Enum KernelMemMap

pub enum KernelMemMap<'a, F, Mem: ComputeMemory<F>> {
    Chunked {
        data: Mem::FSlice<'a>,
        log_min_chunk_size: usize,
    },
    ChunkedMut {
        data: Mem::FSliceMut<'a>,
        log_min_chunk_size: usize,
    },
    Local {
        log_size: usize,
    },
}

A memory mapping specification for a kernel execution.

See ComputeLayer::accumulate_kernels for context on kernel execution.

Variants

Chunked

This maps a chunk of a buffer in global device memory to a read-only kernel buffer.

Fields

data: Mem::FSlice<'a>

log_min_chunk_size: usize

ChunkedMut

This maps a chunk of a mutable buffer in global device memory to a read-write kernel buffer. When the kernel exits, the data in the kernel buffer is written back to the original location.

Fields

data: Mem::FSliceMut<'a>

log_min_chunk_size: usize

Local

This allocates a kernel-local scratchpad buffer. The size specified in the mapping is the total size of all kernel scratchpads. This is so that the kernel’s local scratchpad size scales up proportionally to the size of chunked buffers.

Fields

log_size: usize

Implementations

impl<'a, F, Mem: ComputeMemory<F>> KernelMemMap<'a, F, Mem>

pub fn log_chunks_range(mappings: &[Self]) -> Option<Range<usize>>

Computes a range of possible number of chunks that data can be split into, given a sequence of memory mappings.

pub fn chunks( self, log_chunks: usize, ) -> impl Iterator<Item = KernelMemMap<'a, F, Mem>>