Mesa 26.3.1 Release Notes / 2026-11-18

Mesa 26.3.1 is a bug fix release which fixes bugs found since the 26.3.0 release.

Mesa 26.3.1 implements the OpenGL 4.6 API, but the version reported by glGetString(GL_VERSION) or glGetIntegerv(GL_MAJOR_VERSION) / glGetIntegerv(GL_MINOR_VERSION) depends on the particular driver being used. Some drivers don’t support all the features required in OpenGL 4.6. OpenGL 4.6 is only available if requested at context creation. Compatibility contexts may report a lower version depending on each driver.

Mesa 26.3.1 implements the OpenCL 3.1 API, but the version reported by the CL_DEVICE_VERSION, CL_DEVICE_NUMERIC_VERSION and CL_DEVICE_OPENCL_C_ALL_VERSIONS clGetDeviceInfo queries depends on the particular driver being used.

Mesa 26.3.1 implements the Vulkan 1.4 API, but the version reported by the apiVersion property of the VkPhysicalDeviceProperties struct depends on the particular driver being used.

SHA checksums

TBD.

New features

  • None

Bug fixes

  • Confidential issue #16794

  • Fails to build on macOS < 10.12

  • RadeonSI syncrhonization/barrier issue with Clear+Dispatch (Repro included)

  • V3D: Labwc fails to allocate GBM buffers wider than 4096 pixels unless non-MSAA texture size limit is enabled

  • VC4: Kodi on Pi 3: DRM PRIME / V4L2 fails after Mesa 26.2.6 with avfilter_graph_config: Function not implemented (-38)

  • Windows ARM memory-leak

  • Windows: u_memstream leaks MEM*.tmp files on SPIR-V shader compilation, stalling Lavapipe

  • [ANV] Dragon’s Dogma 2 Dark Arisen Glitch Shadows

  • [OpenMW] Certain shaders cause visual corruption at the bottom of the screen when anti-aliasing is enabled, possibly Z-buffer related

  • anv: compute shader writes to a storage image are partly lost after a render pass cleared the image earlier in the same command buffer on Intel LNL

  • d3d12: NULL dereference in select_shader_variant() when compile_nir() fails

  • d3d12: glClearTexImage through a texture view of another format flushes float32-denormal bit patterns to zero

  • dozen: Driver should fail to enumate anything if requirements are not met

  • dzn: Invalid minImageTransferGranularity specified on graphics/compute queue families

  • dzn: Unaligned block-compressed textures are silently unsupported

  • dzn: external memory buffers cannot be host mapped

  • dzn: maxStorageBufferRange reports 128 MiB, but raw storage buffer views allow 512 MiB

  • dzn: use different bit pattern in device addresses

  • dzn: vkBindBufferMemory fails on memory imported from a host pointer

  • nir_to_dxil has store_def and store_ssa_def and they’re subtly different

  • radeonsi: NULL pipe_memory_object dereference in si_resource_from_memobj()

  • radv/aco: fast-linked GPL pipeline breaks fragment shader interlock on GFX11 (no s_waitcnt_vscnt before the jump to the PSepilog)

  • radv: SIGSEGV in the hang-dump path with RADV_DEBUG=hang (null deref at 0x1a8) — no radv_dumps written

  • radv: VK_EXT_descriptor_heap causes performance slow down in FSR 4.1 workloads

  • radv: attachment shading rate not broadcast to all views

  • radv: fails to decompress when acquiring modifier-tiled images from the foreign queue

  • rusticl/llvmpipe: heap corruption / SIGSEGV in `llvmpipe_set_global_binding` when kernels on two queues are released by one `clWaitForEvents`

  • vkCmdDrawIndexed read access violation in Dozen (in dzn_descriptor_heap_copy)

  • zink/ntv: pack_snorm_2x16 and unpack_snorm_2x16 unimplemented, aborts the process

Changes

Aitor Camacho (1):

  • kk: Use constant address space for load_global_constant_offset

Ali Milhim (3):

  • zink: implement snorm/unorm pack and unpack ops in ntv

  • nir/opt_copy_prop_vars: don’t create a copies array just to kill nothing

  • nir: use temporary deref paths for alias invalidation

Andy Hsu (1):

  • intel/compiler: Handle load_barycentric_at_sample in lower_msaa_config

Arisa Snowbell (1):

  • anv/video: give VDEnc an intra prediction row store for H.264 on Gfx12.5

Ashish Chauhan (2):

  • pvr: fix XFB byte-count draw with unaligned counter offsets

  • pvr: Validate XFB primitive topology at draw

Caio Oliveira (7):

  • nir/intel: Derive push-data divergence from its offset

  • spirv: Move constant handling to its own file

  • util: Initialize memstream outputs on Xbox

  • jay: Use right stride helpers for destination

  • jay: Don’t force alignment for FLAG registers

  • jay: Scalarize indirect push-data loads

  • nir: Track cooperative matrix memory accesses

Calder Young (8):

  • intel: Exclude helper lanes for optimized divergent atomics

  • jay: Fix missing dependency on uses_src_xy when calculating coarse pixel depth

  • jay: Preserve loop header flags when splitting blocks for demote

  • jay: Do not “end blocks” after a HALT_TARGET instruction

  • jay: Make jay_opt_predicate consider HALT to be a jump

  • jay: Stop pointlessly setting uses_vmask to true

  • jay: Lower terminate to demote after jay_populate_prog_data

  • jay: Lower demote to terminate unless the shader uses derivatives

Christian Gmeiner (7):

  • llvmpipe: Track global bindings per context

  • panvk: Correct maxUniformBufferRange for older archs

  • etnaviv: Allocate the planes of an NV12 resource

  • gallium: Add pipe_caps.needs_split_planar_samplers

  • freedreno: Set needs_split_planar_samplers

  • etnaviv/drm: Fail BO creation when no GPU address is left

  • mesa/st: Lower YUV only when the driver needs split samplers

Daniel Drake (1):

  • torx: Depend on idep_mesautilformat for generated format headers

Daniel Schürmann (1):

  • aco/tests: fix GFX11.7 unused source encoding expectations

Dave Airlie (1):

  • gallivm: disable apx feature for now.

David Rosca (7):

  • dri: Move driDestroyConfigs from glx/dri_common and use it

  • glx: Fix driver_configs leak when creating display gpu screen

  • egl/dri2: Fix driver_configs leak when creating display gpu screen

  • egl/wsi/wayland: Fix leaking dmabuf feedback

  • mesa: Handle import fail in ImportMemoryFd/Win32EXT

  • mesa: Handle import fail in ImportSemaphoreFd/Win32EXT

  • mesa: Fix crash in Wait/SignalSemaphoreEXT with null fence

Derek Lesho (4):

  • zink: Use zink uploader for clientptr maps.

  • zink/client_pointer: Flush uploader memory on unmap.

  • zink: Make map/unmap_clientptr_range thread-safe.

  • zink: Respect FLUSH_EXPLICIT on unmap_clientptr.

Dhruv Mark Collins (3):

  • tu/autotune: Order dropping an RP history reference before its reaping

  • tu/autotune: Remove latency tracking entries once their spikes are old

  • tu/autotune: Take one RP history reference per RP decided

Dmitry Baryshkov (11):

  • freedreno: keep the real subgroup size for OpenCL kernels

  • freedreno/ci: drop the fixed a702 uniform texel buffer barrier tests

  • freedreno: round up the NV12 chroma plane size

  • dri: round up the subsampled plane size on dma-buf import

  • freedreno/a5xx: don’t write a magic 1 to HLSQ_CS_CNTL_1

  • freedreno/a5xx: the a509 and a512 register files hold 96 vec4 per wave

  • freedreno/a5xx: correct a505, a506 and a510 entries

  • rusticl: fix the contiguous-rows check in sw_copy

  • ir3: count half registers when deciding to double the thread size

  • ir3: read a5xx SSBO sizes with the 4d resinfo

  • freedreno/a5xx: stop undoing the kernel’s DBG_ECO workarounds

Duncan Brawley (2):

  • pvr: Fix pipeline executable properties segfault

  • pvr: Support pipeline executable properties VkPipelineCreateFlags2 flags

Dylan Baker (11):

  • relnotes: Add SHA sums for 26.3.0

  • docs: update calendar for 26.3.0

  • .pick_status.json: Update to e0856f57b87dd9731a4b738ea739d0b163a31510

  • .pick_status.json: Update to b663f3501c6e42ee92d388f91abe3cf71f78187c

  • .pick_status.json: Update to 360e61e7418a0967590029eebeeceecff46755e2

  • .pick_status.json: Mark 77ed7926bf94c263fcb31f0be7b4e3f9fe1a20e9 as denominated

  • .pick_status.json: Update to beaf3192c2e0a3b4fcc05362c0ca9fc198a4e236

  • .pick_status.json: Mark af46d968ba7798f23b9937f4936439ee36f0dc4c as denominated

  • .pick_status.json: Update to 6aa7eaf09fc1ae2d53adfa8a29fd695278111c62

  • .pick_status.json: Update to 8fae0a4519445af71501dc183b3450d8f407bbf1

  • .pick_status.json: Mark a364749b694b8cef8f8b8dcdc5cd9a61c8751aa0 as backported

Emma Anholt (2):

  • glsl: Fix linking of various FS inout layout qualifiers pre-GLSL-1.50.

  • ir3: Give is_same_type_mov() an option to allow src mods or not.

Erhan Bilgili (1):

  • winsys/amdgpu: Fix UserQ signal point array allocation.

Eric R. Smith (1):

  • panfrost: make sure render_target_mask is non-empty if PLS is enabled

Erik Faye-Lund (11):

  • nir: use dst_access for the store

  • panvk: restore zero-value for flush_id after copy_fbd

  • panvk: add missing flush after copy_fbd

  • spirv: specialize entire constant-data source

  • pan/ci: restore flaky tests

  • pan/ci: remove stale flakes

  • pan/ci: remove stale flakes

  • pan/ci: remove stale flake

  • pan/ci: remove stale flakes

  • pan/ci: remove stale flakes

  • pan/ci: remove stale flake

Fabricio Akio (2):

  • freedreno/a5xx: Program the stream output buffer stride

  • freedreno/a5xx: Implement alpha-to-one

Faith Ekstrand (1):

  • panfrost,nir: Get rid of nir_downgrade_pls_vars()

Francesco Stimola (7):

  • nv50: size surface views by blocks when the block size differs

  • nv50,nvc0: don’t read pipe_surface templates as nv50_surface in clears

  • nv30: don’t read pipe_surface templates as nv30_surface in clears

  • nouveau/winsys: pass the drm fd when deleting subchannel objects

  • nv50: only push constant buffer updates through a current binding

  • nouveau: pass chroma_format_idc to the vp3 H.264 decoder

  • nouveau: tell the vp3 H.264 decoder whether a picture is MBAFF

Francisco Jerez (1):

  • intel/gen: Fix possible corruption of instructions other than standard one/two-source layout during pre-compaction.

Frank Binns (2):

  • docs: drop references to ROGUE_DEBUG & ROGUE_COLOR envvars

  • docs: fix PCO related envvar name

Georg Lehmann (5):

  • aco/opt_postRA: check quad alignment of sgprs

  • aco/opt_postRA: fix split_vector reassign with late kill operands

  • nir/opt_load_store_vectorize: handle load_heap_descriptor

  • nir/opt_large_constants: fix constant deduplication edge case

  • nir/range_analysis: fix 8/16bit ishr uub

Jeongik Cha (2):

  • gfxstream: guard Goldfish sync fd collection in queue submit

  • gfxstream: handle WSI dummy sync objects and reset temporary state

Jesse Natalie (1):

  • d3d12: Fix hash lookup for command signatures

Job Noorman (6):

  • ir3: don’t emit alias.rt for relative const movs

  • ir3: move is_cat[23]_float helper up in their file

  • ir3: only fold fsat to (sat) on cat2/3 float instructions

  • ir3: add has_valid_alias_flags helper

  • ir3: check src flags of aliased collects

  • ir3: don’t create aliases for array srcs

Jose Maria Casanova Crespo (4):

  • broadcom/compiler: align shared memory size to zero-initialization chunk

  • vc4: compute the clip window after the predraw blits

  • v3d/driconf: expose non-MSAA texture limits for labwc

  • v3d: fix the order of the alpha sampler state variants

José Roberto de Souza (1):

  • intel/dev: Update Xe3P PAT entries

Juan A. Suarez Romero (6):

  • broadcom/compiler: write fragment shader scratch from helper invocations

  • v3dv: track memory budget heap usage per process

  • v3dv: apply alpha to coverage with single-sampled rasterization

  • v3dv: use the image depth/stencil format for any Z/S tile buffer access

  • v3dv: update conformance version

  • v3dv: disable early-Z for every job in a suspend/resume chain on V3D 4.2

Lionel Landwerlin (2):

  • anv: fix PIPE_CONTROL with fragment->compute barrier on Xe2+

  • anv: use the actual size for vma allocations

Louis-Francis Ratté-Boulianne (1):

  • vulkan: initialize swapchain private mutex properly

Marek Olšák (4):

  • radv: fix invalid use of radv_image_has_hiz on GFX6-8

  • radv: suspend conditional rendering in CmdResetQueryPool

  • radv: suspend conditional rendering in radv_barrier

  • radv: make buffer clears & copies respect conditional rendering

Matt Arsenault (1):

  • radv/llvm: keep the ring offsets in s[0:1] for GFX11+ compute shaders

Mel Henning (1):

  • nir: Fix incorrect is_ssa replacement

Mike Blumenkrantz (2):

  • glsl: treat volatile as coherent when translating to nir

  • zink: reset attachment storeOp when begin_rendering has no tc info

MoGGuU (1):

  • rocket: allocate separate arrays for each job to avoid double-free

Mohamed Akram (1):

  • glx/apple: fix build on macOS < 10.12

Oliver Clarke (3):

  • hk: fix shader dump use after free

  • hk: remove input attachment offset from tex instructions

  • radv: do layout transition on foreign queue acquire when needed

Olivia Lee (2):

  • pan/va: fix out-of-bounds read propagating swizzles on pseudo instructions

  • panvk: fix null pointer dereference in shader variant helpers

Omar Rashwan (1):

