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).
mma_apple
Apple Silicon MMA implementation for matrix multiply-accumulate operations.
This module provides MMA implementations for Apple M5 GPUs using the simdgroup_matrix hardware instructions (Metal 4.0 / AIR 2.8.0).
Supported operations:
- Float multiply-accumulate: {F16, BF16, F32} inputs, F32 accumulator
- Integer widening multiply-accumulate: {I8, U8} inputs, I32/U32 accumulator
Functions
-
apple_mma_load: Loads a 16x16 matrix fragment for the current simdgroup thread. -
apple_mma_store: Stores a 16x16 matrix fragment from the current simdgroup thread.