Merge remote-tracking branch 'origin/develop' into migx-jit-lib

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp

@@ -63,7 +63,8 @@ __global__ void
             const Block2ETileMap block_2_etile_map,
             index_t NRaw)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemmWelford::template Run<HasMainKBlockLoop>(
@@ -806,7 +807,7 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
         // workspace for welford intermediate mean
         workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
 
-        // workspace for welford intermediate mean
+        // workspace for welford intermediate variance
         workspace_size += gemm_welford_size * sizeof(EMeanVarDataType) + 64;
 
         // workspace for welford intermediate count
@@ -854,7 +855,8 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_multiple_r_xdl_cshuffle.hpp

@@ -60,7 +60,8 @@ __global__ void
             const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -554,7 +555,8 @@ struct DeviceGemmMultipleDMultipleR_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp

@@ -273,7 +273,10 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
               N01_{N01},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              cde_element_op_{cde_element_op}
+              cde_element_op_{cde_element_op},
+              MRaw_{M},
+              NRaw_{N},
+              KRaw_{K}
         {
             a_grid_desc_k0_m_k1_ = DeviceOp::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
             b_grid_desc_k0_n_k1_ = DeviceOp::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
@@ -335,6 +338,11 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         CDEElementwiseOperation cde_element_op_;
+
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
     };
 
     // Invoker
@@ -488,6 +496,85 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
         {
             return false;
         }
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector load of Ds
+            // only support RowMajor for now
+            bool all_valid = true;
+
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                if constexpr(!is_same_v<DLayout, Row>)
+                {
+                    all_valid = false;
+                }
+            });
+
+            if(!all_valid)
+            {
+                return false;
+            }
+
+            // check vector store of E
+            // only support RowMajor for now
+            if constexpr(is_same_v<ELayout, Row>)
+            {
+                if(arg.NRaw_ % CDEShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
 
         return GridwiseOp::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                          arg.b_grid_desc_k0_n_k1_,

include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle.hpp

@@ -51,7 +51,8 @@ __global__ void
                                                 e_grid_desc_mblock_mperblock_nblock_nperblock,
                                             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
@@ -490,7 +491,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp

@@ -239,7 +239,10 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
               N01_{N01},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              c_element_op_{c_element_op}
+              c_element_op_{c_element_op},
+              MRaw_{M},
+              NRaw_{N},
+              KRaw_{K}
         {
             a_grid_desc_k0_m_k1_ =
                 DeviceGemmWmma_CShuffle::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
@@ -276,6 +279,10 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
         AElementwiseOperation a_element_op_;
         BElementwiseOperation b_element_op_;
         CElementwiseOperation c_element_op_;
+        // for checking vector load/store
+        index_t MRaw_;
+        index_t NRaw_;
+        index_t KRaw_;
     };
 
     // Invoker
@@ -417,6 +424,68 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
             return false;
         }
 
+        // check vector load/store
+        {
+            using Row = ck::tensor_layout::gemm::RowMajor;
+            using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+            // check vector load of A
+            if constexpr(is_same_v<ALayout, Row> && ABlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<ALayout, Col> && ABlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.MRaw_ % ABlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector laod of B
+            if constexpr(is_same_v<BLayout, Col> && BBlockTransferSrcVectorDim == 2)
+            {
+                if(arg.KRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else if constexpr(is_same_v<BLayout, Row> && BBlockTransferSrcVectorDim == 1)
+            {
+                // FIXME: not rigorous
+                if(arg.NRaw_ % BBlockTransferSrcScalarPerVector != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+
+            // check vector store of C
+            // only support RowMajor for now
+            if constexpr(is_same_v<CLayout, Row>)
+            {
+                if(arg.NRaw_ % CShuffleBlockTransferScalarPerVector_NPerBlock != 0)
+                {
+                    return false;
+                }
+            }
+            else
+            {
+                return false;
+            }
+        }
+
         return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                            arg.b_grid_desc_k0_n_k1_,
                                            arg.c_grid_desc_m_n_,

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp

@@ -428,7 +428,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
                 return false;
             }
         }
-        else if(ck::get_device_name() == "gfx90a")
+        else if(ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
                            is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp

@@ -574,7 +574,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_layernorm_cshuffle.hpp

@@ -648,7 +648,8 @@ struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp

@@ -114,7 +114,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
         CBlockTransferScalarPerVector_NWaveNPerXDL,
         CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
 
