Repository navigation
[REQ] Single-line array/tile matrix extraction with row indexing #1028
Description
Description
Syntax support for, eg:
tile[i,j,k][1]
arr[i,j][0]
Where tile is a 3D tile of matrices and 1 is the second row of the matrix at i,j,k,
and arr is a 2D array of matrices and 0 is the first row of the matrix at i,j.
Activity
[-][REQ] Single-line tile matrix extraction with row indexing[/-][+][REQ] Single-line array/tile matrix extraction with row indexing[/+]shi-eric commentedon Jul 23, 2026
Related: #1688 covers the same rank-awareness problem in
tile_extract()and chained subscripts for composite-valued tiles. #1688 exercises the rank-2wp.vec3case, while this issue covers applying a matrix-row index after extracting an element from a rank-3 tile. They likely share a fix area, but neither is a duplicate of the other.Add matrix row indexing support (NVIDIAGH-1028)
shi-eric commentedon Aug 20, 2026
This is addressed by 4cd493b.
The change adds rank-aware matrix-row extraction for matrix-dtype tiles, so kernel expressions such as
tile[i, j, k][1]now return the selected row vector. It also supports row writes, negative row indices, and adjoints. The array formarr[i, j][0]was already supported and covered by existing matrix-array extraction tests; both forms requested here were reverified against this commit.Public API behavior
tile[i, j, k][r]now works for matrix-dtype tiles, with row reads and writes across supported tile ranks, negative row indices, and adjoints.arr[i, j][r]remain supported.Test coverage
warp/tests/tile/test_tile_composite_row.pycoverage exercises matrix-row reads and writes for 1Dâ4D tiles, negative row indices, read/write adjoints, and overwrite-gradient clearing;TestTileCompositeRowis registered in the default and debug suites.warp/tests/matrix/test_mat.py::test_mat_array_extractcovers chained row extraction from a 2D matrix-valued array, including gradients.Behavioral verification
The issue's rank-3 tile expression produced the expected matrix row on CUDA, and the rank-2 array expression produced the expected row on both CPU and CUDA.
Add matrix row indexing support (NVIDIAGH-1028)