triton-ascend-case-index-put

What problem does it solve?

This skill demonstrates an optimization for index-based writes (index_put) in Triton-Ascend, reducing memory latency by batching index data into the Unified Buffer (UB) and reusing it inside loops via get_element, thereby avoiding repeated global memory access.

Core Features & Use Cases

• Batch-loads index data into UB before the loop.
• Reuses loaded data inside the loop with tl.get_element.
• Applies to irregular memory access patterns requiring multiple accesses to the same indices.
• Demonstrates performance improvements for Triton-Ascend workflows.

Quick Start

Run the Triton-Ascend case to batch-load indices into UB and reuse them with get_element inside the loop.