cmake list

example/63_layernorm4d_fwd/run_layernorm4d_fwd_example.inc

@@ -46,9 +46,9 @@ int run_layernorm4d_fwd_example()
         {0, W * C, C, 1},
         std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
-                                    save_mean.mDesc.GetStrides().end()},
+                                 save_mean.mDesc.GetStrides().end()},
         {1, 2, 3},
         1e-4,
         x_dev.GetDeviceBuffer(),

include/ck/ck.hpp

@@ -134,6 +134,12 @@
 // inner product using V_DOT with DPP8 modifiers
 #define CK_USE_AMD_V_DOT_DPP8_INLINE_ASM 1
 
+// LDS direct loads using inline assembly
+#define CK_USE_AMD_LDS_DIRECT_LOAD_INLINE_ASM 1
+
+// set stochastic rounding as default for f8 conversions
+#define CK_USE_SR_F8_CONVERSION 1
+
 // block synchronization only s_wait lgkmcnt(0), not vmcnt(0)
 #define CK_EXPERIMENTAL_BLOCK_SYNC_LDS_WITHOUT_SYNC_VMEM 1
 

include/ck/host_utility/device_prop.hpp

@@ -26,7 +26,7 @@ inline std::string get_device_name()
     }
     const std::string raw_name(props.gcnArchName);
 
-    // https://github.com/ROCmSoftwarePlatform/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
+    // https://github.com/ROCm/MIOpen/blob/8498875aef84878e04c1eabefdf6571514891086/src/target_properties.cpp#L40
     static std::map<std::string, std::string> device_name_map = {
         {"Ellesmere", "gfx803"},
         {"Baffin", "gfx803"},

include/ck/stream_config.hpp

@@ -11,6 +11,6 @@ struct StreamConfig
     hipStream_t stream_id_ = nullptr;
     bool time_kernel_      = false;
     int log_level_         = 0;
-    int cold_niters_       = 50;
-    int nrepeat_           = 200;
+    int cold_niters_       = 1;
+    int nrepeat_           = 10;
 };

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

@@ -59,7 +59,9 @@ struct BaseOperator
 
     virtual size_t GetWorkSpaceSize(const BaseArgument*) const { return 0; }
 
-    virtual void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const
+    virtual void SetWorkSpacePointer(BaseArgument* p_arg,
+                                     void* p_workspace,
+                                     const StreamConfig& = StreamConfig{}) const
     {
         assert(p_arg);
         p_arg->p_workspace_ = p_workspace;

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

@@ -376,7 +376,9 @@ struct DeviceBatchNormBwdImpl : public DeviceBatchNormBwd<XDataType,
         return (workspace_size);
     };
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -354,7 +354,9 @@ struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
         return (workspace_size);
     };
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -345,7 +345,9 @@ struct DeviceBatchNormFwdImpl : public DeviceBatchNormFwd<XDataType,
         return (workspace_size);
     };
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -14,6 +14,7 @@
 #include "ck/tensor_operation/gpu/device/device_contraction_multiple_abd.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
@@ -500,22 +501,29 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
             // for sanity check of vector memory access
             for(index_t i = 0; i < NumATensor; ++i)
             {
-                a_mz_stride_[i] = a_ms_ks_strides[i][NumDimM - 1];
-                a_kz_stride_[i] = a_ms_ks_strides[i][NumDimM + NumDimK - 1];
+                as_mz_consecutive_[i] = a_ms_ks_strides[i][NumDimM - 1] == 1;
+                as_kz_consecutive_[i] = a_ms_ks_strides[i][NumDimM + NumDimK - 1] == 1;
+                as_max_read_elems_[i] =
+                    CalculateMaxRead<NumDimM, NumDimK>(a_ms_ks_lengths[i], a_ms_ks_strides[i]);
             }
 
             for(index_t i = 0; i < NumBTensor; ++i)
             {
-                b_nz_stride_[i] = b_ns_ks_strides[i][NumDimN - 1];
-                b_kz_stride_[i] = b_ns_ks_strides[i][NumDimN + NumDimK - 1];
+                bs_nz_consecutive_[i] = b_ns_ks_strides[i][NumDimN - 1] == 1;
+                bs_kz_consecutive_[i] = b_ns_ks_strides[i][NumDimN + NumDimK - 1] == 1;
+                bs_max_read_elems_[i] =
+                    CalculateMaxRead<NumDimN, NumDimK>(b_ns_ks_lengths[i], b_ns_ks_strides[i]);
             }
 
             for(index_t i = 0; i < NumDTensor; ++i)
             {
-                ds_nz_stride_[i] = d_ms_ns_strides[i][NumDimM + NumDimN - 1];
+                ds_nz_consecutive_[i] = d_ms_ns_strides[i][NumDimM + NumDimN - 1] == 1;
+                ds_max_read_elems_[i] =
+                    CalculateMaxRead<NumDimM, NumDimN>(d_ms_ns_lengths[i], d_ms_ns_strides[i]);
             }
 
-            e_nz_stride_ = e_ms_ns_stride[NumDimM + NumDimN - 1];
+            e_nz_consecutive_  = e_ms_ns_stride[NumDimM + NumDimN - 1] == 1;
+            e_max_write_elems_ = CalculateMaxRead<NumDimM, NumDimN>(e_ms_ns_length, e_ms_ns_stride);
         }
 
         // pointers
@@ -545,16 +553,19 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
         BElementwiseOperation b_element_op_;
         CDEElementwiseOperation cde_element_op_;
 
-        // Strides for the last M/N/K dimensions of A/B/Ds/E
-        //   for sanity check of vector load/store
-        std::array<index_t, NumATensor> a_mz_stride_;
-        std::array<index_t, NumATensor> a_kz_stride_;
-
-        std::array<index_t, NumBTensor> b_nz_stride_;
-        std::array<index_t, NumBTensor> b_kz_stride_;
-
-        std::array<index_t, NumDTensor> ds_nz_stride_;
-        index_t e_nz_stride_;
+        // Describe whether the last part of a given dimension of A/B/D/E is consecutive
+        // in the memory or not.
+        std::array<bool, NumATensor> as_mz_consecutive_;
+        std::array<bool, NumATensor> as_kz_consecutive_;
+        std::array<bool, NumBTensor> bs_nz_consecutive_;
+        std::array<bool, NumBTensor> bs_kz_consecutive_;
+        std::array<bool, NumDTensor> ds_nz_consecutive_;
+        bool e_nz_consecutive_;
+
+        std::array<index_t, NumATensor> as_max_read_elems_;
+        std::array<index_t, NumBTensor> bs_max_read_elems_;
+        std::array<index_t, NumDTensor> ds_max_read_elems_;
+        index_t e_max_write_elems_;
     };
 
     // Invoker
@@ -643,73 +654,65 @@ struct DeviceContractionMultipleABD_Xdl_CShuffle
 
         // check vector load/store
         {
-            bool all_valid = true;
-
+            bool valid_as_access = true;
             static_for<0, NumATensor, 1>{}([&](auto i) {
-                // vector memory access of A: could be on M or AK1 dimension
-                if constexpr(ABlockTransferSrcVectorDim == 1)
-                {
-                    if(!(arg.a_mz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I1) %
-                                                             ABlockTransferSrcScalarPerVector ==
-                                                         0))
-                    {
-                        all_valid = false;
-                    }
-                }
-                else
+                const bool valid_a_vector_size =
+                    arg.as_max_read_elems_[i] % ABlockTransferSrcScalarPerVector == 0;
+                const bool valid_a_access_dim_m =
+                    ABlockTransferSrcVectorDim == 1 && arg.as_mz_consecutive_[i];
+                const bool valid_a_access_dim_k =
+                    ABlockTransferSrcVectorDim == 2 && arg.as_kz_consecutive_[i];
+                const bool valid_a_access_dim = valid_a_access_dim_m || valid_a_access_dim_k;
+                if(!(valid_a_vector_size && valid_a_access_dim))
                 {
-                    if(!(arg.a_kz_stride_[i] == 1 && arg.as_grid_desc_ak0_m_ak1_[i].GetLength(I2) %
-                                                             ABlockTransferSrcScalarPerVector ==
-                                                         0))
-                    {
-                        all_valid = false;
-                    }
+                    valid_as_access = false;
                 }
             });
+            if(!valid_as_access)
+            {
+                return false;
+            }
 
