From 959defbb22c61beb7be961778f63af3890bbfb3b Mon Sep 17 00:00:00 2001
From: Anton Gorenko <anton@streamhpc.com>
Date: Mon, 2 Jun 2025 15:51:12 +0500
Subject: [PATCH] Add a special wrapper to use
 DeviceGemmMultipleD_Wmma_CShuffleV3 with old API

ckProfiler uses DeviceGemmMultipleD (tests also call its functions), the wrapper allows to use
DeviceGemmMultipleDSplitK instances there.
---
 .../gpu/device/device_gemm_multiple_d.hpp     | 103 +++++++++++++++++-
 .../gpu/gemm_add_fastgelu.hpp                 |  21 ++++
 2 files changed, 123 insertions(+), 1 deletion(-)

diff --git a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
index ef0b5286ac..3dff1b28c6 100644
--- a/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
+++ b/include/ck/tensor_operation/gpu/device/device_gemm_multiple_d.hpp
@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 #ifndef __HIPCC_RTC__
@@ -149,6 +149,107 @@ struct DeviceGemmMultipleDSplitKBPreShuffle : public BaseOperator
 #endif
 };
 
+/// @brief Wrapper for backward compatibility that allows to use instances of
+///        DeviceGemmMultipleDSplitK in contexts where DeviceGemmMultipleD is expected.
+///
+/// @note  The main area where it can be used is DeviceOperationInstanceFactory::GetInstances().
+///        The only difference between API of DeviceGemmMultipleD and DeviceGemmMultipleDSplitK is
+///        that DeviceGemmMultipleDSplitK::MakeArgumentPointer requires an additional parameter
+///        KBatch which is explicitly passed as 1 by this wrapper.
+template <typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGemmMultipleDSplitKWrapper : public DeviceGemmMultipleD<ALayout,
+                                                                     BLayout,
+                                                                     DsLayout,
+                                                                     ELayout,
+                                                                     ADataType,
+                                                                     BDataType,
+                                                                     DsDataType,
+                                                                     EDataType,
+                                                                     AElementwiseOperation,
+                                                                     BElementwiseOperation,
+                                                                     CDEElementwiseOperation>
+{
+    using DeviceOp = DeviceGemmMultipleDSplitK<ALayout,
+                                               BLayout,
+                                               DsLayout,
+                                               ELayout,
+                                               ADataType,
+                                               BDataType,
+                                               DsDataType,
+                                               EDataType,
+                                               AElementwiseOperation,
+                                               BElementwiseOperation,
+                                               CDEElementwiseOperation>;
+
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+#ifndef __HIPCC_RTC__
+
+    explicit DeviceGemmMultipleDSplitKWrapper(std::unique_ptr<DeviceOp> p_op)
+        : p_op_(std::move(p_op))
+    {
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return p_op_->IsSupportedArgument(p_arg);
+    }
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        std::array<const void*, NumDTensor> p_ds,
+                        void* p_e,
+                        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,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CDEElementwiseOperation cde_element_op) override
+    {
+        return p_op_->MakeArgumentPointer(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          M,
+                                          N,
+                                          K,
+                                          StrideA,
+                                          StrideB,
+                                          StrideDs,
+                                          StrideE,
+                                          1, // KBatch
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return p_op_->MakeInvokerPointer();
+    }
+
+    std::string GetTypeString() const override { return p_op_->GetTypeString(); }
+
+    private:
+    std::unique_ptr<DeviceOp> p_op_;
+
+#endif // __HIPCC_RTC__
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck
diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
index bc92422583..c93e609b7a 100644
--- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
+++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp
@@ -274,6 +274,27 @@ struct DeviceOperationInstanceFactory<
 #endif // CK_ENABLE_FP16
 #endif // CK_USE_XDL
 
+#if defined(CK_USE_WMMA)
+        // Reuse DeviceGemmMultipleDSplitK instances
+        using Wrapper = DeviceGemmMultipleDSplitKWrapper<ALayout,
+                                                         BLayout,
+                                                         ck::Tuple<D0Layout>,
+                                                         ELayout,
+                                                         ADataType,
+                                                         BDataType,
+                                                         ck::Tuple<D0DataType>,
+                                                         EDataType,
+                                                         PassThrough,
+                                                         PassThrough,
+                                                         AddFastGelu>;
+        auto new_op_ptrs =
+            DeviceOperationInstanceFactory<typename Wrapper::DeviceOp>::GetInstances();
+        for(auto& op_ptr : new_op_ptrs)
+        {
+            op_ptrs.emplace_back(std::make_unique<Wrapper>(std::move(op_ptr)));
+        }
+#endif // CK_USE_WMMA
+
         return op_ptrs;
     }
 };