Enum CUDAExecutionProviderCuDNNConvAlgoSearch

pub enum CUDAExecutionProviderCuDNNConvAlgoSearch {
    Exhaustive,
    Heuristic,
    Default,
}

The type of search done for cuDNN convolution algorithms.

Variants

Exhaustive

Expensive exhaustive benchmarking using cudnnFindConvolutionForwardAlgorithmEx. This function will attempt all possible algorithms for cudnnConvolutionForward to find the fastest algorithm. Exhaustive search trades off between memory usage and speed. The first execution of a graph will be slow while possible convolution algorithms are tested.

Heuristic

Lightweight heuristic-based search using cudnnGetConvolutionForwardAlgorithm_v7. Heuristic search sorts available convolution algorithms by expected (based on internal heuristic) relative performance.

Default

Uses the default convolution algorithm, CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM. The default algorithm may not have the best performance depending on specific hardware used. It’s recommended to use Exhaustive or Heuristic to search for a faster algorithm instead. However, Default does have its uses, such as when available memory is tight.

NOTE: This name may be confusing as this is not the default search algorithm for the CUDA EP. The default search algorithm is actually Exhaustive.

Trait Implementations

impl Clone for CUDAExecutionProviderCuDNNConvAlgoSearch

fn clone(&self) -> CUDAExecutionProviderCuDNNConvAlgoSearch

Returns a copy of the value.

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

Performs copy-assignment from source.

impl Debug for CUDAExecutionProviderCuDNNConvAlgoSearch

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

Formats the value using the given formatter.

impl Default for CUDAExecutionProviderCuDNNConvAlgoSearch

fn default() -> Self

Returns the “default value” for a type.