Struct BatchedSpMV 

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

Batched sparse matrix-vector multiplication.

Computes y_i = alpha * A_i * x_i + beta * y_i for a batch of sparse matrices A_i, dense vectors x_i, and output vectors y_i.

The baseline implementation iterates sequentially on the host. For GPU execution, use generate_batched_spmv_ptx to create a single fused kernel that processes all matrices via offset arrays.

Implementations

impl<T: GpuFloat + Add<Output = T> + Mul<Output = T> + AddAssign> BatchedSpMV<T>

pub fn from_device(matrices: &[CsrMatrix<T>]) -> SparseResult<Self>

Create a BatchedSpMV by downloading matrix data from device memory.

Errors

Returns SparseError::InvalidArgument if matrices is empty. Returns SparseError::Cuda on download failure.

pub fn from_host( row_ptrs: Vec<Vec<i32>>, col_indices: Vec<Vec<i32>>, values: Vec<Vec<T>>, rows: Vec<usize>, cols: Vec<usize>, ) -> SparseResult<Self>

Create a BatchedSpMV from host-side CSR arrays.

Errors

Returns SparseError::InvalidArgument on dimension mismatch.

pub fn batch_size(&self) -> usize

Returns the number of matrices in the batch.

pub fn execute( &self, xs: &[Vec<T>], ys: &mut [Vec<T>], alpha: T, beta: T, ) -> SparseResult<()>

Execute the batched SpMV on the host (sequential baseline).

For each matrix A_i: y_i[r] = alpha * sum(A_i[r,c] * x_i[c]) + beta * y_i[r]

Arguments

• xs – Slice of input vectors, one per matrix.
• ys – Slice of output vectors, one per matrix (modified in place).
• alpha – Scalar multiplier for A * x.
• beta – Scalar multiplier for existing y.

Errors

Returns SparseError::DimensionMismatch if slice lengths do not match the batch size, or if individual vector dimensions are wrong.

Trait Implementations

impl<T: Debug + GpuFloat> Debug for BatchedSpMV<T>

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

Formats the value using the given formatter.