Multi-kernel CGEMM (#230)

* Reference CGEMM + test stub

* Format.

* Incomplete simple implementation

* Library instances

* Sketch of tests

* Test fixes.

* Example added

* Cosmetics

* Add elementwise operation kernel and example

* Add comment

* Add template argument of dim . Prepare to support multiple dimension

* Rename example

* Support 1 dimension

* Add static assert

* Add comment

* Second auxiliary buffer added

* Extract pad

* Remove redundant argument

* Support any dimension for elementwise operation

* Remove line

* Let it be the multiple number of CU

* Move thread per block to the parameter of constructor

* Consuming binary ops to do A+B / A-B

* Fix + cosmetics + bf16 test commented out temporarily

* Format

* Enabling bf16 test

* Revert "Enabling bf16 test"

This reverts commit f497e2ba.

* Fix + test reenabled

* fix build

* Revert "fix build"

This reverts commit d7310238

.

* post PR #235 merge fix

* amend

* Single workspace for cgemm + helper

* Perf calc fix

* Review remarks: static_cast

* Review remarks: binary ops templated

* Cleaning

* Removal of instances and their tests

* Review remarks from aosew addressed

* Review remark: unnecessary attribute

* Post-merge fixes

* Restrict 4gemm to PassThrough + bug fix

* Review remarks

* update licence

* change cgemm example to fp16
Co-authored-by: rocking <chunylai@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>
Co-authored-by: Anthony Chang <ac.chang@outlook.com>

example/19_binary_elementwise/broadcast_add_2d_amn_bn.cpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
 #include <iostream>
 #include <cstdlib>
 #include "check_err.hpp"
@@ -17,7 +42,8 @@ using ABDataType             = F16;
 using CDataType              = F16;
 using EltwiseComputeDataType = F32;
 
-using Add = ck::tensor_operation::binary_element_wise::Add;
+using Add = ck::tensor_operation::binary_element_wise::
+    Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
 
 using DeviceElementwiseAddInstance =
     ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
@@ -46,19 +72,19 @@ void host_broadcast2D(
     {
         for(int n = 0; n < N; ++n)
         {
-            ComputeDataType Amn = static_cast<ComputeDataType>(A(m, n));
+            ComputeDataType Amn = ck::type_convert<ComputeDataType>(A(m, n));
             ComputeDataType Cmn = 0;
             if constexpr(broadcastDim == 0)
             {
-                ComputeDataType Bn = static_cast<ComputeDataType>(B(n));
+                ComputeDataType Bn = ck::type_convert<ComputeDataType>(B(n));
                 functor(Cmn, Amn, Bn);
             }
             else
             {
-                ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+                ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
                 functor(Cmn, Amn, Bm);
             }
-            C(m, n) = static_cast<ctype>(Cmn);
+            C(m, n) = ck::type_convert<ctype>(Cmn);
         }
     }
 }

example/19_binary_elementwise/broadcast_add_3d_am_bmnk.cpp

@@ -17,7 +17,8 @@ using ABDataType             = F16;
 using CDataType              = F16;
 using EltwiseComputeDataType = F32;
 
-using Add = ck::tensor_operation::binary_element_wise::Add;
+using Add = ck::tensor_operation::binary_element_wise::
+    Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
 
 using DeviceElementwiseAddInstance =
     ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
@@ -48,11 +49,11 @@ void host_broadcast3D_am_bmnk(HostTensorC& C,
         for(std::size_t n = 0; n < shape[1]; ++n)
             for(std::size_t k = 0; k < shape[2]; ++k)
             {
-                ComputeDataType a_val = static_cast<ComputeDataType>(A(m));
-                ComputeDataType b_val = static_cast<ComputeDataType>(B(m, n, k));
+                ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(m));
+                ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(m, n, k));
                 ComputeDataType c_val = 0;
                 functor(c_val, a_val, b_val);
-                C(m, n, k) = static_cast<ctype>(c_val);
+                C(m, n, k) = ck::type_convert<ctype>(c_val);
             }
 }
 

example/19_binary_elementwise/elementwise_add_1d.cpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
 #include <iostream>
 #include <cstdlib>
 #include "check_err.hpp"
@@ -17,7 +42,8 @@ using ABDataType             = F16;
 using CDataType              = F16;
 using EltwiseComputeDataType = F32;
 
-using Add = ck::tensor_operation::binary_element_wise::Add;
+using Add = ck::tensor_operation::binary_element_wise::
+    Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
 
 using DeviceElementwiseAddInstance =
     ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
@@ -43,11 +69,11 @@ void host_elementwise1D(
 
     for(int m = 0; m < M; ++m)
     {
-        ComputeDataType Am = static_cast<ComputeDataType>(A(m));
-        ComputeDataType Bm = static_cast<ComputeDataType>(B(m));
+        ComputeDataType Am = ck::type_convert<ComputeDataType>(A(m));
+        ComputeDataType Bm = ck::type_convert<ComputeDataType>(B(m));
         ComputeDataType Cm = 0;
         functor(Cm, Am, Bm);
-        C(m) = static_cast<ctype>(Cm);
+        C(m) = ck::type_convert<ctype>(Cm);
     }
 }
 

example/19_binary_elementwise/elementwise_add_4d.cpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
 #include <iostream>
 #include <cstdlib>
 #include "check_err.hpp"
@@ -17,7 +42,8 @@ using ABDataType             = F16;
 using CDataType              = F16;
 using EltwiseComputeDataType = F32;
 
-using Add = ck::tensor_operation::binary_element_wise::Add;
+using Add = ck::tensor_operation::binary_element_wise::
+    Add<EltwiseComputeDataType, EltwiseComputeDataType, EltwiseComputeDataType>;
 
 using DeviceElementwiseAddInstance =
     ck::tensor_operation::device::DeviceBinaryElementwise<ABDataType,
@@ -49,11 +75,11 @@ void host_elementwise4D(HostTensorC& C,
             for(std::size_t h = 0; h < shape[2]; ++h)
                 for(std::size_t w = 0; w < shape[3]; ++w)
                 {
-                    ComputeDataType a_val = static_cast<ComputeDataType>(A(n, c, h, w));
-                    ComputeDataType b_val = static_cast<ComputeDataType>(B(n, c, h, w));
+                    ComputeDataType a_val = ck::type_convert<ComputeDataType>(A(n, c, h, w));
+                    ComputeDataType b_val = ck::type_convert<ComputeDataType>(B(n, c, h, w));
                     ComputeDataType c_val = 0;
                     functor(c_val, a_val, b_val);
-                    C(n, c, h, w) = static_cast<ctype>(c_val);
+                    C(n, c, h, w) = ck::type_convert<ctype>(c_val);
                 }
 }
 

example/22_cgemm/CMakeLists.txt

+add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)

example/22_cgemm/cgemm_xdl_fp16.cpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.hpp>
+
+#include "check_err.hpp"
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "device_tensor.hpp"
+#include "device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_cgemm.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   = F16;
+using BDataType   = F16;
+using CDataType   = F16;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: run kernel # of times (>1)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        exit(0);
+    }
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_real_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_imag_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
+    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
+    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
+    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
+    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
+    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    default:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    auto cgemm = DeviceCGemmInstance{};
+
+    DeviceMem a_m_k_real_device_buf(sizeof(ADataType) * a_m_k_real.mDesc.GetElementSpace());
+    DeviceMem a_m_k_imag_device_buf(sizeof(ADataType) * a_m_k_imag.mDesc.GetElementSpace());
+    DeviceMem b_k_n_real_device_buf(sizeof(BDataType) * b_k_n_real.mDesc.GetElementSpace());
+    DeviceMem b_k_n_imag_device_buf(sizeof(BDataType) * b_k_n_imag.mDesc.GetElementSpace());
+    DeviceMem c_m_n_real_device_buf(sizeof(CDataType) *
+                                    c_m_n_real_device_result.mDesc.GetElementSpace());
+    DeviceMem c_m_n_imag_device_buf(sizeof(CDataType) *
+                                    c_m_n_imag_device_result.mDesc.GetElementSpace());
+    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
+
+    a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
+    a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
+    b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
+    b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    // do GEMM
+    auto invoker = cgemm.MakeInvoker();
+    auto argument =
+        cgemm.MakeArgument(static_cast<ADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
+                           static_cast<ADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(workspace_device_buf.GetDeviceBuffer()),
+                           M,
+                           N,
+                           K,
+                           StrideA,
+                           StrideB,
+                           StrideC,
+                           a_element_op,
+                           b_element_op,
+                           c_element_op);
+
+    if(!cgemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_cgemm with the specified compilation parameters does "
+            "not support this CGEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(8) * M * N * K;
+    std::size_t num_btype =
+        std::size_t(2) *
+        (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, "
+              << cgemm.GetTypeString() << std::endl;
+
+    c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
+    c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CDataType> c_m_n_real_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_imag_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_cgemm   = ReferenceCGemmInstance{};
+        auto ref_invoker = ref_cgemm.MakeInvoker();
+
+        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
+                                                   a_m_k_imag,
+                                                   b_k_n_real,
+                                                   b_k_n_imag,
+                                                   c_m_n_real_host_result,
+                                                   c_m_n_imag_host_result,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        ck::utils::check_err(c_m_n_real_device_result.mData,
+                             c_m_n_real_host_result.mData,
+                             "Verification error: incorrect results in real part!",
+                             1e-2f,
+                             1e-1f);
+        ck::utils::check_err(c_m_n_imag_device_result.mData,
+                             c_m_n_imag_host_result.mData,
+                             "Verification error: incorrect results in imaginary part!",
+                             1e-2f,
+                             1e-1f);
+    }
+
+    return 0;
+}

