Module cuda 

Re-exports

pub use mma::manual::supported_mma_combinations;

pub use mma::manual::supported_scaled_mma_combinations;

Modules

arch

atomic

CUDA’s C++ atomic APIs are a mess of inconsistency. Old-style APIs use weird types, newer std::atomic APIs don’t support many types. So for ops where type support is complex, we lower the atomics directly to PTX. This allows a consistent API across all the different type and vectorization options, and significantly extends the interface that can be accessed from C++.

barrier

binary

builtin

convert

Cuda conversion functions

dialect

mma

packed_ops

plane

ptx

signature

tma

ty