Batched Gemm with multiD (#329)

* add batched_gemm_multiD

* add ds

* rename file

* add batched_gemm_bias example

* add batch_strides into bmm_c_permute

* clean

* rename example_28 to example_29
Co-authored-by: Chao Liu <chao.liu2@amd.com>

example/24_batched_gemm_c_permute/batched_gemm_c_permute_xdl_fp16.cpp

@@ -26,35 +26,36 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
 
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 
-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;
-
-using AElementOp = ck::tensor_operation::element_wise::PassThrough;
-using BElementOp = ck::tensor_operation::element_wise::PassThrough;
-using CElementOp = ck::tensor_operation::element_wise::PassThrough;
-
-// static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 // static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
-static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
 
 // clang-format off
 using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmCPermuteXdl
-//######| ALayout| BLayout| AData| BData| CData| AccData|           A|           B|           C|          GEMM|      Num| 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| 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|              |         |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |   PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|    _NWaveNPerXdl|
-//######|        |        |      |      |      |        |            |            |            |              |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |             |            |                             |                 |
-//      <     Row,     Col,   F16,   F16,   F16,     F32, PassThrough, PassThrough, PassThrough,     MNPadding,        1,   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, 32, 1, 8>,                8>;
-        <     Row,     Col,   F16,   F16,   F16,     F32, PassThrough, PassThrough, PassThrough,     MNKPadding,        1,   256,   128,    64,    32,   8,   8,   32,   32,    2,    1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,            1,           1,               S<1, 32, 1, 8>,                8>;
+//######| ALayout| BLayout| ELayout|     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|
+//######|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   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,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
 // clang-format on
 
 using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
-    ReferenceBatchedGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+    ReferenceBatchedGemm<ADataType, BDataType, EDataType, AElementOp, BElementOp, CDEElementOp>;
 
 int main(int argc, char* argv[])
 {
@@ -62,15 +63,18 @@ int main(int argc, char* argv[])
     int init_method      = 1;
     bool time_kernel     = false;
 
-    const int M = 88;
-    const int N = 64;
-    const int K = 88;
+    const int M = 256;
+    const int N = 128;
+    const int K = 64;
 
     const int stride_A = K;
     const int stride_B = K;
 
-    const int G0 = 1024;
-    const int G1 = 10;
+    const int batch_stride_A = M * K;
+    const int batch_stride_B = K * N;
+
+    const int G0 = 16;
+    const int G1 = 8;
 
     const int batch_count = G0 * G1;
 
@@ -102,21 +106,24 @@ int main(int argc, char* argv[])
                                        std::size_t row,
                                        std::size_t col,
                                        std::size_t stride,
+                                       std::size_t batch_stride,
                                        auto layout) {
         if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
             return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
-                                        std::vector<std::size_t>({row * stride, stride, 1}));
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
         }
         else
         {
             return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
-                                        std::vector<std::size_t>({col * stride, 1, stride}));
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
         }
     };
 
-    Tensor<ADataType> a_g_m_k(f_host_tensor_descriptor(batch_count, M, K, stride_A, ALayout{}));
-    Tensor<BDataType> b_g_k_n(f_host_tensor_descriptor(batch_count, K, N, stride_B, BLayout{}));
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
 
     auto f_host_c_tensor_descriptor = [](std::size_t G0_,
                                          std::size_t G1_,
@@ -131,10 +138,10 @@ int main(int argc, char* argv[])
             std::vector<std::size_t>({stride_G0_, stride_G1_, stride_M_, stride_N_}));
     };
 
-    Tensor<CDataType> c_g0_g1_m_n_host_result(
+    Tensor<EDataType> c_g0_g1_m_n_host_result(
         f_host_c_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
 
-    Tensor<CDataType> c_g0_g1_m_n_device_result(
+    Tensor<EDataType> c_g0_g1_m_n_device_result(
         f_host_c_tensor_descriptor(G0, G1, M, N, stride_G0, stride_G1, stride_M, stride_N));
 
     std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
@@ -156,32 +163,34 @@ int main(int argc, char* argv[])
 
     DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
     DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
-    DeviceMem c_device_buf(sizeof(CDataType) * c_g0_g1_m_n_device_result.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(EDataType) * c_g0_g1_m_n_device_result.mDesc.GetElementSpace());
 
     a_device_buf.ToDevice(a_g_m_k.mData.data());
     b_device_buf.ToDevice(b_g_k_n.mData.data());
 
-    auto a_element_op = AElementOp{};
-    auto b_element_op = BElementOp{};
-    auto c_element_op = CElementOp{};
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
 
     auto gemm    = DeviceGemmInstance{};
     auto invoker = gemm.MakeInvoker();
 
-    // do GEMM
+    // do GEM
     auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
                                       static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
-                                      static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+                                      static_cast<EDataType*>(c_device_buf.GetDeviceBuffer()),
                                       M,
                                       N,
                                       K,
                                       stride_A,
                                       stride_B,
+                                      batch_stride_A,
+                                      batch_stride_B,
                                       batched_gemm_c_permute_desc,
+                                      batch_count,
                                       a_element_op,
                                       b_element_op,
-                                      c_element_op,
-                                      batch_count);
+                                      cde_element_op);
 
     if(!gemm.IsSupportedArgument(argument))
     {
@@ -195,7 +204,7 @@ int main(int argc, char* argv[])
     std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
     std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
                             sizeof(BDataType) * batch_count * K * N +
-                            sizeof(CDataType) * batch_count * M * N;
+                            sizeof(EDataType) * batch_count * M * N;
 
     float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
 
@@ -213,11 +222,11 @@ int main(int argc, char* argv[])
         auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
         auto ref_invoker      = ref_batched_gemm.MakeInvoker();
 
-        Tensor<CDataType> c_g_m_n_host_result = HostTensorDescriptor(
+        Tensor<EDataType> c_g_m_n_host_result = HostTensorDescriptor(
             std::vector<std::size_t>({batch_count, M, N}), std::vector<std::size_t>({M * N, N, 1}));
 
         auto ref_argument = ref_batched_gemm.MakeArgument(
-            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, c_element_op);
+            a_g_m_k, b_g_k_n, c_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
 
         ref_invoker.Run(ref_argument);
 

example/29_batched_gemm_multi_d/CMakeLists.txt

+add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
+add_example_executable(example_batched_gemm_bias_xdl_fp16 batched_gemm_bias_xdl_fp16.cpp)
+

example/29_batched_gemm_multi_d/batched_gemm_bias_xdl_fp16.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#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/device/device_batched_gemm_multi_d_xdl.hpp"
+#include "ck/tensor_operation/gpu/element/binary_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"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+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;
+using Add         = ck::tensor_operation::element_wise::Add;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DDataType        = F16;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiDXdl
+//######| ALayout| BLayout| DELayout|     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|
+//######|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   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,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    const int M = 256 * (rand() % 16 + 1);
+    const int N = 128 * (rand() % 16 + 1);
+    const int K = 64 * (rand() % 16 + 1);
+
+    const int stride_A = K;
+    const int stride_B = K;
+    const int stride_D = 0;
+    const int stride_E = N;
+
+    const int batch_stride_A = M * K;
+    const int batch_stride_B = K * N;
+    const int batch_stride_D = N;
+    const int batch_stride_E = M * N;
+
+    const int batch_count = 16;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    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=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    // GEMM shape
+    auto f_host_tensor_descriptor = [](std::size_t batch_count_,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
+
+    Tensor<DDataType> d_g_m_n(
+        f_host_tensor_descriptor(batch_count, M, N, stride_D, batch_stride_D, DELayout{}));
+
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, DELayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "d_g_m_n: " << d_g_m_n.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_g_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_g_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_g_m_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+    d_device_buf.ToDevice(d_g_m_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {d_device_buf.GetDeviceBuffer()},
+                                      c_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      stride_A,
+                                      stride_B,
+                                      {stride_D},
+                                      stride_E,
+                                      batch_stride_A,
+                                      batch_stride_B,
+                                      {batch_stride_D},
+                                      batch_stride_E,
+                                      batch_count,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    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, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                            sizeof(BDataType) * batch_count * K * N +
+                            sizeof(EDataType) * batch_count * M * N;
+
+    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, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
+
+        using ReferenceBatchedGemmInstance =
+            ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                             BDataType,
+                                                             EDataType,
+                                                             AElementOp,
+                                                             BElementOp,
+                                                             PassThrough>;
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_E, batch_stride_E, DELayout{}));
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int g = 0; g < batch_count; g++)
+        {
+            for(int m = 0; m < M; ++m)
+            {
+                for(int n = 0; n < N; ++n)
+                {
+                    cde_element_op(e_g_m_n_host_result(g, m, n),
+                                   e_g_m_n_host_result(g, m, n),
+                                   d_g_m_n(g, m, n));
+                }
+            }
+        }
+
+        pass = ck::utils::check_err(
+            e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
+    }
+
+    return pass ? 0 : 1;
+}

