c", |
| 877 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x16.c", |
| 878 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc2.c", |
| 879 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20-acc5.c", |
| 880 | "src/f32-raddstoreexpminusmax/gen/neonfma-p5-x20.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 881 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x4.c", |
| 882 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x8.c", |
| 883 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x12.c", |
| 884 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x16.c", |
| 885 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x20.c", |
| 886 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr1recps1fma-x24.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 887 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x4.c", |
| 888 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x8.c", |
| 889 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x12.c", |
| 890 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x16.c", |
| 891 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x20.c", |
| 892 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2fma-x24.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 893 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x4.c", |
| 894 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x8.c", |
| 895 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x12.c", |
| 896 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x16.c", |
| 897 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x20.c", |
| 898 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-nr2recps-x24.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 899 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x4.c", |
| 900 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x8.c", |
| 901 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x12.c", |
| 902 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x16.c", |
| 903 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x20.c", |
| 904 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr1recps1fma-x24.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 905 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x4.c", |
| 906 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x8.c", |
| 907 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x12.c", |
| 908 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x16.c", |
| 909 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x20.c", |
| 910 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2fma-x24.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 911 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x4.c", |
| 912 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x8.c", |
| 913 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x12.c", |
| 914 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x16.c", |
| 915 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x20.c", |
| 916 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-nr2recps-x24.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 917 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x4.c", |
| 918 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x8.c", |
| 919 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x12.c", |
| 920 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x16.c", |
| 921 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x20.c", |
| 922 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr1recps1fma-x24.c", |
| 923 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x4.c", |
| 924 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x8.c", |
| 925 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x12.c", |
| 926 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x16.c", |
| 927 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x20.c", |
| 928 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2fma-x24.c", |
| 929 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x4.c", |
| 930 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x8.c", |
| 931 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x12.c", |
| 932 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x16.c", |
| 933 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x20.c", |
| 934 | "src/f32-sigmoid/gen/neonfma-rr1-p5-nr2recps-x24.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 935 | "src/f32-vmulcaddc/gen/c4-minmax-neonfma-2x.c", |
| 936 | "src/f32-vmulcaddc/gen/c8-minmax-neonfma-2x.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 937 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x4.c", |
| 938 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x8.c", |
| 939 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x12.c", |
| 940 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x16.c", |
| 941 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x20.c", |
| 942 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x24.c", |
| 943 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x28.c", |
| 944 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x32.c", |
| 945 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x36.c", |
| 946 | "src/f32-vsqrt/gen/neonfma-nr1rsqrts1fma1adj-x40.c", |
| 947 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x4.c", |
| 948 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x8.c", |
| 949 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x12.c", |
| 950 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x16.c", |
| 951 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x20.c", |
| 952 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x24.c", |
| 953 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x28.c", |
| 954 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x32.c", |
| 955 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x36.c", |
| 956 | "src/f32-vsqrt/gen/neonfma-nr2fma1adj-x40.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 957 | "src/math/exp-neonfma-rr2-lut64-p2.c", |
| 958 | "src/math/exp-neonfma-rr2-p5.c", |
| 959 | "src/math/expminus-neonfma-rr2-lut64-p2.c", |
| 960 | "src/math/expminus-neonfma-rr2-lut2048-p1.c", |
| 961 | "src/math/expminus-neonfma-rr2-p5.c", |
| 962 | "src/math/expm1minus-neonfma-rr1-lut16-p3.c", |
| 963 | "src/math/expm1minus-neonfma-rr1-p6.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 964 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr1recps1fma.c", |
| 965 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2fma.c", |
| 966 | "src/math/sigmoid-neonfma-rr1-lut64-p2-nr2recps.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 967 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr1recps1fma.c", |
| 968 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2fma.c", |
| 969 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-nr2recps.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 970 | "src/math/sigmoid-neonfma-rr1-p5-nr1recps1fma.c", |
| 971 | "src/math/sigmoid-neonfma-rr1-p5-nr2fma.c", |
| 972 | "src/math/sigmoid-neonfma-rr1-p5-nr2recps.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 973 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr1recps1fma.c", |
| 974 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2fma.c", |
| 975 | "src/math/sigmoid-neonfma-rr2-lut64-p2-nr2recps.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 976 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr1recps1fma.c", |
| 977 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2fma.c", |
| 978 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-nr2recps.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 979 | "src/math/sigmoid-neonfma-rr2-p5-nr1recps1fma.c", |
| 980 | "src/math/sigmoid-neonfma-rr2-p5-nr2fma.c", |
| 981 | "src/math/sigmoid-neonfma-rr2-p5-nr2recps.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 982 | "src/math/sqrt-neonfma-nr1fma.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 983 | "src/math/sqrt-neonfma-nr1rsqrts1fma1adj.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 984 | "src/math/sqrt-neonfma-nr2fma.c", |
| 985 | "src/math/sqrt-neonfma-nr2fma1adj.c", |
| 986 | "src/math/sqrt-neonfma-nr3fma.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 987 | ] |
| 988 | |
| 989 | AARCH64_NEONFMA_UKERNELS = [ |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 990 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x1.c", |
| 991 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x4-neonfma-2x2.c", |
| 992 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x1.c", |
| 993 | "src/f32-conv-hwc/gen/3x3s2p0p1c3x8-neonfma-2x2.c", |
| 994 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x1.c", |
| 995 | "src/f32-conv-hwc/gen/3x3s2p1c3x4-neonfma-2x2.c", |
| 996 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x1.c", |
| 997 | "src/f32-conv-hwc/gen/3x3s2p1c3x8-neonfma-2x2.c", |
| 998 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-neonfma-2x2.c", |
| 999 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc2.c", |
| 1000 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc3.c", |
| 1001 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4-acc4.c", |
| 1002 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-1x4.c", |
| 1003 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4-acc2.c", |
| 1004 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-2x4.c", |
| 1005 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-3x4.c", |
| 1006 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-4x4.c", |
| 1007 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-5x4.c", |
| 1008 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-neonfma-6x4.c", |
| 1009 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc2.c", |
| 1010 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc3.c", |
| 1011 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4-acc4.c", |
| 1012 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-1x4.c", |
| 1013 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4-acc2.c", |
| 1014 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-2x4.c", |
| 1015 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-3x4.c", |
| 1016 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-neonfma-4x4.c", |
| 1017 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc2.c", |
| 1018 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc3.c", |
| 1019 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc4.c", |
| 1020 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4-acc5.c", |
| 1021 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-1x4.c", |
| 1022 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc2.c", |
| 1023 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4-acc3.c", |
| 1024 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-2x4.c", |
| 1025 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4-acc2.c", |
| 1026 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-3x4.c", |
| 1027 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4-acc2.c", |
| 1028 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-4x4.c", |
| 1029 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-neonfma-5x4.c", |
| 1030 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc2.c", |
| 1031 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc3.c", |
| 1032 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc4.c", |
| 1033 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4-acc5.c", |
| 1034 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-1x4.c", |
| 1035 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc2.c", |
| 1036 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4-acc3.c", |
| 1037 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-2x4.c", |
| 1038 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4-acc2.c", |
| 1039 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-neonfma-3x4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1040 | "src/f32-gemm/gen-inc/1x8inc-minmax-neonfma-lane-ld64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1041 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1042 | "src/f32-gemm/gen-inc/4x8inc-minmax-neonfma-lane-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1043 | "src/f32-gemm/gen-inc/5x8inc-minmax-neonfma-lane-ld64.c", |
| 1044 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld64.c", |
| 1045 | "src/f32-gemm/gen-inc/6x8inc-minmax-neonfma-lane-ld128.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1046 | "src/f32-gemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1047 | "src/f32-gemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1048 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
| 1049 | "src/f32-gemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
| 1050 | "src/f32-gemm/gen/5x8-minmax-neonfma-lane-ld64.c", |
| 1051 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1052 | "src/f32-gemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1053 | "src/f32-igemm/gen/1x8-minmax-neonfma-lane-ld64.c", |
| 1054 | "src/f32-igemm/gen/4x2-minmax-neonfma-lane-ld64.c", |
| 1055 | "src/f32-igemm/gen/4x4-minmax-neonfma-lane-ld64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1056 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld64.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1057 | "src/f32-igemm/gen/4x8-minmax-neonfma-lane-ld128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1058 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld64.c", |
| 1059 | "src/f32-igemm/gen/6x8-minmax-neonfma-lane-ld128.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1060 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x4.c", |
| 1061 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x8.c", |
| 1062 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x12.c", |
| 1063 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x16.c", |
| 1064 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x20.c", |
| 1065 | "src/f32-sigmoid/gen/neonfma-rr1-lut64-p2-div-x24.c", |
| 1066 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x4.c", |
| 1067 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x8.c", |
| 1068 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x12.c", |
| 1069 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x16.c", |
| 1070 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x20.c", |
| 1071 | "src/f32-sigmoid/gen/neonfma-rr1-lut2048-p1-div-x24.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1072 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x4.c", |
| 1073 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x8.c", |
| 1074 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x12.c", |
| 1075 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x16.c", |
| 1076 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x20.c", |
| 1077 | "src/f32-sigmoid/gen/neonfma-rr1-p5-div-x24.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1078 | "src/f32-spmm/gen/4x1-minmax-neonfma-pipelined.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1079 | "src/f32-spmm/gen/4x1-minmax-neonfma-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1080 | "src/f32-spmm/gen/4x1-minmax-neonfma.c", |
| 1081 | "src/f32-spmm/gen/4x2-minmax-neonfma.c", |
| 1082 | "src/f32-spmm/gen/4x4-minmax-neonfma.c", |
| 1083 | "src/f32-spmm/gen/8x1-minmax-neonfma-pipelined.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1084 | "src/f32-spmm/gen/8x1-minmax-neonfma-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1085 | "src/f32-spmm/gen/8x1-minmax-neonfma.c", |
| 1086 | "src/f32-spmm/gen/8x2-minmax-neonfma.c", |
| 1087 | "src/f32-spmm/gen/8x4-minmax-neonfma.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1088 | "src/f32-spmm/gen/12x1-minmax-neonfma.c", |
| 1089 | "src/f32-spmm/gen/12x2-minmax-neonfma.c", |
| 1090 | "src/f32-spmm/gen/12x4-minmax-neonfma.c", |
| 1091 | "src/f32-spmm/gen/16x1-minmax-neonfma-pipelined.c", |
| 1092 | "src/f32-spmm/gen/16x1-minmax-neonfma-x2.c", |
| 1093 | "src/f32-spmm/gen/16x1-minmax-neonfma.c", |
| 1094 | "src/f32-spmm/gen/16x2-minmax-neonfma.c", |
| 1095 | "src/f32-spmm/gen/16x4-minmax-neonfma.c", |
| 1096 | "src/f32-spmm/gen/32x1-minmax-neonfma-pipelined.c", |
| 1097 | "src/f32-spmm/gen/32x1-minmax-neonfma-x2.c", |
| 1098 | "src/f32-spmm/gen/32x1-minmax-neonfma.c", |
| 1099 | "src/f32-spmm/gen/32x2-minmax-neonfma.c", |
| 1100 | "src/f32-spmm/gen/32x4-minmax-neonfma.c", |
| 1101 | "src/f32-vbinary/gen/vdiv-minmax-neon-x4.c", |
| 1102 | "src/f32-vbinary/gen/vdiv-minmax-neon-x8.c", |
| 1103 | "src/f32-vbinary/gen/vdivc-minmax-neon-x4.c", |
| 1104 | "src/f32-vbinary/gen/vdivc-minmax-neon-x8.c", |
| 1105 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x4.c", |
| 1106 | "src/f32-vbinary/gen/vrdivc-minmax-neon-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1107 | "src/f32-vsqrt/gen/neon-sqrt-x4.c", |
| 1108 | "src/f32-vsqrt/gen/neon-sqrt-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1109 | "src/math/sigmoid-neonfma-rr1-lut64-p2-div.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1110 | "src/math/sigmoid-neonfma-rr1-lut2048-p1-div.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1111 | "src/math/sigmoid-neonfma-rr1-p5-div.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1112 | "src/math/sigmoid-neonfma-rr2-lut64-p2-div.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1113 | "src/math/sigmoid-neonfma-rr2-lut2048-p1-div.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1114 | "src/math/sigmoid-neonfma-rr2-p5-div.c", |
| 1115 | ] |
| 1116 | |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1117 | NEONV8_UKERNELS = [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1118 | "src/f32-vrnd/gen/vrndd-neonv8-x4.c", |
| 1119 | "src/f32-vrnd/gen/vrndd-neonv8-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1120 | "src/f32-vrnd/gen/vrndne-neonv8-x4.c", |
| 1121 | "src/f32-vrnd/gen/vrndne-neonv8-x8.c", |
| 1122 | "src/f32-vrnd/gen/vrndu-neonv8-x4.c", |
| 1123 | "src/f32-vrnd/gen/vrndu-neonv8-x8.c", |
| 1124 | "src/f32-vrnd/gen/vrndz-neonv8-x4.c", |
| 1125 | "src/f32-vrnd/gen/vrndz-neonv8-x8.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1126 | "src/math/roundd-neonv8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1127 | "src/math/roundne-neonv8.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1128 | "src/math/roundu-neonv8.c", |
| 1129 | "src/math/roundz-neonv8.c", |
| 1130 | ] |
| 1131 | |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1132 | AARCH64_NEONFP16ARITH_UKERNELS = [ |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1133 | "src/f16-clamp/gen/neonfp16arith-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1134 | "src/f16-clamp/gen/neonfp16arith-x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1135 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith-acc2.c", |
| 1136 | "src/f16-dwconv/gen/up8x4-minmax-neonfp16arith.c", |
| 1137 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith-acc2.c", |
| 1138 | "src/f16-dwconv/gen/up8x9-minmax-neonfp16arith.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1139 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith-acc2.c", |
| 1140 | "src/f16-dwconv/gen/up8x25-minmax-neonfp16arith.c", |
| 1141 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith-acc2.c", |
| 1142 | "src/f16-dwconv/gen/up16x4-minmax-neonfp16arith.c", |
| 1143 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith-acc2.c", |
| 1144 | "src/f16-dwconv/gen/up16x9-minmax-neonfp16arith.c", |
| 1145 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith-acc2.c", |
| 1146 | "src/f16-dwconv/gen/up16x25-minmax-neonfp16arith.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1147 | "src/f16-gavgpool/7p7x-minmax-neonfp16arith-c8.c", |
| 1148 | "src/f16-gavgpool/7x-minmax-neonfp16arith-c8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1149 | "src/f16-gemm/gen-inc/1x8inc-minmax-neonfp16arith-ld64.c", |
| 1150 | "src/f16-gemm/gen-inc/1x16inc-minmax-neonfp16arith-ld64.c", |
| 1151 | "src/f16-gemm/gen-inc/4x8inc-minmax-neonfp16arith-ld64.c", |
| 1152 | "src/f16-gemm/gen-inc/4x16inc-minmax-neonfp16arith-ld64.c", |
| 1153 | "src/f16-gemm/gen-inc/6x8inc-minmax-neonfp16arith-ld64.c", |
| 1154 | "src/f16-gemm/gen-inc/6x16inc-minmax-neonfp16arith-ld64.c", |
| 1155 | "src/f16-gemm/gen-inc/8x8inc-minmax-neonfp16arith-ld64.c", |
| 1156 | "src/f16-gemm/gen-inc/8x16inc-minmax-neonfp16arith-ld64.c", |
| 1157 | "src/f16-gemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1158 | "src/f16-gemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1159 | "src/f16-gemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1160 | "src/f16-gemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1161 | "src/f16-gemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1162 | "src/f16-gemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1163 | "src/f16-gemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1164 | "src/f16-gemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1165 | "src/f16-hswish/gen/hswish-neonfp16arith-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1166 | "src/f16-hswish/gen/hswish-neonfp16arith-x16.c", |
| 1167 | "src/f16-igemm/gen/1x8-minmax-neonfp16arith-ld64.c", |
| 1168 | "src/f16-igemm/gen/1x16-minmax-neonfp16arith-ld64.c", |
| 1169 | "src/f16-igemm/gen/4x8-minmax-neonfp16arith-ld64.c", |
| 1170 | "src/f16-igemm/gen/4x16-minmax-neonfp16arith-ld64.c", |
| 1171 | "src/f16-igemm/gen/6x8-minmax-neonfp16arith-ld64.c", |
| 1172 | "src/f16-igemm/gen/6x16-minmax-neonfp16arith-ld64.c", |
| 1173 | "src/f16-igemm/gen/8x8-minmax-neonfp16arith-ld64.c", |
| 1174 | "src/f16-igemm/gen/8x16-minmax-neonfp16arith-ld64.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1175 | "src/f16-prelu/gen/neonfp16arith-2x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1176 | "src/f16-prelu/gen/neonfp16arith-2x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1177 | "src/f16-relu/gen/neonfp16arith-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1178 | "src/f16-relu/gen/neonfp16arith-x16.c", |
| 1179 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1180 | "src/f16-spmm/gen/8x1-minmax-neonfp16arith.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1181 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1182 | "src/f16-spmm/gen/16x1-minmax-neonfp16arith.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1183 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1184 | "src/f16-spmm/gen/24x1-minmax-neonfp16arith.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1185 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith-x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1186 | "src/f16-spmm/gen/32x1-minmax-neonfp16arith.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1187 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x8.c", |
| 1188 | "src/f16-vbinary/gen/vadd-minmax-neonfp16arith-x16.c", |
| 1189 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x8.c", |
| 1190 | "src/f16-vbinary/gen/vaddc-minmax-neonfp16arith-x16.c", |
| 1191 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x8.c", |
| 1192 | "src/f16-vbinary/gen/vdiv-minmax-neonfp16arith-x16.c", |
| 1193 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x8.c", |
| 1194 | "src/f16-vbinary/gen/vdivc-minmax-neonfp16arith-x16.c", |
| 1195 | "src/f16-vbinary/gen/vmax-neonfp16arith-x8.c", |
| 1196 | "src/f16-vbinary/gen/vmax-neonfp16arith-x16.c", |
| 1197 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x8.c", |
| 1198 | "src/f16-vbinary/gen/vmaxc-neonfp16arith-x16.c", |
| 1199 | "src/f16-vbinary/gen/vmin-neonfp16arith-x8.c", |
| 1200 | "src/f16-vbinary/gen/vmin-neonfp16arith-x16.c", |
| 1201 | "src/f16-vbinary/gen/vminc-neonfp16arith-x8.c", |
| 1202 | "src/f16-vbinary/gen/vminc-neonfp16arith-x16.c", |
| 1203 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x8.c", |
| 1204 | "src/f16-vbinary/gen/vmul-minmax-neonfp16arith-x16.c", |
| 1205 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x8.c", |
| 1206 | "src/f16-vbinary/gen/vmulc-minmax-neonfp16arith-x16.c", |
| 1207 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x8.c", |
| 1208 | "src/f16-vbinary/gen/vrdivc-minmax-neonfp16arith-x16.c", |
| 1209 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x8.c", |
| 1210 | "src/f16-vbinary/gen/vrsubc-minmax-neonfp16arith-x16.c", |
| 1211 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x8.c", |
| 1212 | "src/f16-vbinary/gen/vsub-minmax-neonfp16arith-x16.c", |
| 1213 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x8.c", |
| 1214 | "src/f16-vbinary/gen/vsubc-minmax-neonfp16arith-x16.c", |
| 1215 | "src/f16-vmulcaddc/gen/c8-minmax-neonfp16arith-2x.c", |
| 1216 | "src/f16-vmulcaddc/gen/c16-minmax-neonfp16arith-2x.c", |
| 1217 | ] |
| 1218 | |
| 1219 | NEONDOT_UKERNELS = [ |
| 1220 | "src/qs8-gemm/gen/1x8c4-minmax-neondot.c", |
| 1221 | "src/qs8-gemm/gen/1x16c4-minmax-neondot.c", |
| 1222 | "src/qs8-gemm/gen/4x8c4-minmax-neondot.c", |
| 1223 | "src/qs8-gemm/gen/4x16c4-minmax-neondot.c", |
| 1224 | "src/qs8-gemm/gen/6x8c4-minmax-neondot.c", |
| 1225 | "src/qs8-gemm/gen/6x16c4-minmax-neondot.c", |
| 1226 | "src/qs8-gemm/gen/8x8c4-minmax-neondot.c", |
| 1227 | "src/qs8-gemm/gen/8x16c4-minmax-neondot.c", |
| 1228 | "src/qs8-gemm/gen/12x8c4-minmax-neondot.c", |
| 1229 | "src/qs8-igemm/gen/1x8c4-minmax-neondot.c", |
| 1230 | "src/qs8-igemm/gen/1x16c4-minmax-neondot.c", |
| 1231 | "src/qs8-igemm/gen/4x8c4-minmax-neondot.c", |
| 1232 | "src/qs8-igemm/gen/4x16c4-minmax-neondot.c", |
| 1233 | "src/qs8-igemm/gen/6x8c4-minmax-neondot.c", |
| 1234 | "src/qs8-igemm/gen/6x16c4-minmax-neondot.c", |
| 1235 | "src/qs8-igemm/gen/8x8c4-minmax-neondot.c", |
| 1236 | "src/qs8-igemm/gen/8x16c4-minmax-neondot.c", |
| 1237 | "src/qs8-igemm/gen/12x8c4-minmax-neondot.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1238 | ] |
| 1239 | |
| 1240 | SSE_UKERNELS = [ |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1241 | "src/f32-avgpool/9p8x-minmax-sse-c4.c", |
| 1242 | "src/f32-avgpool/9x-minmax-sse-c4.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1243 | "src/f32-clamp/gen/sse-x4.c", |
| 1244 | "src/f32-clamp/gen/sse-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1245 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-1x1.c", |
| 1246 | "src/f32-conv-hwc2chw/3x3s2p1c3x4-sse-2x2.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1247 | "src/f32-dwconv/gen/up4x4-minmax-sse-acc2.c", |
| 1248 | "src/f32-dwconv/gen/up4x4-minmax-sse.c", |
| 1249 | "src/f32-dwconv/gen/up4x9-minmax-sse-acc2.c", |
| 1250 | "src/f32-dwconv/gen/up4x9-minmax-sse.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1251 | "src/f32-dwconv/gen/up4x25-minmax-sse-acc2.c", |
| 1252 | "src/f32-dwconv/gen/up4x25-minmax-sse.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1253 | "src/f32-dwconv/gen/up8x4-minmax-sse-acc2.c", |
| 1254 | "src/f32-dwconv/gen/up8x4-minmax-sse.c", |
| 1255 | "src/f32-dwconv/gen/up8x9-minmax-sse-acc2.c", |
| 1256 | "src/f32-dwconv/gen/up8x9-minmax-sse.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1257 | "src/f32-dwconv/gen/up8x25-minmax-sse-acc2.c", |
| 1258 | "src/f32-dwconv/gen/up8x25-minmax-sse.c", |
| 1259 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc2.c", |
| 1260 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc3.c", |
| 1261 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4-acc4.c", |
| 1262 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-1x4.c", |
| 1263 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4-acc2.c", |
| 1264 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-2x4.c", |
| 1265 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-3x4.c", |
| 1266 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-4x4.c", |
| 1267 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-5x4.c", |
| 1268 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-sse-6x4.c", |
| 1269 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc2.c", |
| 1270 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc3.c", |
| 1271 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4-acc4.c", |
| 1272 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-1x4.c", |
| 1273 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4-acc2.c", |
| 1274 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-2x4.c", |
| 1275 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-3x4.c", |
| 1276 | "src/f32-dwconv2d-chw/gen/3x3s2p1-minmax-sse-4x4.c", |
| 1277 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc2.c", |
| 1278 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc3.c", |
| 1279 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc4.c", |
| 1280 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4-acc5.c", |
| 1281 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-1x4.c", |
| 1282 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc2.c", |
| 1283 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4-acc3.c", |
| 1284 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-2x4.c", |
| 1285 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4-acc2.c", |
| 1286 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-3x4.c", |
| 1287 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4-acc2.c", |
| 1288 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-4x4.c", |
| 1289 | "src/f32-dwconv2d-chw/gen/5x5p2-minmax-sse-5x4.c", |
| 1290 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc2.c", |
| 1291 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc3.c", |
| 1292 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc4.c", |
| 1293 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4-acc5.c", |
| 1294 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-1x4.c", |
| 1295 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc2.c", |
| 1296 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4-acc3.c", |
| 1297 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-2x4.c", |
| 1298 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4-acc2.c", |
| 1299 | "src/f32-dwconv2d-chw/gen/5x5s2p2-minmax-sse-3x4.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1300 | "src/f32-gavgpool-cw/sse-x4.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1301 | "src/f32-gavgpool/7p7x-minmax-sse-c4.c", |
| 1302 | "src/f32-gavgpool/7x-minmax-sse-c4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1303 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-dup.c", |
| 1304 | "src/f32-gemm/gen-inc/1x8inc-minmax-sse-load1.c", |
| 1305 | "src/f32-gemm/gen-inc/1x8s4inc-minmax-sse.c", |
| 1306 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-dup.c", |
| 1307 | "src/f32-gemm/gen-inc/4x8inc-minmax-sse-load1.c", |
| 1308 | "src/f32-gemm/gen-inc/4x8s4inc-minmax-sse.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1309 | "src/f32-gemm/gen/1x8-minmax-sse-dup.c", |
| 1310 | "src/f32-gemm/gen/1x8-minmax-sse-load1.c", |
| 1311 | "src/f32-gemm/gen/1x8s4-minmax-sse.c", |
| 1312 | "src/f32-gemm/gen/4x2c4-minmax-sse.c", |
| 1313 | "src/f32-gemm/gen/4x8-minmax-sse-dup.c", |
| 1314 | "src/f32-gemm/gen/4x8-minmax-sse-load1.c", |
| 1315 | "src/f32-gemm/gen/4x8s4-minmax-sse.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1316 | "src/f32-hswish/gen/hswish-sse-x4.c", |
| 1317 | "src/f32-hswish/gen/hswish-sse-x8.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1318 | "src/f32-ibilinear/gen/sse-c4.c", |
| 1319 | "src/f32-ibilinear/gen/sse-c8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1320 | "src/f32-igemm/gen/1x8-minmax-sse-dup.c", |
| 1321 | "src/f32-igemm/gen/1x8-minmax-sse-load1.c", |
| 1322 | "src/f32-igemm/gen/1x8s4-minmax-sse.c", |
| 1323 | "src/f32-igemm/gen/4x2c4-minmax-sse.c", |
| 1324 | "src/f32-igemm/gen/4x8-minmax-sse-dup.c", |
| 1325 | "src/f32-igemm/gen/4x8-minmax-sse-load1.c", |
| 1326 | "src/f32-igemm/gen/4x8s4-minmax-sse.c", |
| 1327 | "src/f32-maxpool/9p8x-minmax-sse-c4.c", |
| 1328 | "src/f32-pavgpool/9p8x-minmax-sse-c4.c", |
| 1329 | "src/f32-pavgpool/9x-minmax-sse-c4.c", |
| 1330 | "src/f32-ppmm/gen/4x8-minmax-sse.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1331 | "src/f32-prelu/gen/sse-2x4.c", |
| 1332 | "src/f32-prelu/gen/sse-2x8.c", |
| 1333 | "src/f32-relu/gen/sse-x4.c", |
| 1334 | "src/f32-relu/gen/sse-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1335 | "src/f32-rmax/sse.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1336 | "src/f32-spmm/gen/4x1-minmax-sse.c", |
| 1337 | "src/f32-spmm/gen/8x1-minmax-sse.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1338 | "src/f32-spmm/gen/16x1-minmax-sse.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1339 | "src/f32-spmm/gen/32x1-minmax-sse.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1340 | "src/f32-vbinary/gen/vadd-minmax-sse-x4.c", |
| 1341 | "src/f32-vbinary/gen/vadd-minmax-sse-x8.c", |
| 1342 | "src/f32-vbinary/gen/vaddc-minmax-sse-x4.c", |
| 1343 | "src/f32-vbinary/gen/vaddc-minmax-sse-x8.c", |
| 1344 | "src/f32-vbinary/gen/vdiv-minmax-sse-x4.c", |
| 1345 | "src/f32-vbinary/gen/vdiv-minmax-sse-x8.c", |
| 1346 | "src/f32-vbinary/gen/vdivc-minmax-sse-x4.c", |
| 1347 | "src/f32-vbinary/gen/vdivc-minmax-sse-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1348 | "src/f32-vbinary/gen/vmax-sse-x4.c", |
| 1349 | "src/f32-vbinary/gen/vmax-sse-x8.c", |
| 1350 | "src/f32-vbinary/gen/vmaxc-sse-x4.c", |
| 1351 | "src/f32-vbinary/gen/vmaxc-sse-x8.c", |
| 1352 | "src/f32-vbinary/gen/vmin-sse-x4.c", |
| 1353 | "src/f32-vbinary/gen/vmin-sse-x8.c", |
| 1354 | "src/f32-vbinary/gen/vminc-sse-x4.c", |
| 1355 | "src/f32-vbinary/gen/vminc-sse-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1356 | "src/f32-vbinary/gen/vmul-minmax-sse-x4.c", |
| 1357 | "src/f32-vbinary/gen/vmul-minmax-sse-x8.c", |
| 1358 | "src/f32-vbinary/gen/vmulc-minmax-sse-x4.c", |
| 1359 | "src/f32-vbinary/gen/vmulc-minmax-sse-x8.c", |
| 1360 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x4.c", |
| 1361 | "src/f32-vbinary/gen/vrdivc-minmax-sse-x8.c", |
| 1362 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x4.c", |
| 1363 | "src/f32-vbinary/gen/vrsubc-minmax-sse-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1364 | "src/f32-vbinary/gen/vsqrdiff-sse-x4.c", |
| 1365 | "src/f32-vbinary/gen/vsqrdiff-sse-x8.c", |
| 1366 | "src/f32-vbinary/gen/vsqrdiffc-sse-x4.c", |
| 1367 | "src/f32-vbinary/gen/vsqrdiffc-sse-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1368 | "src/f32-vbinary/gen/vsub-minmax-sse-x4.c", |
| 1369 | "src/f32-vbinary/gen/vsub-minmax-sse-x8.c", |
| 1370 | "src/f32-vbinary/gen/vsubc-minmax-sse-x4.c", |
| 1371 | "src/f32-vbinary/gen/vsubc-minmax-sse-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1372 | "src/f32-vlrelu/gen/vlrelu-sse-x4.c", |
