contraction_scale_xdl_fp32.cpp 25.7 KB
Newer Older
Chao Liu's avatar
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
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.

#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>

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

#include "ck/library/utility/check_err.hpp"
#include "ck/library/host_tensor/device_memory.hpp"
#include "ck/library/host_tensor/host_tensor.hpp"
#include "ck/library/host_tensor/host_tensor_generator.hpp"

template <ck::index_t... Is>
using S = ck::Sequence<Is...>;

using F32 = float;

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

using ADataType        = F32;
using BDataType        = F32;
using AccDataType      = F32;
using CShuffleDataType = F32;
Chao Liu's avatar
Chao Liu committed
30
using DsDataType       = ck::Tuple<>;
Chao Liu's avatar
Chao Liu committed
31
32
33
34
35
36
37
38
using EDataType        = F32;

static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;

using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
Chao Liu's avatar
Chao Liu committed
39
using CDEElementOp = ck::tensor_operation::element_wise::Scale;
Chao Liu's avatar
Chao Liu committed
40

Chao Liu's avatar
Chao Liu committed
41
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
Chao Liu's avatar
Chao Liu committed
42
43

// clang-format off
Chao Liu's avatar
Chao Liu committed
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
using DeviceOpInstanceKKNN = ck::tensor_operation::device::
        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| 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| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  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_MWaveMPerXdl| ScalarPerVector|
        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;

using DeviceOpInstanceKNNN = ck::tensor_operation::device::
        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| 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| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  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_MWaveMPerXdl| ScalarPerVector|
        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;

using DeviceOpInstanceMKNN = ck::tensor_operation::device::
        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| 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| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  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_MWaveMPerXdl| ScalarPerVector|
        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4>;

using DeviceOpInstanceMNNN = ck::tensor_operation::device::
        //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle|     DsData| EData|            A|           B|          CDE|           GEMM| NumGemmK| 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| DataType|       Type|  Type|  Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  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_MWaveMPerXdl| ScalarPerVector|
        //#####################################|        |        |        |      |      |        |         |           |      |    Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
        //#####################################|        |        |        |      |      |        |         |           |      |             |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK,   F32,   F32,     F32,      F32, DsDataType,   F32,   AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4>;
Chao Liu's avatar
Chao Liu committed
71
72
// clang-format on

Chao Liu's avatar
Chao Liu committed
73
74
using DeviceOpInstance = DeviceOpInstanceKKNN;

Chao Liu's avatar
Chao Liu committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
// hardcoded for NumDimM == NumDimN == NumDimK == 2
template <ck::index_t NumDimM,
          ck::index_t NumDimN,
          ck::index_t NumDimK,
          typename ADataType,
          typename BDataType,
          typename EDataType,
          typename AccDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CDEElementwiseOperation,
          ck::enable_if_t<NumDimM == 2 && NumDimN == 2 && NumDimK == 2, bool> = false>
struct ReferenceContraction_M2_N2_K2 : public ck::tensor_operation::device::BaseOperator
{
    // Argument
    struct Argument : public ck::tensor_operation::device::BaseArgument
    {
        Argument(const Tensor<ADataType>& a_ms_ks,
93
                 const Tensor<BDataType>& b_ns_ks,
Chao Liu's avatar
Chao Liu committed
94
95
96
97
98
                 Tensor<EDataType>& e_ms_ns,
                 AElementwiseOperation a_element_op,
                 BElementwiseOperation b_element_op,
                 CDEElementwiseOperation cde_element_op)
            : a_ms_ks_{a_ms_ks},
99
              b_ns_ks_{b_ns_ks},
Chao Liu's avatar
Chao Liu committed
100
101
102
103
104
105
106
107
              e_ms_ns_{e_ms_ns},
              a_element_op_{a_element_op},
              b_element_op_{b_element_op},
              cde_element_op_{cde_element_op}
        {
        }

        const Tensor<ADataType>& a_ms_ks_;
108
        const Tensor<BDataType>& b_ns_ks_;
Chao Liu's avatar
Chao Liu committed
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
        Tensor<EDataType>& e_ms_ns_;