example/29_batched_gemm_multi_d/batched_gemm_xdl_fp16.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#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/device/device_batched_gemm_multi_d_xdl.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"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+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;
+
+using ADataType        = F16;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+// static constexpr auto MNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+// static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceBatchedGemmMultiDXdl
+//######| ALayout| BLayout| ELayout|     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|
+//######|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   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,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+// clang-format on
+
+using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
+    ReferenceBatchedGemm<ADataType, BDataType, EDataType, AElementOp, BElementOp, CDEElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    const int M = 256 * (rand() % 16 + 1);
+    const int N = 128 * (rand() % 16 + 1);
+    const int K = 64 * (rand() % 16 + 1);
+
+    const int stride_A = K;
+    const int stride_B = K;
+    const int stride_C = N;
+
+    const int batch_stride_A = M * K;
+    const int batch_stride_B = K * N;
+    const int batch_stride_C = M * N;
+
+    const int batch_count = 16;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    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=n0, 1=yes)\n");
+        exit(0);
+    }
+
+    // GEMM shape
+    auto f_host_tensor_descriptor = [](std::size_t batch_count_,
+                                       std::size_t row,
+                                       std::size_t col,
+                                       std::size_t stride,
+                                       std::size_t batch_stride,
+                                       auto layout) {
+        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, stride, 1}));
+        }
+        else
+        {
+            return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}),
+                                        std::vector<std::size_t>({batch_stride, 1, stride}));
+        }
+    };
+
+    Tensor<ADataType> a_g_m_k(
+        f_host_tensor_descriptor(batch_count, M, K, stride_A, batch_stride_A, ALayout{}));
+    Tensor<BDataType> b_g_k_n(
+        f_host_tensor_descriptor(batch_count, K, N, stride_B, batch_stride_B, BLayout{}));
+
+    Tensor<EDataType> e_g_m_n_device_result(
+        f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+    std::cout << "a_g_m_k: " << a_g_m_k.mDesc << std::endl;
+    std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
+    std::cout << "e_g_m_n: " << e_g_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_g_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a_g_m_k.mDesc.GetElementSpace());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_g_k_n.mDesc.GetElementSpace());
+    DeviceMem c_device_buf(sizeof(EDataType) * e_g_m_n_device_result.mDesc.GetElementSpace());
+
+    a_device_buf.ToDevice(a_g_m_k.mData.data());
+    b_device_buf.ToDevice(b_g_k_n.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(a_device_buf.GetDeviceBuffer(),
+                                      b_device_buf.GetDeviceBuffer(),
+                                      {},
+                                      c_device_buf.GetDeviceBuffer(),
+                                      M,
+                                      N,
+                                      K,
+                                      stride_A,
+                                      stride_B,
+                                      {},
+                                      stride_C,
+                                      batch_stride_A,
+                                      batch_stride_B,
+                                      {},
+                                      batch_stride_C,
+                                      batch_count,
+                                      a_element_op,
+                                      b_element_op,
+                                      cde_element_op);
+
+    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, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop      = std::size_t(2) * batch_count * M * N * K;
+    std::size_t num_btype = sizeof(ADataType) * batch_count * M * K +
+                            sizeof(BDataType) * batch_count * K * N +
+                            sizeof(EDataType) * batch_count * M * N;
+
+    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, "
+              << gemm.GetTypeString() << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(e_g_m_n_device_result.mData.data());
+
+        auto ref_batched_gemm = ReferenceBatchedGemmInstance{};
+        auto ref_invoker      = ref_batched_gemm.MakeInvoker();
+
+        Tensor<EDataType> e_g_m_n_host_result(
+            f_host_tensor_descriptor(batch_count, M, N, stride_C, batch_stride_C, ELayout{}));
+
+        auto ref_argument = ref_batched_gemm.MakeArgument(
+            a_g_m_k, b_g_k_n, e_g_m_n_host_result, a_element_op, b_element_op, cde_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        pass = ck::utils::check_err(
+            e_g_m_n_host_result.mData, e_g_m_n_device_result.mData, "Error: Incorrect results c");
+    }
+
+    return pass ? 0 : 1;
+}

