Mesa 25.2.8 released
The latest release for Mesa 25.2 is out, version 25.2.8, effectively concluding this specific line of development. This means everyone should probably be heading over to the 25.3 branch itself, specifically targeting the more stable and frequently updated 25.2.3 sub-branch moving forward.
This update represents a significant effort from developers worldwide who've focused on ironing out problems in Mesa's codebase. The fixes cover various areas: Aitor Camacho tackled issues with Vulkan command queues and WSI presentations, while Anna Maniscalco specifically addressed legacy graphics information within the Radeon driver.
Elsewhere, Benjamin Cheng bolstered the Radeon driver's video support, ensuring compatibility with a range of encoding standards. Meanwhile, Calder Young worked diligently on Intel hardware bugs, including those affecting fused devices like Intel Arc GPUs, resolving problems related to FP16 sampling, particularly on Broadwell chips.
Christian Gmeiner updated the Meson build system requirements, mandating sysprof-capture-4 for proper operation now. Christoph Pillmayer fixed metadata sorting in NRI, and Daniel Schürmann handled dead dereference instructions in NIR optimization plus legacy Radeon graphics info again.
Over on Gallium3D, Dave Airlie ensured the Dozen driver correctly reports an incompatible status during instance creation failure. David Rosca contributed more fixes to Radeon video capabilities, which touches things like enabling encoding, fixing AV1 issues, and patching up problems with screen content tools and integer motion vector handling.
Compute shaders saw attention too; Emma Anholt plugged a memory leak in the Tianxia driver's compute shader pipeline. Eric Engestrom added SHA sums for release notes clarity, updated pick status JSON files, automatically disabled LLVM support when needed, and put together the release notes for this 25.2.8 version.
Eric R. Smith resolved an assertion failure in Bifrost disassembly on the Panfied driver (I assume you meant 'Pan' here). Erik Faye-Lund improved error handling for format queries in Mesa's main branch and made sure EXT_texture_buffer_object isn't enabled incorrectly, only with certain formats.
Faith Ekstrand fixed a problem with early returns when waiting for sync files using Vulkan DRM. Gert Wollny prevented unnecessary ALU-CF execution on newer Radeon 6000-series cards. For SVGA drivers, Ian Forbes checked for NULL stencil buffers to avoid crashes, and Ian Romanick handled spill allowances (a specific optimization detail) in the Broadwell driver and reduced fp16 instructions in Lavapipe.
Jianxun Zhang disabled BMG modifiers on the driver side, specifically for Intel's Xe2 architecture on the Isl graphics stack. Joshua Simmons fixed an OpCopyLogical issue regarding destination types within VTN drivers (Virtual Texture). Karol Herbst addressed some tricky kernel implementation issues related to Rusticl, including clGetKernelSuggestedLocalWorkSizeKHR and workgroup size limits.
The Panfied driver team, via Dave Airlie's earlier correction, now correctly reports failures. David Rosca also added fixes, which include enabling video encoding support specifically for the Pan hardware component within Mesa, addressing AV1 sequence coding problems, and resolving conflicts related to screen content tools and integer motion vector issues; this last point is typically a graphics term.
Tearing down memory leaks in the compute shader pipeline was Emma Anholt's task on Tianxia. Eric Engestrand contributed SHA hashes for release notes, updated JSON for picking status, auto-disabled LLVM support when needed, and wrote the release notes, demonstrating strong CI/CD integration.
Eric R. Smith fixed an assertion problem in the Bifrost disassembly output on Pan. Erik Faye Lundund improved Gallium3D core error handling for format queries in Mesa, specifically guarding against improper enablement of EXT_texture_buffer_object for RGBA formats to prevent downstream misuse.
Faith Ekstrand made another fix in the Vulkan world, this time related to DRM and sync file waits. Gert Wollny worked on optimization: avoiding unnecessary ALU-CF operations, particularly beneficial on Radeon 6000-series GPUs.
SVGA emulation received a NULL buffer check from Ian Forbes, while Ian Romanick handled variable spilling allowances (and FP16 lowering) within the Broadwell Gallium3D driver implementation and also in Lavapipe. That seems like quite a spread of fixes for him.
Christian Gmeiner's change to the Meson build requirements is worth noting; it now requires sysprof-capture-4, so users might need to check their dependencies. Tapani Pälli took action on Anvil drivers: he fixed metadata sorting issues and also disabled threaded context globally, plus applied specific intel_disable_threaded_context settings for games like Amnesia: The Bunker, which require it.
Sushma Venkatesh Reddy added a workaround layer specifically for Detroit: Become Human within drirc's Anvil driver configuration, preventing potential hangs there. For AMD's drivers, we have fixes from several individuals: Emma Anholt touched Tianxia (compute), Eric Engestrom did his usual comprehensive work across releases and features like SHA sums, etc., and Aitor Camacho fixed Vulkan command queues.
Karol Herbst plugged a specific kernel issue on Freedreno for A6xx hardware. Samuel Pitoiset addressed per-submit capture handling in RGP, specifically within Radeon driver contexts, reducing maxTexelBufferElements to prevent potential problems later.
NVIDIA's Venus hardware emulation saw fixes from Yiwei Zhang: he ensured ring CS and tail updates ordered consistently correctly on that architecture, plus plugged a racy semaphore feedback counter update issue. On the Intel side, Tapani Pälli also generally fixed threaded context issues and enabled specific intel_disable_threaded_context settings for compatibility with games like Amnesia.
Gert Wollny implemented SQTT buffer cleanup on resize. Natalie Vock handled culling flags properly in RADV ray tracing, plus managed VMID reservations during device creation to prevent future conflicts. Vladimír Ćunďo probably could have been here, but he wasn't listed.
Patrick Lerda focused on Radeon 6000-series compatibility, checking read scratch and error filter handling there specifically. Mel Henning fixed bugs within the Zink driver interface (probably Gallium-based?), dealing with screen queue locking issues and managing debug marker interactions properly.
Daniel Schürmann's earlier remedy covered dead code paths in NIR optimization and revisited legacy graphics information on Radeon again, which suggests it might be a persistent issue or needed for older hardware support. And Dave Airlie handled the failure case for creating Dozen driver instances correctly.
The source code for Mesa version 25.2.8 is available for download at this location.
