Open Mesa PanVK Vulkan driver for Android Mali GPUs, talking directly to the vendor mali_kbase kernel driver.
-0091BD?logo=arm&logoColor=white){kind=icon}
Important
🧪 Testers wanted!
- Install PanPlay and play your games.
- After every run, whether the game crashed or not, open the session logs screen and tap Send to cloud (or Share as ZIP).
- Send the link or ZIP to the Telegram testers group, with the game name, your GPU, the FPS you saw and any glitches.
A standalone patch and build layer around pinned upstream Mesa. It produces an
open Mesa PanVK driver that talks directly to Android's proprietary
mali_kbase kernel interface (/dev/mali0).
It is not a Samba-specific or Winlator-specific fork: one driver source/patch stack, with adapter packages around it.
Primary consumers
| # | Consumer | Integration |
|---|---|---|
| 1 | Samba S3 | Android/Bionic, in-process custom Vulkan driver |
| 2 | Bachata S4 | ARM64 glibc Vulkan ICD inside the managed runtime |
| 3 | NativeCode AI | ARM64 glibc PRoot validation and development |
Secondary (only once the primary driver is correct): Winlator or AdrenoTools-style loaders, GameHub or component injectors, and other Android apps that can load an alternate Vulkan ICD.
| App | What it is | Download | |
|---|---|---|---|
 |
PanPlay (apps/panvk-launcher) |
Windows game launcher (Wine + DXVK + built-in X server) with the PanVK driver bundled | PanPlay 1.2.7 · all releases · repo |
 |
PanProbe (apps/panvk-test) |
Vulkan feature/extension info and on-device driver tests | PanProbe 1.2.7 · all releases · repo |
Both apps are tested on the Mali-G615. PanProbe also runs on the other devices in Tested devices. See apps/panvk-launcher/docs for launcher usage.
| Phone | Poco X6 Pro (2311DRK48I, duchamp) |
| SoC | MediaTek Dimensity 8300-Ultra (MT6897) |
| GPU | Mali-G615 MC6, Pan arch v11, CSF frontend |
| Kernel interface | /dev/mali0 (mali_kbase) |
| GPU ID string | Mali-G615 6 cores r1p3 0xB8A3 |
The latest release is 0.1.0 RC4 (release candidate, 2026-10-20). RC4 adds
experimental Mali v7 Bifrost support (jm-v7 001-005). Only the G615 was
retested on RC4 (Poco X6 Pro, PanProbe 39/39); v7 was tested on the Mali-G52
in the Galaxy A06 and A14 (14/38). The rows below are from RC3. All rows marked
2026-10-11 ran the RC3 binary itself (Android .so BuildID 4ec75e25,
driverInfo PanVK-kbase RC3 (Mesa 26.3.0-devel)) with PanProbe 38 tests.
The Galaxy Tab S11 Ultra (v13) was not rerun on RC3; its row is from beta.18
RC1. RC3 adds the jm-v9 001-030 series (v9 JM: Vulkan 1.3, geometry
shaders, tessellation, transform feedback, multiViewport, clip/cull
distances, depth bounds, fill mode, conditional rendering, robustImageAccess2,
GPU timestamps, kbase JM fixes) and the v10 robustImageAccess2 alpha fix-up
on the GPU (csf-v11/184).
