Skip to content

GitLab

"example/01_gemm/gemm_xdl_fp8.cpp" did not exist on "7e9a9d32c7a9259a1bd57b0b461c36d089d26fe8"
Switch branch/tag
Find file Normal viewHistoryPermalink
host_gemm.hpp 4.55 KB
Newer Older
Chao Liu's avatar
GEMM driver and kernel (#29)  
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 
#pragma once
#include "host_tensor.hpp"
#include "gemm_common.hpp"

template <typename AType, typename BType, typename CType>
void host_gemm(const Tensor<AType>& a,
               const Tensor<BType>& b,
               Tensor<CType>& c,
               const GemmMatrixLayout layout)
{
    if(layout == GemmMatrixLayout::MK_KN_MN)
    {
        auto f_mk_kn_mn = [&](auto m, auto n) {
            const int K = a.mDesc.GetLengths()[1];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
            }

            c(m, n) = v;
        };

        make_ParallelTensorFunctor(f_mk_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::MK_NK_MN)
    {
        auto f_mk_nk_mn = [&](auto m, auto n) {
            const int K = a.mDesc.GetLengths()[1];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
            }

            c(m, n) = v;
        };

        make_ParallelTensorFunctor(f_mk_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::KM_KN_MN)
    {
        auto f_km_kn_mn = [&](auto m, auto n) {
            const int K = a.mDesc.GetLengths()[0];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
            }

            c(m, n) = v;
        };

        make_ParallelTensorFunctor(f_km_kn_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::KM_NK_MN)
    {
        auto f_km_nk_mn = [&](auto m, auto n) {
            const int K = a.mDesc.GetLengths()[0];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
            }

            c(m, n) = v;
        };

        make_ParallelTensorFunctor(f_km_nk_mn, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
Chao Liu's avatar
Tweak GEMM kernel (#38)  
Chao Liu committed
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 
    else if(layout == GemmMatrixLayout::MK_KN_NM)
    {
        auto f_mk_kn_nm = [&](auto n, auto m) {
            const int K = a.mDesc.GetLengths()[1];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(k, n));
            }

            c(n, m) = v;
        };

        make_ParallelTensorFunctor(f_mk_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::MK_NK_NM)
    {
        auto f_mk_nk_nm = [&](auto n, auto m) {
            const int K = a.mDesc.GetLengths()[1];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(m, k)) * static_cast<const double>(b(n, k));
            }

            c(n, m) = v;
        };

        make_ParallelTensorFunctor(f_mk_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::KM_KN_NM)
    {
        auto f_km_kn_nm = [&](auto n, auto m) {
            const int K = a.mDesc.GetLengths()[0];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(k, n));
            }

            c(n, m) = v;
        };

        make_ParallelTensorFunctor(f_km_kn_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
    else if(layout == GemmMatrixLayout::KM_NK_NM)
    {
        auto f_km_nk_nm = [&](auto n, auto m) {
            const int K = a.mDesc.GetLengths()[0];

            double v = 0;

            for(int k = 0; k < K; ++k)
            {
                v += static_cast<const double>(a(k, m)) * static_cast<const double>(b(n, k));
            }

            c(n, m) = v;
        };

        make_ParallelTensorFunctor(f_km_nk_nm, c.mDesc.GetLengths()[0], c.mDesc.GetLengths()[1])(
            std::thread::hardware_concurrency());
    }
Chao Liu's avatar
GEMM driver and kernel (#29)  
Chao Liu committed
155
156
157
158
159 
    else
    {
        throw std::runtime_error("wrong! not supported layout");
    }
}