        AElementwiseOperation a_element_op_;
        BElementwiseOperation b_element_op_;
        CDEElementwiseOperation cde_element_op_;
    };

    // Invoker
    struct Invoker : public ck::tensor_operation::device::BaseInvoker
    {
        using Argument = ReferenceContraction_M2_N2_K2::Argument;

        float Run(const Argument& arg)
        {
            auto f_ms_ns = [&](auto m0, auto m1, auto n0, auto n1) {
                const int K0 = arg.a_ms_ks_.mDesc.GetLengths()[2];
                const int K1 = arg.a_ms_ks_.mDesc.GetLengths()[3];

                AccDataType v_acc = 0;

                for(int k0 = 0; k0 < K0; ++k0)
                {
                    for(int k1 = 0; k1 < K1; ++k1)
                    {
                        AccDataType v_a;
                        AccDataType v_b;

                        arg.a_element_op_(
                            v_a, static_cast<const AccDataType>(arg.a_ms_ks_(m0, m1, k0, k1)));
                        arg.b_element_op_(
139
                            v_b, static_cast<const AccDataType>(arg.b_ns_ks_(n0, n1, k0, k1)));
Chao Liu's avatar
Chao Liu committed
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

                        v_acc += v_a * v_b;
                    }
                }

                AccDataType v_c;

                arg.cde_element_op_(v_c, v_acc);

                arg.e_ms_ns_(m0, m1, n0, n1) = v_c;
            };

            make_ParallelTensorFunctor(f_ms_ns,
                                       arg.e_ms_ns_.mDesc.GetLengths()[0],
                                       arg.e_ms_ns_.mDesc.GetLengths()[1],
                                       arg.e_ms_ns_.mDesc.GetLengths()[2],
                                       arg.e_ms_ns_.mDesc.GetLengths()[3])(
                std::thread::hardware_concurrency());

            return 0;
        }

        float Run(const ck::tensor_operation::device::BaseArgument* p_arg,
                  const StreamConfig& /* stream_config */ = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg));
        }
    };

    static constexpr bool IsValidCompilationParameter()
    {
        // TODO: properly implement this check
        return true;
    }

    bool IsSupportedArgument(const ck::tensor_operation::device::BaseArgument*) override
    {
        return true;
    }

    static auto MakeArgument(const Tensor<ADataType>& a_ms_ks,
181
                             const Tensor<BDataType>& b_ns_ks,
Chao Liu's avatar
Chao Liu committed
182
183
184
185
186
                             Tensor<EDataType>& e_ms_ns,
                             AElementwiseOperation a_element_op,
                             BElementwiseOperation b_element_op,
                             CDEElementwiseOperation cde_element_op)
    {
187
        return Argument{a_ms_ks, b_ns_ks, e_ms_ns, a_element_op, b_element_op, cde_element_op};
Chao Liu's avatar
Chao Liu committed
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
    }

    static auto MakeInvoker() { return Invoker{}; }

    virtual std::unique_ptr<ck::tensor_operation::device::BaseInvoker> MakeInvokerPointer()
    {
        return std::make_unique<Invoker>(Invoker{});
    }

    std::string GetTypeString() const override
    {
        auto str = std::stringstream();

        // clang-format off
        str << "ReferenceContraction_M2_N2_K2"
            << std::endl;
        // clang-format on

        return str.str();
    }
};

int main(int argc, char* argv[])
{
    bool do_verification = true;
    int init_method      = 1;
    bool time_kernel     = false;

216
217
218
219
220
221
222
223
224
225
    // A[M0, M1, K0, K1]
    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
    // B[N0, N1, K0, K1]
    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
    // E[M0, M1, N0, N1]
    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};

Chao Liu's avatar
Chao Liu committed
226
    float scale = 1.f;
227
228
229
230
231
232

    if(argc == 1)
    {
        // use default case
    }
    else if(argc == 4)
Chao Liu's avatar
Chao Liu committed
233
234
235
236
237
    {
        do_verification = std::stoi(argv[1]);
        init_method     = std::stoi(argv[2]);
        time_kernel     = std::stoi(argv[3]);
    }
Chao Liu's avatar
Chao Liu committed
238
    else if(argc == 23)
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
    {
        do_verification = std::stoi(argv[1]);
        init_method     = std::stoi(argv[2]);
        time_kernel     = std::stoi(argv[3]);

        const ck::index_t M0 = std::stoi(argv[4]);
        const ck::index_t M1 = std::stoi(argv[5]);

        const ck::index_t N0 = std::stoi(argv[6]);
        const ck::index_t N1 = std::stoi(argv[7]);

        const ck::index_t K0 = std::stoi(argv[8]);
        const ck::index_t K1 = std::stoi(argv[9]);

        a_ms_ks_lengths = {M0, M1, K0, K1};
        a_ms_ks_strides = {
            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};

        b_ns_ks_lengths = {N0, N1, K0, K1};
        b_ns_ks_strides = {
            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};

        e_ms_ns_lengths = {M0, M1, N0, N1};
        e_ms_ns_strides = {
            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};

Chao Liu's avatar
Chao Liu committed
265
        scale = std::stof(argv[26]);
266
    }
Chao Liu's avatar
Chao Liu committed
267
268
269
270
271
    else
    {
        printf("arg1: verification (0=no, 1=yes)\n");
        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
        printf("arg3: time kernel (0=no, 1=yes)\n");
272
273
274
        printf("arg4 to 7: M0, M1, N0, N1, K0, K1\n");
        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
Chao Liu's avatar
Chao Liu committed
275
276
        printf("arg18 to 21: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
        printf("arg22: scale\n");
Chao Liu's avatar
Chao Liu committed
277
278
279
280
281
282
        exit(0);
    }

    Tensor<ADataType> a_ms_ks(
        std::vector<std::size_t>(a_ms_ks_lengths.begin(), a_ms_ks_lengths.end()),
        std::vector<std::size_t>(a_ms_ks_strides.begin(), a_ms_ks_strides.end()));
283
284
285
    Tensor<BDataType> b_ns_ks(
        std::vector<std::size_t>(b_ns_ks_lengths.begin(), b_ns_ks_lengths.end()),
        std::vector<std::size_t>(b_ns_ks_strides.begin(), b_ns_ks_strides.end()));
Chao Liu's avatar
Chao Liu committed
286
287
288
289
290
291
292
293
    Tensor<EDataType> e_ms_ns_host_result(
        std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
        std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
    Tensor<EDataType> e_ms_ns_device_result(
        std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
        std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));

    std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
294
    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
Chao Liu's avatar
Chao Liu committed
295
296
297
298
299
300
301
    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;

    switch(init_method)
    {
    case 0: break;
    case 1:
        a_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
302
        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
Chao Liu's avatar
Chao Liu committed
303
        break;
304
    default:
Chao Liu's avatar
Chao Liu committed
305
        a_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
306
        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
Chao Liu's avatar
Chao Liu committed
307
308
309
        break;
    }

310
311
312
    DeviceMem a_device_buf(sizeof(ADataType) * a_ms_ks.mDesc.GetElementSpace());
    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpace());
    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpace());
Chao Liu's avatar
Chao Liu committed
313

314
315
    a_device_buf.ToDevice(a_ms_ks.mData.data());
    b_device_buf.ToDevice(b_ns_ks.mData.data());
Chao Liu's avatar
Chao Liu committed
316
317

    // set zero
318
    e_device_buf.SetZero();
Chao Liu's avatar
Chao Liu committed
319
320
321

    auto a_element_op   = AElementOp{};
    auto b_element_op   = BElementOp{};
Chao Liu's avatar
Chao Liu committed
322
    auto cde_element_op = CDEElementOp{scale};
Chao Liu's avatar
Chao Liu committed
323
324

    // device operation
325
326
327
328
    auto op       = DeviceOpInstance{};
    auto invoker  = op.MakeInvoker();
    auto argument = op.MakeArgument(a_device_buf.GetDeviceBuffer(),
                                    b_device_buf.GetDeviceBuffer(),
Chao Liu's avatar
Chao Liu committed
329
                                    std::array<const void*, 0>{},
330
331
332
333
334
                                    e_device_buf.GetDeviceBuffer(),
                                    a_ms_ks_lengths,
                                    a_ms_ks_strides,
                                    b_ns_ks_lengths,
                                    b_ns_ks_strides,
Chao Liu's avatar
Chao Liu committed
335
336
                                    std::array<std::vector<ck::index_t>, 0>{},
                                    std::array<std::vector<ck::index_t>, 0>{},
337
338
339
340
341
                                    e_ms_ns_lengths,
                                    e_ms_ns_strides,
                                    a_element_op,
                                    b_element_op,
                                    cde_element_op);
Chao Liu's avatar
Chao Liu committed
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366

    if(!op.IsSupportedArgument(argument))
    {
        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;

        return 0;
    }

    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});

    ck::index_t M = std::accumulate(e_ms_ns_lengths.begin(),
                                    e_ms_ns_lengths.begin() + NumDimM,
                                    ck::index_t{1},
                                    std::multiplies<ck::index_t>{});

    ck::index_t N = std::accumulate(e_ms_ns_lengths.begin() + NumDimM,
                                    e_ms_ns_lengths.begin() + NumDimM + NumDimN,
                                    ck::index_t{1},
                                    std::multiplies<ck::index_t>{});

    ck::index_t K = std::accumulate(a_ms_ks_lengths.begin() + NumDimM,
                                    a_ms_ks_lengths.begin() + NumDimM + NumDimK,
                                    ck::index_t{1},
                                    std::multiplies<ck::index_t>{});

Chao Liu's avatar
Chao Liu committed
367
368
369
    std::size_t flop = std::size_t(2) * M * N * K;
    std::size_t num_btype =
        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
Chao Liu's avatar
Chao Liu committed
370
371
372
373
374
375
376
377

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

    float gb_per_sec = num_btype / 1.E6 / ave_time;

    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
              << op.GetTypeString() << std::endl;

378
    e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
Chao Liu's avatar
Chao Liu committed
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400

    if(do_verification)
    {
        Tensor<CShuffleDataType> c_ms_ns_host_result(
            std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
            std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));

        using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
                                                                  NumDimN,
                                                                  NumDimK,
                                                                  ADataType,
                                                                  BDataType,
                                                                  CShuffleDataType,
                                                                  AccDataType,
                                                                  AElementOp,
                                                                  BElementOp,
                                                                  PassThrough>;

        auto ref_gemm    = ReferenceOpInstance{};
        auto ref_invoker = ref_gemm.MakeInvoker();

        auto ref_argument = ref_gemm.MakeArgument(
401
            a_ms_ks, b_ns_ks, c_ms_ns_host_result, a_element_op, b_element_op, PassThrough{});
Chao Liu's avatar
Chao Liu committed
402
403
404
405
406
407
408
409
410
411
412
413

        ref_invoker.Run(ref_argument);

        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
        {
            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
            {
                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
                {
                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
                    {
                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
Chao Liu's avatar
Chao Liu committed
414
                                       c_ms_ns_host_result(m0, m1, n0, n1));
Chao Liu's avatar
Chao Liu committed
415
416
417
418
419
420
421
422
423
424
                    }
                }
            }
        }

        return ck::utils::check_err(e_ms_ns_device_result.mData, e_ms_ns_host_result.mData) ? 0 : 1;
    }

    return 0;
}