Merge branch 'develop' into lwpck-471

include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp

@@ -465,7 +465,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if 0
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "

include/ck/tensor_operation/gpu/device/impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp

@@ -400,6 +400,7 @@ struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
+#if DEBUG_LOG
             {
                 std::cout << "num_batches_of_GEMM = " << arg.num_subbatches_ << std::endl;
                 std::cout << "a_grid_desc_k0_m_k1{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
@@ -413,6 +414,7 @@ struct DeviceConv3dFwdXdl_Input_N_Di_Hi_Wi_C_Weight_K_Z_Y_X_C_Output_N_Do_Ho_Wo_
                 std::cout << "c_grid_desc_m_n{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,

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

@@ -1272,6 +1272,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
             float ave_time = 0;
             for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
             {
+#if DEBUG_LOG
                 {
                     std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
                               << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
@@ -1304,6 +1305,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
                               << arg.c_grid_desc_m0_m10_m11_n0_n10_n11_container_[i].GetLength(I5)
                               << " ) " << std::endl;
                 }
+#endif
 
                 if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                 arg.b_grid_desc_k0_n_k1_container_[i],

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

@@ -1274,6 +1274,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
             float ave_time = 0;
             for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
             {
+#if DEBUG_LOG
                 {
                     std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
                               << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
@@ -1310,6 +1311,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
                               << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I7)
                               << " ) " << std::endl;
                 }
+#endif
 
                 if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                 arg.b_grid_desc_k0_n_k1_container_[i],

include/ck/tensor_operation/gpu/device/device_elementwise_2d.hpp → include/ck/tensor_operation/gpu/device/impl/device_elementwise_2d_impl.hpp

@@ -8,7 +8,7 @@
 
 #include "ck/utility/math.hpp"
 #include "ck/utility/sequence.hpp"
-#include "ck/tensor_operation/gpu/device/device_elementwise_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_elementwise_2d.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 
@@ -26,10 +26,10 @@ template <typename InDataTypeTuple,
           index_t NPerThread,
           typename InScalarPerVectorSeq,
           typename OutScalarPerVectorSeq>
-struct DeviceElementwise : public DeviceElementwiseBase<InDataTypeTuple,
-                                                        OutDataTypeTuple,
-                                                        ElementwiseOperation,
-                                                        NumDim_m + NumDim_n>
+struct DeviceElementwise2dImpl : public DeviceElementwise<InDataTypeTuple,
+                                                          OutDataTypeTuple,
+                                                          ElementwiseOperation,
+                                                          NumDim_m + NumDim_n>
 {
     static constexpr index_t NumDim = NumDim_m + NumDim_n;
 

include/ck/tensor_operation/gpu/device/impl/device_elementwise.hpp → include/ck/tensor_operation/gpu/device/impl/device_elementwise_impl.hpp

@@ -8,7 +8,7 @@
 
 #include "ck/utility/math.hpp"
 #include "ck/utility/sequence.hpp"
-#include "ck/tensor_operation/gpu/device/device_elementwise_base.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_elementwise_1d.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 
@@ -25,8 +25,8 @@ template <typename InDataTypeTuple,
           index_t MPerThread,
           typename InScalarPerVectorSeq,
           typename OutScalarPerVectorSeq>
-struct DeviceElementwise
-    : public DeviceElementwiseBase<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>
+struct DeviceElementwiseImpl
+    : public DeviceElementwise<InDataTypeTuple, OutDataTypeTuple, ElementwiseOperation, NumDim>
 {
     static constexpr int NumInput  = InDataTypeTuple::Size();
     static constexpr int NumOutput = OutDataTypeTuple::Size();

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

@@ -270,18 +270,18 @@ struct DeviceElementwiseNormalizationImpl
                  const std::vector<index_t> reduceDims,
                  XElementwiseOperation x_elementwise_op,
                  YElementwiseOperation y_elementwise_op,
-                 AccDataType epsilon,
+                 double epsilon,
                  const std::array<const void*, NumInput> in_dev_buffers,
                  const GammaDataType* p_gamma,
                  const BetaDataType* p_beta,
                  YDataType* p_y)
-            : epsilon_(epsilon),
-              p_gamma_(p_gamma),
+            : p_gamma_(p_gamma),
               p_beta_(p_beta),
               p_y_(p_y),
               x_elementwise_op_(x_elementwise_op),
               y_elementwise_op_(y_elementwise_op)
         {
+            epsilon_ = static_cast<AccDataType>(epsilon);
 
             Lengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(lengths, reduceDims);
             for(int i = 0; i < NumInput; i++)
@@ -543,7 +543,7 @@ struct DeviceElementwiseNormalizationImpl
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
                         const std::vector<index_t> reduceDims,
-                        AccDataType epsilon,
+                        double epsilon,
                         const std::array<const void*, NumInput> in_dev_buffers,
                         const void* p_gamma,
                         const void* p_beta,

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

@@ -327,6 +327,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_k0_m0_m1_k1_{"
                           << arg.a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
@@ -341,6 +342,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
                 std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
             }
+#endif
 
             if(!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_multiple_d_xdl_cshuffle.hpp

@@ -431,9 +431,6 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
             const index_t grid_size =
                 arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
 
-            const auto K =
-                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
-
             auto launch_kernel = [&](auto has_main_k_block_loop) {
                 constexpr bool has_main_loop = has_main_k_block_loop.value;
 
@@ -471,6 +468,8 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
                                               arg.block_2_etile_map_);
             };
 
+            const auto K = arg.a_grid_desc_m_k_.GetLength(I1);
+
             if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
             {
                 return launch_kernel(integral_constant<bool, true>{});

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

@@ -510,7 +510,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperatio
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if 0
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
                           << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
@@ -525,8 +525,8 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<0, ReduceOperatio
                 std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
 
-                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0) << "}"
-                          << std::endl;
+                std::cout << "arg.reduce_grid_desc_m_{ " << arg.reduce_grid_desc_m_.GetLength(I0)
+                          << "}" << std::endl;
             }
 #endif
 

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

@@ -310,7 +310,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if 0
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "

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

@@ -459,7 +459,7 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if 0
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
                           << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
@@ -486,7 +486,6 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
 
             const index_t grid_size =
                 arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
-
             const auto K =
                 arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
 

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

@@ -514,7 +514,7 @@ struct DeviceGemmLayerNorm_Xdl_CShuffle : public BaseOperator
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-#if 0
+#if DEBUG_LOG
             {
                 std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
                           << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "