gemm_split_k.cpp 8.72 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.

ltqin's avatar
ltqin committed
4
5
6
#include <iostream>
#include <initializer_list>
#include <cstdlib>
Chao Liu's avatar
Chao Liu committed
7
8
9
10
11
12

#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"

13
14
#include "ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp"

Chao Liu's avatar
Chao Liu committed
15
#include "ck/library/utility/check_err.hpp"
16
17
18
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
19
#include "ck/library/utility/literals.hpp"
Chao Liu's avatar
Chao Liu committed
20
21
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"

22
#include "ck/library/utility/host_gemm.hpp"
ltqin's avatar
ltqin committed
23

Chao Liu's avatar
Chao Liu committed
24
enum struct GemmMatrixLayout
ltqin's avatar
ltqin committed
25
26
27
28
29
30
31
32
33
34
35
36
{
    MK_KN_MN, // 0
    MK_NK_MN, // 1
    KM_KN_MN, // 2
    KM_NK_MN, // 3
};

template <typename T>
static bool check_out(const Tensor<T>& ref, const Tensor<T>& result)
{
    float max_diff = 1e-6;

37
    for(std::size_t i = 0; i < ref.mData.size(); ++i)
ltqin's avatar
ltqin committed
38
39
40
41
42
43
44
45
46
47
48
    {
        float diff = std::abs(double(ref.mData[i]) - double(result.mData[i]));
        if(max_diff < diff)
        {
            return false;
        }
    }

    return true;
}

49
struct gemmArgs
ltqin's avatar
ltqin committed
50
{
Chao Liu's avatar
Chao Liu committed
51
    GemmMatrixLayout layout;
52
53
54
55
56
57
58
59
    int M;
    int N;
    int K;
    int StrideA;
    int StrideB;
    int StrideC;
    int KBatch;
};
ltqin's avatar
ltqin committed
60

61
62
int test_gemm(const gemmArgs& args)
{
63
64
65
66
67
    using Row = ck::tensor_layout::gemm::RowMajor;
    using Col = ck::tensor_layout::gemm::ColumnMajor;

    using PassThrough = ck::tensor_operation::element_wise::PassThrough;

ltqin's avatar
ltqin committed
68
69
    bool a_row_major, b_row_major, c_row_major;

70
    switch(args.layout)
ltqin's avatar
ltqin committed
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
    {
    case GemmMatrixLayout::MK_KN_MN:
        a_row_major = true;
        b_row_major = true;
        c_row_major = true;
        break;
    case GemmMatrixLayout::MK_NK_MN:
        a_row_major = true;
        b_row_major = false;
        c_row_major = true;
        break;
    case GemmMatrixLayout::KM_KN_MN:
        a_row_major = false;
        b_row_major = true;
        c_row_major = true;
        break;
    case GemmMatrixLayout::KM_NK_MN:
        a_row_major = false;
        b_row_major = false;
        c_row_major = true;
        break;
    default: printf("not supported layout"); return 1;
    }

    auto f_host_tensor_descriptor =
        [](std::size_t row, std::size_t col, std::size_t stride, bool row_major) {
97
98
            using namespace ck::literals;

ltqin's avatar
ltqin committed
99
100
            if(row_major)
            {
101
                return HostTensorDescriptor({row, col}, {stride, 1_uz});
ltqin's avatar
ltqin committed
102
103
104
            }
            else
            {
105
                return HostTensorDescriptor({row, col}, {1_uz, stride});
ltqin's avatar
ltqin committed
106
107
108
            }
        };

109
110
    Tensor<float> a_m_k(f_host_tensor_descriptor(args.M, args.K, args.StrideA, a_row_major));
    Tensor<float> b_k_n(f_host_tensor_descriptor(args.K, args.N, args.StrideB, b_row_major));
111
112
113
114
    Tensor<float> c_m_n_host_result(
        f_host_tensor_descriptor(args.M, args.N, args.StrideC, c_row_major));
    Tensor<float> c_m_n_device_result(
        f_host_tensor_descriptor(args.M, args.N, args.StrideC, c_row_major));
ltqin's avatar
ltqin committed
115
116

    // init data
117
    std::size_t num_thread = 1;
ltqin's avatar
ltqin committed
118
119
120
121
122
123
124
125
126
127
128
129
    a_m_k.GenerateTensorValue(GeneratorTensor_2<float>{-5, 5}, num_thread);
    b_k_n.GenerateTensorValue(GeneratorTensor_2<float>{-5, 5}, num_thread);
    // set zero to c_device_buf
    c_m_n_device_result.GenerateTensorValue(GeneratorTensor_0<float>{}, num_thread);

    host_gemm_mk_kn_mn(a_m_k,
                       b_k_n,
                       c_m_n_host_result,
                       ck::tensor_operation::element_wise::PassThrough{},
                       ck::tensor_operation::element_wise::PassThrough{},
                       ck::tensor_operation::element_wise::PassThrough{});