| 1373 | "src/f32-vlrelu/gen/vlrelu-sse-x8.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1374 | "src/f32-vmulcaddc/gen/c4-minmax-sse-2x.c", |
| 1375 | "src/f32-vmulcaddc/gen/c8-minmax-sse-2x.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1376 | "src/f32-vsqrt/gen/sse-sqrt-x4.c", |
| 1377 | "src/f32-vsqrt/gen/sse-sqrt-x8.c", |
| 1378 | "src/f32-vunary/gen/vabs-sse-x4.c", |
| 1379 | "src/f32-vunary/gen/vabs-sse-x8.c", |
| 1380 | "src/f32-vunary/gen/vneg-sse-x4.c", |
| 1381 | "src/f32-vunary/gen/vneg-sse-x8.c", |
| 1382 | "src/f32-vunary/gen/vsqr-sse-x4.c", |
| 1383 | "src/f32-vunary/gen/vsqr-sse-x8.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1384 | "src/math/roundd-sse-addsub.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1385 | "src/math/roundne-sse-addsub.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1386 | "src/math/roundu-sse-addsub.c", |
| 1387 | "src/math/roundz-sse-addsub.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1388 | "src/math/sqrt-sse-hh1mac.c", |
| 1389 | "src/math/sqrt-sse-nr1mac.c", |
| 1390 | "src/math/sqrt-sse-nr2mac.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1391 | "src/x32-fill/sse.c", |
| 1392 | "src/x32-packx/x4-sse.c", |
| 1393 | "src/x32-pad/sse.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1394 | ] |
| 1395 | |
| 1396 | SSE2_UKERNELS = [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1397 | "src/f32-argmaxpool/4x-sse2-c4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1398 | "src/f32-argmaxpool/9p8x-sse2-c4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1399 | "src/f32-argmaxpool/9x-sse2-c4.c", |
| 1400 | "src/f32-prelu/gen/sse2-2x4.c", |
| 1401 | "src/f32-prelu/gen/sse2-2x8.c", |
| 1402 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1403 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8-acc2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1404 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x8.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1405 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc2.c", |
| 1406 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1407 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x12.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1408 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc2.c", |
| 1409 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1410 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1411 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc2.c", |
| 1412 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1413 | "src/f32-raddstoreexpminusmax/gen/sse2-p5-x20.c", |
| 1414 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x4.c", |
| 1415 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x8.c", |
| 1416 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x12.c", |
| 1417 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x16.c", |
| 1418 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x20.c", |
| 1419 | "src/f32-sigmoid/gen/sse2-lut64-p2-div-x24.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1420 | "src/f32-sigmoid/gen/sse2-p5-div-x4.c", |
| 1421 | "src/f32-sigmoid/gen/sse2-p5-div-x8.c", |
| 1422 | "src/f32-sigmoid/gen/sse2-p5-div-x12.c", |
| 1423 | "src/f32-sigmoid/gen/sse2-p5-div-x16.c", |
| 1424 | "src/f32-sigmoid/gen/sse2-p5-div-x20.c", |
| 1425 | "src/f32-sigmoid/gen/sse2-p5-div-x24.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1426 | "src/f32-vlrelu/gen/vlrelu-sse2-x4.c", |
| 1427 | "src/f32-vlrelu/gen/vlrelu-sse2-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1428 | "src/f32-vrnd/gen/vrndd-sse2-x4.c", |
| 1429 | "src/f32-vrnd/gen/vrndd-sse2-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1430 | "src/f32-vrnd/gen/vrndne-sse2-x4.c", |
| 1431 | "src/f32-vrnd/gen/vrndne-sse2-x8.c", |
| 1432 | "src/f32-vrnd/gen/vrndu-sse2-x4.c", |
| 1433 | "src/f32-vrnd/gen/vrndu-sse2-x8.c", |
| 1434 | "src/f32-vrnd/gen/vrndz-sse2-x4.c", |
| 1435 | "src/f32-vrnd/gen/vrndz-sse2-x8.c", |
| 1436 | "src/math/exp-sse2-rr2-lut64-p2.c", |
| 1437 | "src/math/exp-sse2-rr2-p5.c", |
| 1438 | "src/math/expm1minus-sse2-rr2-lut16-p3.c", |
| 1439 | "src/math/expm1minus-sse2-rr2-p6.c", |
| 1440 | "src/math/expminus-sse2-rr2-p5.c", |
| 1441 | "src/math/roundd-sse2-cvt.c", |
| 1442 | "src/math/roundne-sse2-cvt.c", |
| 1443 | "src/math/roundu-sse2-cvt.c", |
| 1444 | "src/math/roundz-sse2-cvt.c", |
| 1445 | "src/math/sigmoid-sse2-rr2-lut64-p2-div.c", |
| 1446 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr1.c", |
| 1447 | "src/math/sigmoid-sse2-rr2-lut64-p2-nr2.c", |
| 1448 | "src/math/sigmoid-sse2-rr2-p5-div.c", |
| 1449 | "src/math/sigmoid-sse2-rr2-p5-nr1.c", |
| 1450 | "src/math/sigmoid-sse2-rr2-p5-nr2.c", |
| 1451 | "src/qs8-dwconv/gen/up8x9-minmax-sse2-mul16.c", |
| 1452 | "src/qs8-dwconv/gen/up16x9-minmax-sse2-mul16.c", |
| 1453 | "src/qs8-dwconv/gen/up24x9-minmax-sse2-mul16.c", |
| 1454 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c8-acc2.c", |
| 1455 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c16-acc2.c", |
| 1456 | "src/qs8-gavgpool/gen/7p7x-minmax-sse2-c24-acc2.c", |
| 1457 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c8-acc2.c", |
| 1458 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c16-acc2.c", |
| 1459 | "src/qs8-gavgpool/gen/7x-minmax-sse2-c24-acc2.c", |
| 1460 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld64.c", |
| 1461 | "src/qs8-gemm/gen/1x4c2-minmax-sse2-ld128.c", |
| 1462 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse2.c", |
| 1463 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld64.c", |
| 1464 | "src/qs8-gemm/gen/1x4c8-minmax-sse2-ld128.c", |
| 1465 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse2.c", |
| 1466 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld64.c", |
| 1467 | "src/qs8-gemm/gen/2x4c8-minmax-sse2-ld128.c", |
| 1468 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse2.c", |
| 1469 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld64.c", |
| 1470 | "src/qs8-gemm/gen/3x4c8-minmax-sse2-ld128.c", |
| 1471 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse2.c", |
| 1472 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 1473 | "src/qs8-gemm/gen/4x4c2-minmax-sse2-ld128.c", |
| 1474 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse2.c", |
| 1475 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld64.c", |
| 1476 | "src/qs8-igemm/gen/1x4c2-minmax-sse2-ld128.c", |
| 1477 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld64.c", |
| 1478 | "src/qs8-igemm/gen/1x4c8-minmax-sse2-ld128.c", |
| 1479 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld64.c", |
| 1480 | "src/qs8-igemm/gen/2x4c8-minmax-sse2-ld128.c", |
| 1481 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld64.c", |
| 1482 | "src/qs8-igemm/gen/3x4c8-minmax-sse2-ld128.c", |
| 1483 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld64.c", |
| 1484 | "src/qs8-igemm/gen/4x4c2-minmax-sse2-ld128.c", |
| 1485 | "src/qs8-requantization/fp32-sse2.c", |
| 1486 | "src/qs8-requantization/precise-sse2.c", |
| 1487 | "src/qs8-requantization/q31-sse2.c", |
| 1488 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x8.c", |
| 1489 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x16.c", |
| 1490 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x24.c", |
| 1491 | "src/qs8-vadd/gen/minmax-sse2-mul16-ld64-x32.c", |
| 1492 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c", |
| 1493 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x16.c", |
| 1494 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x24.c", |
| 1495 | "src/qs8-vaddc/gen/minmax-sse2-mul16-ld64-x32.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1496 | "src/qu8-avgpool/9p8x-minmax-sse2-c8.c", |
| 1497 | "src/qu8-avgpool/9x-minmax-sse2-c8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1498 | "src/qu8-dwconv/up8x9-minmax-sse2.c", |
| 1499 | "src/qu8-gavgpool/7p7x-minmax-sse2-c8.c", |
| 1500 | "src/qu8-gavgpool/7x-minmax-sse2-c8.c", |
| 1501 | "src/qu8-gemm/2x4c8-minmax-sse2.c", |
| 1502 | "src/qu8-gemm/4x4c2-minmax-sse2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1503 | "src/qu8-igemm/4x4c2-minmax-sse2.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1504 | "src/qu8-requantization/fp32-sse2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1505 | "src/qu8-requantization/precise-sse2.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1506 | "src/qu8-requantization/q31-sse2.c", |
| 1507 | "src/qu8-vadd/minmax-sse2.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1508 | "src/u8-clamp/sse2-x64.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1509 | "src/u8-maxpool/9p8x-minmax-sse2-c16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1510 | "src/u8-rmax/sse2.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1511 | "src/x8-zip/x2-sse2.c", |
| 1512 | "src/x8-zip/x3-sse2.c", |
| 1513 | "src/x8-zip/x4-sse2.c", |
| 1514 | "src/x8-zip/xm-sse2.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1515 | "src/x32-unpool/sse2.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1516 | "src/x32-zip/x2-sse2.c", |
| 1517 | "src/x32-zip/x3-sse2.c", |
| 1518 | "src/x32-zip/x4-sse2.c", |
| 1519 | "src/x32-zip/xm-sse2.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1520 | ] |
| 1521 | |
| 1522 | SSSE3_UKERNELS = [ |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1523 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc2.c", |
| 1524 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc3.c", |
| 1525 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4-acc4.c", |
| 1526 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-1x4.c", |
| 1527 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4-acc2.c", |
| 1528 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-2x4.c", |
| 1529 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-3x4.c", |
| 1530 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-4x4.c", |
| 1531 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-5x4.c", |
| 1532 | "src/f32-dwconv2d-chw/gen/3x3p1-minmax-ssse3-6x4.c", |
| 1533 | "src/qs8-dwconv/gen/up8x9-minmax-ssse3-mul16.c", |
| 1534 | "src/qs8-dwconv/gen/up16x9-minmax-ssse3-mul16.c", |
| 1535 | "src/qs8-dwconv/gen/up24x9-minmax-ssse3-mul16.c", |
| 1536 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c8-acc2.c", |
| 1537 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c16-acc2.c", |
| 1538 | "src/qs8-gavgpool/gen/7p7x-minmax-ssse3-c24-acc2.c", |
| 1539 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c8-acc2.c", |
| 1540 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c16-acc2.c", |
| 1541 | "src/qs8-gavgpool/gen/7x-minmax-ssse3-c24-acc2.c", |
| 1542 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| 1543 | "src/qs8-gemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| 1544 | "src/qs8-gemm/gen/1x4c2-xw-minmax-ssse3.c", |
| 1545 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| 1546 | "src/qs8-gemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| 1547 | "src/qs8-gemm/gen/1x4c8-xw-minmax-ssse3.c", |
| 1548 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld64.c", |
| 1549 | "src/qs8-gemm/gen/2x4c8-minmax-ssse3-ld128.c", |
| 1550 | "src/qs8-gemm/gen/2x4c8-xw-minmax-ssse3.c", |
| 1551 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld64.c", |
| 1552 | "src/qs8-gemm/gen/3x4c8-minmax-ssse3-ld128.c", |
| 1553 | "src/qs8-gemm/gen/3x4c8-xw-minmax-ssse3.c", |
| 1554 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 1555 | "src/qs8-gemm/gen/4x4c2-minmax-ssse3-ld128.c", |
| 1556 | "src/qs8-gemm/gen/4x4c2-xw-minmax-ssse3.c", |
| 1557 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld64.c", |
| 1558 | "src/qs8-igemm/gen/1x4c2-minmax-ssse3-ld128.c", |
| 1559 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld64.c", |
| 1560 | "src/qs8-igemm/gen/1x4c8-minmax-ssse3-ld128.c", |
| 1561 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld64.c", |
| 1562 | "src/qs8-igemm/gen/2x4c8-minmax-ssse3-ld128.c", |
| 1563 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld64.c", |
| 1564 | "src/qs8-igemm/gen/3x4c8-minmax-ssse3-ld128.c", |
| 1565 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld64.c", |
| 1566 | "src/qs8-igemm/gen/4x4c2-minmax-ssse3-ld128.c", |
| 1567 | "src/qs8-requantization/precise-ssse3.c", |
| 1568 | "src/qs8-requantization/q31-ssse3.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1569 | "src/qu8-requantization/precise-ssse3.c", |
| 1570 | "src/qu8-requantization/q31-ssse3.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1571 | ] |
| 1572 | |
| 1573 | SSE41_UKERNELS = [ |
| 1574 | "src/f32-prelu/gen/sse41-2x4.c", |
| 1575 | "src/f32-prelu/gen/sse41-2x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1576 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x4.c", |
| 1577 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x8.c", |
| 1578 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x12.c", |
| 1579 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x16.c", |
| 1580 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x20.c", |
| 1581 | "src/f32-sigmoid/gen/sse41-lut64-p2-div-x24.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1582 | "src/f32-sigmoid/gen/sse41-p5-div-x4.c", |
| 1583 | "src/f32-sigmoid/gen/sse41-p5-div-x8.c", |
| 1584 | "src/f32-sigmoid/gen/sse41-p5-div-x12.c", |
| 1585 | "src/f32-sigmoid/gen/sse41-p5-div-x16.c", |
| 1586 | "src/f32-sigmoid/gen/sse41-p5-div-x20.c", |
| 1587 | "src/f32-sigmoid/gen/sse41-p5-div-x24.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1588 | "src/f32-vlrelu/gen/vlrelu-sse41-x4.c", |
| 1589 | "src/f32-vlrelu/gen/vlrelu-sse41-x8.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1590 | "src/f32-vrnd/gen/vrndd-sse41-x4.c", |
| 1591 | "src/f32-vrnd/gen/vrndd-sse41-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1592 | "src/f32-vrnd/gen/vrndne-sse41-x4.c", |
| 1593 | "src/f32-vrnd/gen/vrndne-sse41-x8.c", |
| 1594 | "src/f32-vrnd/gen/vrndu-sse41-x4.c", |
| 1595 | "src/f32-vrnd/gen/vrndu-sse41-x8.c", |
| 1596 | "src/f32-vrnd/gen/vrndz-sse41-x4.c", |
| 1597 | "src/f32-vrnd/gen/vrndz-sse41-x8.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1598 | "src/math/roundd-sse41.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1599 | "src/math/roundne-sse41.c", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 1600 | "src/math/roundu-sse41.c", |
| 1601 | "src/math/roundz-sse41.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1602 | "src/qs8-dwconv/gen/up8x9-minmax-sse41-mul16.c", |
| 1603 | "src/qs8-dwconv/gen/up16x9-minmax-sse41-mul16.c", |
| 1604 | "src/qs8-dwconv/gen/up24x9-minmax-sse41-mul16.c", |
| 1605 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c8-acc2.c", |
| 1606 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c16-acc2.c", |
| 1607 | "src/qs8-gavgpool/gen/7p7x-minmax-sse41-c24-acc2.c", |
| 1608 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c8-acc2.c", |
| 1609 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c16-acc2.c", |
| 1610 | "src/qs8-gavgpool/gen/7x-minmax-sse41-c24-acc2.c", |
| 1611 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld64.c", |
| 1612 | "src/qs8-gemm/gen/1x4c2-minmax-sse41-ld128.c", |
| 1613 | "src/qs8-gemm/gen/1x4c2-xw-minmax-sse41.c", |
| 1614 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld64.c", |
| 1615 | "src/qs8-gemm/gen/1x4c8-minmax-sse41-ld128.c", |
| 1616 | "src/qs8-gemm/gen/1x4c8-xw-minmax-sse41.c", |
| 1617 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld64.c", |
| 1618 | "src/qs8-gemm/gen/2x4c8-minmax-sse41-ld128.c", |
| 1619 | "src/qs8-gemm/gen/2x4c8-xw-minmax-sse41.c", |
| 1620 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld64.c", |
| 1621 | "src/qs8-gemm/gen/3x4c8-minmax-sse41-ld128.c", |
| 1622 | "src/qs8-gemm/gen/3x4c8-xw-minmax-sse41.c", |
| 1623 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 1624 | "src/qs8-gemm/gen/4x4c2-minmax-sse41-ld128.c", |
| 1625 | "src/qs8-gemm/gen/4x4c2-xw-minmax-sse41.c", |
| 1626 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld64.c", |
| 1627 | "src/qs8-igemm/gen/1x4c2-minmax-sse41-ld128.c", |
| 1628 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld64.c", |
| 1629 | "src/qs8-igemm/gen/1x4c8-minmax-sse41-ld128.c", |
| 1630 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld64.c", |
| 1631 | "src/qs8-igemm/gen/2x4c8-minmax-sse41-ld128.c", |
| 1632 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld64.c", |
| 1633 | "src/qs8-igemm/gen/3x4c8-minmax-sse41-ld128.c", |
| 1634 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld64.c", |
| 1635 | "src/qs8-igemm/gen/4x4c2-minmax-sse41-ld128.c", |
| 1636 | "src/qs8-requantization/fp32-sse4.c", |
| 1637 | "src/qs8-requantization/precise-sse4.c", |
| 1638 | "src/qs8-requantization/q31-sse4.c", |
| 1639 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x8.c", |
| 1640 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x16.c", |
| 1641 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x24.c", |
| 1642 | "src/qs8-vadd/gen/minmax-sse41-mul16-ld64-x32.c", |
| 1643 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x8.c", |
| 1644 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x16.c", |
| 1645 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x24.c", |
| 1646 | "src/qs8-vadd/gen/minmax-sse41-mul32-ld32-x32.c", |
| 1647 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c", |
| 1648 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x16.c", |
| 1649 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x24.c", |
| 1650 | "src/qs8-vaddc/gen/minmax-sse41-mul16-ld64-x32.c", |
| 1651 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x8.c", |
| 1652 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x16.c", |
| 1653 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x24.c", |
| 1654 | "src/qs8-vaddc/gen/minmax-sse41-mul32-ld32-x32.c", |
| 1655 | "src/qu8-requantization/precise-sse4.c", |
| 1656 | "src/qu8-requantization/q31-sse4.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1657 | ] |
| 1658 | |
| 1659 | AVX_UKERNELS = [ |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1660 | "src/f32-clamp/gen/avx-x8.c", |
| 1661 | "src/f32-clamp/gen/avx-x16.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1662 | "src/f32-dwconv/gen/up8x4-minmax-avx-acc2.c", |
| 1663 | "src/f32-dwconv/gen/up8x4-minmax-avx.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1664 | "src/f32-dwconv/gen/up8x9-minmax-avx-acc2.c", |
| 1665 | "src/f32-dwconv/gen/up8x9-minmax-avx.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1666 | "src/f32-dwconv/gen/up8x25-minmax-avx-acc2.c", |
| 1667 | "src/f32-dwconv/gen/up8x25-minmax-avx.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1668 | "src/f32-dwconv/gen/up16x4-minmax-avx-acc2.c", |
| 1669 | "src/f32-dwconv/gen/up16x4-minmax-avx.c", |
| 1670 | "src/f32-dwconv/gen/up16x9-minmax-avx-acc2.c", |
| 1671 | "src/f32-dwconv/gen/up16x9-minmax-avx.c", |
| 1672 | "src/f32-dwconv/gen/up16x25-minmax-avx-acc2.c", |
| 1673 | "src/f32-dwconv/gen/up16x25-minmax-avx.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1674 | "src/f32-gemm/gen-inc/1x8inc-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1675 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx-broadcast.c", |
| 1676 | "src/f32-gemm/gen-inc/3x16inc-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1677 | "src/f32-gemm/gen-inc/4x8inc-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1678 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1679 | "src/f32-gemm/gen-inc/5x8inc-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1680 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1681 | "src/f32-gemm/gen-inc/6x8inc-minmax-avx-broadcast.c", |
| 1682 | "src/f32-gemm/gen-inc/7x8inc-minmax-avx-broadcast.c", |
| 1683 | "src/f32-gemm/gen/1x8-minmax-avx-broadcast.c", |
| 1684 | "src/f32-gemm/gen/1x16-minmax-avx-broadcast.c", |
| 1685 | "src/f32-gemm/gen/3x16-minmax-avx-broadcast.c", |
| 1686 | "src/f32-gemm/gen/4x8-minmax-avx-broadcast.c", |
| 1687 | "src/f32-gemm/gen/4x16-minmax-avx-broadcast.c", |
| 1688 | "src/f32-gemm/gen/5x8-minmax-avx-broadcast.c", |
| 1689 | "src/f32-gemm/gen/5x16-minmax-avx-broadcast.c", |
| 1690 | "src/f32-gemm/gen/6x8-minmax-avx-broadcast.c", |
| 1691 | "src/f32-gemm/gen/7x8-minmax-avx-broadcast.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1692 | "src/f32-hswish/gen/hswish-avx-x8.c", |
| 1693 | "src/f32-hswish/gen/hswish-avx-x16.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1694 | "src/f32-igemm/gen/1x8-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1695 | "src/f32-igemm/gen/1x16-minmax-avx-broadcast.c", |
| 1696 | "src/f32-igemm/gen/3x16-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1697 | "src/f32-igemm/gen/4x8-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1698 | "src/f32-igemm/gen/4x16-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1699 | "src/f32-igemm/gen/5x8-minmax-avx-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1700 | "src/f32-igemm/gen/5x16-minmax-avx-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1701 | "src/f32-igemm/gen/6x8-minmax-avx-broadcast.c", |
| 1702 | "src/f32-igemm/gen/7x8-minmax-avx-broadcast.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 1703 | "src/f32-prelu/gen/avx-2x8.c", |
| 1704 | "src/f32-prelu/gen/avx-2x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1705 | "src/f32-relu/gen/avx-x8.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1706 | "src/f32-relu/gen/avx-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1707 | "src/f32-rmax/avx.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1708 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x8.c", |
| 1709 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x16.c", |
| 1710 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x24.c", |
| 1711 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x32.c", |
| 1712 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x40.c", |
| 1713 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x48.c", |
| 1714 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x56.c", |
| 1715 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x64.c", |
| 1716 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x72.c", |
| 1717 | "src/f32-sigmoid/gen/avx-rr2-p5-div-x80.c", |
| 1718 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x8.c", |
| 1719 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x16.c", |
| 1720 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x24.c", |
| 1721 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x32.c", |
| 1722 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x40.c", |
| 1723 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x48.c", |
| 1724 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x56.c", |
| 1725 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x64.c", |
| 1726 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x72.c", |
| 1727 | "src/f32-sigmoid/gen/avx-rr2-p5-nr2-x80.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1728 | "src/f32-vbinary/gen/vadd-minmax-avx-x8.c", |
| 1729 | "src/f32-vbinary/gen/vadd-minmax-avx-x16.c", |
| 1730 | "src/f32-vbinary/gen/vaddc-minmax-avx-x8.c", |
| 1731 | "src/f32-vbinary/gen/vaddc-minmax-avx-x16.c", |
| 1732 | "src/f32-vbinary/gen/vdiv-minmax-avx-x8.c", |
| 1733 | "src/f32-vbinary/gen/vdiv-minmax-avx-x16.c", |
| 1734 | "src/f32-vbinary/gen/vdivc-minmax-avx-x8.c", |
| 1735 | "src/f32-vbinary/gen/vdivc-minmax-avx-x16.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1736 | "src/f32-vbinary/gen/vmax-avx-x8.c", |
| 1737 | "src/f32-vbinary/gen/vmax-avx-x16.c", |
| 1738 | "src/f32-vbinary/gen/vmaxc-avx-x8.c", |
| 1739 | "src/f32-vbinary/gen/vmaxc-avx-x16.c", |
| 1740 | "src/f32-vbinary/gen/vmin-avx-x8.c", |
| 1741 | "src/f32-vbinary/gen/vmin-avx-x16.c", |
| 1742 | "src/f32-vbinary/gen/vminc-avx-x8.c", |
| 1743 | "src/f32-vbinary/gen/vminc-avx-x16.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1744 | "src/f32-vbinary/gen/vmul-minmax-avx-x8.c", |
| 1745 | "src/f32-vbinary/gen/vmul-minmax-avx-x16.c", |
| 1746 | "src/f32-vbinary/gen/vmulc-minmax-avx-x8.c", |
| 1747 | "src/f32-vbinary/gen/vmulc-minmax-avx-x16.c", |
| 1748 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x8.c", |
| 1749 | "src/f32-vbinary/gen/vrdivc-minmax-avx-x16.c", |
| 1750 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x8.c", |
| 1751 | "src/f32-vbinary/gen/vrsubc-minmax-avx-x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1752 | "src/f32-vbinary/gen/vsqrdiff-avx-x8.c", |
| 1753 | "src/f32-vbinary/gen/vsqrdiff-avx-x16.c", |
| 1754 | "src/f32-vbinary/gen/vsqrdiffc-avx-x8.c", |
| 1755 | "src/f32-vbinary/gen/vsqrdiffc-avx-x16.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1756 | "src/f32-vbinary/gen/vsub-minmax-avx-x8.c", |
| 1757 | "src/f32-vbinary/gen/vsub-minmax-avx-x16.c", |
| 1758 | "src/f32-vbinary/gen/vsubc-minmax-avx-x8.c", |
| 1759 | "src/f32-vbinary/gen/vsubc-minmax-avx-x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1760 | "src/f32-vlrelu/gen/vlrelu-avx-x8.c", |
| 1761 | "src/f32-vlrelu/gen/vlrelu-avx-x16.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1762 | "src/f32-vrnd/gen/vrndd-avx-x8.c", |
| 1763 | "src/f32-vrnd/gen/vrndd-avx-x16.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1764 | "src/f32-vrnd/gen/vrndne-avx-x8.c", |
| 1765 | "src/f32-vrnd/gen/vrndne-avx-x16.c", |
| 1766 | "src/f32-vrnd/gen/vrndu-avx-x8.c", |
| 1767 | "src/f32-vrnd/gen/vrndu-avx-x16.c", |
| 1768 | "src/f32-vrnd/gen/vrndz-avx-x8.c", |
| 1769 | "src/f32-vrnd/gen/vrndz-avx-x16.c", |
| 1770 | "src/f32-vscale/avx-x32.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1771 | "src/f32-vsqrt/gen/avx-sqrt-x8.c", |
| 1772 | "src/f32-vsqrt/gen/avx-sqrt-x16.c", |
| 1773 | "src/f32-vunary/gen/vabs-avx-x8.c", |
| 1774 | "src/f32-vunary/gen/vabs-avx-x16.c", |
| 1775 | "src/f32-vunary/gen/vneg-avx-x8.c", |
| 1776 | "src/f32-vunary/gen/vneg-avx-x16.c", |
| 1777 | "src/f32-vunary/gen/vsqr-avx-x8.c", |
| 1778 | "src/f32-vunary/gen/vsqr-avx-x16.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1779 | "src/math/exp-avx-rr2-p5.c", |
| 1780 | "src/math/expm1minus-avx-rr2-lut4-p4-perm.c", |
| 1781 | "src/math/expm1minus-avx-rr2-lut16-p3.c", |
| 1782 | "src/math/expm1minus-avx-rr2-p6.c", |
| 1783 | "src/math/sigmoid-avx-rr2-lut64-p2-div.c", |
| 1784 | "src/math/sigmoid-avx-rr2-p5-div.c", |
| 1785 | "src/math/sigmoid-avx-rr2-p5-nr1.c", |
| 1786 | "src/math/sigmoid-avx-rr2-p5-nr2.c", |
| 1787 | ] |
| 1788 | |
| 1789 | XOP_UKERNELS = [ |
| 1790 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld64.c", |
| 1791 | "src/qs8-gemm/gen/1x4c2-minmax-xop-ld128.c", |
| 1792 | "src/qs8-gemm/gen/1x4c2-xw-minmax-xop.c", |
| 1793 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld64.c", |
| 1794 | "src/qs8-gemm/gen/1x4c8-minmax-xop-ld128.c", |
| 1795 | "src/qs8-gemm/gen/1x4c8-xw-minmax-xop.c", |
| 1796 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld64.c", |
| 1797 | "src/qs8-gemm/gen/2x4c8-minmax-xop-ld128.c", |
| 1798 | "src/qs8-gemm/gen/2x4c8-xw-minmax-xop.c", |
| 1799 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld64.c", |
| 1800 | "src/qs8-gemm/gen/3x4c8-minmax-xop-ld128.c", |
| 1801 | "src/qs8-gemm/gen/3x4c8-xw-minmax-xop.c", |
| 1802 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld64.c", |
| 1803 | "src/qs8-gemm/gen/4x4c2-minmax-xop-ld128.c", |
| 1804 | "src/qs8-gemm/gen/4x4c2-xw-minmax-xop.c", |
| 1805 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld64.c", |
| 1806 | "src/qs8-igemm/gen/1x4c2-minmax-xop-ld128.c", |
| 1807 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld64.c", |
| 1808 | "src/qs8-igemm/gen/1x4c8-minmax-xop-ld128.c", |
| 1809 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld64.c", |
| 1810 | "src/qs8-igemm/gen/2x4c8-minmax-xop-ld128.c", |
| 1811 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld64.c", |
| 1812 | "src/qs8-igemm/gen/3x4c8-minmax-xop-ld128.c", |
| 1813 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld64.c", |
| 1814 | "src/qs8-igemm/gen/4x4c2-minmax-xop-ld128.c", |
| 1815 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x8.c", |
| 1816 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x16.c", |
| 1817 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x24.c", |
| 1818 | "src/qs8-vadd/gen/minmax-xop-mul32-ld32-x32.c", |
| 1819 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x8.c", |
| 1820 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x16.c", |
| 1821 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x24.c", |
| 1822 | "src/qs8-vaddc/gen/minmax-xop-mul32-ld32-x32.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1823 | ] |
| 1824 | |
| 1825 | FMA3_UKERNELS = [ |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1826 | "src/f32-dwconv/gen/up8x4-minmax-fma3-acc2.c", |
| 1827 | "src/f32-dwconv/gen/up8x4-minmax-fma3.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1828 | "src/f32-dwconv/gen/up8x9-minmax-fma3-acc2.c", |
| 1829 | "src/f32-dwconv/gen/up8x9-minmax-fma3.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1830 | "src/f32-dwconv/gen/up8x25-minmax-fma3-acc2.c", |
| 1831 | "src/f32-dwconv/gen/up8x25-minmax-fma3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1832 | "src/f32-dwconv/gen/up16x4-minmax-fma3-acc2.c", |
| 1833 | "src/f32-dwconv/gen/up16x4-minmax-fma3.c", |
| 1834 | "src/f32-dwconv/gen/up16x9-minmax-fma3-acc2.c", |
| 1835 | "src/f32-dwconv/gen/up16x9-minmax-fma3.c", |
| 1836 | "src/f32-dwconv/gen/up16x25-minmax-fma3-acc2.c", |
| 1837 | "src/f32-dwconv/gen/up16x25-minmax-fma3.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1838 | "src/f32-gemm/gen-inc/1x8inc-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1839 | "src/f32-gemm/gen-inc/1x16inc-minmax-fma3-broadcast.c", |
| 1840 | "src/f32-gemm/gen-inc/1x16s4inc-minmax-fma3-broadcast.c", |
| 1841 | "src/f32-gemm/gen-inc/3x16inc-minmax-fma3-broadcast.c", |
| 1842 | "src/f32-gemm/gen-inc/3x16s4inc-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1843 | "src/f32-gemm/gen-inc/4x8inc-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1844 | "src/f32-gemm/gen-inc/4x16inc-minmax-fma3-broadcast.c", |
| 1845 | "src/f32-gemm/gen-inc/4x16s4inc-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1846 | "src/f32-gemm/gen-inc/5x8inc-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1847 | "src/f32-gemm/gen-inc/5x16inc-minmax-fma3-broadcast.c", |
| 1848 | "src/f32-gemm/gen-inc/5x16s4inc-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1849 | "src/f32-gemm/gen-inc/6x8inc-minmax-fma3-broadcast.c", |
| 1850 | "src/f32-gemm/gen-inc/7x8inc-minmax-fma3-broadcast.c", |
| 1851 | "src/f32-gemm/gen-inc/8x8inc-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1852 | "src/f32-gemm/gen/1x8-minmax-fma3-broadcast.c", |
| 1853 | "src/f32-gemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1854 | "src/f32-gemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1855 | "src/f32-gemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1856 | "src/f32-gemm/gen/3x16s4-minmax-fma3-broadcast.c", |
| 1857 | "src/f32-gemm/gen/4x8-minmax-fma3-broadcast.c", |
| 1858 | "src/f32-gemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1859 | "src/f32-gemm/gen/4x16s4-minmax-fma3-broadcast.c", |
| 1860 | "src/f32-gemm/gen/5x8-minmax-fma3-broadcast.c", |
| 1861 | "src/f32-gemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1862 | "src/f32-gemm/gen/5x16s4-minmax-fma3-broadcast.c", |
| 1863 | "src/f32-gemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1864 | "src/f32-gemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1865 | "src/f32-gemm/gen/8x8-minmax-fma3-broadcast.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1866 | "src/f32-hswish/gen/hswish-fma3-x8.c", |
| 1867 | "src/f32-hswish/gen/hswish-fma3-x16.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1868 | "src/f32-igemm/gen/1x8-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1869 | "src/f32-igemm/gen/1x16-minmax-fma3-broadcast.c", |
| 1870 | "src/f32-igemm/gen/1x16s4-minmax-fma3-broadcast.c", |
| 1871 | "src/f32-igemm/gen/3x16-minmax-fma3-broadcast.c", |
| 1872 | "src/f32-igemm/gen/3x16s4-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1873 | "src/f32-igemm/gen/4x8-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1874 | "src/f32-igemm/gen/4x16-minmax-fma3-broadcast.c", |
| 1875 | "src/f32-igemm/gen/4x16s4-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1876 | "src/f32-igemm/gen/5x8-minmax-fma3-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1877 | "src/f32-igemm/gen/5x16-minmax-fma3-broadcast.c", |
| 1878 | "src/f32-igemm/gen/5x16s4-minmax-fma3-broadcast.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 1879 | "src/f32-igemm/gen/6x8-minmax-fma3-broadcast.c", |
| 1880 | "src/f32-igemm/gen/7x8-minmax-fma3-broadcast.c", |
| 1881 | "src/f32-igemm/gen/8x8-minmax-fma3-broadcast.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1882 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x8.c", |
| 1883 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x16.c", |
| 1884 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x24.c", |
| 1885 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x32.c", |
| 1886 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x40.c", |
| 1887 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x48.c", |
| 1888 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x56.c", |
| 1889 | "src/f32-vsqrt/gen/fma3-nr1fma1adj-x64.c", |
| 1890 | "src/math/sqrt-fma3-nr1fma.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 1891 | "src/math/sqrt-fma3-nr1fma1adj.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1892 | "src/math/sqrt-fma3-nr2fma.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1893 | ] |
| 1894 | |
| 1895 | AVX2_UKERNELS = [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1896 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1897 | "src/f32-raddexpminusmax/gen/avx2-p5-x64-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1898 | "src/f32-raddexpminusmax/gen/avx2-p5-x64.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1899 | "src/f32-raddexpminusmax/gen/avx2-p5-x72-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1900 | "src/f32-raddexpminusmax/gen/avx2-p5-x72.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1901 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1902 | "src/f32-raddexpminusmax/gen/avx2-p5-x80-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1903 | "src/f32-raddexpminusmax/gen/avx2-p5-x80.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1904 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1905 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1906 | "src/f32-raddexpminusmax/gen/avx2-p5-x96-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1907 | "src/f32-raddexpminusmax/gen/avx2-p5-x96.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1908 | "src/f32-raddextexp/gen/avx2-p5-x64-acc2.c", |
| 1909 | "src/f32-raddextexp/gen/avx2-p5-x64-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1910 | "src/f32-raddextexp/gen/avx2-p5-x64.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1911 | "src/f32-raddextexp/gen/avx2-p5-x72-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1912 | "src/f32-raddextexp/gen/avx2-p5-x72.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1913 | "src/f32-raddextexp/gen/avx2-p5-x80-acc2.c", |
| 1914 | "src/f32-raddextexp/gen/avx2-p5-x80-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1915 | "src/f32-raddextexp/gen/avx2-p5-x80.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1916 | "src/f32-raddextexp/gen/avx2-p5-x96-acc2.c", |
| 1917 | "src/f32-raddextexp/gen/avx2-p5-x96-acc3.c", |
| 1918 | "src/f32-raddextexp/gen/avx2-p5-x96-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1919 | "src/f32-raddextexp/gen/avx2-p5-x96.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1920 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc2.c", |
| 1921 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1922 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x64.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1923 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1924 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x72.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1925 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc2.c", |
| 1926 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1927 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x80.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1928 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc2.c", |
| 1929 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc3.c", |
| 1930 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1931 | "src/f32-raddstoreexpminusmax/gen/avx2-p5-x96.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1932 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x8.c", |
| 1933 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x16.c", |
| 1934 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x24.c", |
| 1935 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x32.c", |
| 1936 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x40.c", |
| 1937 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x48.c", |
| 1938 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x56.c", |
| 1939 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x64.c", |
| 1940 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x72.c", |
| 1941 | "src/f32-sigmoid/gen/avx2-rr1-p5-div-x80.c", |
| 1942 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x8.c", |
| 1943 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x16.c", |
| 1944 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x24.c", |
| 1945 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x32.c", |
| 1946 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x40.c", |
| 1947 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x48.c", |
| 1948 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x56.c", |
| 1949 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x64.c", |
| 1950 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x72.c", |
| 1951 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr1fma-x80.c", |
| 1952 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x8.c", |
| 1953 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x16.c", |
| 1954 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x24.c", |
| 1955 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x32.c", |
| 1956 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x40.c", |
| 1957 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x48.c", |
| 1958 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x56.c", |
| 1959 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x64.c", |
| 1960 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x72.c", |
| 1961 | "src/f32-sigmoid/gen/avx2-rr1-p5-nr2fma-x80.c", |
| 1962 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x8.c", |
| 1963 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x16.c", |
| 1964 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x24.c", |
| 1965 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x32.c", |
| 1966 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x40.c", |
| 1967 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x48.c", |
| 1968 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x56.c", |
| 1969 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x64.c", |
| 1970 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x72.c", |
| 1971 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x80.c", |
| 1972 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x88.c", |
| 1973 | "src/f32-vscaleexpminusmax/gen/avx2-p5-x96.c", |
| 1974 | "src/f32-vscaleextexp/gen/avx2-p5-x8.c", |
| 1975 | "src/f32-vscaleextexp/gen/avx2-p5-x16.c", |
| 1976 | "src/f32-vscaleextexp/gen/avx2-p5-x24.c", |
| 1977 | "src/f32-vscaleextexp/gen/avx2-p5-x32.c", |
| 1978 | "src/f32-vscaleextexp/gen/avx2-p5-x40.c", |
| 1979 | "src/f32-vscaleextexp/gen/avx2-p5-x48.c", |
| 1980 | "src/f32-vscaleextexp/gen/avx2-p5-x56.c", |
| 1981 | "src/f32-vscaleextexp/gen/avx2-p5-x64.c", |
| 1982 | "src/f32-vscaleextexp/gen/avx2-p5-x72.c", |
| 1983 | "src/f32-vscaleextexp/gen/avx2-p5-x80.c", |
| 1984 | "src/f32-vscaleextexp/gen/avx2-p5-x88.c", |
| 1985 | "src/f32-vscaleextexp/gen/avx2-p5-x96.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1986 | "src/math/exp-avx2-rr2-lut8-p3-perm.c", |
| 1987 | "src/math/exp-avx2-rr2-lut8-p4-perm.c", |
| 1988 | "src/math/exp-avx2-rr2-p5.c", |
| 1989 | "src/math/expminus-avx2-rr2-p5.c", |
| 1990 | "src/math/expm1minus-avx2-rr1-lut4-p4-perm.c", |
| 1991 | "src/math/expm1minus-avx2-rr1-lut8-p4-perm.c", |
| 1992 | "src/math/expm1minus-avx2-rr1-lut16-p3-gather.c", |
| 1993 | "src/math/expm1minus-avx2-rr1-p6.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1994 | "src/math/extexp-avx2-p5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 1995 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-div.c", |
| 1996 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr1fma.c", |
| 1997 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma.c", |
| 1998 | "src/math/sigmoid-avx2-rr1-lut64-p2-gather-nr2fma1adj.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 1999 | "src/math/sigmoid-avx2-rr1-p5-div.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2000 | "src/math/sigmoid-avx2-rr1-p5-nr1fma.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2001 | "src/math/sigmoid-avx2-rr1-p5-nr2fma.c", |
| 2002 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-div.c", |
| 2003 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr1fma.c", |
| 2004 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma.c", |
| 2005 | "src/math/sigmoid-avx2-rr2-lut64-p2-gather-nr2fma1adj.c", |
| 2006 | "src/math/sigmoid-avx2-rr2-p5-div.c", |
| 2007 | "src/math/sigmoid-avx2-rr2-p5-nr1fma.c", |
| 2008 | "src/math/sigmoid-avx2-rr2-p5-nr2fma.c", |
| 2009 | "src/qs8-dwconv/gen/up8x9-minmax-avx2-mul32.c", |
| 2010 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul16.c", |
| 2011 | "src/qs8-dwconv/gen/up16x9-minmax-avx2-mul32.c", |
| 2012 | "src/qs8-dwconv/gen/up24x9-minmax-avx2-mul32.c", |
| 2013 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul16.c", |
| 2014 | "src/qs8-dwconv/gen/up32x9-minmax-avx2-mul32.c", |
| 2015 | "src/qs8-gemm/gen/1x8c8-minmax-avx2.c", |
| 2016 | "src/qs8-gemm/gen/1x8c8-xw-minmax-avx2.c", |
| 2017 | "src/qs8-gemm/gen/2x8c8-minmax-avx2.c", |
| 2018 | "src/qs8-gemm/gen/2x8c8-xw-minmax-avx2.c", |
| 2019 | "src/qs8-gemm/gen/3x8c8-minmax-avx2.c", |
| 2020 | "src/qs8-gemm/gen/3x8c8-xw-minmax-avx2.c", |
| 2021 | "src/qs8-igemm/gen/1x8c8-minmax-avx2.c", |
| 2022 | "src/qs8-igemm/gen/2x8c8-minmax-avx2.c", |
| 2023 | "src/qs8-igemm/gen/3x8c8-minmax-avx2.c", |
| 2024 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2025 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2026 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2027 | "src/qs8-vadd/gen/minmax-avx2-mul32-ld64-x32.c", |
| 2028 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x8.c", |
| 2029 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x16.c", |
| 2030 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x24.c", |
| 2031 | "src/qs8-vaddc/gen/minmax-avx2-mul32-ld64-x32.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2032 | ] |
| 2033 | |
| 2034 | AVX512F_UKERNELS = [ |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 2035 | "src/f32-clamp/gen/avx512f-x16.c", |
| 2036 | "src/f32-clamp/gen/avx512f-x32.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2037 | "src/f32-dwconv/gen/up16x4-minmax-avx512f-acc2.c", |
| 2038 | "src/f32-dwconv/gen/up16x4-minmax-avx512f.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2039 | "src/f32-dwconv/gen/up16x9-minmax-avx512f-acc2.c", |
| 2040 | "src/f32-dwconv/gen/up16x9-minmax-avx512f.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2041 | "src/f32-dwconv/gen/up16x25-minmax-avx512f-acc2.c", |
| 2042 | "src/f32-dwconv/gen/up16x25-minmax-avx512f.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2043 | "src/f32-dwconv/gen/up32x4-minmax-avx512f-acc2.c", |
| 2044 | "src/f32-dwconv/gen/up32x4-minmax-avx512f.c", |
| 2045 | "src/f32-dwconv/gen/up32x9-minmax-avx512f-acc2.c", |
| 2046 | "src/f32-dwconv/gen/up32x9-minmax-avx512f.c", |
| 2047 | "src/f32-dwconv/gen/up32x25-minmax-avx512f-acc2.c", |
| 2048 | "src/f32-dwconv/gen/up32x25-minmax-avx512f.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2049 | "src/f32-gemm/gen-inc/1x16inc-minmax-avx512f-broadcast.c", |
| 2050 | "src/f32-gemm/gen-inc/4x16inc-minmax-avx512f-broadcast.c", |
| 2051 | "src/f32-gemm/gen-inc/5x16inc-minmax-avx512f-broadcast.c", |
| 2052 | "src/f32-gemm/gen-inc/6x16inc-minmax-avx512f-broadcast.c", |
| 2053 | "src/f32-gemm/gen-inc/7x16inc-minmax-avx512f-broadcast.c", |
| 2054 | "src/f32-gemm/gen-inc/8x16inc-minmax-avx512f-broadcast.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2055 | "src/f32-gemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2056 | "src/f32-gemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2057 | "src/f32-gemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2058 | "src/f32-gemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2059 | "src/f32-gemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2060 | "src/f32-gemm/gen/8x16-minmax-avx512f-broadcast.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2061 | "src/f32-hswish/gen/hswish-avx512f-x16.c", |
| 2062 | "src/f32-hswish/gen/hswish-avx512f-x32.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2063 | "src/f32-igemm/gen/1x16-minmax-avx512f-broadcast.c", |
| 2064 | "src/f32-igemm/gen/4x16-minmax-avx512f-broadcast.c", |
| 2065 | "src/f32-igemm/gen/5x16-minmax-avx512f-broadcast.c", |
| 2066 | "src/f32-igemm/gen/6x16-minmax-avx512f-broadcast.c", |
| 2067 | "src/f32-igemm/gen/7x16-minmax-avx512f-broadcast.c", |
| 2068 | "src/f32-igemm/gen/8x16-minmax-avx512f-broadcast.c", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 2069 | "src/f32-prelu/gen/avx512f-2x16.c", |
| 2070 | "src/f32-prelu/gen/avx512f-2x32.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2071 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2072 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2073 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x128.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2074 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2075 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x144.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2076 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2077 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2078 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x160.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2079 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2080 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2081 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2082 | "src/f32-raddexpminusmax/gen/avx512f-p5-scalef-x192.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2083 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2084 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2085 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x128.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2086 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2087 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x144.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2088 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2089 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2090 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x160.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2091 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2092 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2093 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2094 | "src/f32-raddextexp/gen/avx512f-p5-scalef-x192.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2095 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc2.c", |
| 2096 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128-acc4.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2097 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x128.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2098 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144-acc3.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2099 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x144.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2100 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc2.c", |
| 2101 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160-acc5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2102 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x160.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2103 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc2.c", |
| 2104 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc3.c", |
| 2105 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192-acc6.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2106 | "src/f32-raddstoreexpminusmax/gen/avx512f-p5-scalef-x192.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2107 | "src/f32-relu/gen/avx512f-x16.c", |
| 2108 | "src/f32-relu/gen/avx512f-x32.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2109 | "src/f32-rmax/avx512f.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2110 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x16.c", |
| 2111 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x32.c", |
| 2112 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x48.c", |
| 2113 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x64.c", |
| 2114 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x80.c", |
| 2115 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x96.c", |
| 2116 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x112.c", |
| 2117 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-div-x128.c", |
| 2118 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x16.c", |
| 2119 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x32.c", |
| 2120 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x48.c", |
| 2121 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x64.c", |
| 2122 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x80.c", |
| 2123 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x96.c", |
| 2124 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x112.c", |
| 2125 | "src/f32-sigmoid/gen/avx512f-rr1-lut16-p3-perm-scalef-nr1fma-x128.c", |
| 2126 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x16.c", |
| 2127 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x32.c", |
| 2128 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x48.c", |
| 2129 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x64.c", |
| 2130 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x80.c", |
| 2131 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x96.c", |
| 2132 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x112.c", |
| 2133 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-div-x128.c", |
| 2134 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x16.c", |
| 2135 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x32.c", |
| 2136 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x48.c", |
| 2137 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x64.c", |
| 2138 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x80.c", |
| 2139 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x96.c", |
| 2140 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x112.c", |
| 2141 | "src/f32-sigmoid/gen/avx512f-rr1-p5-scalef-nr1fma-x128.c", |
| 2142 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x16.c", |
| 2143 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x32.c", |
| 2144 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x48.c", |
| 2145 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x64.c", |
| 2146 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x80.c", |
| 2147 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x96.c", |
| 2148 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x112.c", |
