add an example of customized type convert - bfp16_rtn (#869)

* add an example of customized bfp16_rtn

* fixed threadwise_copy

---------
Co-authored-by: Jing Zhang <jizha@amd.com>

example/01_gemm/CMakeLists.txt

@@ -40,6 +40,9 @@ endif()
 if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
   add_example_executable(example_gemm_xdl_bf16 gemm_xdl_bf16.cpp)
   add_dependencies(example_gemm_xdl example_gemm_xdl_bf16)
+
+  add_example_executable(example_gemm_xdl_bf16_rtn gemm_xdl_bf16_rtn.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_bf16_rtn)
 endif()
 
 if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)

example/01_gemm/gemm_xdl_bf16_rtn.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/utility/type_convert.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = float;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = ck::tensor_operation::element_wise::ConvertBF16RTN;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
+// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           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|        Type|         DataType| 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, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    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 ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+#include "run_gemm_example.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

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

@@ -39,6 +39,12 @@ struct PassThrough
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<half_t, float>(half_t& y, const float& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
     {

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp

@@ -104,13 +104,13 @@ struct ThreadwiseTensorSliceTransfer_v6r1
 
             // apply pointwise operation
             static_for<0, ScalarPerVector, 1>{}([&](auto i) {
-                SrcData v;
+                DstData v;
 
                 // apply element-wise operation
                 element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
 
                 // apply type convert
-                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+                dst_vector_container.template AsType<DstData>()(i) = v;
             });
 
             const bool is_dst_valid =

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

@@ -92,11 +92,11 @@ struct ReferenceGemm : public device::BaseOperator
                         ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
                 }
 
-                AccDataType v_c;
+                CDataType v_c;
 
                 arg.c_element_op_(v_c, v_acc);
 
-                arg.c_m_n_(m, n) = ck::type_convert<CDataType>(v_c);
+                arg.c_m_n_(m, n) = v_c;
             };
 
             make_ParallelTensorFunctor(