| Device | SoC | GPU (deviceName) | gpu_id | Arch | Android / kernel / kbase | Tested on | Driver build | PanProbe | PanPlay D3D11 cube |
|---|---|---|---|---|---|---|---|---|---|
| Poco X6 Pro (reference) | Dimensity 8300-Ultra (MT6897) | Mali-G615 MC6 | 0xb8a31030 |
v11 | 16 / 6.1 custom / CSF 1.21 | Own device, 2026-10-11 | RC3 | 38/38 | OK on beta.18 RC1; NFS: Most Wanted runs without stutter on a warm shader cache |
| Lenovo Tab TB336FU | MediaTek (mt8755) | Mali-G57 MC2 | 0x90930010 |
v9 (JM) | 16 / 5.15 / JM 11.38 | Own device, 2026-10-11 | RC3 | 38/38 (RC2 and earlier: 2/17) | OK: D3D11 FL11_0 used (max FL11_1), ~70-76 FPS; D3D8/9/10 cubes OK. NFS Most Wanted races at ~18 FPS (GPU-bound, no faults) |
| Pixel 7 | Tensor G2 (GS201) | Mali-G710 MC7 | 0xa8620004 |
v10 | 13 / 5.10.157 / CSF 1.14 | Firebase Test Lab, 2026-10-11 | RC3 | 38/38 | OK on beta.18 RC1 |
| Pixel 7a | Tensor G2 (GS201) | Mali-G710 MC7 | 0xa8620004 |
v10 | 13 / 5.10.149 / CSF 1.14 | Firebase Test Lab, 2026-10-11 | RC3 | 38/38 | n/a |
| motorola edge 40 neo | MT6879 | Mali-G610 MC3 | 0xa8670000 |
v10 | 14 / 5.10.218 / CSF 1.18 | Firebase Test Lab, 2026-10-11 | RC3 | 38/38 | 47 FPS on beta.18 RC1 |
| Pixel 8 | Tensor G3 | Mali-G715 MC7 | 0xb8a21020 |
v11 | 16 / 6.1.145 / CSF 1.38 | Firebase Test Lab, 2026-10-11 | RC3 | 38/38 | 48 FPS on beta.18 RC1 |
| Pixel 9 | Tensor G4 | Mali-G715 MC7 | 0xb8a21030 |
v11 | 16 / 6.1.145 / CSF 1.38 | Firebase Test Lab, 2026-10-11 | RC3 | 38/38 | n/a |
| Galaxy Tab S10 Ultra (SM-X920) | Dimensity 9300+ (MT6989) | Mali-G720-Immortalis MC12 | 0xc8700000 |
v12 | 14 / 6.1.75 / CSF 1.21 | Samsung Remote Test Lab, 2026-10-11 | RC3 | 38/38 (gs_viewport_depth 20/20) |
OK: FL11_0, ~47 FPS, no GPU fault or DEVICE_LOST. beta.18 RC1: NFS: Most Wanted race 81.7 FPS |
| Galaxy Tab S11 Ultra (SM-X930) | MT6991 | Mali-G925-Immortalis MC12 | 0xd8300015 |
v13 | 16 / 6.6.102 / CSF 1.30 | Samsung Remote Test Lab, 2026-10-07 | beta.18-dev+combined (not rerun on RC3) | 37/37 | 48.1 FPS |
| Xiaomi 2511FPC34G (user report) | MT6899 | Mali-G720 MC8 | 0xc8700010 |
v12 | n/a | Tester, beta.17 | beta.17 (release) | 35-36/37, gs_viewport_depth fails intermittently |
n/a |
On the motorola edge 40 neo, kbase rejects MEM_ALLOC_EX with ENOTTY, and
the MEM_ALLOC fallback from patch 110 works. More tester data (88 records
from 19 devices, up to beta.15) is in
docs/universal/DEVICES.md. v9 details:
worklogs/v9-jm/PROGRESS.md; RC3 validation:
worklogs/g615-dxvk/PROGRESS.md.
