init of grouped_gemm

composable_kernel/include/config.hpp

@@ -171,5 +171,12 @@ enum ActivTypeEnum_t
 using index_t      = int32_t;
 using long_index_t = int64_t;
 
+struct gemm_desc
+{
+    ck::index_t M, N, K;
+    ck::index_t StrideA, StrideB, StrideC;
+    ck::index_t OffsetA, OffsetB, OffsetC;
+};
+
 } // namespace ck
 #endif

device_operation/include/device_gemm.hpp

@@ -59,6 +59,23 @@ struct DeviceGemm : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGroupedGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
+                                                              const void* p_b,
+                                                              void* p_c,
+                                                              std::vector<gemm_desc> gemm_shapes,
+                                                              AElementwiseOperation a_element_op,
+                                                              BElementwiseOperation b_element_op,
+                                                              CElementwiseOperation c_element_op,
+                                                              ck::index_t KBatch = 1) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
 template <typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation>

example/12_grouped_gemm_xdl/README.md

+# Instructions for ```gemm_xdl``` Example
+
+## Docker script
+```bash
+docker run                                                                   \
+-it                                                                          \
+--rm                                                                         \
+--privileged                                                                 \
+--group-add sudo                                                             \
+-w /root/workspace                                                           \
+-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace                                \
+rocm/tensorflow:rocm4.3.1-tf2.6-dev                                          \
+/bin/bash
+```
+
+## Build ```gemm_xdl```
+```bash
+mkdir build && cd build
+```
+
+```bash
+# Need to specify target ID, example below is gfx908
+cmake                                                                  \
+-D BUILD_DEV=OFF                                                       \
+-D CMAKE_BUILD_TYPE=Release                                            \
+-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 "   \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc                              \
+-D CMAKE_PREFIX_PATH=/opt/rocm                                         \
+..
+```
+
+```bash
+ make -j gemm_xdl
+```
+
+## Run ```gemm_xdl```
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: run kernel # of times (>1)
+./example/gemm_xdl 0 1 5
+```
+
+Result (MI100 @ 1087Mhz, 133.5TFlops peak FP16)
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m_k1_{512, 3840, 8}
+arg.b_grid_desc_k0_n_k1_{512, 4096, 8}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up
+Start running 5 times...
+Perf: 1.19685 ms, 107.657 TFlops, 78.8501 GB/s
+```

example/12_grouped_gemm_xdl/grouped_gemm_xdl.cpp

+#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"
+#include "device_grouped_gemm_xdl.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.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   = 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_t::Default;
+// static constexpr auto GemmMNPadding =
+// ck::tensor_operation::device::GemmSpecialization_t::MNPadding;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdl
+//######| 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|         |
+//######|      |      |      |        |        |        |        |            |            |            |              |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |
+        <   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>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AElementOp, BElementOp, CElementOp>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    int nrepeat          = 5;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = 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: run kernel # of times (>1)\n");
+        exit(0);
+    }
+
+    int group_count = 1;
+
+    // GEMM shape
+    std::vector<ck::gemm_desc> gemm_shapes;
+
+    int A_size = 0, B_size = 0, C_size = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        int M = 256;
+        int N = 512;
+        int K = 1024;
+
+        gemm_shapes.push_back({M, N, K, K, K, N, A_size, B_size, C_size});
+
+        A_size += M * K;
+        B_size += N * K;
+        C_size += M * N;
+    }
+
+    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<CDataType>> c_host_tensors;
+    std::vector<Tensor<CDataType>> c_device_tensors;
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].K, gemm_shapes[i].N, gemm_shapes[i].StrideB, BLayout{})));
+        c_host_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+        c_device_tensors.push_back(Tensor<CDataType>(f_host_tensor_descriptor(
+            gemm_shapes[i].M, gemm_shapes[i].N, gemm_shapes[i].StrideC, CLayout{})));
+
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+    }
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        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});
+            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});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    DeviceMem a_tensors_device_buf(sizeof(ADataType) * A_size);
+    DeviceMem b_tensors_device_buf(sizeof(BDataType) * B_size);
+    DeviceMem c_tensors_device_buf(sizeof(CDataType) * C_size);
+
+    std::vector<ADataType> a_tensors_data, b_tensors_data, c_tensors_data;
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        a_tensors_data.insert(
+            a_tensors_data.end(), a_tensors[i].mData.begin(), a_tensors[i].mData.end());
+        b_tensors_data.insert(
+            b_tensors_data.end(), b_tensors[i].mData.begin(), b_tensors[i].mData.end());
+    }
+
+    a_tensors_device_buf.ToDevice(a_tensors_data.data());
+    b_tensors_device_buf.ToDevice(b_tensors_data.data());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_tensors_device_buf.GetDeviceBuffer()),
+                          static_cast<BDataType*>(b_tensors_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_tensors_device_buf.GetDeviceBuffer()),
+                          gemm_shapes,
+                          a_element_op,
+                          b_element_op,
+                          c_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, nrepeat);
+
+    c_tensors_data.resize(C_size);
+
+    c_tensors_device_buf.FromDevice(c_tensors_data.data());
+
+    for(int i = 0; i < gemm_shapes.size(); i++)
+    {
+        memcpy(c_device_tensors[i].mData.data(),
+               c_tensors_data.data() + gemm_shapes[i].OffsetC,
+               c_device_tensors[i].mData.size() * sizeof(CDataType));
+    }
+
+    if(do_verification)
+    {
+        for(int i = 0; i < gemm_shapes.size(); i++)
+        {
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            check_error(c_host_tensors[i], c_device_tensors[i]);
+        }
+    }
+#if 0
+    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{}));
+    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{}));
+
+    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:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+        b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+    }
+
+    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());
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    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,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_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, nrepeat);
+
+    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(CDataType) * 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;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        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);
+
+        check_error(c_m_n_host_result, c_m_n_device_result);
+    }
+#endif
+
+    return 0;
+}

example/CMakeLists.txt

@@ -24,8 +24,10 @@ set(GEMM_XDL_ALPHA_BETA_SOURCE 8_gemm_xdl_alpha_beta/gemm_xdl_alpha_beta.cpp)
 set(CONV2D_FWD_XDL_INT8_SOURCE 9_conv2d_fwd_xdl_int8/conv2d_fwd_xdl_int8.cpp)
 set(CONV3D_FWD_XDL_SOURCE 10_conv3d_fwd_xdl/conv3d_fwd_xdl.cpp)
 set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp)
+set(GROUPED_GEMM_XDL_SOURCE 12_grouped_gemm_xdl/grouped_gemm_xdl.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
+add_executable(grouped_gemm_xdl ${GROUPED_GEMM_XDL_SOURCE})
 add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
 add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
 add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
@@ -38,6 +40,7 @@ add_executable(conv3d_fwd_xdl ${CONV3D_FWD_XDL_SOURCE})
 add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
+target_link_libraries(grouped_gemm_xdl PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)