-            // vector memory access of B: could be on N or BK1 dimension
+            bool valid_bs_access = true;
             static_for<0, NumBTensor, 1>{}([&](auto i) {
-                if constexpr(BBlockTransferSrcVectorDim == 1)
+                const bool valid_b_vector_size =
+                    arg.bs_max_read_elems_[i] % BBlockTransferSrcScalarPerVector == 0;
+                const bool valid_b_access_dim_n =
+                    BBlockTransferSrcVectorDim == 1 && arg.bs_nz_consecutive_[i];
+                const bool valid_b_access_dim_k =
+                    BBlockTransferSrcVectorDim == 2 && arg.bs_kz_consecutive_[i];
+                const bool valid_b_access_dim = valid_b_access_dim_n || valid_b_access_dim_k;
+                if(!(valid_b_vector_size && valid_b_access_dim))
                 {
-                    if(!(arg.b_nz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I1) %
-                                                             BBlockTransferSrcScalarPerVector ==
-                                                         0))
-                    {
-                        all_valid = false;
-                    }
-                }
-                else
-                {
-                    if(!(arg.b_kz_stride_[i] == 1 && arg.bs_grid_desc_bk0_n_bk1_[i].GetLength(I2) %
-                                                             BBlockTransferSrcScalarPerVector ==
-                                                         0))
-                    {
-                        all_valid = false;
-                    }
+                    valid_bs_access = false;
                 }
             });
+            if(!valid_bs_access)
+            {
+                return false;
+            }
 
-            // check vector load of Ds
+            bool valid_ds_access = true;
             static_for<0, NumDTensor, 1>{}([&](auto i) {
-                if(!(arg.ds_nz_stride_[i] == 1 &&
-                     arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
-                             CDEBlockTransferScalarPerVector_NPerBlock ==
-                         0))
+                const bool valid_d_vector_size =
+                    arg.ds_max_read_elems_[i] % CDEBlockTransferScalarPerVector_NPerBlock == 0;
+                // Vector read of Ds is always on N dimension.
+                const bool valid_d_access_dim = arg.ds_nz_consecutive_[i];
+                if(!(valid_d_vector_size && valid_d_access_dim))
                 {
-                    all_valid = false;
+                    valid_ds_access = false;
                 }
             });
-
-            // vector memory access of E: always on NPerBlock dimension
-            if(!(arg.e_nz_stride_ == 1 &&
-                 arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
-                         CDEBlockTransferScalarPerVector_NPerBlock ==
-                     0))
+            if(!valid_ds_access)
             {
-                all_valid = false;
+                return false;
             }
 
-            if(!all_valid)
+            const bool valid_e_vector_size =
+                arg.e_max_write_elems_ % CDEBlockTransferScalarPerVector_NPerBlock == 0;
+            // Vector write of E is always on N dimension.
+            const bool valid_e_access_dim = arg.e_nz_consecutive_;
+            if(!(valid_e_vector_size && valid_e_access_dim))
             {
                 return false;
             }

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

@@ -13,6 +13,7 @@
 #include "ck/tensor_operation/gpu/device/device_contraction_multiple_d.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_utils.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
@@ -183,7 +184,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
             return generate_tuple([&](auto i) { return vec[i]; }, num);
         };
 
-        const auto a_ms_ns_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
+        const auto a_ms_ks_lengths = to_tuple(a_ms_ks_lengths_vec, Number<NumDimM + NumDimK>{});
         const auto a_ms_ks_strides = to_tuple(a_ms_ks_strides_vec, Number<NumDimM + NumDimK>{});
 
         // dimension Ids for M0, M1, ...
@@ -194,14 +195,14 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
             typename arithmetic_sequence_gen<NumDimM, NumDimM + NumDimK, 1>::type{};
 
         // lengths for M0, M1, ...
-        const auto mLengths = get_container_subset(a_ms_ns_lengths, mDimIds);
+        const auto mLengths = get_container_subset(a_ms_ks_lengths, mDimIds);
 
         // lengths for K0, K1, ...
-        const auto kLengths = get_container_subset(a_ms_ns_lengths, kDimIds);
+        const auto kLengths = get_container_subset(a_ms_ks_lengths, kDimIds);
 
         // naive tensor A[M0, M1, M2, ..., K0, K1, K2...]
         const auto a_grid_desc_ms_ks =
-            make_naive_tensor_descriptor(a_ms_ns_lengths, a_ms_ks_strides);
+            make_naive_tensor_descriptor(a_ms_ks_lengths, a_ms_ks_strides);
 
         // transformed tensor A[MRaw = M0 * M1 * M2 * ... , KRaw = K0 * K1 * K2 * ...]
         const auto a_grid_desc_mraw_kraw = transform_tensor_descriptor(
@@ -383,7 +384,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                  const void* p_b_grid,
                  std::array<const void*, NumDTensor> p_ds_grid,
                  void* p_e_grid,
-                 const std::vector<index_t>& a_ms_ns_lengths,
+                 const std::vector<index_t>& a_ms_ks_lengths,
                  const std::vector<index_t>& a_ms_ks_strides,
                  const std::vector<index_t>& b_ns_ks_lengths,
                  const std::vector<index_t>& b_ns_ks_strides,
@@ -398,7 +399,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
               p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
               p_ds_grid_{},
               p_e_grid_{static_cast<EDataType*>(p_e_grid)},
-              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(a_ms_ns_lengths, a_ms_ks_strides)},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(a_ms_ks_lengths, a_ms_ks_strides)},
               b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(b_ns_ks_lengths, b_ns_ks_strides)},
               ds_grid_desc_m_n_{},
               e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N(e_ms_ns_lengths, e_ms_ns_strides)},
@@ -411,13 +412,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
               block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
               a_element_op_{a_element_op},
               b_element_op_{b_element_op},
-              cde_element_op_{cde_element_op},
-              a_mz_stride_{},
-              a_kz_stride_{},
-              b_nz_stride_{},
-              b_kz_stride_{},
-              ds_nz_stride_{},
-              e_nz_stride_{}
+              cde_element_op_{cde_element_op}
         {
             // populate pointer, batch stride, desc for Ds
             static_for<0, NumDTensor, 1>{}([&](auto i) {
@@ -448,18 +443,26 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
             }
 
             // for sanity check of vector memory access
-            a_mz_stride_ = a_ms_ks_strides[NumDimM - 1];
-            a_kz_stride_ = a_ms_ks_strides[NumDimM + NumDimK - 1];
+            a_mz_consecutive_ = a_ms_ks_strides[NumDimM - 1] == 1;
+            a_kz_consecutive_ = a_ms_ks_strides[NumDimM + NumDimK - 1] == 1;
+            a_max_read_elems_ =
+                CalculateMaxRead<NumDimM, NumDimK>(a_ms_ks_lengths, a_ms_ks_strides);
 
-            b_nz_stride_ = b_ns_ks_strides[NumDimN - 1];
-            b_kz_stride_ = b_ns_ks_strides[NumDimN + NumDimK - 1];
+            b_nz_consecutive_ = b_ns_ks_strides[NumDimN - 1] == 1;
+            b_kz_consecutive_ = b_ns_ks_strides[NumDimN + NumDimK - 1] == 1;
+            b_max_read_elems_ =
+                CalculateMaxRead<NumDimN, NumDimK>(b_ns_ks_lengths, b_ns_ks_strides);
 
             for(index_t i = 0; i < NumDTensor; ++i)
             {
-                ds_nz_stride_[i] = ds_ms_ns_strides[i][NumDimM + NumDimN - 1];
+                ds_nz_consecutive_[i] = ds_ms_ns_strides[i][NumDimM + NumDimN - 1] == 1;
+                ds_max_read_elems_[i] =
+                    CalculateMaxRead<NumDimM, NumDimN>(ds_ms_ns_lengths[i], ds_ms_ns_strides[i]);
             }
 
-            e_nz_stride_ = e_ms_ns_strides[NumDimM + NumDimN - 1];
+            e_nz_consecutive_ = e_ms_ns_strides[NumDimM + NumDimN - 1] == 1;
+            e_max_write_elems_ =
+                CalculateMaxRead<NumDimM, NumDimN>(e_ms_ns_lengths, e_ms_ns_strides);
         }
 
         void Print() const
@@ -499,15 +502,19 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
         BElementwiseOperation b_element_op_;
         CDEElementwiseOperation cde_element_op_;
 
-        // Strides for the last M/N/K dimensions of A/B/Ds/E
-        //   for sanity check of vector load/store
-        index_t a_mz_stride_;
-        index_t a_kz_stride_;
-        index_t b_nz_stride_;
-        index_t b_kz_stride_;
-        std::array<index_t, NumDTensor> ds_nz_stride_;
-        index_t e_mz_stride_;
-        index_t e_nz_stride_;
+        // Describe whether the last part of a given dimension of A/B/D/E is consecutive
+        // in the memory or not.
+        bool a_mz_consecutive_;
+        bool a_kz_consecutive_;
+        bool b_nz_consecutive_;
+        bool b_kz_consecutive_;
+        std::array<bool, NumDTensor> ds_nz_consecutive_;
+        bool e_nz_consecutive_;
+
+        index_t a_max_read_elems_;
+        index_t b_max_read_elems_;
+        std::array<index_t, NumDTensor> ds_max_read_elems_;
+        index_t e_max_write_elems_;
     };
 
