Lodestar.Gpu 0.1.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

TiledSparseDenseProduct

The product of a resident CSR matrix with a dense block, tiled through shared memory.

public sealed class TiledSparseDenseProduct

Example — the shape a caller writes.

using Lodestar.Gpu.Compute;

using var context = GpuContext.Create(preferCpu: true);
using var matrix = DeviceSparseMatrix.Upload(
    context, [0, 1, 2], [0, 1], [2.0, 3.0], rowCount: 2, columnCount: 2);
var kernel = new TiledSparseDenseProduct(context);

double[] product = kernel.Multiply(matrix, [1.0, 0.0, 0.0, 1.0], width: 2);

double first = product[0];  // => 2
double last = product[3];  // => 3

Members — one page each.

Member	What it does
TiledSparseDenseProduct.Multiply	Computes matrix · block

Remarks — one group per row and column tile. The group loads a tile of the row's stored values and their column indices into shared memory, so a row's non-zeros are read from global memory once per group rather than once per thread.

Accumulation walks the row in stored order — the order CsrMatrix.Multiply walks it — so the two agree far inside the tolerance the tests assert. A looser agreement would mean the orders had diverged, which is why the tests compare at 1e-6 and the real agreement is much tighter.

Double precision, because the CPU operand is. bench/README.md predicted that would cost the gate on a consumer card; it does not, because the kernel is bound by memory bandwidth rather than by the double-precision units. It clears at 8.66× and 18.9× at vectorizer sizes and misses below them, which is where the crossing point belongs.

Build this once. Loading compiles, and a cold launch measures ILGPU's compiler.

Applies to — net10.0, netstandard2.1.

See also — DeviceSparseMatrix, DeviceDenseBlock, the namespace index.