Struct Csr5Matrix 

pub struct Csr5Matrix<T: GpuFloat> { /* private fields */ }

A sparse matrix in CSR5 format, stored on GPU.

CSR5 augments the standard CSR representation with tile-based metadata for load-balanced GPU SpMV. The non-zero elements are logically divided into tiles of CSR5_TILE_WIDTH elements each. Each tile maps to one warp during SpMV execution.

Implementations

impl<T: GpuFloat> Csr5Matrix<T>

pub fn from_csr_host( rows: u32, cols: u32, row_ptr: &[i32], col_idx: &[i32], values: &[T], ) -> SparseResult<Self>

Constructs a CSR5 matrix from host-side CSR arrays.

This performs the CSR-to-CSR5 conversion on the host, computing tile pointers and descriptors, then uploads everything to GPU memory.

Arguments

• rows – Number of rows.
• cols – Number of columns.
• row_ptr – CSR row pointer (length rows + 1).
• col_idx – CSR column indices (length nnz).
• values – CSR values (length nnz).

Errors

Returns SparseError::InvalidFormat if inputs are inconsistent. Returns SparseError::Cuda on GPU allocation/transfer failure.

pub fn from_csr(csr: &CsrMatrix<T>) -> SparseResult<Self>

Constructs a CSR5 matrix from an existing GPU-resident CSR matrix.

Downloads CSR data to host, computes tile metadata, then re-uploads everything.

Errors

Returns SparseError::Cuda on GPU transfer/allocation failure.

pub fn rows(&self) -> u32

Returns the number of rows.

pub fn cols(&self) -> u32

Returns the number of columns.

pub fn nnz(&self) -> u32

Returns the total non-zero count.

pub fn num_tiles(&self) -> u32

Returns the number of tiles.

pub fn row_ptr(&self) -> &DeviceBuffer<i32>

Returns a reference to the CSR row pointer device buffer.

pub fn col_idx(&self) -> &DeviceBuffer<i32>

Returns a reference to the CSR column index device buffer.

pub fn values(&self) -> &DeviceBuffer<T>

Returns a reference to the values device buffer.

pub fn tile_ptr(&self) -> &DeviceBuffer<u32>

Returns a reference to the tile pointer device buffer.

pub fn tile_desc(&self) -> &DeviceBuffer<TileDescriptor>

Returns a reference to the tile descriptor device buffer.

pub fn calibrator(&self) -> &DeviceBuffer<T>

Returns a reference to the calibrator device buffer.

pub fn tile_metadata_to_host( &self, ) -> SparseResult<(Vec<u32>, Vec<TileDescriptor>)>

Downloads tile metadata to host for inspection.

Errors

Returns SparseError::Cuda on transfer failure.