Understanding HIP and WMMA Intrinsics

This project is a personal exploration of HIP programming and the RDNA3 Wave Matrix Multiply-Accumulate (WMMA) intrinsic. The primary goal was to deepen my understanding of the WMMA intrinsic and extend the fixed-size example provided in the GPUOpen tutorial to support arbitrary matrix dimensions. While this project is primarily for personal learning, it may also serve as a helpful reference for others interested in exploring the WMMA intrinsic.

Note: The WMMA intrinsic is specific to RDNA3 GPUs for now, so running this project requires an RDNA3-compatible GPU. A future feature may include testing this implementation on RDNA4 hardware when it becomes available. For production-grade GPU matrix multiplication, it is highly recommended to use rocWMMA, which provides a robust and optimized abstraction over the WMMA functionality.

Objective

This project aims to:

1. Provide a simple example of HIP programming and WMMA usage for GPU-accelerated computation.
2. Extend beyond the fixed-size example in the GPUOpen tutorial by supporting arbitrary matrix dimensions (¨M, N, K¨).
3. Enhance understanding of the WMMA intrinsic's mechanics, especially around data loading and storing.

Features

• Matrix Multiplication with WMMA Intrinsic: Demonstrates how to use HIP-specific WMMA (various implementations).
• Support for Arbitrary Sizes: Goes beyond the fixed-size (16x16) example, allowing users to experiment with any matrix dimensions.
• Shared Memory and WMMA Comparison: Runs both a shared memory kernel and a WMMA kernel for performance comparison.
• Verification Mode: Compares GPU results with CPU reference computations to ensure correctness.

Future Plans

If time permits, the following enhancements are planned:

1. Further optimisations of WMMA HGEMM kernel: Explore other optimisations to further improve WMMA HGEMM kernel.
2. Explore WMMA in Other Kernel Types: Investigate the use of WMMA intrinsics in other GPU workloads beyond matrix multiplication.
3. Test on RDNA4: Extend the implementation to test and validate the WMMA intrinsic on future RDNA4 hardware.

How to Build and Run

Prerequisites

• AMD ROCm installed with HIP support.
• CMake version 3.10 or higher.
• AMD RDNA3 GPU

Steps

1. Clone the repository and navigate to its root directory:

   git clone https://github.com/AJcodes/hip_wmma_samples.git
   cd hip_wmma_samples

2. Build the project:

   mkdir build
   cd build
   cmake ..
   make

3. Run the executable:

   ./hgemm

Usage

The program outputs performance metrics for all kernels, such as:

GEMM Kernel Type: XXX
----------------------------------------------------------------------------------
Kernel execution time for sizes (128, 128, 128): X.XX ms
...
Kernel execution time for sizes (1024, 1024, 1024): X.XX ms
...
Kernel execution time for sizes (4096, 4096, 4096): X.XX ms
----------------------------------------------------------------------------------

You can modify the matrix dimensions by changing M, N, and K variables in main.cpp.

Key Insights

• This project emphasizes understanding the mechanics of HIP programming and RDNA3 WMMA intrinsics, particularly how to handle data loading and storage effectively.
• It is intended as a learning tool and not as an optimized implementation for production use.

Acknowledgments

This project was inspired by:

• The GPUOpen RDNA3 WMMA Tutorial.
• The excellent work on rocWMMA, which should be used for real-world applications involving WMMA.