Enum sql_ops::partition::gpu_radix_partition::GpuHistogramAlgorithm

pub enum GpuHistogramAlgorithm {
    Chunked,
    Contiguous,
}

Specifies the histogram algorithm that computes the partition offsets.

Variants

Chunked

Chunked partitions, that are computed on the GPU.

Chunked computes a separate set of partitions per thread block. Tuples of the resulting partitions are thus distributed among all chunks.

It was originally introduced for NUMA locality by Schuh et al. in “An Experimental Comparison of Thirteen Relational Equi-Joins in Main Memory”.

On GPUs, it has two main benefits. First, thread blocks don’t communicate to compute the histogram, and can avoid global synchronization. Second, the offsets are smaller, and potentially we can use 32-bit integers instead of 64-bit integers when caching them in shared memory during the partitioning phase.

Contiguous

Contiguous partitions, that are computed on the GPU.

Contiguous computes the “normal” partition layout. Each resulting partition is laid out contiguously in memory.

Note that this algorithm does not work on pre-Pascal GPUs, because it requires cooperative launch capability to perform grid synchronization.

Trait Implementations

impl Clone for GpuHistogramAlgorithm

fn clone(&self) -> GpuHistogramAlgorithm

Returns a copy of the value.

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

Performs copy-assignment from source.

impl Debug for GpuHistogramAlgorithm

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

Formats the value using the given formatter.

impl From<CpuHistogramAlgorithm> for GpuHistogramAlgorithm

fn from(algo: CpuHistogramAlgorithm) -> Self

Performs the conversion.

impl From<GpuHistogramAlgorithm> for HistogramAlgorithmType

fn from(algo: GpuHistogramAlgorithm) -> Self

Performs the conversion.

impl Copy for GpuHistogramAlgorithm