add bias example

example/15_grouped_gemm/CMakeLists.txt

 add_example_executable(example_grouped_gemm_xdl_fp16 grouped_gemm_xdl_fp16.cpp)
+add_example_executable(example_grouped_gemm_bias_xdl_fp16 grouped_gemm_bias_xdl_fp16.cpp)

example/15_grouped_gemm/grouped_gemm_bias_xdl_fp16.cpp

+// 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/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_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_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 D0DataType       = F16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Col;
+using ELayout = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
+    // clang-format off
+//######| 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
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    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);
+    }
+
+    int group_count = rand() % 16 + 1;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<const void*> p_a, p_b;
+    std::vector<std::vector<const void*>> p_ds;
+    std::vector<void*> p_c;
+
+    gemm_descs.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = 256 + 256 * i;
+        int N = 128 + 128 * i;
+        int K = 64 + 64 * i;
+
+        int stride_A = K;
+        int stride_B = K;
+        int stride_C = N;
+
+        std::vector<ck::index_t> stride_Ds = {0};
+
+        gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, stride_Ds});
+    }
+
+    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}));
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> e_host_tensors;
+    std::vector<Tensor<EDataType>> e_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d0_tensors.reserve(group_count);
+    e_host_tensors.reserve(group_count);
+    e_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, d0_tensors_device,
+        e_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d0_tensors_device.reserve(group_count);
+    e_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
+        d0_tensors.push_back(Tensor<D0DataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_Ds_[0], ELayout{})));
+        e_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+        e_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << e_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
+
+        switch(init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            d0_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            d0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        }
+    }
+
+    for(std::size_t i = 0; i < gemm_descs.size(); i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpace()));
+        b_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpace()));
+        d0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSpace()));
+        e_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSpace()));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+
+        p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
+        p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
+        p_ds.push_back({d0_tensors_device[i]->GetDeviceBuffer()});
+        p_c.push_back(e_tensors_device[i]->GetDeviceBuffer());
+    }
+
+    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(
+        p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+    DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
+
+    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});
+
+    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)
+    {
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data());
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      e_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < gemm_descs[i].M_; ++m)
+            {
+                for(int n = 0; n < gemm_descs[i].N_; ++n)
+                {
+                    cde_element_op(
+                        e_host_tensors[i](m, n), e_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            pass &= ck::utils::check_err(e_device_tensors[i].mData, e_host_tensors[i].mData);
+        }
+    }
+
+    return pass ? 0 : 1;
+}

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

@@ -66,7 +66,7 @@ __global__ void
     GridwiseGemm::template Run<HasMainKBlockLoop>(
         gemm_desc_ptr[group_id].a_ptr_,
         gemm_desc_ptr[group_id].b_ptr_,
-        ck::Tuple<>{},
+        gemm_desc_ptr[group_id].ds_ptr_,
         gemm_desc_ptr[group_id].e_ptr_,
         p_shared,
         a_element_op,
@@ -74,9 +74,7 @@ __global__ void
         c_element_op,
         gemm_desc_ptr[group_id].a_grid_desc_k0_m_k1_,
         gemm_desc_ptr[group_id].b_grid_desc_k0_n_k1_,
-        ck::StaticallyIndexedArray<
-            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
-            0>{},
+        gemm_desc_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
         gemm_desc_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
         gemm_desc_ptr[group_id].block_2_ctile_map_);
 #else
@@ -354,7 +352,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         }
     }
 
-    static auto MakeCGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
     {
         const auto c_grid_desc_mraw_nraw = [&]() {
             if constexpr(is_same<tensor_layout::gemm::RowMajor, DELayout>::value)
@@ -414,7 +412,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
 
     using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
     using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
-    using EGridDesc_M_N       = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
+    using EGridDesc_M_N       = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
 
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle<
@@ -571,7 +569,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
                     DeviceGroupedGemmXdl::MakeBGridDescriptor_BK0_N_BK1(K, N, StrideB);
 
                 const auto e_grid_desc_m_n_ =
-                    DeviceGroupedGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC);
+                    DeviceGroupedGemmXdl::MakeEGridDescriptor_M_N(M, N, StrideC);
 
                 const index_t grid_size_grp =
                     GroupedGemmBlock2ETileMap(e_grid_desc_m_n_, 0)
@@ -599,23 +597,20 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
                         ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of
                                                                          // different
 
-                    typename GridwiseGemm::DsGridPointer p_ds_grid_;
+                    typename GridwiseGemm::DsGridPointer p_ds_grid_{};
 
-                    if constexpr(NumDTensor > 0)
-                    {
-                        static_for<0, NumDTensor, 1>{}([&](auto j) {
-                            using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+                    static_for<0, NumDTensor, 1>{}([&](auto j) {
+                        using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
 
-                            p_ds_grid_(i) = static_cast<const DDataType*>(p_Ds[i][j]);
+                        p_ds_grid_(j) = static_cast<const DDataType*>(p_Ds[i][j]);
 
-                            const auto d_grid_desc_m_n = GridwiseGemm::MakeEGridDescriptor_M_N(
-                                M, N, gemm_descs[i].stride_Ds_[j]);
+                        const auto d_grid_desc_m_n = DeviceGroupedGemmXdl::MakeEGridDescriptor_M_N(
+                            M, N, gemm_descs[i].stride_Ds_[j]);
 
-                            ds_grid_desc_mblock_mperblock_nblock_nperblock_(j) =
-                                GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                                    d_grid_desc_m_n);
-                        });
-                    }
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock_(j) =
+                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                d_grid_desc_m_n);
+                    });
 
                     gemm_desc_kernel_arg_.push_back(
                         GemmBiasTransKernelArg{a_grid_desc_k0_m_k1_,