Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
device_direct_convolution_1.cuh 3.42 KB
Newer Older
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
1 
#pragma once
Chao Liu's avatar
rename  
Chao Liu committed
2 
#include "gridwise_direct_convolution_1.cuh"
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
3
4 

template <class T, class InDesc, class WeiDesc, class OutDesc>
Chao Liu's avatar
rename  
Chao Liu committed
5 
void device_direct_convolution_1(
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
6
7
8
9
10
11
12
13
14
15
16
17
18 
    InDesc, const Tensor<T>& in, WeiDesc, const Tensor<T>& wei, OutDesc, Tensor<T>& out)
{
    std::size_t data_sz = sizeof(T);
    DeviceMem in_device_buf(data_sz * in.mDesc.GetElementSpace());
    DeviceMem wei_device_buf(data_sz * wei.mDesc.GetElementSpace());
    DeviceMem out_device_buf(data_sz * out.mDesc.GetElementSpace());

    int num_thread = std::thread::hardware_concurrency();

    in_device_buf.ToDevice(in.mData.data());
    wei_device_buf.ToDevice(wei.mData.data());
    out_device_buf.ToDevice(out.mData.data());
Chao Liu's avatar
refactor direct  
Chao Liu committed
19
20
21
22 
    constexpr auto I0 = Number<0>{};
    constexpr auto I1 = Number<1>{};
    constexpr auto I2 = Number<2>{};
    constexpr auto I3 = Number<3>{};
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
23
24
25
26
27
28 

    constexpr auto in_desc          = InDesc{};
    constexpr auto wei_desc         = WeiDesc{};
    constexpr auto out_desc         = OutDesc{};
    constexpr unsigned OutTileSizeH = 2;
    constexpr unsigned OutTileSizeW = 2;
Chao Liu's avatar
rename  
Chao Liu committed
29
30 
    constexpr unsigned NPerBlock    = 1;
    constexpr unsigned KPerBlock    = 4;
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
31 
    constexpr unsigned CPerBlock    = 2;
Chao Liu's avatar
rename  
Chao Liu committed
32 
    constexpr unsigned YPerBlock    = 8;
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
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 
    constexpr unsigned XPerBlock    = 16;

    constexpr unsigned NBlockOpLen0 = 1;
    constexpr unsigned NBlockOpLen1 = 1;
    constexpr unsigned NBlockOpLen2 = 4;
    constexpr unsigned NBlockOpLen3 = 32;

    constexpr unsigned BlockSize = 128;

    constexpr unsigned GridSize = (out_desc.GetLength(I0) / NPerBlock) *
                                  (out_desc.GetLength(I1) / KPerBlock) *
                                  (out_desc.GetLength(I2) / (OutTileSizeH * YPerBlock)) *
                                  (out_desc.GetLength(I3) / (OutTileSizeW * XPerBlock));

    dim3 block_dim(BlockSize);
    dim3 grid_dim(GridSize);

    printf("%s: BlockSize %u, GridSize %u \n", __func__, BlockSize, GridSize);

    cudaEvent_t start, stop;
    float elapsedTime;

    cudaEventCreate(&start);
    cudaEventRecord(start, 0);
Chao Liu's avatar
refactor  
Chao Liu committed
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74 
    gridwise_direct_convolution_1<T,
                                  InDesc,
                                  WeiDesc,
                                  OutDesc,
                                  OutTileSizeH,
                                  OutTileSizeW,
                                  NPerBlock,
                                  KPerBlock,
                                  CPerBlock,
                                  YPerBlock,
                                  XPerBlock,
                                  NBlockOpLen0,
                                  NBlockOpLen1,
                                  NBlockOpLen2,
                                  NBlockOpLen3,
                                  BlockSize,
                                  GridSize>
Chao Liu's avatar
faster: output skip LDS  
Chao Liu committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91 
        <<<grid_dim, block_dim>>>(InDesc{},
                                  static_cast<T*>(in_device_buf.GetDeviceBuffer()),
                                  WeiDesc{},
                                  static_cast<T*>(wei_device_buf.GetDeviceBuffer()),
                                  OutDesc{},
                                  static_cast<T*>(out_device_buf.GetDeviceBuffer()));

    cudaEventCreate(&stop);
    cudaEventRecord(stop, 0);
    cudaEventSynchronize(stop);

    cudaEventElapsedTime(&elapsedTime, start, stop);
    printf("Elapsed time : %f ms\n", elapsedTime);

    checkCudaErrors(cudaGetLastError());
    out_device_buf.FromDevice(out.mData.data());
}