Skip to content

GitLab

Switch branch/tag
Find file Normal viewHistoryPermalink
threadwise_tensor_op.cuh 4.8 KB
Newer Older
Chao Liu's avatar
add 2nd version of blockwise_tensor_op  
Chao Liu committed
1
2
3 
#pragma once
#include "constant_tensor_descriptor.cuh"
Chao Liu's avatar
faster  
Chao Liu committed
4 
#define THREADWISE_TENSOR_OP_METHOD 0
Chao Liu's avatar
improved blockwise_tensor_op  
Chao Liu committed
5
6 

#if THREADWISE_TENSOR_OP_METHOD == 0
Chao Liu's avatar
add 2nd version of blockwise_tensor_op  
Chao Liu committed
7 
template <class TFloat, class SrcDesc, class DstDesc, class F>
Chao Liu's avatar
faster  
Chao Liu committed
8
9
10
11
12
13 
__device__ void threadwise_4d_tensor_op(SrcDesc,
                                        TFloat* const __restrict__ p_src,
                                        DstDesc,
                                        TFloat* __restrict__ p_dst,
                                        F f,
                                        bool flag = false)
Chao Liu's avatar
add 2nd version of blockwise_tensor_op  
Chao Liu committed
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32 
{
    constexpr auto I0 = Index<0>{};
    constexpr auto I1 = Index<1>{};
    constexpr auto I2 = Index<2>{};
    constexpr auto I3 = Index<3>{};

    constexpr auto src_desc = SrcDesc{};
    constexpr auto dst_desc = DstDesc{};

    static_assert(is_same<decltype(src_desc.GetLengths()), decltype(dst_desc.GetLengths())>::value);

#if 0
    if(threadIdx.x == 0)
    {
        print_ConstantTensorDescriptor(src_desc);
        print_ConstantTensorDescriptor(dst_desc);
    }
#endif
Chao Liu's avatar
faster  
Chao Liu committed
33
34
35
36
37 
#if 1
    if(flag && threadIdx.x != 0)
        return;
#endif
Chao Liu's avatar
add 2nd version of blockwise_tensor_op  
Chao Liu committed
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 
    for(unsigned did0 = 0; did0 < src_desc.GetLength(I0); ++did0)
    {
        for(unsigned did1 = 0; did1 < src_desc.GetLength(I1); ++did1)
        {
            for(unsigned did2 = 0; did2 < src_desc.GetLength(I2); ++did2)
            {
                for(unsigned did3 = 0; did3 < src_desc.GetLength(I3); ++did3)
                {
                    const unsigned sindex =
                        src_desc.GetStride(I0) * did0 + src_desc.GetStride(I1) * did1 +
                        src_desc.GetStride(I2) * did2 + src_desc.GetStride(I3) * did3;

                    const unsigned dindex =
                        dst_desc.GetStride(I0) * did0 + dst_desc.GetStride(I1) * did1 +
                        dst_desc.GetStride(I2) * did2 + dst_desc.GetStride(I3) * did3;

                    f(p_src[sindex], p_dst[dindex]);

#if 0
                    if(threadIdx.x == 0)
                    {
                        printf("threadwise_4d_tensor_op: 1: thread id %u, \t"
                               "sindex %u, p_src[sindex] %f, \t"
                               "dindex %u, p_dst[dindex] %f\n",
                               threadIdx.x,
                               sindex,
                               p_src[sindex],
                               dindex,
                               p_dst[dindex]);
                    }
#endif
                }
            }
        }
    }
}
Chao Liu's avatar
improved blockwise_tensor_op  
Chao Liu committed
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 
#endif

#if THREADWISE_TENSOR_OP_METHOD == 1
template <class TFloat, class SrcDesc, class DstDesc, class F>
__device__ void threadwise_4d_tensor_op(
    SrcDesc, TFloat* const __restrict__ p_src, DstDesc, TFloat* __restrict__ p_dst, F f)
{
    constexpr auto I0 = Index<0>{};
    constexpr auto I1 = Index<1>{};
    constexpr auto I2 = Index<2>{};
    constexpr auto I3 = Index<3>{};

    constexpr auto src_desc = SrcDesc{};
    constexpr auto dst_desc = DstDesc{};

    static_assert(is_same<decltype(src_desc.GetLengths()), decltype(dst_desc.GetLengths())>::value);

#if 0
    if(threadIdx.x == 0)
    {
        print_ConstantTensorDescriptor(src_desc);
        print_ConstantTensorDescriptor(dst_desc);
    }
#endif

    unsigned sindex = 0;
    unsigned dindex = 0;

    for(unsigned did0 = 0; did0 < src_desc.GetLength(I0); ++did0)
    {
        for(unsigned did1 = 0; did1 < src_desc.GetLength(I1); ++did1)
        {
            for(unsigned did2 = 0; did2 < src_desc.GetLength(I2); ++did2)
            {
                for(unsigned did3 = 0; did3 < src_desc.GetLength(I3); ++did3)
                {
                    f(p_src[sindex], p_dst[dindex]);

#if 0
                    if(threadIdx.x == 0)
                    {
                        printf("threadwise_4d_tensor_op: 1: thread id %u, \t"
                               "sindex %u, p_src[sindex] %f, \t"
                               "dindex %u, p_dst[dindex] %f\n",
                               threadIdx.x,
                               sindex,
                               p_src[sindex],
                               dindex,
                               p_dst[dindex]);
                    }
#endif
                    sindex += src_desc.GetStride(I3);
                    dindex += dst_desc.GetStride(I3);
                }

                sindex += src_desc.GetStride(I2) - src_desc.GetLength(I3) * src_desc.GetStride(I3);
                dindex += dst_desc.GetStride(I2) - dst_desc.GetLength(I3) * dst_desc.GetStride(I3);
            }

            sindex += src_desc.GetStride(I1) - src_desc.GetLength(I2) * src_desc.GetStride(I2);
            dindex += dst_desc.GetStride(I1) - dst_desc.GetLength(I2) * dst_desc.GetStride(I2);
        }

        sindex += src_desc.GetStride(I0) - src_desc.GetLength(I1) * src_desc.GetStride(I1);
        dindex += dst_desc.GetStride(I0) - dst_desc.GetLength(I1) * dst_desc.GetStride(I1);
    }
}
#endif