example/CMakeLists.txt

@@ -48,10 +48,11 @@ add_subdirectory(11_conv2d_bwd_weight)
 add_subdirectory(12_reduce)
 add_subdirectory(13_pool2d_fwd)
 add_subdirectory(14_gemm_xdl_requant_relu_requant)
-add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(15_grouped_gemm)
 add_subdirectory(16_gemm_reduce)
+add_subdirectory(17_convnd_bwd_data_xdl)
 add_subdirectory(18_batched_gemm_reduce)
 add_subdirectory(19_binary_elementwise)
 add_subdirectory(20_convnd_bwd_weight_xdl)
 add_subdirectory(21_gemm_layernorm)
+add_subdirectory(22_cgemm)

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

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#pragma once
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceCGemm : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a_real,
+                                                              const void* p_a_imag,
+                                                              const void* p_b_real,
+                                                              const void* p_b_imag,
+                                                              void* p_c_real,
+                                                              void* p_c_imag,
+                                                              void* p_workspace,
+                                                              ck::index_t M,
+                                                              ck::index_t N,
+                                                              ck::index_t K,
+                                                              ck::index_t StrideA,
+                                                              ck::index_t StrideB,
+                                                              ck::index_t StrideC,
+                                                              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;
+    virtual std::size_t GetWorkspaceSize(index_t MRaw,
+                                         index_t NRaw,
+                                         index_t KRaw,
+                                         index_t StrideA,
+                                         index_t StrideB,
+                                         index_t StrideC)     = 0;
+};
+
+template <typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+using DeviceCGemmPtr = std::unique_ptr<
+    DeviceCGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -60,8 +60,8 @@ template <
     index_t CThreadTransferDstScalarPerVector,
     enable_if_t<
         is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
-            is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
-            is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
         bool> = false>
 struct DeviceGemmDl
     : public DeviceGemm<AElementwiseOperation, BElementwiseOperation, CElementwiseOperation>

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
 #pragma once
 #include "data_type.hpp"
 
@@ -5,14 +30,22 @@ namespace ck {
 namespace tensor_operation {
 namespace binary_element_wise {
 
-struct Add
+template <typename Y, typename X1, typename X2>
+struct Add;
+
+template <>
+struct Add<double, double, double>
 {
     __host__ __device__ constexpr void
     operator()(double& dst, const double& src1, const double& src2) const
     {
         dst = src1 + src2;
     }
+};
 
+template <>
+struct Add<float, float, float>
+{
     __host__ __device__ constexpr void
     operator()(float& dst, const float& src1, const float& src2) const
     {
@@ -20,6 +53,75 @@ struct Add
     }
 };
 
+template <>
+struct Add<half_t, half_t, half_t>
+{
+    __host__ __device__ constexpr void
+    operator()(half_t& dst, const half_t& src1, const half_t& src2) const
+    {
+        dst = src1 + src2;
+    }
+};
+
+template <>
+struct Add<bhalf_t, bhalf_t, bhalf_t>
+{
+    __host__ __device__ constexpr void
+    operator()(bhalf_t& dst, const bhalf_t& src1, const bhalf_t& src2) const
+    {
+        const float x1 = ck::type_convert<float>(src1);
+        const float x2 = ck::type_convert<float>(src2);
+        const float y  = x1 + x2;
+        dst            = ck::type_convert<bhalf_t>(y);
+    }
+};
+
+template <typename Y, typename X1, typename X2>
+struct Substract;
+
+template <>
+struct Substract<double, double, double>
+{
+    __host__ __device__ constexpr void
+    operator()(double& dst, const double& src1, const double& src2) const
+    {
+        dst = src1 - src2;
+    }
+};
+
+template <>
+struct Substract<float, float, float>
+{
+    __host__ __device__ constexpr void
+    operator()(float& dst, const float& src1, const float& src2) const
+    {
+        dst = src1 - src2;
+    }
+};
+
+template <>
+struct Substract<half_t, half_t, half_t>
+{
+    __host__ __device__ constexpr void
+    operator()(half_t& dst, const half_t& src1, const half_t& src2) const
+    {
+        dst = src1 - src2;
+    }
+};
+
+template <>
+struct Substract<bhalf_t, bhalf_t, bhalf_t>
+{
+    __host__ __device__ constexpr void
+    operator()(bhalf_t& dst, const bhalf_t& src1, const bhalf_t& src2) const
+    {
+        const float x1 = ck::type_convert<float>(src1);
+        const float x2 = ck::type_convert<float>(src2);
+        const float y  = x1 - x2;
+        dst            = ck::type_convert<bhalf_t>(y);
+    }
+};
+
 } // namespace binary_element_wise
 } // namespace tensor_operation
 } // namespace ck

