For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo module

tile_loader

TileLoader module for efficient tile loading in GPU matrix multiplication.

This module provides utilities for loading matrix tiles from global memory to shared memory using two different mechanisms:

1. TMA (Tensor Memory Accelerator): Hardware-accelerated loads that can efficiently transfer 2D tiles with multicast support for multi-block clusters.

2. cp.async: Software-based asynchronous copy instructions with manual bounds checking and swizzling for optimal shared memory access patterns.

The TileLoader struct abstracts these loading mechanisms to provide a unified interface for the matmul kernel's producer threads.

Structs

• ​CPAsyncBarrierHandler: The cp.async barrier handler: noop on prepare, arrives on complete.
• ​TileLoaderCPAsync: Software-based tile loader using cp.async instructions.
• ​TileLoaderTMA: TMA-based tile loader for hardware-accelerated memory transfers.
• ​TMABarrierHandler: TMA barrier handler: sets expected bytes on prepare, noop on complete.

Traits

• ​BarrierHandler: Handles barrier lifecycle for different transfer mechanisms.
• ​TileLoader: Base trait for tile loading mechanisms in matrix multiplication.

Functions

• ​async_copy_with_bound_check: Helper function for cp.async with boundary checking.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit