IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Mojo function

attn_tmem_addr

def attn_tmem_addr[cluster_size: Int]() -> UInt32

Returns the TMEM base, 0, as a value ptxas cannot fold.

A literal 0 makes ptxas rebuild every TMEM operand per tcgen05.mma (an IMAD.MOV plus an R2UR per use) instead of keeping one uniform register; inline asm cannot hide it, since ptxas folds the PTX immediate. A CTA's cluster rank is below cluster_size, so shifting it right by log2_ceil(cluster_size) is a run-time 0 that costs an S2R and a shift. Inline asm hides the rank from LLVM, which folds cluster.ctarank to 0 under nvvm.cluster_dim = (1, 1, 1), and ptxas folds neither the rank nor the shift. %cluster_ctaid.y is also 0 in these 1-D clusters, but ptxas computes it as %ctaid.y modulo the cluster shape: three constant-bank loads and a reciprocal division.

Parameters:

  • ​cluster_size (Int): CTAs per cluster, or 1 when launched without a cluster.

Returns:

UInt32

Was this page helpful?