Expand description
GhostFlow CUDA Backend - Real GPU Acceleration
This module provides real CUDA GPU acceleration when compiled with the cuda feature.
Without the feature, it provides CPU fallback implementations.
Re-exports§
pub use device::CudaDevice;pub use device::DeviceGuard;pub use device::get_all_devices;pub use device::select_best_device;pub use memory::GpuMemoryPool;pub use memory::GpuTensor;pub use memory::get_global_gpu_pool;pub use stream::CudaStream;pub use stream::CudaEvent;pub use stream::CudaTimer;pub use tensor::CudaTensor;pub use error::CudaError;pub use error::CudaResult;pub use blas::CuBlas;
Modules§
- blas
- cuBLAS wrapper for linear algebra operations
- device
- CUDA device management - Real Implementation
- error
- CUDA error types
- ffi
- CUDA FFI bindings - Real CUDA Runtime API
- kernels
- CUDA kernel definitions
- memory
- GPU memory management
- ops
- CUDA tensor operations
- stream
- CUDA streams for async execution - Real Implementation
- tensor
- CUDA tensor type - Real GPU tensor implementation
Functions§
- cuda_
version - Get CUDA version
- current_
device - Get current CUDA device ID
- device_
count - Get number of available CUDA devices
- empty_
cache - Empty CUDA cache (free cached memory)
- init
- Initialize CUDA runtime
- is_
available - Check if CUDA is available at runtime
- memory_
info - Memory info for current device
- set_
device - Set current CUDA device
- synchronize
- Synchronize all CUDA operations on current device