gemm_xdl_fp16.cpp 16.3 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
Chao Liu's avatar
Chao Liu committed
14
#include "device_gemm_xdl.hpp"
rocking5566's avatar
rocking5566 committed
15
#include "device_gemm_xdl_c_shuffle.hpp"
Chao Liu's avatar
Chao Liu committed
16
#include "element_wise_operation.hpp"
Chao Liu's avatar
Chao Liu committed
17
#include "reference_gemm.hpp"
Chao Liu's avatar
Chao Liu committed
18
#include "gemm_specialization.hpp"
Chao Liu's avatar
Chao Liu committed
19

Chao Liu's avatar
Chao Liu committed
20
21
22
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;

Chao Liu's avatar
Chao Liu committed
23
24
25
26
27
28
29
30
using F16 = ck::half_t;
using F32 = float;

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

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

Chao Liu's avatar
Chao Liu committed
31
32
33
34
35
36
37
38
39
using ADataType   = ck::half_t;
using BDataType   = ck::half_t;
using CDataType   = ck::half_t;
using AccDataType = float;

using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;

Chao Liu's avatar
Chao Liu committed
40
41
42
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
Chao Liu's avatar
Chao Liu committed
43

Chao Liu's avatar
Chao Liu committed
44

Chao Liu's avatar
Chao Liu committed
45
// clang-format off
46
#if 0
Chao Liu's avatar
Chao Liu committed
47
48
49
50
51
52
53
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|      Num|
//######|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|
//######|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|         |
//######|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |
//    [256, 128, 4, 8], 1 stage, 2 occupancy
        <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1>;
54
55
#elif 1
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
56
57
58
59
60
//######|AData| BData| CData| AccData| Shuffle| ALayout| BLayout| CLayout|           A|           B|           C| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|     CBlockTransferClusterLengths|  CBlockTransfer|
//######| Type|  Type|  Type|    Type|    Data|        |        |        | Elementwise| Elementwise| Elementwise|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//######|     |      |      |        |    Type|        |        |        |   Operation|   Operation|   Operation|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl|   _NWaveNPerXdl|
//######|     |      |      |        |        |        |        |        |            |            |            |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |
        <  F16,   F16,   F16,     F32,     F16,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,           1,           1,             S<1, 1, 32, 1, 1, 8>,               8>;
Chao Liu's avatar
Chao Liu committed
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
#elif 0
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout|           A|           B|           C|          GEMM| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|      Num|
//######|  Type|  Type|  Type|    Type|        |        |        | Elementwise| Elementwise| Elementwise|Spacialization|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| SrcDstVectorDim|       DstScalar| Prefetch|
//######|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|              |      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector|         |
//######|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |
//    [128, 144, 8, 8], 1 stage, 1 occupancy, bounded by LDS size
//     99 TFlops, 120 blocks (1024x2160x3840)
//     99 TFlops, 960 blocks (4096x4320x3840)
        <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   144,     8,  8,   16,   16,    2,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        1>;
//    [128, 144, 4, 8], 1 stage, 2 occupancy,
//     92 TFlops, 120 blocks (1024x2160x3840)
//    120 TFlops, 240 blocks (1024x4320x3840)
//    128 TFlops, 960 blocks (4096x4320x3840)
//      <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,   128,   144,     4,  8,   16,   16,    2,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        1>;
//    [ 64, 144, 8, 8], 1 stage, 2 occupancy/
//     96 TFlops, 240 blocks (1024x2160x3840)
//     96 TFlops, 480 blocks (1024x4320x3840)
//     99 TFlops,1920 blocks (4096x4320x3840)
//      <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   144,     8,  8,   16,   16,    1,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        1>;
//    [ 64, 144, 8, 8], 2 stage, 2 occupancy
//     93 TFlops
//      <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   144,     8,  8,   16,   16,    1,    9,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<8,  8, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        2>;
//    [ 64, 144, 4, 8], 1 stage, 2 occupancy
//     87 TFlops
//      <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   144,     4,  8,   16,   16,    1,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        1>;
//    [ 64, 144, 4, 8], 2 stage, 2 occupancy
//     85 TFlops
//      <   F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   GemmDefault,   256,    64,   144,     4,  8,   16,   16,    1,    9,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 4>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,      true,               7,               1,        2>;
#endif
Chao Liu's avatar
Chao Liu committed
91
92
// clang-format on

Chao Liu's avatar
Chao Liu committed
93
94
using ReferenceGemmInstance = ck::tensor_operation::host::
    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
95
96
97

int main(int argc, char* argv[])
{
Chao Liu's avatar
Chao Liu committed
98
99
100
    bool do_verification = 0;
    int init_method      = 0;
    int nrepeat          = 5;
101
102
103
104
105
106
107
108
109
110

    // GEMM shape
    ck::index_t M = 3840;
    ck::index_t N = 4096;
    ck::index_t K = 4096;

    ck::index_t StrideA = 4096;
    ck::index_t StrideB = 4096;
    ck::index_t StrideC = 4096;

Chao Liu's avatar
Chao Liu committed
111
112
    if(argc == 4)
    {
113
114
115
        do_verification = std::stoi(argv[1]);
        init_method     = std::stoi(argv[2]);
        nrepeat         = std::stoi(argv[3]);
Chao Liu's avatar
Chao Liu committed
116
117
118
119
120
121
122
123
124
125
    }
    else if(argc == 10)
    {
        do_verification = std::stoi(argv[1]);
        init_method     = std::stoi(argv[2]);
        nrepeat         = std::stoi(argv[3]);

        M = std::stoi(argv[4]);
        N = std::stoi(argv[5]);
        K = std::stoi(argv[6]);
126

Chao Liu's avatar
Chao Liu committed
127
128
129
130
131
132
133
134
135
136
137
138
        StrideA = std::stoi(argv[7]);
        StrideB = std::stoi(argv[8]);
        StrideC = std::stoi(argv[9]);
    }
    else
    {
        printf("arg1: verification (0=no, 1=yes)\n");
        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
        printf("arg3: run kernel # of times (>1)\n");
        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
        exit(0);
    }
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155

    auto f_host_tensor_descriptor =
        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
            {
                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
                                            std::vector<std::size_t>({stride, 1}));
            }
            else
            {
                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
                                            std::vector<std::size_t>({1, stride}));
            }
        };

    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
156
157
    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
158
159
160
161
162
163
164
165
166

    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;

    switch(init_method)
    {
    case 0: break;
    case 1:
167
168
        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
169
        break;
170
    case 2:
171
172
        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
173
174
175
176
        break;
    default:
        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
177
178
179
180
181
182
183
184
185
    }

    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());

    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
    b_k_n_device_buf.ToDevice(b_k_n.mData.data());

Chao Liu's avatar
Chao Liu committed
186
187
188
189
    auto a_element_op = AElementOp{};
    auto b_element_op = BElementOp{};
    auto c_element_op = CElementOp{};

190
    // do GEMM
Chao Liu's avatar
Chao Liu committed
191
    auto gemm     = DeviceGemmInstance{};
192
193
194
195
196
197
198
199
200
    auto invoker  = gemm.MakeInvoker();
    auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
                                      static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
                                      static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
                                      M,
                                      N,
                                      K,
                                      StrideA,
                                      StrideB,
Chao Liu's avatar
Chao Liu committed
201
                                      StrideC,
Chao Liu's avatar
Chao Liu committed
202
203
204
                                      a_element_op,
                                      b_element_op,
                                      c_element_op);
205
206
207
208
209
210
211
212
213
214
215
216

    if(!gemm.IsSupportedArgument(argument))
    {
        throw std::runtime_error(
            "wrong! device_gemm with the specified compilation parameters does "
            "not support this GEMM problem");
    }

    float ave_time = invoker.Run(argument, nrepeat);

    std::size_t flop = std::size_t(2) * M * N * K;
    std::size_t num_btype =
Chao Liu's avatar
Chao Liu committed
217
        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
218
219
220
221
222

    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;

    float gb_per_sec = num_btype / 1.E6 / ave_time;

Chao Liu's avatar
Chao Liu committed
223
224
    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
              << gemm.GetTypeString() << std::endl;
225
226
227
228
229

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

    if(do_verification)
    {
Chao Liu's avatar
Chao Liu committed
230
231
232
233
234
235
236
        auto ref_gemm    = ReferenceGemmInstance{};
        auto ref_invoker = ref_gemm.MakeInvoker();

        auto ref_argument = ref_gemm.MakeArgument(
            a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);

        ref_invoker.Run(ref_argument);
237
238
239

        check_error(c_m_n_host_result, c_m_n_device_result);
    }
Jianfeng Yan's avatar
Jianfeng Yan committed
240
241

    return 0;
242
}