MathDx Tile APIs#
The MathDx Tile APIs provide selected MathDx operations inside CUDA Tile kernels.
Use a single call for an FFT or a factorization inside a Tile kernel.
tileiras inlines the corresponding SIMT implementation into that kernel.
No launch boundary or round trip to global memory separates the tile code from the MathDx routine.
Two interfaces provide the same functionality. The C++ wrappers form groups by the library that implements each operation. Python provides three modules for FFT, linear algebra, and compression.
Requirements, installation, and two worked kernels in both languages.
How the interop works, plus the CTA size, tile layouts, and the conventions every MathDx Tile API shares.
Choosing a configuration, batching, fusion, and how to measure.
Mirrored C++ and Python example suites, with a learning path.
Warning
MathDx Tile is compatible with CUDA 13.5 only.
Note
MathDx Tile is a preview feature built on the TileIR SIMT preview mode. Coverage, performance, and API spellings may change between releases.