gluon-pipeline-opt

Community

Hide memory latency with dual-stage prefetch.

Authorleonling-ll
Version1.0.0
Installs0

System Documentation

What problem does it solve?

Two-stage pipeline optimizations to hide global memory latency and LDS read latency in Gluon GEMM kernels on CDNA3 (gfx942) and CDNA4 (gfx950) GPUs, enabling higher MFMA throughput and better occupancy.

Core Features & Use Cases

  • Stage 1: Global prefetch with double buffering to overlap data loads with compute across CDNA3/CDNA4 variants.
  • Stage 2: Local prefetch to overlap ds_read with MFMA, architecture-agnostic.
  • Use Case: Optimizing a Gluon GEMM kernel on MI300X/MI350 class GPUs to reduce vmcnt/lgkmcnt stalls and improve MFMA efficiency.

Quick Start

Apply Stage 1 to your kernel, verify correctness, then consider Stage 2 to further hide remaining stalls.

Dependency Matrix

Required Modules

None required

Components

Standard package

💻 Claude Code Installation

Recommended: Let Claude install automatically. Simply copy and paste the text below to Claude Code.

Please help me install this Skill:
Name: gluon-pipeline-opt
Download link: https://github.com/leonling-ll/claude-skills/archive/main.zip#gluon-pipeline-opt

Please download this .zip file, extract it, and install it in the .claude/skills/ directory.
View Source Repository

Agent Skills Search Helper

Install a tiny helper to your Agent, search and equip skill from 471,000+ vetted skills library on demand.