Release notes#

Versioning#

MathDx Tile is versioned and released with libmathdx: the mathdx_tile-<os>-<arch>-<version> archive, the nvidia-libmathdx-cu13 wheel, and the libmathdx inside them share one version. A release may add a fourth digit, such as 0.5.0.1. Headers and the CMake package report three. libmathdx changes are in the libmathdx release notes.

Compatibility#

Each MathDx Tile release pairs with the CUDA toolkit in its row and, for the Python API, with the nvmath-python, cuda-tile, and cuda-tile-preview releases.

MathDx Tile

nvmath-python

cuda-tile and cuda-tile-preview

CUDA toolkit

0.5.0

1.0.1 or newer

1.8.1 or newer, below 1.9

13.5

0.5.0#

  • First release of the MathDx Tile APIs, as a preview feature built on the TileIR SIMT preview mode.

  • New functionalities:

    • cuFFTDx: complex and real forward and inverse FFTs inside a Tile kernel, and suggested_tile_config for measured configurations. See cuFFTDx in Tile and FFT cutile-python API.

    • cuBLASDx and cuSolverDx: Cholesky, LU, QR, and LQ factorizations and their solves, triangular and tridiagonal solves, Hermitian and generalized Hermitian eigensolvers, SVD, and least squares. See cuBLASDx in Tile, cuSolverDx in Tile and Linear Algebra cutile-python API.

    • nvCOMPDx: LZ4 and ANS compression and decompression with the size and alignment helpers the buffers require. See nvCOMPDx in Tile and Compression cutile-python API.

    • CUDA Tile C++ headers and the mathdx_tile CMake package, with one imported target per library. See Getting Started.

    • The nvmath.tile cuTile Python modules fft, linalg, and compress, shipped in the nvidia-libmathdx-cu13 wheel.

    • Mirrored C++ and Python example suites. See Examples.