| Arch | Frontend | Status | Evidence |
|---|---|---|---|
| v7 | JM | Experimental from RC4 | RC4 (jm-v7 001-005): no env gate, GPU variant ignored, no MEM_JIT_INIT, growable tiler heap, atom stride 56/64 check, ION fallback. Mali-G52 MC2 r1 (Galaxy A06, 0x74021000) and G52 MC1 (Galaxy A14, 0x72120000): device creation works, PanProbe 14/38. GS, XFB, tessellation, robustness2 and MDI are in patches/jm-v7-wip, not in RC4. v6 takes the same path but is untested |
| v9 | JM | Supported from RC3 (RC2 and earlier: experimental, partly broken) | RC3 (jm-v9 001-030): Vulkan 1.3 and the DXVK FL11_0/FL11_1 features. G57 MC2 passes PanProbe 38/38, the D3D8-D3D11 cubes and NFS Most Wanted. G68/G77/G78/G78AE have model rows but are untested. RC2 and earlier: 2/17 real PanProbe passes, Vulkan 1.1 |
| v10 | CSF | Works on RC3 | G610 MC3 and G710 MC7 (Pixel 7, Pixel 7a) pass 38/38. G710 needs the model row from csf-v11/140, so beta.17 does not recognise it |
| v11 | CSF | Release target; works on RC3 | G615 MC6 (reference) and G715 MC7 (Pixel 8, Pixel 9) pass 38/38 |
| v12 | CSF | Works on RC3 | Immortalis-G720 MC12 passes 38/38 (gs_viewport_depth 20/20). The G720 MC8 user report has an intermittent gs_viewport_depth failure on beta.17 (open, untested since) |
| v13 | CSF | Works on beta.18 RC1; not rerun on RC3 | Immortalis-G925 MC12 passes 37/37 on beta.18 RC1 (beta.17: 34/36) |
| v14 | CSF | Built, untested | G1-Ultra (0xe8800010) matches its model row, but vkCreateDevice failed on beta.15 (MEM_ALLOC_EX ENOTTY). The 110 fallback is not yet tested on v14 |
- Missing gpu_id rows: RC2 and earlier have no row for the G77
(
PAN_PROD_ID(9, 0, 0));jm-v9/006and021add G77, G78 and G78AE in RC3. G510, G620, G625 and plain G725 have never been seen, so their gpu_id values are unconfirmed. The G710 row only exists from csf-v11/140 (beta.18 RC1). - v9 chunk caps: gpu_prerast and tessellated draws on v9 are unrolled in a fixed number of chunks (direct tess 1024 iterations, indirect tess 32, direct restart strips 1024 chunks, other indirect draws 8). Larger draws are cut short, with a one-time warning in logcat.
- v9 viewport ordering: with multiViewport and different depth ranges per viewport, v9 draws once per viewport, so primitives are reordered by viewport. CSF keeps the original order.
- v14: G1-Ultra needs a rerun with beta.17 or later to check the
MEM_ALLOC_EXfallback. - G720 MC8: intermittent
gs_viewport_depthfailure (PROGRESS item 40, not started).
The Mali-G615 (Mesa PAN_ARCH v11, CSF frontend; Arm's 4th generation
Valhall, announced 2022) on the reference device above is the validated
release target. From 0.1.0 RC3, v9 to v13 are supported: the v9 JM
Mali-G57, G610 and G710 (v10), G615 and G715 (v11) and Immortalis-G720 (v12)
pass PanProbe 38/38 on RC3, and Immortalis-G925 (v13) passed 37/37 on beta.18
RC1 (see Tested devices). v14 and the untested GPUs below
are built but not validated. Arm's marketing generations (Utgard, Midgard,
Bifrost, Valhall 1st to 4th gen, 5th Gen, G1) and Mesa PAN_ARCH numbers are
different schemes; the full chronological GPU list, mappings, frontends and
upstream driver status are in
docs/MALI-GPU-ARCHITECTURES.md.
Status key:
- ✅ Supported: validated on a device.
- 🧪 Tested: PanProbe 38/38 on RC3 (G925: 37/37 on beta.18 RC1) and the D3D11 cube on beta.18 or later; less coverage than the G615.
⚠️ Tested, known issue: run on a device by us or a user; works, but with an open issue.- 🔨 Built, untested: compiled into the universal ICD and recognised by the Mesa model table, but never run on that GPU.
- 📋 TODO: a profile exists in
profiles/and the port is planned. - ❔ Possible, not tried: Mesa has a backend for this arch, but no profile or device test exists here.
- ❌ Not possible: no PanVK (Vulkan) backend exists for the arch.
One row per Mesa arch. Each GPU carries its own status mark.
| Mesa arch | Arm family / generation | Frontend | GPUs (status per GPU) | Notes |
|---|---|---|---|---|
| n/a | Pre-Utgard (fixed function) | n/a | ❌ Mali-55, ❌ Mali-110 | No programmable shaders |
| n/a (Lima) | Utgard | n/a | ❌ Mali-200, ❌ Mali-300, ❌ Mali-400 MP, ❌ Mali-450 MP, ❌ Mali-470 MP | GLES 2 only (Lima), no Vulkan |
| v4 | Midgard 1st to 3rd gen | JM | ❌ Mali-T604, ❌ T658, ❌ T622, ❌ T624, ❌ T628, ❌ T678, ❌ T720 | No PanVK backend |
| v5 | Midgard 3rd/4th gen | JM | ❌ Mali-T760, ❌ T820, ❌ T830, ❌ T860, ❌ T880 | No PanVK backend |
| v6 | Bifrost 1st/2nd gen | JM | ❌ Mali-G71, ❔ G72 | Mesa marks G71 unsupported; G72 experimental upstream |
| v7 | Bifrost 1st to 3rd gen | JM | 🧪 Mali-G52, ❔ G31, ❔ G51, ❔ G76 | Experimental from RC4 (jm-v7 001-005). G52 MC2 (Galaxy A06) and G52 MC1 (Galaxy A14): PanProbe 14/38. GS/XFB/tessellation still in patches/jm-v7-wip |
| v9 | Valhall 1st/2nd gen | JM | ✅ Mali-G57 (MC2), 🔨 G68, 🔨 G77, 🔨 G78, 🔨 G78AE | Supported from RC3. The jm-v9 001-030 series implements Vulkan 1.3, GS, tessellation, XFB, multiViewport, depth bounds, fill mode, conditional rendering, robustImageAccess2 and timestamps, and adds model rows for G77 (0x9000), G78 (0x9002) and G78AE (0x9005); G68 (0x9004) comes from upstream. G57 MC2: PanProbe 38/38, D3D8-D3D11 cubes and NFS Most Wanted on RC3. Other v9 parts not validated. RC2 and earlier: experimental, PanProbe 1-2/17, Vulkan 1.1 |
| v10 | Valhall 3rd gen | CSF | 🧪 Mali-G610, 🧪 G710, 🔨 G310, ❔ G510 | G610 MC3 (motorola edge 40 neo) and G710 MC7 (Pixel 7, Pixel 7a) pass PanProbe 38/38 on RC3 (Firebase Test Lab, 2026-10-11). G710 gets its model row in csf-v11/140 (beta.18 RC1). G310 is built, untested; G510 needs a tester's gpu_id |
| v11 | Valhall 4th gen | CSF | ✅ Mali-G615, 🧪 G715, ❔ Immortalis-G715 | G615 validated on Poco X6 Pro (Dimensity 8300). G715 MC7 (Pixel 8, Pixel 9) passes PanProbe 38/38 on RC3. Immortalis-G715 is untested |
| v12 | 5th Gen | CSF | 🧪 Immortalis-G720, |
Immortalis-G720 MC12 (Galaxy Tab S10 Ultra) passes PanProbe 36/36 on beta.17, 37/37 on beta.18-dev and 38/38 on RC3. beta.18 names it by core count (131). A user's G720 MC8 scores 35-36/37 on beta.17 (intermittent gs_viewport_depth, open). G620 is built, untested |
| v13 | 5th Gen | CSF | 🧪 Immortalis-G925, 🔨 Mali-G725, 🔨 G625 | Immortalis-G925 MC12 (Galaxy Tab S11 Ultra) passes PanProbe 37/37 on beta.18-dev (beta.17: 34/36, fixed by 130 and 132). G725 and G625 share the product id and are built, untested |
| v14 | 5th Gen, G1 series | CSF | 🔨 Mali G1-Ultra, 🔨 G1-Premium, 🔨 G1-Pro | Built, untested (beta.13 universal ICD). All three are in the Mesa model table; experimental upstream |
| v15 (unconfirmed) | G2 series | CSF | ❌ Mali G2-Ultra NX, ❌ G2-Premium NX (rumoured), ❌ G2-Pro NX (rumoured) | Not in Mesa yet |
"Possible" means Mesa has code for the arch. It does not mean the arch works here: each one still needs a kbase profile, its kbase frontend path (JM or CSF) and device validation. The driver has a job-manager kbase path for v9, and from RC4 an experimental one for v7 (Bifrost); v6 takes the same path but is untested.
sources.lock exact Mesa commit + reference repos (no moving branches)
profiles/ per-GPU Kbase profiles (arch/gpu-id/frontend/uapi)
patches/ qualified patch families (never one unqualified blob)
meson/ cross files (android-aarch64, linux-aarch64-native)
scripts/ fetch / patch / build / package / validate / release
tests/ kbase-probe, vulkan-smoke, compute, offscreen, ahb,
android-surface, sync, android-loader-app, dxvk-vkd3d
docs/ architecture, build, portability, matrix, profiles,
app-compat, release, Kbase sparse feasibility,
Mali GPU architectures (MALI-GPU-ARCHITECTURES.md)
validation/ G615 capability dumps, DXVK/vkd3d profiles and matrix
.github/workflows/ build / release / source-drift
See docs/BUILD-POCO-X6-PRO.md.
./scripts/fetch-mesa.sh
./scripts/apply-patches.sh --profile g615-v11-csf
./scripts/build-android.sh --profile g615-v11-csf
./scripts/validate-binary.sh --abi android dist/android-g615-v11-csf/libvulkan_panfrost.so
./scripts/package-android-adpkg.sh --profile g615-v11-csfTiers: dev (build only) < alpha (compute/offscreen) < beta (WSI +
app-loader on one profile) < rc (primary consumer) < stable (multi-device
- soak). First Poco build is never
stable. Releases are immutable; a patch change creates a new release even if the Mesa SHA is unchanged.
Latest: panvk-kbase-v0.1.0-rc4
and panvk-kbase-v0.1.0-rc4-lowmem
(release candidates: experimental v7 Bifrost; only the G615 was retested, 39/39). The RC4 binaries were
recovered from the reference device; their exact source was not committed. The last release tested on every GPU is
panvk-kbase-v0.1.0-rc3
(Mesa 5a07217f plus the committed series up to csf-v11/184 and jm-v9 030; v9 to v13 supported,
PanProbe 38/38 on G57, G610, G710, G615, G715 and Immortalis-G720; v13 not rerun on RC3). Tags no longer carry the g615 prefix; older releases keep theirs. Each
release ships the Android and glibc drivers, an .adpkg package, an EMULATOR
zip, the test APK and screenshots. Full history: CHANGELOG.md.
