References#
CUDA Tile#
Writing Tile Kernels [1] — the Tile programming model, for both CUDA Tile C++ and cuTile Python. The reference for everything MathDx Tile builds on:
ct.load/ct.store,partition_view, tile operations, and the interop limitations described in MathDx Tile APIs.CUDA Tile C++ API Reference [3] — the C++ API in full. General Principles [4] defines what a tile extent may be, including the integral-constant wrappers and the
_icliterals that Achieving High Performance refers to.cuTile Python [5] — the Python API in full, including its own autotuning guide [6].
MathDx#
MathDx Tile exposes a subset of the MathDx device libraries. Their classic device APIs remain the reference for what each routine computes. cuRANDDx has no Tile API.
Python#
nvmath-python device APIs [13] — the Python bindings to the classic MathDx device APIs, the non-Tile counterpart to cutile-python APIs