     // Invoker
@@ -616,65 +623,47 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                           (BBlockTransferSrcVectorDim == 1 || BBlockTransferSrcVectorDim == 2),
                       "wrong!");
 
-        // vector memory access of A: could be on M or AK1 dimension
-        if constexpr(ABlockTransferSrcVectorDim == 1)
+        const bool valid_a_vector_size =
+            arg.a_max_read_elems_ % ABlockTransferSrcScalarPerVector == 0;
+        const bool valid_a_access_dim_m = ABlockTransferSrcVectorDim == 1 && arg.a_mz_consecutive_;
+        const bool valid_a_access_dim_k = ABlockTransferSrcVectorDim == 2 && arg.a_kz_consecutive_;
+        const bool valid_a_access_dim   = valid_a_access_dim_m || valid_a_access_dim_k;
+        if(!(valid_a_vector_size && valid_a_access_dim))
         {
-            if(!(arg.a_mz_stride_ == 1 &&
-                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) % ABlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
-        }
-        else
-        {
-            if(!(arg.a_kz_stride_ == 1 &&
-                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
+            return false;
         }
 
-        // vector memory access of B: could be on N or BK1 dimension
-        if constexpr(BBlockTransferSrcVectorDim == 1)
-        {
-            if(!(arg.b_nz_stride_ == 1 &&
-                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) % BBlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
-        }
-        else
+        const bool valid_b_vector_size =
+            arg.b_max_read_elems_ % BBlockTransferSrcScalarPerVector == 0;
+        const bool valid_b_access_dim_n = BBlockTransferSrcVectorDim == 1 && arg.b_nz_consecutive_;
+        const bool valid_b_access_dim_k = BBlockTransferSrcVectorDim == 2 && arg.b_kz_consecutive_;
+        const bool valid_b_access_dim   = valid_b_access_dim_n || valid_b_access_dim_k;
+        if(!(valid_b_vector_size && valid_b_access_dim))
         {
-            if(!(arg.b_kz_stride_ == 1 &&
-                 arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) % BBlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
+            return false;
         }
 
-        // vector memory access of Ds: always on NPerBlock dimension
-        bool valid_d_access = true;
-
+        bool valid_ds_access = true;
         static_for<0, NumDTensor, 1>{}([&](auto i) {
-            if(!(arg.ds_nz_stride_[i] == 1 &&
-                 arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_[i].GetLength(I3) %
-                         CDEBlockTransferScalarPerVector_NPerBlock ==
-                     0))
+            const bool valid_d_vector_size =
+                arg.ds_max_read_elems_[i] % CDEBlockTransferScalarPerVector_NPerBlock == 0;
+            // Vector read of Ds is always on N dimension.
+            const bool valid_d_access_dim = arg.ds_nz_consecutive_[i];
+            if(!(valid_d_vector_size && valid_d_access_dim))
             {
-                valid_d_access = false;
+                valid_ds_access = false;
             }
         });
-
-        if(valid_d_access == false)
+        if(!valid_ds_access)
         {
             return false;
         }
 
