CudaTensorOps.Transfer.ixx File Reference

CUDA tensor transfer operations partition. More...

#include <cuda_runtime.h>
#include <memory>
#include <format>
#include <stdexcept>
#include <source_location>
#include <cstring>
#include "Kernels/Transfer.Copy.h"
import Cuda.Error;
import Dnn.Tensor;
import Dnn.TensorDataTypeMap;
import Dnn.TensorDataType;
import Compute.IExecutionContext;
import Compute.CudaManagedMemoryResource;
import Dnn.TensorDataTypeTraits;
import Compute.CudaTensorDataType;
import Dnn.ITensor;
import Compute.CudaPinnedMemoryResource;
import Compute.CudaDeviceMemoryResource;
import Compute.ExecutionContext;
import Compute.DeviceType;
import Serialization.Tensor;

Classes
struct	Mila::Dnn::Compute::Cuda::TransferOps
	CUDA specialization of TensorOps for tensor transfer operations. More...

Namespaces
namespace	Mila
	Mila main API namespace.
namespace	Mila::Dnn
namespace	Mila::Dnn::Compute
namespace	Mila::Dnn::Compute::Cuda

Detailed Description

CUDA tensor transfer operations partition.

Provides tensor transfer operations using ExecutionContext for stream management. TensorOps work with tensor data directly (via data()) and accept optional ExecutionContext for explicit stream control with zero-overhead borrowing semantics.

Implementation strategy:

• Raw pointer semantics for ExecutionContext (non-owning borrow)
• Automatic fallback to default CUDA stream when no context provided
• Stream-based asynchronous execution for pipeline optimization
• Automatic type conversion using CUDA kernels
• Memory-efficient staging for host-device transfers with conversion