-    using Argument = typename GridwiseGemm::Argument;
+    using Argument              = typename GridwiseGemm::Argument;
+    using DefaultBlock2CTileMap = typename GridwiseGemm::DefaultBlock2CTileMap;
 
     // Invoker
     struct Invoker : public BaseInvoker
@@ -138,8 +139,9 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     "setting");
             }
 
+            const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(karg);
+            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0           = karg.K0;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
@@ -152,7 +154,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                         hipMemset(karg.p_c_grid, 0, karg.M * karg.N * sizeof(CDataType)));
 
                 ave_time = launch_and_time_kernel(
-                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
+                    stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg, b2c_map);
             };
 
             if(has_main_k0_block_loop)
@@ -162,7 +164,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              true,
-                                                             InMemoryDataOperationEnum::Set>;
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -171,7 +174,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              true,
-                                                             InMemoryDataOperationEnum::AtomicAdd>;
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -183,7 +187,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              false,
-                                                             InMemoryDataOperationEnum::Set>;
+                                                             InMemoryDataOperationEnum::Set,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }
@@ -192,7 +197,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
                     const auto kernel =
                         kernel_gemm_xdlops_v2r4r2_simplified<GridwiseGemm,
                                                              false,
-                                                             InMemoryDataOperationEnum::AtomicAdd>;
+                                                             InMemoryDataOperationEnum::AtomicAdd,
+                                                             DefaultBlock2CTileMap>;
 
                     Run(kernel);
                 }

include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp

@@ -37,7 +37,8 @@ __global__ void
             const BElementwiseOperation b_element_op,
             const CDEElementwiseOperation cde_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();
@@ -703,7 +704,8 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
+        if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
+             ck::get_device_name() == "gfx940"))
         {
             return false;
         }

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -130,7 +130,8 @@ __global__ void
             const Block2ETileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_xdl_cshuffle.hpp

@@ -78,7 +78,8 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp

@@ -155,7 +155,8 @@ __global__ void
             const Block2ETileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -810,7 +811,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
                 return false;
             }
         }
-        else if(get_device_name() == "gfx90a")
+        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
                            is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp

@@ -135,7 +135,8 @@ __global__ void
             const Block2ETileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -684,7 +685,7 @@ struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
                 return false;
             }
         }
-        else if(get_device_name() == "gfx90a")
+        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940")
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
                            is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))

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

@@ -38,7 +38,8 @@ __global__ void
                                 const BElementwiseOperation b_element_op,
                                 const CDEElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp

@@ -10,8 +10,7 @@
 #include "ck/tensor_operation/gpu/device/device_normalization.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_normalization_selector.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -20,6 +19,10 @@ namespace tensor_operation {
 namespace device {
 
 // Y = Normalization(X, Beta, Gamma)
+// M: Invarient length
+// K: Reduce length (Calculate mean and variance along K dimension)
+// eg. Length = [N, C, H, W], reduce dim = [C, H, W]
+// Then, M = N, K = C * H * W
 template <typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
@@ -68,7 +71,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
 
     static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
                                     const std::vector<index_t>& inStrides,
-                                    int blkGroupSize,
                                     int numBlockTileIteration)
     {
         constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
@@ -117,10 +119,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
         const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
 
-        const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
         const auto inPad_M =
             math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
-        const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
+        const auto inPad_K = K_BlockTileSize * numBlockTileIteration - reduceLength;
 
         auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
             in_grid_desc_m_k,
@@ -132,7 +133,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         return (in_grid_desc_m_k_padded);
     };
 
-    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
 
     struct Argument : public BaseArgument
     {
@@ -162,26 +163,22 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
             gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
             betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
 
-            long_index_t invariant_total_length;
-            long_index_t reduce_total_length;
+            long_index_t invariant_length;
+            long_index_t reduce_length;
 
-            std::tie(invariant_total_length, reduce_total_length) =
+            std::tie(invariant_length, reduce_length) =
                 get_2d_lengths<Rank, NumReduceDim>(Lengths_);
 
-            blkGroupSize_          = 1;
-            numBlockTileIteration_ = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
+            numBlockTileIteration_ = math::integer_divide_ceil(reduce_length, K_BlockTileSize);
 
-            gridSize_ = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
-                        M_BlockTileSize * blkGroupSize_;
+            gridSize_ = math::integer_divide_ceil(invariant_length, M_BlockTileSize);
 
-            x_grid_desc_m_k_ =
-                MakeSrc2dDescriptor(Lengths_, xStrides_, blkGroupSize_, numBlockTileIteration_);
+            x_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, xStrides_, numBlockTileIteration_);
             gamma_grid_desc_m_k_ =
-                MakeSrc2dDescriptor(Lengths_, gammaStrides_, blkGroupSize_, numBlockTileIteration_);
+                MakeSrc2dDescriptor(Lengths_, gammaStrides_, numBlockTileIteration_);
             beta_grid_desc_m_k_ =
-                MakeSrc2dDescriptor(Lengths_, betaStrides_, blkGroupSize_, numBlockTileIteration_);
-            y_grid_desc_m_k_ =
-                MakeSrc2dDescriptor(Lengths_, yStrides_, blkGroupSize_, numBlockTileIteration_);
+                MakeSrc2dDescriptor(Lengths_, betaStrides_, numBlockTileIteration_);
+            y_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, yStrides_, numBlockTileIteration_);
 
             isSweeponce_ =
                 x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
@@ -202,7 +199,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
 
         YElementwiseOperation y_elementwise_op_;
 
-        int blkGroupSize_;
         int numBlockTileIteration_;
         size_t gridSize_;
 
@@ -286,6 +282,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
 
                 if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
                     return false;
+
+                if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
+                    return false;
             };
         }
         else
@@ -295,12 +294,12 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
 
             if(p_arg_->Lengths_[Rank - 1] % XSrcVectorSize != 0)
                 return false;
-        };
 
-        if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
-        {
-            return false;
-        }
+            if(p_arg_->Lengths_[Rank - 1] % YDstVectorSize != 0)
+            {
+                return false;
+            }
+        };
 
         // if fastest dim is not reduced
         if constexpr(GammaSrcVectorDim == 0)

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

@@ -56,6 +56,12 @@ struct PassThrough
         y = type_convert<bhalf_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<bhalf_t, half_t>(bhalf_t& y, const half_t& x) const
+    {
+        y = type_convert<bhalf_t>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
     {
@@ -86,6 +92,23 @@ struct UnaryConvert
     }
 };
 
+struct ConvertBF16RTN
+{
+    // convert to bf16 using round to nearest (rtn)
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        // check Y datatype
+        static_assert(is_same<Y, bhalf_t>::value, "Data type is not supported by this operation!");
+
+        // check X datatype
+        static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = bf16_convert_rtn<Y>(x);
+    }
+};
+
 struct Scale
 {
     __host__ __device__ Scale(float scale) : scale_(scale) {}

include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp

@@ -587,4 +587,52 @@ struct OffsettedBlockToCTileMap
     index_t block_start_;
 };
 
+/**
+ * @brief      Simple tile mapping which creates 3D grid of block of threads.
+ *
+ * @paragraph  Description
+ *             This Block-to-C-tile-map creates a 3D grid (n_blocks, m_blocks, z_blocks) of thread
+ *             blocks. The first 2D are regular 2D tiles created by division of output GEMM
+ *             dimenions by corresponding tile size. The third dimension (Z) is a k-split dimension,
+ *             which denotes the number of blocks we use to divide work on GEMM K dimension onto.
+ *
+ * @tparam     MPerBlock  Output block tile size in M dimension.
+ * @tparam     NPerBlock  Output block tile size in N dimension.
+ */
+template <index_t MPerBlock, index_t NPerBlock>
+struct BlockToCTileMap_3DGrid_KSplit
+{
+
+    __host__ __device__ BlockToCTileMap_3DGrid_KSplit() = default;
+
+    __host__ __device__ constexpr auto
+    CalculateGridSize(index_t M, index_t N, index_t k_split) const
+    {
+        // Create 3D grid
+        const auto M0 = math::integer_divide_ceil(M, MPerBlock);
+        const auto N0 = math::integer_divide_ceil(N, NPerBlock);
+
+        return std::make_tuple(N0, M0, k_split);
+    }
+
+    template <typename TopIdx>
+    __device__ constexpr auto CalculateBottomIndex(const TopIdx&) const
+    {
+        return make_tuple(blockIdx.z, blockIdx.y, blockIdx.x);
+    }
+
+    template <typename CTileIdx, typename CTileDim>
+    __host__ __device__ bool ValidCTileIndex(const CTileIdx& /* c_tile_idx */,
+                                             const CTileDim& /* c_tile_dim */) const
+    {
+        return true; // always valid provided that user gets grid size from CalculateGridSize()
+    }
+
+    template <typename CGridDesc_M_N>
+    __host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
+    {
+        return true;
+    }
+};
+
 } // namespace ck