-        // vector memory access of E: always on NPerBlock dimension
-        if(!(arg.e_nz_stride_ == 1 &&
-             arg.e_grid_desc_mblock_mperblock_nblock_nperblock_.GetLength(I3) %
-                     CDEBlockTransferScalarPerVector_NPerBlock ==
-                 0))
+        const bool valid_e_vector_size =
+            arg.e_max_write_elems_ % CDEBlockTransferScalarPerVector_NPerBlock == 0;
+        // Vector write of E is always on N dimension.
+        const bool valid_e_access_dim = arg.e_nz_consecutive_;
+        if(!(valid_e_vector_size && valid_e_access_dim))
         {
             return false;
         }
@@ -692,7 +681,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                              const void* p_b,
                              std::array<const void*, NumDTensor> p_ds,
                              void* p_e,
-                             const std::vector<index_t>& a_ms_ns_lengths,
+                             const std::vector<index_t>& a_ms_ks_lengths,
                              const std::vector<index_t>& a_ms_ks_strides,
                              const std::vector<index_t>& b_ns_ks_lengths,
                              const std::vector<index_t>& b_ns_ks_strides,
@@ -708,7 +697,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                         p_b,
                         p_ds,
                         p_e,
-                        a_ms_ns_lengths,
+                        a_ms_ks_lengths,
                         a_ms_ks_strides,
                         b_ns_ks_lengths,
                         b_ns_ks_strides,
@@ -729,7 +718,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                         const void* p_b,
                         std::array<const void*, NumDTensor> p_ds,
                         void* p_e,
-                        const std::vector<index_t>& a_ms_ns_lengths,
+                        const std::vector<index_t>& a_ms_ks_lengths,
                         const std::vector<index_t>& a_ms_ks_strides,
                         const std::vector<index_t>& b_ns_ks_lengths,
                         const std::vector<index_t>& b_ns_ks_strides,
@@ -745,7 +734,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
                                           p_b,
                                           p_ds,
                                           p_e,
-                                          a_ms_ns_lengths,
+                                          a_ms_ks_lengths,
                                           a_ms_ks_strides,
                                           b_ns_ks_lengths,
                                           b_ns_ks_strides,

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cassert>
+#include <sstream>
+#include <vector>
+
+#include "ck/ck.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/**
+ * Calculates the maximum number of subsequent elements of the fast changing dimension
+ * that are consecutive in memory.
+ *
+ * Example:
+ *   NumDimM = 2, NumDimK = 3
+ *   A shape =   [  2,   3,  4, 5, 6]
+ *   A strides = [360, 120, 30, 6, 1]
+ *                |  M   |  |   K  |
+ *   It follows from strides that K is FCD and all the subsequent elements of K are consecutive
+ *   in memory.
+ *   But if strides were [360, 120, 6, 24, 1], then only 6 subsequent elements of K would be
+ *   consecutive in memory.
+ *
+ * Assumes that the dimensions are split into two groups of `NumDim1` and `NumDim2` dimensions.
+ */
+template <index_t NumDim1, index_t NumDim2>
+auto CalculateMaxRead(const std::vector<index_t>& lengths, const std::vector<index_t>& strides)
+{
+    if(lengths.size() != NumDim1 + NumDim2)
+    {
+        std::ostringstream err;
+        err << "Incorrect number of lengths in " << __FILE__ << ":" << __LINE__
+            << ", in function: " << __func__;
+        throw std::runtime_error(err.str());
+    }
+    if(strides.size() != NumDim1 + NumDim2)
+    {
+        std::ostringstream err;
+        err << "Incorrect number of strides in " << __FILE__ << ":" << __LINE__
+            << ", in function: " << __func__;
+        throw std::runtime_error(err.str());
+    }
+
+    // Determine the beginning and end idx of the group representing the FCD.
+    index_t begin_idx, end_idx;
+    if(strides[NumDim1 - 1] == 1)
+    {
+        begin_idx = 0;
+        end_idx   = NumDim1 - 1;
+    }
+    else if(strides[NumDim1 + NumDim2 - 1] == 1)
+    {
+        begin_idx = NumDim1;
+        end_idx   = NumDim1 + NumDim2 - 1;
+    }
+    else
+    {
+        // The dimension consecutive in memory is not the last dimension of any group, so only
+        // one element can be read/written at once.
+        return 1;
+    }
+
+    index_t consecutive_stride = 1;
+    for(index_t dim_idx = end_idx; dim_idx >= begin_idx; --dim_idx)
+    {
+        if(strides[dim_idx] == consecutive_stride)
+        {
+            consecutive_stride *= lengths[dim_idx];
+        }
+        else
+        {
+            break;
+        }
+    }
+    const index_t max_subsequent_elems = consecutive_stride;
+    return max_subsequent_elems;
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_elementwise_3d_impl.hpp

@@ -13,6 +13,7 @@
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 
 #include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/stream_utility.hpp"
 
 namespace ck {
@@ -292,6 +293,12 @@ struct DeviceElementwise3dImpl : public DeviceElementwise<InDataTypeTuple,
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
     {
+        if((ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
+            ck::get_device_name() == "gfx942"))
+        {
+            return false;
+        }
+
         const Argument* pArg = dynamic_cast<const Argument*>(p_arg);
 
         if(pArg == nullptr)

