Can this skill help me convert existing CUDA projects?

Yes, it provides specific mapping tables for CUDA-to-HIP equivalents and instructions on using tools like hipify-perl for automated code conversion.

How do I handle AMD-specific wavefront sizes?

The skill emphasizes using the 'warpSize' built-in rather than hardcoding values, as AMD typically uses a wavefront of 64 threads compared to NVIDIA's 32-thread warp.

What is the main benefit of using HIP over CUDA?

HIP (Heterogeneous-compute Interface for Portability) allows you to write code that is syntactically similar to CUDA but can be compiled to run on both AMD and NVIDIA GPUs, providing better hardware flexibility.

Does it support the latest AMD GPU architectures?

It includes support for various architectures including RDNA2 (gfx1030), RDNA3 (gfx1100), CDNA2 (gfx90a), and CDNA3 (gfx942) with specific compiler offload flags.

AMD ROCm & HIP Development

Name: AMD ROCm & HIP Development
Author: datathings

bydatathings

•

Data Science & ML

Develops high-performance GPU kernels and manages AMD ROCm compute stacks using HIP for cross-platform portability.

This skill provides comprehensive guidance for developers working within the AMD ROCm ecosystem, specifically focusing on HIP (Heterogeneous-compute Interface for Portability). It assists in writing optimized GPU kernels, integrating high-performance libraries like rocBLAS and rocFFT, and porting existing CUDA codebases to run efficiently on AMD hardware. Whether you are building CMake projects for CDNA/RDNA architectures or debugging with rocGDB, this skill ensures best practices for memory management, occupancy optimization, and asynchronous execution are followed.

Key Features

01HIP kernel development and syntax guidance for .hip and .cpp files

02CUDA-to-HIP automated and manual porting support using hipify tools

037 GitHub stars

04Build system configuration for CMake and Makefile targeting specific AMD GPU architectures

05GPU hardware profiling and debugging using rocProfiler and rocGDB

06Integration of optimized ROCm libraries including rocBLAS, rocFFT, and rocSOLVER

Use Cases

01Optimizing GPU memory throughput and occupancy for RDNA3 and CDNA3 architectures

02Implementing custom high-performance HIP kernels for scientific simulations

03Porting a CUDA-based machine learning library to run on AMD Instinct or Radeon GPUs

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add datathings/marketplace rocm

For use in Claude.ai and ChatGPT

Download Skill