triton.experimental.gluon.language.nvidia.blackwell.tma.async_scatter

triton.experimental.gluon.language.nvidia.blackwell.tma.async_scatter(tensor_desc, x_offsets, y_offset, src, cache_policy=None, _semantic=None)

Asynchronously scatter elements from shared memory to global memory using TMA.

Parameters:

  • tensor_desc (tensor_descriptor) – The tensor descriptor.

  • x_offsets (tensor) – 1D tensor of X offsets.

  • y_offset (int) – Scalar Y offset.

  • src (tensor_memory_descriptor) – The source data, must be in NVMMASharedLayout.

  • cache_policy – A compile-time CachePolicy containing only an L2 eviction policy. None leaves the policy unspecified. Cache policies are performance hints and do not affect synchronization.