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

@@ -821,7 +821,9 @@ struct DeviceGemmMultipleDLayernorm_Xdl_CShuffle
         return (workspace_size);
     };
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -380,7 +380,9 @@ struct DeviceGemm_Xdl_CShuffle_LdsDirectLoad : public DeviceGemm<ALayout,
             << " LoopScheduler: "
             << LoopSchedToString[LoopSched] << ", "
             << "PipelineVersion: "
-            << PipelineVersionToString[PipelineVer];
+            << PipelineVersionToString[PipelineVer] << ", "
+            << "Prefetch: "
+            << NumGemmKPrefetchStage;
         // clang-format on
 
         return str.str();

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

@@ -226,7 +226,9 @@ struct DeviceGemmXdlStreamK : public DeviceGemmStreamK<ALayout,
         }
     }
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -357,15 +357,17 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
         return out_gemmm_gemmn_desc;
     }
 
+    // Shape of Ds and E must be aligned. Strides can be different.
+    // Pass e_g_n_k_wos_lengths for logical broadcast.
     static auto MakeDsGridDescriptor_M_N(
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
         const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
     {
         return generate_tuple(
             [&](auto i) {
                 using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
-                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(e_g_n_k_wos_lengths,
                                                                   ds_g_n_k_wos_strides[i]);
             },
             Number<NumDTensor>{});
@@ -569,7 +571,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
 
                 // D desc
                 ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
-                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+                    e_g_n_k_wos_lengths, ds_g_n_k_wos_strides[i]);
             });
             compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
 
@@ -916,8 +918,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                          is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
                          is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
                          is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
-                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
-                         is_same_v<DLayout, ctc::G_K>)
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
             {
                 const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
 
@@ -925,6 +926,27 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
                 {
                     valid = false;
                 }
+
+                if constexpr(is_same_v<DLayout, ctc::G_K>)
+                {
+                    // G and K must be the same
+                    if(arg.ds_g_n_k_wos_lengths_[i][0] != arg.e_g_n_k_wos_lengths_[0] ||
+                       arg.ds_g_n_k_wos_lengths_[i][2] != arg.e_g_n_k_wos_lengths_[2])
+                    {
+                        valid = false;
+                    }
+                }
+                else
+                {
+                    // E and D must have the same shape
+                    for(index_t d = 0; d < NDimSpatial + 3; d++)
+                    {
+                        if(arg.ds_g_n_k_wos_lengths_[i][d] != arg.e_g_n_k_wos_lengths_[d])
+                        {
+                            valid = false;
+                        }
+                    }
+                }
             }
             else
             {

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

@@ -631,8 +631,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
                          is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
                          is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
                          is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
-                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
-                         is_same_v<DLayout, ctc::G_K>)
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
             {
                 const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
 

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

@@ -817,12 +817,15 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
         return arg.group_count_ * sizeof(GroupedGemmKernelArgument<NumDTensor>);
     }
 
-    void SetWorkSpacePointer(BaseArgument* p_arg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* p_arg,
+                             void* p_workspace,
+                             const StreamConfig& stream_config = StreamConfig{}) const override
     {
         auto p_arg_          = dynamic_cast<Argument*>(p_arg);
         p_arg_->p_workspace_ = p_workspace;
 
-        hip_check_error(hipMemset(p_workspace, 0, GetWorkSpaceSize(p_arg)));
+        hip_check_error(
+            hipMemsetAsync(p_workspace, 0, GetWorkSpaceSize(p_arg), stream_config.stream_id_));
     }
 
     static void SetKBatch(Argument& arg, index_t k_batch) { arg.UpdateKBatch(k_batch); }

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

@@ -577,7 +577,9 @@ struct DeviceNormalizationFwdSplitKImpl : public DeviceNormalizationFwd<XDataTyp
         return (workspace_size);
     };
 
-    void SetWorkSpacePointer(BaseArgument* pArg, void* p_workspace) const override
+    void SetWorkSpacePointer(BaseArgument* pArg,
+                             void* p_workspace,
+                             const StreamConfig& = StreamConfig{}) const override
     {
         Argument* pArg_ = dynamic_cast<Argument*>(pArg);
 

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

@@ -308,12 +308,6 @@ struct GNDHWK : public BaseTensorLayout
     static constexpr const char* name = "GNDHWK";
 };
 
-// for output bias
-struct GK : public BaseTensorLayout
-{
-    static constexpr const char* name = "GK";
-};
-
 // output tensor
 // packed NWGK/NHWGK/NDHWGK
 struct NWGK : public BaseTensorLayout