Enum numa_gpu::runtime::cuda::CudaTransferStrategy

pub enum CudaTransferStrategy {
    PageableCopy,
    PinnedCopy,
    LazyPinnedCopy,
    Coherence,
}

Specify the CUDA transfer strategy.

Defines which strategy with which to transfer data from main-memory to device memory.

A Prefetch strategy is not defined, because cuda_prefetch_async() requires unified memory allocated with cuda_malloc_managed(). Thus, regular memory that is allocated with the system allocator cannot be used in conjunction with prefetching. As unified memory is more specific than the general system memory, prefetching is handled separately using IntoCudaIterator.

Variants

PageableCopy

Copy directly from pageable memory.

Transfers data using cuda_memcopy() directly from the specified memory location. No intermediate steps are performed.

This strategy can also transfer memory that is pinned by the user before the transfer.

PinnedCopy

Copy using an intermediate, pinned buffer.

The transfer first copies data from its original location into a pinned buffer. Then data is transferred using cuda_memcpy().

In principle, this strategy can also tranfer from a pinned memory location. However, the PageableCopy strategy would avoid the additional overhead incurred by the intermediate buffer.

LazyPinnedCopy

Pin the memory in-place and the copy.

The transfer first pins the memory in-place using cuda_host_register(). Then the data is transferred using cuda_memcpy().

Don’t use this strategy with a pinned memory location, as calling cuda_host_register() on a pinned location is probably undefined behavior.

Coherence

Access the memory in-place without any copies.

Requires that the GPU has cache-coherent access to main-memory. E.g., POWER9 and Tesla V100 with NVLink 2.0.

Trait Implementations

impl Clone for CudaTransferStrategy

fn clone(&self) -> CudaTransferStrategy

Returns a copy of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for CudaTransferStrategy

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Copy for CudaTransferStrategy