library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2022 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#pragma once
+#include <iostream>
+#include <sstream>
+#include "device_base.hpp"
+#include "host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+// FIXME: support arbitrary elementwise operation for A/B/C
+template <
+    typename ADataType,
+    typename BDataType,
+    typename CDataType,
+    typename AElementwiseOperation,
+    typename BElementwiseOperation,
+    typename CElementwiseOperation,
+    enable_if_t<
+        is_same_v<AElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<BElementwiseOperation, ck::tensor_operation::element_wise::PassThrough> &&
+            is_same_v<CElementwiseOperation, ck::tensor_operation::element_wise::PassThrough>,
+        bool> = false>
+struct ReferenceCGemm : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        Argument(const Tensor<ADataType>& a_m_k_real,
+                 const Tensor<ADataType>& a_m_k_imag,
+                 const Tensor<BDataType>& b_k_n_real,
+                 const Tensor<BDataType>& b_k_n_imag,
+                 Tensor<CDataType>& c_m_n_real,
+                 Tensor<CDataType>& c_m_n_imag,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+            : a_m_k_real_{a_m_k_real},
+              a_m_k_imag_{a_m_k_imag},
+              b_k_n_real_{b_k_n_real},
+              b_k_n_imag_{b_k_n_imag},
+              c_m_n_real_{c_m_n_real},
+              c_m_n_imag_{c_m_n_imag},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              c_element_op_{c_element_op}
+        {
+        }
+
+        const Tensor<ADataType>& a_m_k_real_;
+        const Tensor<ADataType>& a_m_k_imag_;
+        const Tensor<BDataType>& b_k_n_real_;
+        const Tensor<BDataType>& b_k_n_imag_;
+        Tensor<CDataType>& c_m_n_real_;
+        Tensor<CDataType>& c_m_n_imag_;
+
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CElementwiseOperation c_element_op_;
+    };
+
+    // Invoker
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceCGemm::Argument;
+
+        float Run(const Argument& arg)
+        {
+            const std::size_t K = arg.a_m_k_real_.mDesc.GetLengths()[1];
+
+            if(K != arg.a_m_k_imag_.mDesc.GetLengths()[1])
+            {
+                throw std::runtime_error("wrong! Incompatible real and imag sizes in CGEMM");
+            }
+
+            auto f_mk_kn_mn_real = [&](auto m, auto n) {
+                float v_c_real = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_real += v_a_real * v_b_real - v_a_imag * v_b_imag;
+                }
+
+                arg.c_m_n_real_(m, n) = v_c_real;
+            };
+
+            auto f_mk_kn_mn_imag = [&](auto m, auto n) {
+                float v_c_imag = 0;
+
+                for(std::size_t k = 0; k < K; ++k)
+                {
+                    float v_a_real = ck::type_convert<float>(arg.a_m_k_real_(m, k));
+                    float v_a_imag = ck::type_convert<float>(arg.a_m_k_imag_(m, k));
+                    float v_b_real = ck::type_convert<float>(arg.b_k_n_real_(k, n));
+                    float v_b_imag = ck::type_convert<float>(arg.b_k_n_imag_(k, n));
+
+                    v_c_imag += v_a_real * v_b_imag + v_a_imag * v_b_real;
+                }
+
+                arg.c_m_n_imag_(m, n) = v_c_imag;
+            };
+
+            make_ParallelTensorFunctor(f_mk_kn_mn_real,
+                                       arg.c_m_n_real_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_real_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+            make_ParallelTensorFunctor(f_mk_kn_mn_imag,
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[0],
+                                       arg.c_m_n_imag_.mDesc.GetLengths()[1])(
+                std::thread::hardware_concurrency());
+
+            return 0;
+        }
+
+        float Run(const 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 device::BaseArgument*) override { return true; }
+
+    static auto MakeArgument(const Tensor<ADataType>& a_m_k_real,
+                             const Tensor<ADataType>& a_m_k_imag,
+                             const Tensor<BDataType>& b_k_n_real,
+                             const Tensor<BDataType>& b_k_n_imag,
+                             Tensor<CDataType>& c_m_n_real,
+                             Tensor<CDataType>& c_m_n_imag,
+                             AElementwiseOperation a_element_op,
+                             BElementwiseOperation b_element_op,
+                             CElementwiseOperation c_element_op)
+    {
+        return Argument{a_m_k_real,
+                        a_m_k_imag,
+                        b_k_n_real,
+                        b_k_n_imag,
+                        c_m_n_real,
+                        c_m_n_imag,
+                        a_element_op,
+                        b_element_op,
+                        c_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceCGemm"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck