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

masked_load_apple

Apple M5 (AGX3) hardware edge-masked vector loads.

build_edge_mask computes a per-lane in-bounds mask; the *edge_masked_load family does a predicated vector load that zeroes masked-off lanes, letting a kernel vectorize a boundary tile without a scalar remainder loop or an out-of-bounds read. The three entry points differ only in address-space resolution:

  • edge_masked_load — the pointer's own space (must be GLOBAL or SHARED).
  • gmem_edge_masked_load — reinterprets any pointer as device (global).
  • smem_edge_masked_load — reinterprets any pointer as threadgroup (shared).

Apple M5 only: the AGX3 intrinsics don't exist elsewhere, so callers dispatch on compute_capability() == 5.

Functions​

  • ​build_edge_mask: Computes the per-lane edge mask for a vector access (Apple M5 only).
  • ​edge_masked_load: Loads a predicated (edge-masked) vector in the pointer's own address space (Apple M5 only).
  • ​gmem_edge_masked_load: Loads an edge-masked vector, treating ptr as device (global) memory (Apple M5 only).
  • ​smem_edge_masked_load: Loads an edge-masked vector, treating ptr as threadgroup (shared) memory (Apple M5 only).
View source