Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
device_implicit_gemm_convolution.cuh 4.29 KB
Newer Older
Chao Liu's avatar
adding implicit gemm  
Chao Liu committed
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 
#pragma once
#include "gridwise_implicit_gemm_convolution.cuh"

template <class T, class InDesc, class WeiDesc, class OutDesc>
void device_implicit_gemm_convolution(
    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());

    constexpr auto I0 = Number<0>{};
    constexpr auto I1 = Number<1>{};
    constexpr auto I2 = Number<2>{};
    constexpr auto I3 = Number<3>{};

    constexpr auto in_desc  = InDesc{};
    constexpr auto wei_desc = WeiDesc{};
    constexpr auto out_desc = OutDesc{};

#if 1
    constexpr unsigned OutTileSizeH = 2;
    constexpr unsigned OutTileSizeW = 2;
    constexpr unsigned NPerBlock    = 2;
    constexpr unsigned KPerBlock    = 32;
    constexpr unsigned CPerBlock    = 4;
    constexpr unsigned YPerBlock    = 1;
    constexpr unsigned XPerBlock    = 16;

    constexpr unsigned NPerThread = 2;
    constexpr unsigned KPerThread = 4;
    constexpr unsigned CPerThread = 2;

    constexpr unsigned BlockSize = 128;
#elif 0
    constexpr unsigned OutTileSizeH = 2;
    constexpr unsigned OutTileSizeW = 2;
    constexpr unsigned NPerBlock    = 2;
    constexpr unsigned KPerBlock    = 32;
    constexpr unsigned CPerBlock    = 4;
    constexpr unsigned YPerBlock    = 1;
    constexpr unsigned XPerBlock    = 27;

    constexpr unsigned NPerThread = 2;
    constexpr unsigned KPerThread = 4;
    constexpr unsigned CPerThread = 2;

    constexpr unsigned BlockSize = 216;
#elif 0
    constexpr unsigned OutTileSizeH = 2;
    constexpr unsigned OutTileSizeW = 2;
    constexpr unsigned NPerBlock    = 2;
    constexpr unsigned KPerBlock    = 32;
    constexpr unsigned CPerBlock    = 4;
    constexpr unsigned YPerBlock    = 1;
    constexpr unsigned XPerBlock    = 32;

    constexpr unsigned NPerThread = 2;
    constexpr unsigned KPerThread = 4;
    constexpr unsigned CPerThread = 2;

    constexpr unsigned BlockSize = 256;
#endif

    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);

    gridwise_implicit_gemm_convolution<T,
                                       InDesc,
                                       WeiDesc,
                                       OutDesc,
                                       OutTileSizeH,
                                       OutTileSizeW,
                                       NPerBlock,
                                       KPerBlock,
                                       CPerBlock,
                                       YPerBlock,
                                       XPerBlock,
                                       NPerThread,
                                       KPerThread,
                                       CPerThread,
                                       BlockSize,
                                       GridSize>
        <<<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());
}