Mesa 26.1 Brings Unified Memory Support to AMD APUs with Rusticl

After nearly a year in development, Mesa 26.1 will include a critical fix for AMD APU users leveraging the open-source Rusticl OpenCL driver. The improvement, authored by Karol Herbst of Red Hat, addresses a fundamental limitation in how RadeonSI reported memory capabilities on systems with integrated Radeon graphics.

The Unified Memory Problem

AMD APUs present a unique challenge for graphics drivers. Unlike discrete GPUs with dedicated VRAM, APUs share system memory between the CPU and integrated GPU. This unified memory architecture should theoretically offer significant performance advantages, but only if the driver properly exposes this capability to applications.

Prior to this fix, RadeonSI incorrectly reported memory characteristics on APUs, leading to suboptimal OpenCL performance. Applications couldn't properly leverage the large system memory available alongside the small dedicated VRAM, resulting in unnecessary memory transfers and reduced efficiency.

Technical Details of the Fix

The solution involves a one-line change to RadeonSI that correctly identifies APU configurations and reports unified memory status. Additionally, the fix implements CL_KERNEL_GLOBAL_WORK_SIZE for custom devices using Rusticl, and addresses global memory reporting on mixed memory devices.

This improvement is particularly significant because it enables proper OpenCL 3.0 functionality on AMD APUs without requiring users to adopt ROCm, AMD's proprietary driver stack. For many users, especially those in Linux distributions that prioritize open-source drivers, this represents a major step forward.

Performance Implications

With unified memory properly reported, applications can now make intelligent decisions about memory allocation and data movement. This should result in:

• Reduced CPU-GPU data transfers
• Better memory utilization on systems with limited VRAM
• Improved performance in OpenCL workloads
• More accurate work size calculations for kernels

The timing is particularly relevant as Mesa 26.1 enters its stabilization phase ahead of the expected release later this quarter.

Context and Timeline

Interestingly, this merge request was authored in July 2024 but only landed in Mesa 26.1-devel on April 13, 2026. This delay highlights the careful review process Mesa undergoes, particularly for changes affecting core driver functionality.

For users running current Mesa versions on AMD APUs, this fix will be available once Mesa 26.1 releases, likely in the coming weeks. Those eager to test can track the development branch or wait for the stable release.

The improvement underscores the ongoing maturation of open-source graphics drivers for AMD hardware. As Rusticl continues to evolve, users can expect increasingly competitive OpenCL performance without proprietary dependencies.

What This Means for Users:

• APU owners will see better OpenCL performance in supported applications
• The fix enables proper OpenCL 3.0 feature utilization
• No need to switch to ROCm for basic OpenCL functionality
• Particularly beneficial for systems with limited dedicated VRAM

This change represents another step in AMD's open-source graphics driver ecosystem, providing users with more options and better performance on integrated graphics platforms.