| Release | Highlights |
|---|---|
| 0.1.0 RC4 | Experimental v7 Bifrost (jm-v7 001-005: no env gate, no MEM_JIT_INIT, growable tiler heap, atom stride check, ION fallback); G52 MC2/MC1 (Galaxy A06/A14) PanProbe 14/38. Mesa 5a07217f, API 29 build. driverInfo RC4. Android BuildID 6654fada. Retested on G615 only: PanProbe 39/39. Six launcher zips (adrenotools, Winlator, cmod, Ludashi, bannerlator, GameHub). Binary recovered from the reference device; exact source not committed. |
| 0.1.0 RC4 lowmem | Separate RC4 build (Mesa tree git-5437602824), same v7 support. Lowmem BC mode on by default (binary-patched; the original build needed PANVK_BC_MODE=lowmem; PANVK_BC_MODE=exact turns it off): BC1-BC5/BC7 use about half to a third of the memory (BC7 10.9 → 5.5 MB, BC1 8.2 → 2.7 MB at 2048²), sampling about 11% faster, BC1-BC5 slightly lossy (39-50 dB vs RC4). Known issue: BC image-to-image copy and raw BC readback fail with lowmem on. All other tests match RC4 (direct run, G615). .so SHA-256 c395db09, BuildID aff7a07f. Launcher zips repackaged from the RC4 layouts. Exact source not committed. |
| 0.1.0 RC3 lowmem | RC3 plus csf-v11/185: BC1 RGB (no alpha) and BC7 shadows are stored as AFRC at 8 bpp instead of 12/16 bpp. About 25% less memory for BC1 RGB and 33% less for BC7; sampling speed is unchanged (not faster); quality is lower (about 48 dB against the exact decode, RC3 was 57-64 dB; 0.1-0.5 dB loss against the source). BC1 RGBA, BC2 and BC3 keep the RC3 rates. BC7 with alpha at 8 bpp is untested. PANVK_BC_AFRC=0 restores exact decode; PANVK_BC_AFRC=24 or 32 forces that rate for all BC formats (RC3 used 24 for BC1/BC2 and 32 for BC3/BC7). driverInfo RC3-lowmem. Android BuildID 9d975491. Retested on G615 only: PanProbe 38/38, bc_decode device checks, NFS Most Wanted with the DXVK HUD. PanProbe 1.2.7 bundles it; PanPlay 1.2.7 still bundles RC3. |
| 0.1.0 RC3 | v9 to v13 supported. v9 JM (jm-v9 001-030): Vulkan 1.3, GS, tessellation, XFB, multiViewport, clip/cull, depth bounds, fill mode, conditional rendering, robustImageAccess2, GPU timestamps, G77/G78/G78AE rows and kbase JM fixes; G57 MC2 PanProbe 38/38, D3D8-D3D11 cubes, NFS Most Wanted. v10 GPU-emulated robustImageAccess2 alpha fix (184; G610 and G710 38/38 on Firebase). driverInfo RC3. Android BuildID 4ec75e25. PanProbe 38/38 on G57 (v9), G610, G710 x2 (v10), G615, G715 x2 (v11) and Immortalis-G720 MC12 (v12, gs_viewport_depth 20/20); v13 G925 not rerun (37/37 on RC1). |
| 0.1.0 RC2 | robustImageAccess2 on v10 and later (181): G615 PanProbe 38/38, CTS robustness2 image 4332/0, D3D12CreateDevice OK. Mali-G77 model row (jm-v9/006, variant 0 from a tester log, unproven on G77). driverInfo beta.18-rc2. Tag prefix g615 dropped (panvk-kbase-v). RC2 binary not yet device-tested. |
| beta.18 RC1 | First release tested on v10, v11, v12 and v13 hardware: PanProbe 37/37 on G610, G710, G615, G715, Immortalis-G720 and Immortalis-G925. BCn rework: BC textures decode into a compact tiled AFBC/AFRC shadow, about 2x faster bandwidth-bound sampling and 38-65% less shadow memory (150, 151, 180; AFRC is lossy, PANVK_BC_AFRC=0 for exact). Shader compile stalls: faster register allocation, NFS: Most Wanted pipelines 3.6 s to about 0.3 s, identical output (152-154, 160-164). Mesa disk shader cache on by default on Android (170), each Android property looked up once (171). robustBufferAccess2 on v10 (122), v13 vertex stores (130, 132), core-count GPU names (131), G710 model row (140). Known: G720 MC8 gs_viewport_depth intermittent (untested on beta.18), v14 untested. |
