Development build. This page describes main, not a released package. The latest published Lodestar.Gpu is 0.2.0 — read its documentation.

Home › Gpu › GPU kernels

Uploads a CSR matrix to the accelerator.

public static DeviceSparseMatrix Upload(GpuContext context, ReadOnlySpan<int> rowPointers, ReadOnlySpan<int> columnIndices, ReadOnlySpan<double> values, int rowCount, int columnCount)

Parameters — context is the accelerator to upload to. rowPointers holds rowCount + 1 offsets into the other two. columnIndices holds one column per stored value, ascending inside a row. values holds the stored values in the same order. rowCount and columnCount are the matrix's shape; columnCount is not derived from the indices, because a matrix may have trailing columns no row stores anything in.

Returns — DeviceSparseMatrix, owning three device buffers the caller disposes.

Exceptions — ArgumentNullException when context is null; ObjectDisposedException when context was disposed; ArgumentOutOfRangeException when a dimension is below 1; ArgumentException when the three arrays do not describe one CSR matrix — rowPointers not starting at 0, decreasing, or ending anywhere but at the value count, or a column index outside [0, columnCount).

Example — the shape a caller writes.

using Lodestar.Gpu.Compute;

using var context = GpuContext.Create(preferCpu: true);
using var matrix = DeviceSparseMatrix.Upload(
    context, [0, 2, 3], [0, 1, 1], [1.0, 2.0, 3.0], rowCount: 2, columnCount: 2);

int stored = matrix.NonZeroCount;  // => 3

Remarks — the shapes and the structure are checked because each has its own failure. rowPointers of the wrong length silently shifts every row; columnIndices and values of different lengths reads past one of them; and a final offset that disagrees with the value count means the last row is truncated without anything noticing. The kernel indexes device memory with the offsets and the columns unchecked, so a first offset past 0 drops stored values, and an offset that steps back or a column outside the matrix reads another buffer (#898). The structure is checked in one pass per array, once per upload.

Column indices ascending inside a row is a precondition, not a check. The kernel relies on it only to read contiguous memory, so an unsorted row produces the same arithmetic more slowly rather than a wrong answer — but it is what CsrMatrix guarantees and what a caller building the arrays by hand has to preserve.

A matrix storing nothing is accepted, as CsrMatrix accepts it: a block of documents holding only stop words is the ordinary one, and its product is zeros. If an allocation fails, the buffers already on the device are released before the exception leaves (#1265).

Applies to — net10.0, netstandard2.1.

See also — DeviceSparseMatrix.