profile_gemm_add_add_fastgelu.cpp 6.14 KB
Newer Older
Chao Liu's avatar
Chao Liu committed
1
2
3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.

4
5
6
7
8
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>

Chao Liu's avatar
Chao Liu committed
9
#include "profiler/include/profile_gemm_add_add_fastgelu_impl.hpp"
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

int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{
    enum struct MatrixLayout
    {
        MK_KN_MN_MN_MN, // 0
        MK_NK_MN_MN_MN, // 1
        KM_KN_MN_MN_MN, // 2
        KM_NK_MN_MN_MN, // 3
        MK_KN_NM_MN_MN, // 4
        MK_NK_NM_MN_MN, // 5
        KM_KN_NM_MN_MN, // 6
        KM_NK_NM_MN_MN, // 7
    };

    enum struct MatrixDataType
    {
        F32_F32_F32_F32_F32,      // 0
        F16_F16_F16_F16_F16,      // 1
        BF16_BF16_BF16_BF16_BF16, // 2
        INT8_INT8_INT8_INT8_INT8, // 3
    };

    if(argc != 16)
    {
        // clang-format off
        printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
        printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
        printf("                     1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
        printf("                     2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
        printf("                     3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
        printf("arg4: verification (0: no; 1: yes)\n");
        printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
        printf("arg6: print tensor value (0: no; 1: yes)\n");
        printf("arg7: time kernel (0=no, 1=yes)\n");
        printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
        // clang-format on
        exit(1);
    }

    const auto data_type       = static_cast<MatrixDataType>(std::stoi(argv[2]));
    const auto layout          = static_cast<MatrixLayout>(std::stoi(argv[3]));
    const bool do_verification = std::stoi(argv[4]);
    const int init_method      = std::stoi(argv[5]);
    const bool do_log          = std::stoi(argv[6]);
    const bool time_kernel     = std::stoi(argv[7]);

    const int M = std::stoi(argv[8]);
    const int N = std::stoi(argv[9]);
    const int K = std::stoi(argv[10]);

    const int StrideA  = std::stoi(argv[11]);
    const int StrideB  = std::stoi(argv[12]);
    const int StrideD0 = std::stoi(argv[13]);
    const int StrideD1 = std::stoi(argv[14]);
    const int StrideE  = std::stoi(argv[15]);

    using F16 = ck::half_t;
    using F32 = float;

    using Row = ck::tensor_layout::gemm::RowMajor;
    using Col = ck::tensor_layout::gemm::ColumnMajor;

    auto profile = [&](auto a_type,
                       auto b_type,
                       auto acc_type,
                       auto d0_type,
                       auto d1_type,
                       auto e_type,
                       auto a_layout,
                       auto b_layout,
                       auto d0_layout,
                       auto d1_layout,
                       auto e_layout) {
        using ADataType   = decltype(a_type);
        using BDataType   = decltype(b_type);
        using AccDataType = decltype(acc_type);
        using D0DataType  = decltype(d0_type);
        using D1DataType  = decltype(d1_type);
        using EDataType   = decltype(e_type);

        using ALayout  = decltype(a_layout);
        using BLayout  = decltype(b_layout);
        using D0Layout = decltype(d0_layout);
        using D1Layout = decltype(d1_layout);
        using ELayout  = decltype(e_layout);

        const int DefaultStrideA  = ck::is_same_v<ALayout, Row> ? K : M;
        const int DefaultStrideB  = ck::is_same_v<BLayout, Row> ? N : K;
        const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
        const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
        const int DefaultStrideE  = ck::is_same_v<ELayout, Row> ? N : M;

        return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
                                                                BDataType,
                                                                AccDataType,
                                                                D0DataType,
                                                                D1DataType,
                                                                EDataType,
                                                                ALayout,
                                                                BLayout,
                                                                D0Layout,
                                                                D1Layout,
                                                                ELayout>(
            do_verification,
            init_method,
            do_log,
            time_kernel,
            M,
            N,
            K,
            (StrideA < 0) ? DefaultStrideA : StrideA,
            (StrideB < 0) ? DefaultStrideB : StrideB,
            (StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
            (StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
            (StrideE < 0) ? DefaultStrideE : StrideE);
    };

    if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
    {
        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
    }
    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
            layout == MatrixLayout::MK_NK_MN_MN_MN)
    {
        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
    }
    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
            layout == MatrixLayout::KM_KN_MN_MN_MN)
    {
        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
    }
    else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
            layout == MatrixLayout::KM_NK_MN_MN_MN)
    {
        return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
    }
    else
    {
        std::cout << "this data_type & layout is not implemented" << std::endl;

        return 0;
    }
}