| 2149 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-div-x128.c", |
| 2150 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x16.c", |
| 2151 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x32.c", |
| 2152 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x48.c", |
| 2153 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x64.c", |
| 2154 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x80.c", |
| 2155 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x96.c", |
| 2156 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x112.c", |
| 2157 | "src/f32-sigmoid/gen/avx512f-rr2-lut32-p2-perm2-scalef-nr1fma-x128.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2158 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x16.c", |
| 2159 | "src/f32-vbinary/gen/vadd-minmax-avx512f-x32.c", |
| 2160 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x16.c", |
| 2161 | "src/f32-vbinary/gen/vaddc-minmax-avx512f-x32.c", |
| 2162 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x16.c", |
| 2163 | "src/f32-vbinary/gen/vdiv-minmax-avx512f-x32.c", |
| 2164 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x16.c", |
| 2165 | "src/f32-vbinary/gen/vdivc-minmax-avx512f-x32.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2166 | "src/f32-vbinary/gen/vmax-avx512f-x16.c", |
| 2167 | "src/f32-vbinary/gen/vmax-avx512f-x32.c", |
| 2168 | "src/f32-vbinary/gen/vmaxc-avx512f-x16.c", |
| 2169 | "src/f32-vbinary/gen/vmaxc-avx512f-x32.c", |
| 2170 | "src/f32-vbinary/gen/vmin-avx512f-x16.c", |
| 2171 | "src/f32-vbinary/gen/vmin-avx512f-x32.c", |
| 2172 | "src/f32-vbinary/gen/vminc-avx512f-x16.c", |
| 2173 | "src/f32-vbinary/gen/vminc-avx512f-x32.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2174 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x16.c", |
| 2175 | "src/f32-vbinary/gen/vmul-minmax-avx512f-x32.c", |
| 2176 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x16.c", |
| 2177 | "src/f32-vbinary/gen/vmulc-minmax-avx512f-x32.c", |
| 2178 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x16.c", |
| 2179 | "src/f32-vbinary/gen/vrdivc-minmax-avx512f-x32.c", |
| 2180 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x16.c", |
| 2181 | "src/f32-vbinary/gen/vrsubc-minmax-avx512f-x32.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2182 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x16.c", |
| 2183 | "src/f32-vbinary/gen/vsqrdiff-avx512f-x32.c", |
| 2184 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x16.c", |
| 2185 | "src/f32-vbinary/gen/vsqrdiffc-avx512f-x32.c", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2186 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x16.c", |
| 2187 | "src/f32-vbinary/gen/vsub-minmax-avx512f-x32.c", |
| 2188 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x16.c", |
| 2189 | "src/f32-vbinary/gen/vsubc-minmax-avx512f-x32.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2190 | "src/f32-vlrelu/gen/vlrelu-avx512f-x16.c", |
| 2191 | "src/f32-vlrelu/gen/vlrelu-avx512f-x32.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2192 | "src/f32-vrnd/gen/vrndd-avx512f-x16.c", |
| 2193 | "src/f32-vrnd/gen/vrndd-avx512f-x32.c", |
| 2194 | "src/f32-vrnd/gen/vrndne-avx512f-x16.c", |
| 2195 | "src/f32-vrnd/gen/vrndne-avx512f-x32.c", |
| 2196 | "src/f32-vrnd/gen/vrndu-avx512f-x16.c", |
| 2197 | "src/f32-vrnd/gen/vrndu-avx512f-x32.c", |
| 2198 | "src/f32-vrnd/gen/vrndz-avx512f-x16.c", |
| 2199 | "src/f32-vrnd/gen/vrndz-avx512f-x32.c", |
| 2200 | "src/f32-vscale/avx512f-x64.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2201 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x16.c", |
| 2202 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x32.c", |
| 2203 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x48.c", |
| 2204 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x64.c", |
| 2205 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x80.c", |
| 2206 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x96.c", |
| 2207 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x112.c", |
| 2208 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x128.c", |
| 2209 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x144.c", |
| 2210 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x160.c", |
| 2211 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x176.c", |
| 2212 | "src/f32-vscaleexpminusmax/gen/avx512f-p5-scalef-x192.c", |
| 2213 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x16.c", |
| 2214 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x32.c", |
| 2215 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x48.c", |
| 2216 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x64.c", |
| 2217 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x80.c", |
| 2218 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x96.c", |
| 2219 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x112.c", |
| 2220 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x128.c", |
| 2221 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x144.c", |
| 2222 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x160.c", |
| 2223 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x176.c", |
| 2224 | "src/f32-vscaleextexp/gen/avx512f-p5-scalef-x192.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2225 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x16.c", |
| 2226 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x32.c", |
| 2227 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x48.c", |
| 2228 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x64.c", |
| 2229 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x80.c", |
| 2230 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x96.c", |
| 2231 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x112.c", |
| 2232 | "src/f32-vsqrt/gen/avx512f-nr1fma1adj-x128.c", |
| 2233 | "src/f32-vunary/gen/vabs-avx512f-x16.c", |
| 2234 | "src/f32-vunary/gen/vabs-avx512f-x32.c", |
| 2235 | "src/f32-vunary/gen/vneg-avx512f-x16.c", |
| 2236 | "src/f32-vunary/gen/vneg-avx512f-x32.c", |
| 2237 | "src/f32-vunary/gen/vsqr-avx512f-x16.c", |
| 2238 | "src/f32-vunary/gen/vsqr-avx512f-x32.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2239 | "src/math/exp-avx512f-rr2-lut16-p3-perm-scalef.c", |
| 2240 | "src/math/exp-avx512f-rr2-lut16-p3-perm.c", |
| 2241 | "src/math/exp-avx512f-rr2-lut32-p2-perm2-scalef.c", |
| 2242 | "src/math/exp-avx512f-rr2-lut32-p2-perm2.c", |
| 2243 | "src/math/exp-avx512f-rr2-p5-scalef.c", |
| 2244 | "src/math/exp-avx512f-rr2-p5.c", |
| 2245 | "src/math/expm1minus-avx512f-rr1-lut16-p3-perm.c", |
| 2246 | "src/math/expm1minus-avx512f-rr1-p6.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2247 | "src/math/extexp-avx512f-p5.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2248 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-div.c", |
| 2249 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma.c", |
| 2250 | "src/math/sigmoid-avx512f-rr1-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2251 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-div.c", |
| 2252 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma.c", |
| 2253 | "src/math/sigmoid-avx512f-rr1-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| 2254 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-div.c", |
| 2255 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma.c", |
| 2256 | "src/math/sigmoid-avx512f-rr1-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2257 | "src/math/sigmoid-avx512f-rr1-p5-scalef-div.c", |
| 2258 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma.c", |
| 2259 | "src/math/sigmoid-avx512f-rr1-p5-scalef-nr1fma1adj.c", |
| 2260 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-div.c", |
| 2261 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma.c", |
| 2262 | "src/math/sigmoid-avx512f-rr2-lut16-p3-perm-scalef-nr1fma1adj.c", |
| 2263 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-div.c", |
| 2264 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma.c", |
| 2265 | "src/math/sigmoid-avx512f-rr2-lut32-p2-perm2-scalef-nr1fma1adj.c", |
| 2266 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-div.c", |
| 2267 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma.c", |
| 2268 | "src/math/sigmoid-avx512f-rr2-lut64-p2-gather-scalef-nr1fma1adj.c", |
| 2269 | "src/math/sigmoid-avx512f-rr2-p5-scalef-div.c", |
| 2270 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma.c", |
| 2271 | "src/math/sigmoid-avx512f-rr2-p5-scalef-nr1fma1adj.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2272 | "src/math/sqrt-avx512f-nr1fma.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2273 | "src/math/sqrt-avx512f-nr1fma1adj.c", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2274 | "src/math/sqrt-avx512f-nr2fma.c", |
| 2275 | ] |
| 2276 | |
| 2277 | AVX512SKX_UKERNELS = [ |
| 2278 | "src/qs8-dwconv/gen/up16x9-minmax-avx512skx-mul32.c", |
| 2279 | "src/qs8-dwconv/gen/up32x9-minmax-avx512skx-mul32.c", |
| 2280 | "src/qs8-gemm/gen/1x16c8-minmax-avx512skx.c", |
| 2281 | "src/qs8-gemm/gen/2x16c8-minmax-avx512skx.c", |
| 2282 | "src/qs8-gemm/gen/3x16c8-minmax-avx512skx.c", |
| 2283 | "src/qs8-gemm/gen/4x16c8-minmax-avx512skx.c", |
| 2284 | "src/qs8-igemm/gen/1x16c8-minmax-avx512skx.c", |
| 2285 | "src/qs8-igemm/gen/2x16c8-minmax-avx512skx.c", |
| 2286 | "src/qs8-igemm/gen/3x16c8-minmax-avx512skx.c", |
| 2287 | "src/qs8-igemm/gen/4x16c8-minmax-avx512skx.c", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2288 | ] |
| 2289 | |
| 2290 | AARCH32_ASM_UKERNELS = [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2291 | "src/f32-gemm/4x4-aarch32-vfp-ld64.S", |
| 2292 | "src/f32-gemm/4x4-minmax-aarch32-vfp-ld64.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2293 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2294 | "src/f32-gemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2295 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2296 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2297 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2298 | "src/f32-gemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2299 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a53.S", |
| 2300 | "src/f32-igemm/4x8-minmax-aarch32-neon-cortex-a55.S", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2301 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a7.S", |
| 2302 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-cortex-a75.S", |
| 2303 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-ld64.S", |
| 2304 | "src/f32-igemm/gen/4x8-minmax-aarch32-neon-pld-cortex-a75.S", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2305 | ] |
| 2306 | |
| 2307 | AARCH64_ASM_UKERNELS = [ |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2308 | "src/f16-gemm/gen-inc/1x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2309 | "src/f16-gemm/gen-inc/1x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2310 | "src/f16-gemm/gen-inc/4x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2311 | "src/f16-gemm/gen-inc/4x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2312 | "src/f16-gemm/gen-inc/6x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2313 | "src/f16-gemm/gen-inc/6x16inc-minmax-aarch64-neonfp16arith-ld32.S", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2314 | "src/f16-gemm/gen-inc/8x8inc-minmax-aarch64-neonfp16arith-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2315 | "src/f16-gemm/gen/1x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2316 | "src/f16-gemm/gen/1x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2317 | "src/f16-gemm/gen/4x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2318 | "src/f16-gemm/gen/4x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2319 | "src/f16-gemm/gen/6x8-minmax-aarch64-neonfp16arith-ld64.S", |
| 2320 | "src/f16-gemm/gen/6x16-minmax-aarch64-neonfp16arith-ld32.S", |
| 2321 | "src/f16-gemm/gen/8x8-minmax-aarch64-neonfp16arith-ld64.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2322 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma-cortex-a55.S", |
| 2323 | "src/f32-dwconv/up4x9-minmax-aarch64-neonfma.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2324 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2325 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2326 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2327 | "src/f32-gemm/gen-inc/1x8inc-minmax-aarch64-neonfma-ld64.S", |
| 2328 | "src/f32-gemm/gen-inc/1x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2329 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2330 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
| 2331 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2332 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2333 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2334 | "src/f32-gemm/gen-inc/4x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2335 | "src/f32-gemm/gen-inc/4x12inc-minmax-aarch64-neonfma-cortex-a53.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2336 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
| 2337 | "src/f32-gemm/gen-inc/5x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
| 2338 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a53.S", |
| 2339 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a55.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2340 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a57.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2341 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a73.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2342 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2343 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld64.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2344 | "src/f32-gemm/gen-inc/6x8inc-minmax-aarch64-neonfma-ld128.S", |
| 2345 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2346 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2347 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2348 | "src/f32-gemm/gen/1x8-minmax-aarch64-neonfma-ld64.S", |
| 2349 | "src/f32-gemm/gen/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2350 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2351 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2352 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2353 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2354 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld64.S", |
| 2355 | "src/f32-gemm/gen/4x8-minmax-aarch64-neonfma-ld128.S", |
| 2356 | "src/f32-gemm/gen/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2357 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2358 | "src/f32-gemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2359 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2360 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2361 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2362 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2363 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2364 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld64.S", |
| 2365 | "src/f32-gemm/gen/6x8-minmax-aarch64-neonfma-ld128.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2366 | "src/f32-igemm/1x8-minmax-aarch64-neonfma-cortex-a53.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2367 | "src/f32-igemm/1x12-minmax-aarch64-neonfma-cortex-a53.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2368 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2369 | "src/f32-igemm/4x8-minmax-aarch64-neonfma-cortex-a55.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2370 | "src/f32-igemm/4x12-minmax-aarch64-neonfma-cortex-a53.S", |
| 2371 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a53.S", |
| 2372 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a55.S", |
| 2373 | "src/f32-igemm/6x8-minmax-aarch64-neonfma-cortex-a73.S", |
| 2374 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2375 | "src/f32-igemm/gen/1x8-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2376 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2377 | "src/f32-igemm/gen/4x8-minmax-aarch64-neonfma-cortex-a75.S", |
| 2378 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2379 | "src/f32-igemm/gen/5x8-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2380 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a57.S", |
| 2381 | "src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-cortex-a75.S", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2382 | "src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S", |
| 2383 | "src/qs8-gemm/4x16c4-aarch64-neondot-cortex-a55.S", |
| 2384 | "src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2385 | ] |
| 2386 | |
| 2387 | cc_defaults { |
| 2388 | name: "xnnpack_internal_default", |
| 2389 | vendor_available: true, |
| 2390 | sdk_version: "current", |
| 2391 | local_include_dirs: [ |
| 2392 | "include", |
| 2393 | "src", |
| 2394 | ], |
| 2395 | cflags: [ |
| 2396 | "-std=c99", |
| 2397 | "-DXNN_LOG_LEVEL=2", |
| 2398 | "-Wno-unused-parameter", |
| 2399 | "-Wno-missing-field-initializers", |
| 2400 | "-Wno-pointer-arith", |
Miao Wang | ef75bc9 | 2020-07-25 09:15:12 -0700 | [diff] [blame] | 2401 | "-Wno-implicit-function-declaration", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2402 | ], |
| 2403 | stl: "libc++_static", |
| 2404 | } |
| 2405 | |
| 2406 | cc_library_static { |
| 2407 | name: "xnnpack_tables", |
| 2408 | defaults: ["xnnpack_internal_default"], |
| 2409 | srcs: TABLE_SRCS, |
| 2410 | } |
| 2411 | |
| 2412 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2413 | name: "xnnpack_logging_utils", |
| 2414 | defaults: ["xnnpack_internal_default"], |
| 2415 | srcs: LOGGING_SRCS, |
| 2416 | header_libs: [ |
| 2417 | "fp16_headers", |
| 2418 | ], |
| 2419 | static_libs: [ |
| 2420 | "libclog", |
| 2421 | "libpthreadpool", |
| 2422 | ], |
| 2423 | } |
| 2424 | |
| 2425 | cc_library_static { |
| 2426 | name: "xnnpack_memory_planner", |
| 2427 | defaults: ["xnnpack_internal_default"], |
| 2428 | srcs: [ |
| 2429 | "src/memory-planner.c", |
| 2430 | ], |
| 2431 | cflags: [ |
| 2432 | "-DXNN_ENABLE_MEMOPT=1", |
| 2433 | ], |
| 2434 | header_libs: [ |
| 2435 | "fp16_headers", |
| 2436 | ], |
| 2437 | static_libs: [ |
| 2438 | "libpthreadpool", |
| 2439 | "xnnpack_logging_utils", |
| 2440 | ], |
| 2441 | } |
| 2442 | |
| 2443 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2444 | name: "xnnpack_im2col", |
| 2445 | defaults: ["xnnpack_internal_default"], |
| 2446 | srcs: [ |
| 2447 | "src/im2col.c", |
| 2448 | ], |
| 2449 | } |
| 2450 | |
| 2451 | cc_library_static { |
| 2452 | name: "xnnpack_indirection", |
| 2453 | defaults: ["xnnpack_internal_default"], |
| 2454 | srcs: [ |
| 2455 | "src/indirection.c", |
| 2456 | ], |
| 2457 | header_libs: [ |
| 2458 | "fp16_headers", |
| 2459 | "fxdiv_headers", |
| 2460 | ], |
| 2461 | static_libs: [ |
| 2462 | "libpthreadpool", |
| 2463 | ], |
| 2464 | } |
| 2465 | |
| 2466 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2467 | name: "xnnpack_packing", |
| 2468 | defaults: ["xnnpack_internal_default"], |
| 2469 | srcs: [ |
| 2470 | "src/packing.c", |
| 2471 | ], |
| 2472 | header_libs: [ |
| 2473 | "fp16_headers", |
| 2474 | "fxdiv_headers", |
| 2475 | ], |
| 2476 | static_libs: [ |
| 2477 | "libpthreadpool", |
| 2478 | ], |
| 2479 | } |
| 2480 | |
| 2481 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2482 | name: "xnnpack_operator_run", |
| 2483 | defaults: ["xnnpack_internal_default"], |
| 2484 | srcs: [ |
| 2485 | "src/operator-run.c", |
| 2486 | ], |
| 2487 | cflags: [ |
| 2488 | "-Wno-vla", |
| 2489 | ], |
| 2490 | header_libs: [ |
| 2491 | "fp16_headers", |
| 2492 | "fxdiv_headers", |
| 2493 | ], |
| 2494 | static_libs: [ |
| 2495 | "libclog", |
| 2496 | "libpthreadpool", |
| 2497 | ], |
| 2498 | } |
| 2499 | |
| 2500 | cc_library_static { |
| 2501 | name: "xnnpack_operators", |
| 2502 | defaults: ["xnnpack_internal_default"], |
| 2503 | srcs: OPERATOR_SRCS + [ |
| 2504 | "src/memory.c", |
| 2505 | "src/operator-delete.c", |
| 2506 | ], |
| 2507 | header_libs: [ |
| 2508 | "fp16_headers", |
| 2509 | "fxdiv_headers", |
| 2510 | ], |
| 2511 | static_libs: [ |
| 2512 | "libclog", |
| 2513 | "libpthreadpool", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 2514 | "xnnpack_logging_utils", |
| 2515 | "xnnpack_packing", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2516 | ], |
| 2517 | whole_static_libs: [ |
| 2518 | "xnnpack_indirection", |
| 2519 | ], |
| 2520 | } |
| 2521 | |
| 2522 | cc_library_static { |
| 2523 | name: "xnnpack_scalar_ukernels", |
| 2524 | defaults: ["xnnpack_internal_default"], |
| 2525 | srcs: SCALAR_UKERNELS, |
| 2526 | header_libs: [ |
| 2527 | "fp16_headers", |
| 2528 | "fxdiv_headers", |
| 2529 | ], |
| 2530 | static_libs: [ |
| 2531 | "libpthreadpool", |
| 2532 | "xnnpack_tables", |
| 2533 | ], |
| 2534 | } |
| 2535 | |
| 2536 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2537 | name: "xnnpack_psimd_accmath_ukernels", |
| 2538 | defaults: ["xnnpack_internal_default"], |
| 2539 | srcs: PSIMD_ACCMATH_UKERNELS, |
| 2540 | cflags: [ |
| 2541 | "-O3", |
| 2542 | ], |
| 2543 | arch: { |
| 2544 | arm: { |
| 2545 | cflags: [ |
| 2546 | "-marm", |
| 2547 | "-mfpu=neon", |
| 2548 | ], |
| 2549 | }, |
| 2550 | }, |
| 2551 | header_libs: [ |
| 2552 | "fp16_headers", |
| 2553 | "psimd_headers", |
| 2554 | ], |
| 2555 | static_libs: [ |
| 2556 | "libpthreadpool", |
| 2557 | "xnnpack_tables", |
| 2558 | ], |
| 2559 | } |
| 2560 | |
| 2561 | cc_library_static { |
| 2562 | name: "xnnpack_neon_ukernels", |
| 2563 | defaults: ["xnnpack_internal_default"], |
| 2564 | arch: { |
| 2565 | arm: { |
| 2566 | srcs: NEON_UKERNELS, |
| 2567 | cflags: [ |
| 2568 | "-marm", |
| 2569 | "-mfpu=neon", |
| 2570 | ], |
| 2571 | }, |
| 2572 | arm64: { |
| 2573 | srcs: NEON_UKERNELS, |
| 2574 | }, |
| 2575 | x86: { enabled: false, }, |
| 2576 | x86_64: { enabled: false, }, |
| 2577 | }, |
| 2578 | header_libs: [ |
| 2579 | "fp16_headers", |
| 2580 | ], |
| 2581 | static_libs: [ |
| 2582 | "libpthreadpool", |
| 2583 | "xnnpack_tables", |
| 2584 | ], |
| 2585 | } |
| 2586 | |
| 2587 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2588 | name: "xnnpack_neondot_ukernels", |
| 2589 | defaults: ["xnnpack_internal_default"], |
| 2590 | arch: { |
| 2591 | arm: { |
| 2592 | srcs: NEONDOT_UKERNELS, |
| 2593 | cflags: [ |
| 2594 | "-marm", |
| 2595 | "-march=armv8.2-a+dotprod", |
| 2596 | "-mfpu=neon-fp-armv8", |
| 2597 | ], |
| 2598 | }, |
| 2599 | arm64: { |
| 2600 | srcs: NEONDOT_UKERNELS, |
| 2601 | cflags: [ |
| 2602 | "-march=armv8.2-a+dotprod", |
| 2603 | ], |
| 2604 | }, |
| 2605 | x86: { enabled: false, }, |
| 2606 | x86_64: { enabled: false, }, |
| 2607 | }, |
| 2608 | header_libs: [ |
| 2609 | "fp16_headers", |
| 2610 | ], |
| 2611 | static_libs: [ |
| 2612 | "libpthreadpool", |
| 2613 | "xnnpack_tables", |
| 2614 | ], |
| 2615 | } |
| 2616 | |
| 2617 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2618 | name: "xnnpack_neonfma_ukernels", |
| 2619 | defaults: ["xnnpack_internal_default"], |
| 2620 | arch: { |
| 2621 | arm: { |
| 2622 | srcs: NEONFMA_UKERNELS, |
| 2623 | cflags: [ |
| 2624 | "-marm", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2625 | "-march=armv7-a", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2626 | "-mfpu=neon-vfpv4", |
| 2627 | ], |
| 2628 | }, |
| 2629 | arm64: { |
| 2630 | srcs: NEONFMA_UKERNELS + AARCH64_NEONFMA_UKERNELS, |
| 2631 | }, |
| 2632 | x86: { enabled: false, }, |
| 2633 | x86_64: { enabled: false, }, |
| 2634 | }, |
| 2635 | header_libs: [ |
| 2636 | "fp16_headers", |
| 2637 | ], |
| 2638 | static_libs: [ |
| 2639 | "libpthreadpool", |
| 2640 | "xnnpack_tables", |
| 2641 | ], |
| 2642 | } |
| 2643 | |
| 2644 | cc_library_static { |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2645 | name: "xnnpack_neonv8_ukernels", |
| 2646 | defaults: ["xnnpack_internal_default"], |
| 2647 | arch: { |
| 2648 | arm: { |
| 2649 | srcs: NEONV8_UKERNELS, |
| 2650 | cflags: [ |
| 2651 | "-marm", |
| 2652 | "-march=armv8-a", |
| 2653 | "-mfpu=neon-fp-armv8", |
| 2654 | ], |
| 2655 | }, |
| 2656 | arm64: { |
| 2657 | srcs: NEONV8_UKERNELS, |
| 2658 | }, |
| 2659 | x86: { enabled: false, }, |
| 2660 | x86_64: { enabled: false, }, |
| 2661 | }, |
| 2662 | header_libs: [ |
| 2663 | "fp16_headers", |
| 2664 | ], |
| 2665 | static_libs: [ |
| 2666 | "libpthreadpool", |
| 2667 | "xnnpack_tables", |
| 2668 | ], |
| 2669 | } |
| 2670 | |
| 2671 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2672 | name: "xnnpack_neonfp16arith_ukernels", |
| 2673 | defaults: ["xnnpack_internal_default"], |
| 2674 | arch: { |
| 2675 | arm: { enabled: false, }, |
| 2676 | arm64: { |
| 2677 | srcs: AARCH64_NEONFP16ARITH_UKERNELS, |
| 2678 | cflags: [ |
| 2679 | "-march=armv8.2-a+fp16", |
| 2680 | ], |
| 2681 | }, |
| 2682 | x86: { enabled: false, }, |
| 2683 | x86_64: { enabled: false, }, |
| 2684 | }, |
| 2685 | header_libs: [ |
| 2686 | "fp16_headers", |
| 2687 | ], |
| 2688 | static_libs: [ |
| 2689 | "libpthreadpool", |
| 2690 | "xnnpack_tables", |
| 2691 | ], |
| 2692 | } |
| 2693 | |
| 2694 | cc_library_static { |
| 2695 | name: "xnnpack_asm_ukernels", |
| 2696 | defaults: ["xnnpack_internal_default"], |
| 2697 | arch: { |
| 2698 | arm: { |
| 2699 | srcs: AARCH32_ASM_UKERNELS, |
| 2700 | }, |
| 2701 | arm64: { |
| 2702 | srcs: AARCH64_ASM_UKERNELS, |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2703 | clang_asflags: [ |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2704 | "-march=armv8.2-a+fp16+dotprod", |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 2705 | ], |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2706 | }, |
| 2707 | x86: { enabled: false, }, |
| 2708 | x86_64: { enabled: false, }, |
| 2709 | }, |
| 2710 | } |
| 2711 | |
| 2712 | cc_library_static { |
| 2713 | name: "xnnpack_sse2_ukernels", |
| 2714 | defaults: ["xnnpack_internal_default"], |
| 2715 | arch: { |
| 2716 | arm: { enabled: false, }, |
| 2717 | arm64: { enabled: false, }, |
| 2718 | x86: { |
| 2719 | srcs: SSE_UKERNELS + SSE2_UKERNELS, |
| 2720 | cflags: [ |
| 2721 | "-msse2", |
| 2722 | ], |
| 2723 | }, |
| 2724 | x86_64: { |
| 2725 | srcs: SSE_UKERNELS + SSE2_UKERNELS, |
| 2726 | cflags: [ |
| 2727 | "-msse2", |
| 2728 | ], |
| 2729 | }, |
| 2730 | }, |
| 2731 | header_libs: [ |
| 2732 | "fp16_headers", |
| 2733 | ], |
| 2734 | static_libs: [ |
| 2735 | "libpthreadpool", |
| 2736 | "xnnpack_tables", |
| 2737 | ], |
| 2738 | } |
| 2739 | |
| 2740 | cc_library_static { |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 2741 | name: "xnnpack_ssse3_ukernels", |
| 2742 | defaults: ["xnnpack_internal_default"], |
| 2743 | arch: { |
| 2744 | arm: { enabled: false, }, |
| 2745 | arm64: { enabled: false, }, |
| 2746 | x86: { |
| 2747 | srcs: SSSE3_UKERNELS, |
| 2748 | cflags: [ |
| 2749 | "-mssse3", |
| 2750 | ], |
| 2751 | }, |
| 2752 | x86_64: { |
| 2753 | srcs: SSSE3_UKERNELS, |
| 2754 | cflags: [ |
| 2755 | "-mssse3", |
| 2756 | ], |
| 2757 | }, |
| 2758 | }, |
| 2759 | header_libs: [ |
| 2760 | "fp16_headers", |
| 2761 | ], |
| 2762 | static_libs: [ |
| 2763 | "libpthreadpool", |
| 2764 | "xnnpack_tables", |
| 2765 | ], |
| 2766 | } |
| 2767 | |
| 2768 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2769 | name: "xnnpack_sse41_ukernels", |
| 2770 | defaults: ["xnnpack_internal_default"], |
| 2771 | arch: { |
| 2772 | arm: { enabled: false, }, |
| 2773 | arm64: { enabled: false, }, |
| 2774 | x86: { |
| 2775 | srcs: SSE41_UKERNELS, |
| 2776 | cflags: [ |
| 2777 | "-msse4.1", |
| 2778 | ], |
| 2779 | }, |
| 2780 | x86_64: { |
| 2781 | srcs: SSE41_UKERNELS, |
| 2782 | cflags: [ |
| 2783 | "-msse4.1", |
| 2784 | ], |
| 2785 | }, |
| 2786 | }, |
| 2787 | header_libs: [ |
| 2788 | "fp16_headers", |
| 2789 | ], |
| 2790 | static_libs: [ |
| 2791 | "libpthreadpool", |
| 2792 | "xnnpack_tables", |
| 2793 | ], |
| 2794 | } |
| 2795 | |
| 2796 | cc_library_static { |
| 2797 | name: "xnnpack_avx_ukernels", |
| 2798 | defaults: ["xnnpack_internal_default"], |
| 2799 | arch: { |
| 2800 | arm: { enabled: false, }, |
| 2801 | arm64: { enabled: false, }, |
| 2802 | x86: { |
| 2803 | srcs: AVX_UKERNELS, |
| 2804 | cflags: [ |
| 2805 | "-mavx", |
| 2806 | ], |
| 2807 | }, |
| 2808 | x86_64: { |
| 2809 | srcs: AVX_UKERNELS, |
| 2810 | cflags: [ |
| 2811 | "-mavx", |
| 2812 | ], |
| 2813 | }, |
| 2814 | }, |
| 2815 | header_libs: [ |
| 2816 | "fp16_headers", |
| 2817 | ], |
| 2818 | static_libs: [ |
| 2819 | "libpthreadpool", |
| 2820 | "xnnpack_tables", |
| 2821 | ], |
| 2822 | } |
| 2823 | |
| 2824 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2825 | name: "xnnpack_xop_ukernels", |
| 2826 | defaults: ["xnnpack_internal_default"], |
| 2827 | arch: { |
| 2828 | arm: { enabled: false, }, |
| 2829 | arm64: { enabled: false, }, |
| 2830 | x86: { |
| 2831 | srcs: XOP_UKERNELS, |
| 2832 | cflags: [ |
| 2833 | "-mxop", |
| 2834 | ], |
| 2835 | }, |
| 2836 | x86_64: { |
| 2837 | srcs: XOP_UKERNELS, |
| 2838 | cflags: [ |
| 2839 | "-mxop", |
| 2840 | ], |
| 2841 | }, |
| 2842 | }, |
| 2843 | header_libs: [ |
| 2844 | "fp16_headers", |
| 2845 | ], |
| 2846 | static_libs: [ |
| 2847 | "libpthreadpool", |
| 2848 | "xnnpack_tables", |
| 2849 | ], |
| 2850 | } |
| 2851 | |
| 2852 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2853 | name: "xnnpack_fma3_ukernels", |
| 2854 | defaults: ["xnnpack_internal_default"], |
| 2855 | arch: { |
| 2856 | arm: { enabled: false, }, |
| 2857 | arm64: { enabled: false, }, |
| 2858 | x86: { |
| 2859 | srcs: FMA3_UKERNELS, |
| 2860 | cflags: [ |
| 2861 | "-mfma", |
| 2862 | ], |
| 2863 | }, |
| 2864 | x86_64: { |
| 2865 | srcs: FMA3_UKERNELS, |
| 2866 | cflags: [ |
| 2867 | "-mfma", |
| 2868 | ], |
| 2869 | }, |
| 2870 | }, |
| 2871 | header_libs: [ |
| 2872 | "fp16_headers", |
| 2873 | ], |
| 2874 | static_libs: [ |
| 2875 | "libpthreadpool", |
| 2876 | "xnnpack_tables", |
| 2877 | ], |
| 2878 | } |
| 2879 | |
| 2880 | cc_library_static { |
| 2881 | name: "xnnpack_avx2_ukernels", |
| 2882 | defaults: ["xnnpack_internal_default"], |
| 2883 | arch: { |
| 2884 | arm: { enabled: false, }, |
| 2885 | arm64: { enabled: false, }, |
| 2886 | x86: { |
| 2887 | srcs: AVX2_UKERNELS, |
| 2888 | cflags: [ |
| 2889 | "-mfma", |
| 2890 | "-mavx2", |
| 2891 | ], |
| 2892 | }, |
| 2893 | x86_64: { |
| 2894 | srcs: AVX2_UKERNELS, |
| 2895 | cflags: [ |
| 2896 | "-mfma", |
| 2897 | "-mavx2", |
| 2898 | ], |
| 2899 | }, |
| 2900 | }, |
| 2901 | header_libs: [ |
| 2902 | "fp16_headers", |
| 2903 | ], |
| 2904 | static_libs: [ |
| 2905 | "libpthreadpool", |
| 2906 | "xnnpack_tables", |
| 2907 | ], |
| 2908 | } |
| 2909 | |
| 2910 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2911 | name: "xnnpack_avx512skx_ukernels", |
| 2912 | defaults: ["xnnpack_internal_default"], |
| 2913 | arch: { |
| 2914 | arm: { enabled: false, }, |
| 2915 | arm64: { enabled: false, }, |
| 2916 | x86: { |
| 2917 | srcs: AVX512SKX_UKERNELS, |
| 2918 | cflags: [ |
| 2919 | "-mavx512f", |
| 2920 | "-mavx512cd", |
| 2921 | "-mavx512bw", |
| 2922 | "-mavx512dq", |
| 2923 | "-mavx512vl", |
| 2924 | ], |
| 2925 | }, |
| 2926 | x86_64: { |
| 2927 | srcs: AVX512SKX_UKERNELS, |
| 2928 | cflags: [ |
| 2929 | "-mavx512f", |
| 2930 | "-mavx512cd", |
| 2931 | "-mavx512bw", |
| 2932 | "-mavx512dq", |
| 2933 | "-mavx512vl", |
| 2934 | ], |
| 2935 | }, |
| 2936 | }, |
| 2937 | header_libs: [ |
| 2938 | "fp16_headers", |
| 2939 | ], |
| 2940 | static_libs: [ |
| 2941 | "libpthreadpool", |
| 2942 | "xnnpack_tables", |
| 2943 | ], |
| 2944 | } |
| 2945 | |
| 2946 | cc_library_static { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2947 | name: "xnnpack_avx512f_ukernels", |
| 2948 | defaults: ["xnnpack_internal_default"], |
| 2949 | arch: { |
| 2950 | arm: { enabled: false, }, |
| 2951 | arm64: { enabled: false, }, |
| 2952 | x86: { |
| 2953 | srcs: AVX512F_UKERNELS, |
| 2954 | cflags: [ |
| 2955 | "-mavx512f", |
| 2956 | ], |
| 2957 | }, |
| 2958 | x86_64: { |
| 2959 | srcs: AVX512F_UKERNELS, |
| 2960 | cflags: [ |
| 2961 | "-mavx512f", |
| 2962 | ], |
| 2963 | }, |
| 2964 | }, |
| 2965 | header_libs: [ |
| 2966 | "fp16_headers", |
| 2967 | ], |
| 2968 | static_libs: [ |
| 2969 | "libpthreadpool", |
| 2970 | "xnnpack_tables", |
| 2971 | ], |
| 2972 | } |
| 2973 | |
| 2974 | cc_library_static { |
| 2975 | name: "xnnpack_ukernels", |
| 2976 | defaults: ["xnnpack_internal_default"], |
| 2977 | arch: { |
| 2978 | arm: { |
| 2979 | whole_static_libs: [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2980 | "xnnpack_psimd_accmath_ukernels", |
| 2981 | "xnnpack_neon_ukernels", |
| 2982 | "xnnpack_neonfma_ukernels", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2983 | "xnnpack_neonv8_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2984 | "xnnpack_neondot_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2985 | "xnnpack_asm_ukernels", |
| 2986 | ], |
| 2987 | }, |
| 2988 | arm64: { |
| 2989 | whole_static_libs: [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2990 | "xnnpack_psimd_accmath_ukernels", |
| 2991 | "xnnpack_neon_ukernels", |
| 2992 | "xnnpack_neonfma_ukernels", |
Miao Wang | c0aa11a | 2020-06-10 13:41:26 -0700 | [diff] [blame] | 2993 | "xnnpack_neonv8_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2994 | "xnnpack_neonfp16arith_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 2995 | "xnnpack_neondot_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 2996 | "xnnpack_asm_ukernels", |
| 2997 | ], |
| 2998 | }, |
| 2999 | x86: { |
| 3000 | whole_static_libs: [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3001 | "xnnpack_psimd_accmath_ukernels", |
| 3002 | "xnnpack_sse2_ukernels", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 3003 | "xnnpack_ssse3_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3004 | "xnnpack_sse41_ukernels", |
| 3005 | "xnnpack_avx_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3006 | "xnnpack_xop_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3007 | "xnnpack_fma3_ukernels", |
| 3008 | "xnnpack_avx2_ukernels", |
| 3009 | "xnnpack_avx512f_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3010 | "xnnpack_avx512skx_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3011 | ], |
| 3012 | }, |
| 3013 | x86_64: { |
| 3014 | whole_static_libs: [ |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3015 | "xnnpack_psimd_accmath_ukernels", |
| 3016 | "xnnpack_sse2_ukernels", |
Miao Wang | 2534c2f | 2020-03-16 11:58:04 -0700 | [diff] [blame] | 3017 | "xnnpack_ssse3_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3018 | "xnnpack_sse41_ukernels", |
| 3019 | "xnnpack_avx_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3020 | "xnnpack_xop_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3021 | "xnnpack_fma3_ukernels", |
| 3022 | "xnnpack_avx2_ukernels", |
| 3023 | "xnnpack_avx512f_ukernels", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3024 | "xnnpack_avx512skx_ukernels", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3025 | ], |
| 3026 | }, |
| 3027 | }, |
| 3028 | whole_static_libs: [ |
| 3029 | "xnnpack_scalar_ukernels", |
| 3030 | "xnnpack_tables", |
| 3031 | ], |
| 3032 | } |
| 3033 | |
| 3034 | cc_library_static { |
| 3035 | name: "libXNNPACK", |
| 3036 | defaults: ["xnnpack_internal_default"], |
| 3037 | export_include_dirs: ["include"], |
| 3038 | srcs: [ |
| 3039 | "src/init.c", |
| 3040 | "src/runtime.c", |
| 3041 | "src/subgraph.c", |
| 3042 | "src/tensor.c", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3043 | ] + SUBGRAPH_SRCS, |
| 3044 | header_libs: [ |
| 3045 | "fp16_headers", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3046 | ], |
| 3047 | whole_static_libs: [ |
| 3048 | "libclog", |
| 3049 | "libcpuinfo", |
| 3050 | "libpthreadpool", |
| 3051 | "xnnpack_ukernels", |
| 3052 | "xnnpack_operator_run", |
| 3053 | "xnnpack_operators", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3054 | "xnnpack_logging_utils", |
| 3055 | "xnnpack_memory_planner", |
| 3056 | "xnnpack_packing", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3057 | ], |
| 3058 | } |
| 3059 | |
| 3060 | // Tests and benchmarks |
| 3061 | cc_defaults { |
| 3062 | name: "xnnpack_tests_default", |
| 3063 | vendor_available: true, |
| 3064 | stl: "libc++_static", |
| 3065 | local_include_dirs: [ |
| 3066 | "bench", |
| 3067 | "models", |
| 3068 | "test", |
| 3069 | "src", |
| 3070 | ], |
| 3071 | cflags: [ |
Miao Wang | 400e404 | 2020-04-17 10:15:59 -0700 | [diff] [blame] | 3072 | "-Wno-unused-function", |
| 3073 | "-Wno-unused-parameter", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3074 | "-Wno-unused-private-field", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3075 | ], |
| 3076 | header_libs: [ |
| 3077 | "fp16_headers", |
| 3078 | ], |
| 3079 | static_libs: [ |
| 3080 | "libXNNPACK", |
| 3081 | "libpthreadpool", |
| 3082 | "libgmock", |
| 3083 | ], |
| 3084 | shared_libs: [ |
| 3085 | "liblog", |
| 3086 | ], |
| 3087 | } |
| 3088 | |
| 3089 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3090 | name: "xnnpack_mobilenet_v1_fp32", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3091 | defaults: ["xnnpack_tests_default"], |
| 3092 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3093 | "models/fp32-mobilenet-v1.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3094 | ], |
| 3095 | } |
| 3096 | |
| 3097 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3098 | name: "xnnpack_qs8_mobilenet_v1", |
| 3099 | defaults: ["xnnpack_tests_default"], |
| 3100 | srcs: [ |
| 3101 | "models/qs8-mobilenet-v1.cc", |
| 3102 | ], |
| 3103 | } |
| 3104 | |
| 3105 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3106 | name: "xnnpack_mobilenet_v1_fp16", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3107 | defaults: ["xnnpack_tests_default"], |
| 3108 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3109 | "models/fp16-mobilenet-v1.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3110 | ], |
| 3111 | } |
| 3112 | |
| 3113 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3114 | name: "xnnpack_qs8_mobilenet_v2", |
| 3115 | defaults: ["xnnpack_tests_default"], |
| 3116 | srcs: [ |
| 3117 | "models/qs8-mobilenet-v2.cc", |
| 3118 | ], |
| 3119 | } |
| 3120 | |
| 3121 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3122 | name: "xnnpack_mobilenet_v2_fp32", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3123 | defaults: ["xnnpack_tests_default"], |
| 3124 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3125 | "models/fp32-mobilenet-v2.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3126 | ], |
| 3127 | } |
| 3128 | |
| 3129 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3130 | name: "xnnpack_mobilenet_v2_fp16", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3131 | defaults: ["xnnpack_tests_default"], |
| 3132 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3133 | "models/fp16-mobilenet-v2.cc", |
| 3134 | ], |
| 3135 | } |
| 3136 | |
| 3137 | |
| 3138 | cc_library_static { |
| 3139 | name: "xnnpack_mobilenet_v3_large_fp32", |
| 3140 | defaults: ["xnnpack_tests_default"], |
| 3141 | srcs: [ |
| 3142 | "models/fp32-mobilenet-v3-large.cc", |
| 3143 | ], |
| 3144 | } |
| 3145 | |
| 3146 | cc_library_static { |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3147 | name: "xnnpack_mobilenet_v3_large_fp16", |
| 3148 | defaults: ["xnnpack_tests_default"], |
| 3149 | srcs: [ |
| 3150 | "models/fp16-mobilenet-v3-large.cc", |
| 3151 | ], |
| 3152 | } |
| 3153 | |
| 3154 | cc_library_static { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3155 | name: "xnnpack_mobilenet_v3_small_fp32", |
| 3156 | defaults: ["xnnpack_tests_default"], |
| 3157 | srcs: [ |
| 3158 | "models/fp32-mobilenet-v3-small.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3159 | ], |
| 3160 | } |
| 3161 | |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3162 | cc_library_static { |
| 3163 | name: "xnnpack_mobilenet_v3_small_fp16", |
| 3164 | defaults: ["xnnpack_tests_default"], |
| 3165 | srcs: [ |
| 3166 | "models/fp16-mobilenet-v3-small.cc", |
| 3167 | ], |
| 3168 | } |
| 3169 | |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3170 | cc_benchmark { |
| 3171 | name: "xnnpack_end2end_bench", |
| 3172 | defaults: ["xnnpack_tests_default"], |
| 3173 | srcs: [ |
| 3174 | "bench/end2end.cc", |
| 3175 | "bench/utils.cc", |
| 3176 | ], |
| 3177 | cflags: [ |
| 3178 | "-Wno-unused-result" |
| 3179 | ], |
| 3180 | static_libs: [ |
| 3181 | "libcpuinfo", |
| 3182 | "libgoogle-benchmark", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3183 | "xnnpack_qs8_mobilenet_v1", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3184 | "xnnpack_mobilenet_v1_fp32", |
| 3185 | "xnnpack_mobilenet_v1_fp16", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3186 | "xnnpack_qs8_mobilenet_v2", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3187 | "xnnpack_mobilenet_v2_fp32", |
| 3188 | "xnnpack_mobilenet_v2_fp16", |
| 3189 | "xnnpack_mobilenet_v3_large_fp32", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3190 | "xnnpack_mobilenet_v3_large_fp16", |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3191 | "xnnpack_mobilenet_v3_small_fp32", |
Miao Wang | 5eea831 | 2020-12-07 09:12:40 -0800 | [diff] [blame^] | 3192 | "xnnpack_mobilenet_v3_small_fp16", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3193 | ], |
| 3194 | } |
| 3195 | |
| 3196 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3197 | name: "xnnpack_abs_nc_test", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3198 | defaults: ["xnnpack_tests_default"], |
| 3199 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3200 | "test/abs-nc.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3201 | ], |
| 3202 | test_suites: [ |
| 3203 | "general-tests", |
| 3204 | ], |
| 3205 | } |
| 3206 | |
| 3207 | cc_test { |
| 3208 | name: "xnnpack_add_nd_test", |
| 3209 | defaults: ["xnnpack_tests_default"], |
| 3210 | srcs: [ |
| 3211 | "test/add-nd.cc", |
| 3212 | ], |
| 3213 | test_suites: [ |
| 3214 | "general-tests", |
| 3215 | ], |
| 3216 | } |
| 3217 | |
| 3218 | cc_test { |
| 3219 | name: "xnnpack_argmax_pooling_nhwc_test", |
| 3220 | defaults: ["xnnpack_tests_default"], |
| 3221 | srcs: [ |
| 3222 | "test/argmax-pooling-nhwc.cc", |
| 3223 | ], |
| 3224 | test_suites: [ |
| 3225 | "general-tests", |
| 3226 | ], |
| 3227 | } |
| 3228 | |
| 3229 | cc_test { |
| 3230 | name: "xnnpack_average_pooling_nhwc_test", |
| 3231 | defaults: ["xnnpack_tests_default"], |
| 3232 | srcs: [ |
| 3233 | "test/average-pooling-nhwc.cc", |
| 3234 | ], |
| 3235 | test_suites: [ |
| 3236 | "general-tests", |
| 3237 | ], |
| 3238 | } |
| 3239 | |
| 3240 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3241 | name: "xnnpack_bankers_rounding_nc_test", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3242 | defaults: ["xnnpack_tests_default"], |
| 3243 | srcs: [ |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3244 | "test/bankers-rounding-nc.cc", |
| 3245 | ], |
| 3246 | test_suites: [ |
| 3247 | "general-tests", |
| 3248 | ], |
| 3249 | } |
| 3250 | |
| 3251 | cc_test { |
| 3252 | name: "xnnpack_ceiling_nc_test", |
| 3253 | defaults: ["xnnpack_tests_default"], |
| 3254 | srcs: [ |
| 3255 | "test/ceiling-nc.cc", |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3256 | ], |
| 3257 | test_suites: [ |
| 3258 | "general-tests", |
| 3259 | ], |
| 3260 | } |
| 3261 | |
| 3262 | cc_test { |
| 3263 | name: "xnnpack_channel_shuffle_nc_test", |
| 3264 | defaults: ["xnnpack_tests_default"], |
| 3265 | srcs: [ |
| 3266 | "test/channel-shuffle-nc.cc", |
| 3267 | ], |
| 3268 | test_suites: [ |
| 3269 | "general-tests", |
| 3270 | ], |
| 3271 | } |
| 3272 | |
| 3273 | cc_test { |
| 3274 | name: "xnnpack_clamp_nc_test", |
| 3275 | defaults: ["xnnpack_tests_default"], |
| 3276 | srcs: [ |
| 3277 | "test/clamp-nc.cc", |
| 3278 | ], |
| 3279 | test_suites: [ |
| 3280 | "general-tests", |
| 3281 | ], |
| 3282 | } |
| 3283 | |
| 3284 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3285 | name: "xnnpack_constant_pad_nd_test", |
| 3286 | defaults: ["xnnpack_tests_default"], |
| 3287 | srcs: [ |
| 3288 | "test/constant-pad-nd.cc", |
| 3289 | ], |
| 3290 | test_suites: [ |
| 3291 | "general-tests", |
| 3292 | ], |
| 3293 | } |
| 3294 | |
| 3295 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3296 | name: "xnnpack_convolution_nhwc_test", |
| 3297 | defaults: ["xnnpack_tests_default"], |
| 3298 | srcs: [ |
| 3299 | "test/convolution-nhwc.cc", |
| 3300 | ], |
| 3301 | test_suites: [ |
| 3302 | "general-tests", |
| 3303 | ], |
| 3304 | } |
| 3305 | |
| 3306 | cc_test { |
| 3307 | name: "xnnpack_convolution_nchw_test", |
| 3308 | defaults: ["xnnpack_tests_default"], |
| 3309 | srcs: [ |
| 3310 | "test/convolution-nchw.cc", |
| 3311 | ], |
| 3312 | test_suites: [ |
| 3313 | "general-tests", |
| 3314 | ], |
| 3315 | } |
| 3316 | |
| 3317 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3318 | name: "xnnpack_copy_nc_test", |
| 3319 | defaults: ["xnnpack_tests_default"], |
| 3320 | srcs: [ |
| 3321 | "test/copy-nc.cc", |
| 3322 | ], |
| 3323 | test_suites: [ |
| 3324 | "general-tests", |
| 3325 | ], |
| 3326 | } |
| 3327 | |
| 3328 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3329 | name: "xnnpack_deconvolution_nhwc_test", |
| 3330 | defaults: ["xnnpack_tests_default"], |
| 3331 | srcs: [ |
| 3332 | "test/deconvolution-nhwc.cc", |
| 3333 | ], |
| 3334 | test_suites: [ |
| 3335 | "general-tests", |
| 3336 | ], |
| 3337 | } |
| 3338 | |
| 3339 | cc_test { |
| 3340 | name: "xnnpack_divide_nd_test", |
| 3341 | defaults: ["xnnpack_tests_default"], |
| 3342 | srcs: [ |
| 3343 | "test/divide-nd.cc", |
| 3344 | ], |
| 3345 | test_suites: [ |
| 3346 | "general-tests", |
| 3347 | ], |
| 3348 | } |
| 3349 | |
| 3350 | cc_test { |
| 3351 | name: "xnnpack_fully_connected_nc_test", |
| 3352 | defaults: ["xnnpack_tests_default"], |
| 3353 | srcs: [ |
| 3354 | "test/fully-connected-nc.cc", |
| 3355 | ], |
| 3356 | test_suites: [ |
| 3357 | "general-tests", |
| 3358 | ], |
| 3359 | } |
| 3360 | |
| 3361 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3362 | name: "xnnpack_floor_nc_test", |
| 3363 | defaults: ["xnnpack_tests_default"], |
| 3364 | srcs: [ |
| 3365 | "test/floor-nc.cc", |
| 3366 | ], |
| 3367 | test_suites: [ |
| 3368 | "general-tests", |
| 3369 | ], |
| 3370 | } |
| 3371 | |
| 3372 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3373 | name: "xnnpack_global_average_pooling_nwc_test", |
| 3374 | defaults: ["xnnpack_tests_default"], |
| 3375 | srcs: [ |
| 3376 | "test/global-average-pooling-nwc.cc", |
| 3377 | ], |
| 3378 | test_suites: [ |
| 3379 | "general-tests", |
| 3380 | ], |
| 3381 | } |
| 3382 | |
| 3383 | cc_test { |
| 3384 | name: "xnnpack_global_average_pooling_ncw_test", |
| 3385 | defaults: ["xnnpack_tests_default"], |
| 3386 | srcs: [ |
| 3387 | "test/global-average-pooling-ncw.cc", |
| 3388 | ], |
| 3389 | test_suites: [ |
| 3390 | "general-tests", |
| 3391 | ], |
| 3392 | } |
| 3393 | |
| 3394 | cc_test { |
| 3395 | name: "xnnpack_hardswish_nc_test", |
| 3396 | defaults: ["xnnpack_tests_default"], |
| 3397 | srcs: [ |
| 3398 | "test/hardswish-nc.cc", |
| 3399 | ], |
| 3400 | test_suites: [ |
| 3401 | "general-tests", |
| 3402 | ], |
| 3403 | } |
| 3404 | |
| 3405 | cc_test { |
| 3406 | name: "xnnpack_leaky_relu_nc_test", |
| 3407 | defaults: ["xnnpack_tests_default"], |
| 3408 | srcs: [ |
| 3409 | "test/leaky-relu-nc.cc", |
| 3410 | ], |
| 3411 | test_suites: [ |
| 3412 | "general-tests", |
| 3413 | ], |
| 3414 | } |
| 3415 | |
| 3416 | cc_test { |
| 3417 | name: "xnnpack_max_pooling_nhwc_test", |
| 3418 | defaults: ["xnnpack_tests_default"], |
| 3419 | srcs: [ |
| 3420 | "test/max-pooling-nhwc.cc", |
| 3421 | ], |
| 3422 | test_suites: [ |
| 3423 | "general-tests", |
| 3424 | ], |
| 3425 | } |
| 3426 | |
| 3427 | cc_test { |
| 3428 | name: "xnnpack_maximum_nd_test", |
| 3429 | defaults: ["xnnpack_tests_default"], |
| 3430 | srcs: [ |
| 3431 | "test/maximum-nd.cc", |
| 3432 | ], |
| 3433 | test_suites: [ |
| 3434 | "general-tests", |
| 3435 | ], |
| 3436 | } |
| 3437 | |
| 3438 | cc_test { |
| 3439 | name: "xnnpack_minimum_nd_test", |
| 3440 | defaults: ["xnnpack_tests_default"], |
| 3441 | srcs: [ |
| 3442 | "test/minimum-nd.cc", |
| 3443 | ], |
| 3444 | test_suites: [ |
| 3445 | "general-tests", |
| 3446 | ], |
| 3447 | } |
| 3448 | |
| 3449 | cc_test { |
| 3450 | name: "xnnpack_multiply_nd_test", |
| 3451 | defaults: ["xnnpack_tests_default"], |
| 3452 | srcs: [ |
| 3453 | "test/multiply-nd.cc", |
| 3454 | ], |
| 3455 | test_suites: [ |
| 3456 | "general-tests", |
| 3457 | ], |
| 3458 | } |
| 3459 | |
| 3460 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3461 | name: "xnnpack_negate_nc_test", |
| 3462 | defaults: ["xnnpack_tests_default"], |
| 3463 | srcs: [ |
| 3464 | "test/negate-nc.cc", |
| 3465 | ], |
| 3466 | test_suites: [ |
| 3467 | "general-tests", |
| 3468 | ], |
| 3469 | } |
| 3470 | |
| 3471 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3472 | name: "xnnpack_prelu_nc_test", |
| 3473 | defaults: ["xnnpack_tests_default"], |
| 3474 | srcs: [ |
| 3475 | "test/prelu-nc.cc", |
| 3476 | ], |
| 3477 | test_suites: [ |
| 3478 | "general-tests", |
| 3479 | ], |
| 3480 | } |
| 3481 | |
| 3482 | cc_test { |
| 3483 | name: "xnnpack_resize_bilinear_nhwc_test", |
| 3484 | defaults: ["xnnpack_tests_default"], |
| 3485 | srcs: [ |
| 3486 | "test/resize-bilinear-nhwc.cc", |
| 3487 | ], |
| 3488 | test_suites: [ |
| 3489 | "general-tests", |
| 3490 | ], |
| 3491 | } |
| 3492 | |
| 3493 | cc_test { |
| 3494 | name: "xnnpack_sigmoid_nc_test", |
| 3495 | defaults: ["xnnpack_tests_default"], |
| 3496 | srcs: [ |
| 3497 | "test/sigmoid-nc.cc", |
| 3498 | ], |
| 3499 | test_suites: [ |
| 3500 | "general-tests", |
| 3501 | ], |
| 3502 | } |
| 3503 | |
| 3504 | cc_test { |
| 3505 | name: "xnnpack_softmax_nc_test", |
| 3506 | defaults: ["xnnpack_tests_default"], |
| 3507 | srcs: [ |
| 3508 | "test/softmax-nc.cc", |
| 3509 | ], |
| 3510 | test_suites: [ |
| 3511 | "general-tests", |
| 3512 | ], |
| 3513 | } |
| 3514 | |
| 3515 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3516 | name: "xnnpack_square_nc_test", |
| 3517 | defaults: ["xnnpack_tests_default"], |
| 3518 | srcs: [ |
| 3519 | "test/square-nc.cc", |
| 3520 | ], |
| 3521 | test_suites: [ |
| 3522 | "general-tests", |
| 3523 | ], |
| 3524 | } |
| 3525 | |
| 3526 | cc_test { |
| 3527 | name: "xnnpack_square_root_nc_test", |
| 3528 | defaults: ["xnnpack_tests_default"], |
| 3529 | srcs: [ |
| 3530 | "test/square-root-nc.cc", |
| 3531 | ], |
| 3532 | test_suites: [ |
| 3533 | "general-tests", |
| 3534 | ], |
| 3535 | } |
| 3536 | |
| 3537 | cc_test { |
| 3538 | name: "xnnpack_square_difference_nd_test", |
| 3539 | defaults: ["xnnpack_tests_default"], |
| 3540 | srcs: [ |
| 3541 | "test/squared-difference-nd.cc", |
| 3542 | ], |
| 3543 | test_suites: [ |
| 3544 | "general-tests", |
| 3545 | ], |
| 3546 | } |
| 3547 | |
| 3548 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3549 | name: "xnnpack_subtract_nd_test", |
| 3550 | defaults: ["xnnpack_tests_default"], |
| 3551 | srcs: [ |
| 3552 | "test/subtract-nd.cc", |
| 3553 | ], |
| 3554 | test_suites: [ |
| 3555 | "general-tests", |
| 3556 | ], |
| 3557 | } |
| 3558 | |
| 3559 | cc_test { |
Miao Wang | 86f5fbe | 2020-07-24 11:16:10 -0700 | [diff] [blame] | 3560 | name: "xnnpack_truncation_nc_test", |
| 3561 | defaults: ["xnnpack_tests_default"], |
| 3562 | srcs: [ |
| 3563 | "test/truncation-nc.cc", |
| 3564 | ], |
| 3565 | test_suites: [ |
| 3566 | "general-tests", |
| 3567 | ], |
| 3568 | } |
| 3569 | |
| 3570 | cc_test { |
Miao Wang | e999347 | 2020-02-10 15:00:10 -0800 | [diff] [blame] | 3571 | name: "xnnpack_unpooling_nhwc_test", |
| 3572 | defaults: ["xnnpack_tests_default"], |
| 3573 | srcs: [ |
| 3574 | "test/unpooling-nhwc.cc", |
| 3575 | ], |
| 3576 | test_suites: [ |
| 3577 | "general-tests", |
| 3578 | ], |
| 3579 | } |