130
131
132
    DeviceMem a_device_buf(sizeof(float) * a_m_k.mDesc.GetElementSpaceSize());
    DeviceMem b_device_buf(sizeof(float) * b_k_n.mDesc.GetElementSpaceSize());
    DeviceMem c_device_buf(sizeof(float) * c_m_n_device_result.mDesc.GetElementSpaceSize());
ltqin's avatar
ltqin committed
133
134
135
136
137

    a_device_buf.ToDevice(a_m_k.mData.data());
    b_device_buf.ToDevice(b_k_n.mData.data());
    c_device_buf.ToDevice(c_m_n_device_result.mData.data());

138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
    auto test = [&](auto a_layout, auto b_layout, auto c_layout) {
        bool success = false;

        using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<decltype(a_layout),
                                                                        decltype(b_layout),
                                                                        decltype(c_layout),
                                                                        float,
                                                                        float,
                                                                        float,
                                                                        PassThrough,
                                                                        PassThrough,
                                                                        PassThrough>;

        const auto gemm_ptrs =
            ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
                DeviceOp>::GetInstances();

        for(auto& gemm_ptr : gemm_ptrs)
        {
            auto argument_ptr =
                gemm_ptr->MakeArgumentPointer(static_cast<float*>(a_device_buf.GetDeviceBuffer()),
                                              static_cast<float*>(b_device_buf.GetDeviceBuffer()),
                                              static_cast<float*>(c_device_buf.GetDeviceBuffer()),
                                              args.M,
                                              args.N,
                                              args.K,
                                              args.StrideA,
                                              args.StrideB,
                                              args.StrideC,
                                              ck::tensor_operation::element_wise::PassThrough{},
                                              ck::tensor_operation::element_wise::PassThrough{},
                                              ck::tensor_operation::element_wise::PassThrough{},
                                              args.KBatch);

            auto invoker_ptr = gemm_ptr->MakeInvokerPointer();

            if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
            {
                invoker_ptr->Run(argument_ptr.get());

                c_device_buf.FromDevice(c_m_n_device_result.mData.data());

                if(!check_out(c_m_n_host_result, c_m_n_device_result))
                {
                    success = false;
                    break;
                }
                success = true;
            }
        }

        return success;
    };

    bool success = false;
ltqin's avatar
ltqin committed
193

194
    if(args.layout == GemmMatrixLayout::MK_KN_MN)
ltqin's avatar
ltqin committed
195
    {
196
        success = test(Row{}, Row{}, Row{});
ltqin's avatar
ltqin committed
197
    }
198
    else if(args.layout == GemmMatrixLayout::MK_NK_MN)
ltqin's avatar
ltqin committed
199
    {
200
        success = test(Row{}, Col{}, Row{});
ltqin's avatar
ltqin committed
201
    }
202
    else if(args.layout == GemmMatrixLayout::KM_KN_MN)
ltqin's avatar
ltqin committed
203
    {
204
        success = test(Col{}, Row{}, Row{});
ltqin's avatar
ltqin committed
205
206
207
    }
    else
    {
208
        success = test(Col{}, Col{}, Row{});
ltqin's avatar
ltqin committed
209
210
    }

211
    auto error_code = 0;
ltqin's avatar
ltqin committed
212
213
214
215
216
217
218
    if(success)
    {
        std::cout << "test split k : Pass" << std::endl;
    }
    else
    {
        std::cout << "test split k: Fail " << std::endl;
219
        error_code = -1; // test needs to report failure
220
221
222
223
224
225
226
227
228
    }
    return error_code;
}

int main(int argc, char* argv[])
{
    std::vector<gemmArgs> test_cases;
    if(argc == 1)
    {
Anthony Chang's avatar
Anthony Chang committed
229
230
        test_cases = {{GemmMatrixLayout::MK_KN_MN, 1024, 1024, 1024, 1024, 1024, 1024, 2},
                      {GemmMatrixLayout::MK_KN_MN, 1024, 1024, 1024, 1024, 1024, 1024, 8}};
231
232
233
    }
    else if(argc == 9)
    {
Chao Liu's avatar
Chao Liu committed
234
        const auto layout = static_cast<GemmMatrixLayout>(std::stoi(argv[1]));
235

236
237
238
        const int M = std::stoi(argv[2]);
        const int N = std::stoi(argv[3]);
        const int K = std::stoi(argv[4]);
239

240
241
242
243
244
        const int StrideA = std::stoi(argv[5]);
        const int StrideB = std::stoi(argv[6]);
        const int StrideC = std::stoi(argv[7]);
        const int KBatch  = std::stoi(argv[8]);
        test_cases        = {{layout, M, N, K, StrideA, StrideB, StrideC, KBatch}};
245
246
247
248
249
250
251
252
253
254
    }
    else
    {
        printf("arg1: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
        printf("arg2 to 7: M, N, K, StrideA, StrideB, StrideC KBatch\n");
        return -1;
    }
Anthony Chang's avatar
Anthony Chang committed
255
    bool error = false;
256
    for(const auto& kinder : test_cases)
257
    {
Anthony Chang's avatar
Anthony Chang committed
258
        error |= test_gemm(kinder);
ltqin's avatar
ltqin committed
259
    }
Anthony Chang's avatar
Anthony Chang committed
260
    return error ? 1 : 0;
ltqin's avatar
ltqin committed
261
}