Struct DeviceCsrMatrix 

pub struct DeviceCsrMatrix<T: FvmScalar, B: AcceleratorBackend> {
    pub row_offsets: B::Buffer<u32>,
    pub column_indices: B::Buffer<u32>,
    pub values: B::Buffer<T>,
    pub row_count: usize,
    pub column_count: usize,
    pub workspace: Option<B::Buffer<u8>>,
}

A validated CSR matrix with device-resident structure and values.

Fields

row_offsets: B::Buffer<u32>

Zero-based row offsets, with row_count + 1 entries.

column_indices: B::Buffer<u32>

Zero-based column indices.

values: B::Buffer<T>

Nonzero values in row-major CSR order.

row_count: usize

Number of represented rows.

column_count: usize

Required input vector length.

workspace: Option<B::Buffer<u8>>

Reusable implementation-specific SpMV scratch allocation.

Implementations

impl<T: FvmScalar, B: AcceleratorBackend> DeviceCsrMatrix<T, B>

pub fn from_pattern( backend: &B, pattern: &CsrPattern, values: &[T], ) -> Result<Self, AcceleratorError>

Build a square matrix from a symbolic closure-DOF pattern.

pub fn from_global_pattern( backend: &B, pattern: &GlobalCsrPattern, values: &[T], ) -> Result<Self, AcceleratorError>

Build a compact-row matrix from a globally numbered pattern.

Output rows follow pattern.rows; input columns retain their global numbers, so column_count is one greater than the largest column.

impl<T: FvmScalar> DeviceCsrMatrix<T, CpuBackend>

pub fn spmv( &self, alpha: T, x: &<CpuBackend as AcceleratorBackend>::Buffer<T>, beta: T, y: &mut <CpuBackend as AcceleratorBackend>::Buffer<T>, ) -> Result<(), AcceleratorError>

Compute y = alpha * A * x + beta * y in deterministic row order.