example/CMakeLists.txt

@@ -47,3 +47,4 @@ add_subdirectory(25_gemm_bias_c_permute)
 add_subdirectory(26_contraction)
 add_subdirectory(27_layernorm)
 add_subdirectory(28_grouped_gemm_bias)
+add_subdirectory(29_batched_gemm_multi_d)
\ No newline at end of file

include/ck/tensor_operation/gpu/device/device_batched_gemm_c_permute.hpp

@@ -14,9 +14,15 @@ struct BatchedGemmCPermuteDesc
     ck::index_t stride_G0_, stride_G1_, stride_M_, stride_N_;
 };
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename CElementwiseOperation>
+          typename CDEElementwiseOperation>
 struct DeviceBatchedGemmCPermute : public BaseOperator
 {
     virtual std::unique_ptr<BaseArgument>
@@ -28,20 +34,36 @@ struct DeviceBatchedGemmCPermute : public BaseOperator
                         index_t K,
                         index_t stride_A,
                         index_t stride_B,
+                        index_t batch_stride_A,
+                        index_t batch_stride_B,
                         BatchedGemmCPermuteDesc batched_gemm_c_permute_desc,
+                        index_t BatchCount,
                         AElementwiseOperation a_element_op,
                         BElementwiseOperation b_element_op,
-                        CElementwiseOperation c_element_op,
-                        ck::index_t BatchCount) = 0;
+                        CDEElementwiseOperation c_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <typename AElementwiseOperation,
+template <typename ALayout,
+          typename BLayout,
+          typename DELayout,
+          typename ADataType,
+          typename BDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
           typename BElementwiseOperation,
-          typename CElementwiseOperation>
-using DeviceBatchedGemmCPermutePtr = std::unique_ptr<
-    DeviceBatchedGemmCPermute<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+          typename CDEElementwiseOperation>
+using DeviceBatchedGemmCPermutePtr =
+    std::unique_ptr<DeviceBatchedGemmCPermute<ALayout,
+                                              BLayout,
+                                              DELayout,
+                                              ADataType,
+                                              BDataType,
+                                              EDataType,
+                                              AElementwiseOperation,
+                                              BElementwiseOperation,
+                                              CDEElementwiseOperation>>;
 
 } // namespace device
 } // namespace tensor_operation

include/ck/tensor_operation/gpu/device/device_batched_gemm_multi_d.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceBatchedGemmMultiD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        std::array<ck::index_t, NumDTensor> StrideDs,
+                        ck::index_t StrideE,
+                        ck::index_t BatchStrideA,
+                        ck::index_t BatchStrideB,
+                        std::array<ck::index_t, NumDTensor> BatchStrideDs,
+                        ck::index_t BatchStrideE,
+                        ck::index_t Batch,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck