File size: 7,039 Bytes
be903e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 | // Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2020 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.
#include "deepcopy_vulkan.h"
#include "layer_shader_type.h"
namespace ncnn {
DeepCopy_vulkan::DeepCopy_vulkan()
{
support_vulkan = true;
support_image_storage = true;
pipeline_deepcopy = 0;
pipeline_deepcopy_pack4 = 0;
pipeline_deepcopy_pack8 = 0;
}
int DeepCopy_vulkan::create_pipeline(const Option& opt)
{
const Mat& shape = bottom_shapes.empty() ? Mat() : bottom_shapes[0];
const Mat& out_shape = top_shapes.empty() ? Mat() : top_shapes[0];
int elempack = 1;
if (shape.dims == 1) elempack = opt.use_shader_pack8 && shape.w % 8 == 0 ? 8 : shape.w % 4 == 0 ? 4 : 1;
if (shape.dims == 2) elempack = opt.use_shader_pack8 && shape.h % 8 == 0 ? 8 : shape.h % 4 == 0 ? 4 : 1;
if (shape.dims == 3) elempack = opt.use_shader_pack8 && shape.c % 8 == 0 ? 8 : shape.c % 4 == 0 ? 4 : 1;
int out_elempack = 1;
if (out_shape.dims == 1) out_elempack = opt.use_shader_pack8 && out_shape.w % 8 == 0 ? 8 : out_shape.w % 4 == 0 ? 4 : 1;
if (out_shape.dims == 2) out_elempack = opt.use_shader_pack8 && out_shape.h % 8 == 0 ? 8 : out_shape.h % 4 == 0 ? 4 : 1;
if (out_shape.dims == 3) out_elempack = opt.use_shader_pack8 && out_shape.c % 8 == 0 ? 8 : out_shape.c % 4 == 0 ? 4 : 1;
size_t elemsize;
size_t out_elemsize;
if (opt.use_fp16_storage)
{
elemsize = elempack * 2u;
out_elemsize = out_elempack * 2u;
}
else if (opt.use_fp16_packed)
{
elemsize = elempack == 1 ? 4u : elempack * 2u;
out_elemsize = out_elempack == 1 ? 4u : out_elempack * 2u;
}
else
{
elemsize = elempack * 4u;
out_elemsize = out_elempack * 4u;
}
Mat shape_packed;
if (shape.dims == 1) shape_packed = Mat(shape.w / elempack, (void*)0, elemsize, elempack);
if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
Mat out_shape_packed;
if (out_shape.dims == 1) out_shape_packed = Mat(out_shape.w / out_elempack, (void*)0, out_elemsize, out_elempack);
if (out_shape.dims == 2) out_shape_packed = Mat(out_shape.w, out_shape.h / out_elempack, (void*)0, out_elemsize, out_elempack);
if (out_shape.dims == 3) out_shape_packed = Mat(out_shape.w, out_shape.h, out_shape.c / out_elempack, (void*)0, out_elemsize, out_elempack);
std::vector<vk_specialization_type> specializations(0 + 5);
specializations[0 + 0].i = shape_packed.dims;
specializations[0 + 1].i = shape_packed.w;
specializations[0 + 2].i = shape_packed.h;
specializations[0 + 3].i = shape_packed.c;
specializations[0 + 4].i = shape_packed.cstep;
Mat local_size_xyz;
if (out_shape_packed.dims == 1)
{
local_size_xyz.w = std::min(64, out_shape_packed.w);
local_size_xyz.h = 1;
local_size_xyz.c = 1;
}
if (out_shape_packed.dims == 2)
{
local_size_xyz.w = std::min(8, out_shape_packed.w);
local_size_xyz.h = std::min(8, out_shape_packed.h);
local_size_xyz.c = 1;
}
if (out_shape_packed.dims == 3)
{
local_size_xyz.w = std::min(4, out_shape_packed.w);
local_size_xyz.h = std::min(4, out_shape_packed.h);
local_size_xyz.c = std::min(4, out_shape_packed.c);
}
// pack1
if (shape.dims == 0 || elempack == 1)
{
pipeline_deepcopy = new Pipeline(vkdev);
pipeline_deepcopy->set_optimal_local_size_xyz(local_size_xyz);
pipeline_deepcopy->create(LayerShaderType::deepcopy, opt, specializations);
}
// pack4
if (shape.dims == 0 || elempack == 4)
{
pipeline_deepcopy_pack4 = new Pipeline(vkdev);
pipeline_deepcopy_pack4->set_optimal_local_size_xyz(local_size_xyz);
pipeline_deepcopy_pack4->create(LayerShaderType::deepcopy_pack4, opt, specializations);
}
// pack8
if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
{
pipeline_deepcopy_pack8 = new Pipeline(vkdev);
pipeline_deepcopy_pack8->set_optimal_local_size_xyz(local_size_xyz);
pipeline_deepcopy_pack8->create(LayerShaderType::deepcopy_pack8, opt, specializations);
}
return 0;
}
int DeepCopy_vulkan::destroy_pipeline(const Option& /*opt*/)
{
delete pipeline_deepcopy;
pipeline_deepcopy = 0;
delete pipeline_deepcopy_pack4;
pipeline_deepcopy_pack4 = 0;
delete pipeline_deepcopy_pack8;
pipeline_deepcopy_pack8 = 0;
return 0;
}
int DeepCopy_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute& cmd, const Option& opt) const
{
int elempack = bottom_blob.elempack;
top_blob.create_like(bottom_blob, opt.blob_vkallocator);
if (top_blob.empty())
return -100;
std::vector<VkMat> bindings(2);
bindings[0] = bottom_blob;
bindings[1] = top_blob;
std::vector<vk_constant_type> constants(5);
constants[0].i = bottom_blob.dims;
constants[1].i = bottom_blob.w;
constants[2].i = bottom_blob.h;
constants[3].i = bottom_blob.c;
constants[4].i = bottom_blob.cstep;
const Pipeline* pipeline = elempack == 8 ? pipeline_deepcopy_pack8
: elempack == 4 ? pipeline_deepcopy_pack4
: pipeline_deepcopy;
cmd.record_pipeline(pipeline, bindings, constants, top_blob);
return 0;
}
int DeepCopy_vulkan::forward(const VkImageMat& bottom_blob, VkImageMat& top_blob, VkCompute& cmd, const Option& opt) const
{
int elempack = bottom_blob.elempack;
top_blob.create_like(bottom_blob, opt.blob_vkallocator);
if (top_blob.empty())
return -100;
std::vector<VkImageMat> bindings(2);
bindings[0] = bottom_blob;
bindings[1] = top_blob;
std::vector<vk_constant_type> constants(5);
constants[0].i = bottom_blob.dims;
constants[1].i = bottom_blob.w;
constants[2].i = bottom_blob.h;
constants[3].i = bottom_blob.c;
constants[4].i = 0; //bottom_blob.cstep;
const Pipeline* pipeline = elempack == 8 ? pipeline_deepcopy_pack8
: elempack == 4 ? pipeline_deepcopy_pack4
: pipeline_deepcopy;
cmd.record_pipeline(pipeline, bindings, constants, top_blob);
return 0;
}
} // namespace ncnn
|