  • iris: fix double free in iris_bufmgr

Patrick Lerda (1):

  • r600: fix shaders.indexing.tmp_array.vertexid related issue

Paulo Zanoni (2):

  • vulkan: export vk_queue_stop_submit_thread()

  • anv: extract and call anv_queue_drain() during anv_DestroyDevice()

Pavel Ondračka (1):

  • r300: skip output modifier absolute or negated inputs

Pierre-Eric Pelloux-Prayer (2):

  • gallium/ddebug: implement resource_release

  • radeonsi: fix debug dump of buffer descriptors

Raphtaliyah (2):

  • radv: Mask device coherent memory types if the feature is not enabled

  • loader: fix drm device memory management issues

Rhys Perry (2):

  • nir/uub: fix imul by constant with overflow

  • aco: consider p_jump_to_epilog as a POPS end export

Rusi Genev (3):

  • v3d: Use the initial sample mask for centroid interpolation

  • v3d: Fix interpolateAtOffset without MSAA

  • v3d: Fix centroid interpolation at offsets

Samuel Pitoiset (7):

  • radv: fix copying VRS rates for multiview or layered rendering on RDNA2

  • spirv: fix memory semantics for image loads/stores

  • aco: make sure SSBO SMEM loads respect barriers

  • nir/lower_memory_model: don’t skip readonly accesses

  • radv: split copies after optimizing memcpy

  • radv: do not try to dump VBOs without a VS with RADV_DEBUG=hang

  • radv: fix emitting push constants with ESO

Serdar Kocdemir (1):

  • util: avoid Windows memstream temp file leaks

Simon Perretta (1):

  • pco: fence before mutex release ops

Tapani Pälli (1):

  • anv/dric: add workarounds for Dragon’s Dogma 2

Timothy Arceri (5):

  • ac,radv: move the pipe-misaligned metadata check to common code

  • radeonsi: invalidate L2 before shaders read pipe-misaligned HTILE

  • radeonsi: decompress resident depth textures rendered to after becoming resident

  • radeonsi: flush the IB on NGG -> legacy transitions on Navi21 too

  • radeonsi: write back L2 after DCC fast clears if RBs and L2 aren’t coherent

Valentine Burley (6):

  • tu: Lower terminate to demote in fragment shaders

  • tu: Advertise VK_KHR_shader_maximal_reconvergence

  • panfrost: Remove redundant perfetto header include

  • tu: Disable LRZ for imported depth images

  • Revert “tu: Reject VK_FORMAT_S8_UINT AHB import”

  • vulkan/android: Force linear for depth AHBs when compression disabled

Vinson Lee (2):

  • kopper: fix unused and uninitialized warnings on macOS

  • util/u_trace: Fix UtilPerfTraceTest.EndOfFrameOnce unit test on macOS

Yiwei Zhang (2):

  • anv: allow re-importing BOs with mismatched device address flags

  • zink: gate conditional rendering buffer usage behind the extension

Zan Dobersek (1):

  • freedreno: update SP_DBG_ECO_CNTL magic reg value on A740

Zmicier Prybysh (1):

  • ethosu: Fix same-size RESIZE_BILINEAR case

alexandre minette (1):

  • freedreno/ir3: use implicit LOD zero for a3xx multisample fetches

hamb3r (18):

  • dzn: Handle missing variable descriptor count layout support output

  • d3d12: Preserve raw texel bits in clear_texture

  • dzn: Report 512 MiB for maxStorageBufferRange

  • dzn: Pass modified resource desc in dzn_BindBufferMemory2

  • dzn: Enumerate zero devices when runtime requirements are not met

  • dzn: Handle descriptor heap allocation failure in cmd_buffer_update_heaps

  • dzn: Preserve host-visible memory types for host-imported buffers

  • dzn: Set minImageTransferGranularity to (1, 1, 1)

  • dzn: Apply all dependencies in CmdWaitEvents2

  • d3d12: Clean up failed shader variant allocations in compile_nir()

  • d3d12: Hook up DXIL logger in compile_nir()

  • d3d12: Handle shader variant compilation failures

  • dzn: Fix resource flags for MSAA and concurrent linear images

  • dzn: Handle D3D12 image allocation info failures

  • microsoft/compiler: Unify DXIL SSA definition storage

  • microsoft/compiler: Fix memory leaks on nir_to_dxil() error paths

  • dzn: Set bits 52-63 of buffer device addresses

  • dzn: Fix query result zero-fill copy sizes

mmes (1):

  • jay: set SLICE_REPACK operands to stride 4

mozixun (1):

  • zink: don’t charge a batch the whole slab for a suballocated resource

osman.yagci (2):

  • panfrost: handle a NULL batch in the remaining callers

  • mesa: don’t advertise APIs that are not built