triton.experimental.gluon.language.amd.cdna4.compute_efficient_padded_shared_layout

triton.experimental.gluon.language.amd.cdna4.compute_efficient_padded_shared_layout()

Compute an efficient padded shared layout that avoids bank conflicts.

Parameters:

  • dot_operand_layout (DotOperandLayout) – The layout for the dot operand that will be copied to shared memory with padding. Must have an AMDMFMALayout v4 (CDNA4) parent.

  • shape (List[int]) – Shared memory tile shape for the dot operand — [BM, BK] for operand A or [BK, BN] for operand B.

  • dtype (dtype) – Element type of the tensor that will live in this shared memory allocation (e.g. ttgl.float16, ttgl.float8e4nv). Only types with bitwidth in {4, 8, 16} are supported. For packed fp4 (two values per byte), pass ttgl.uint8 — at the LDS level 4-bit shares the 8-bit padding pattern.

  • is_k_contig (bool) – K is the contiguous dim in shared memory.

Returns:

The layout, or None if the input falls outside the supported set. Common reasons for None: k_width not in {4, 8, 16}; element bitwidth not in {4, 8, 16}; or an MFMA instruction shape or kWidth combination not handled by the underlying algorithm.