| beta.17 | Correctness and reach. G615: GS draws with primitive restart no longer fault the GPU (120); zero depth-range viewports keep exact depth (121); depth bounds keeps FPK/early ZS while off (116); FS gl_PrimitiveID survives viewport runs (115); VMR secondaries (114); prerast fan splits (112); SetEvent/ResetEvent never return DEVICE_LOST (111); X11 software present over MIT-SHM (117). Universal: BC emulation unless all BC formats are native (108), v12+ viewport depth (109), kbase CSF uAPI layouts and 16K pages (110), vertex stores on v13/v14 (119), vendor-neutral gralloc mapper (android/014). PanProbe 36/36 x3; CTS 32339 pass / 115 fail, sync + memory gate unchanged. Fixes for G610/G720/G925/G1-Ultra and stock ROMs untested on hardware. |
| beta.16 | Removes the remaining CPU waits on the v11 kbase path. Software WSI present no longer blocks the app's submit thread until the GPU finishes the frame (106). Up to three retired tiler heaps can be in flight, so vkQueueSubmit no longer stalls 35-55 ms on heap backpressure (107). NFS: Most Wanted race: 85.5 to 90.6 fps (DXVK HUD), p99 frame time 22 to 16 ms. NFS is now limited by Wine, not the GPU. CTS sync + memory gate: same 56 failures; PanProbe 17/17. |
| beta.15 | Faster kbase submission on v11: no CPU graphics drain on tiler heap renewal (102), same-queue semaphores waited on the GPU (103), next tiler heap created on a worker thread (104). NFS: Most Wanted on G615 goes from 24 to 38-40 fps, p99 frame time 102 to 58 ms, >50 ms frames 650 to ~80. PANVK_DEBUG=trace works on kbase again (101, 105). CTS sync + memory: no new failures; PanProbe 17/17. v9 still experimental. Fallout 4 hangs before loading the driver (also on beta.14). |
| beta.14 | Adds EXPERIMENTAL, partly broken Mali v9 (Valhall JM: G57/G68/G77/G78 class) support to the universal ICD (jm-v9 series: kbase JM atom submission, v9 backend ported from FristOneRR-Panvk-Source, Vulkan 1.1 reporting). Tested only on Mali-G57 MC2: CTS api.smoke 4/6, simple_draw 4/4, synchronization.basic 21 pass/8 not supported, PanProbe 1/17. v10/v12/v13/v14 untested; v11 was tested on beta.13. driverInfo reads PanVK-kbase beta.14. |
| beta.13 | One universal Android ICD for v10-v14. Only v11 (G615) is tested; v10/v12/v13/v14 are built but untested, and testers are needed. The kbase path now admits v14 without PAN_I_WANT_A_BROKEN_VULKAN_DRIVER; v13 was already admitted. Fixes PanProbe gs_viewport_depth case A, which regressed in beta.11 (100). On kbase, gpu_prerast arenas are committed up front again. GPU-fault growth lost the first GS draw. The shared-TLS memory fix from 097 stays. The kbase CS register-count fallback is now 128 on v12+; 96 was too small. driverInfo now reads PanVK-kbase beta.13 (Mesa 26.3.0-devel ...), shown in the DXVK HUD and PanProbe. |
| beta.12 | One universal Android ICD for v10, v11 and v12. Only v11 (G615) is tested; v10 and v12 are built but untested, and testers are needed. The release notes give the command that reports your GPU ID. |
| beta.11 | Fixes VK_ERROR_DEVICE_LOST on tiler heap OOM (098). Fixes the memory blow-up from per-pool TLS and eagerly committed prerast arenas (097). |
| beta.10 | Fast and correct 32-bit WoW64 games: placed maps now map the BO's dma-buf again at the requested address (091) instead of a shadow copy. Need for Speed Most Wanted (DXVK D3D9) went from 0.5 fps to 39-66 fps, with clean HUD and text. Also GPU chunking of large and indirect prerast draws (094, 096), the tessellation conditional-state fix (093), and the sample count for attachment-less secondaries (095). |
| beta.9 | depthBounds (092), shaderOutputViewportIndex from VS/TES (090), per-viewport depth clamp/clip (089), CSF event-memory sync words (085). |
| beta.8 | X11 surfaces (VK_KHR_xlib_surface, VK_KHR_xcb_surface) for Wine/Proton launchers that display through Termux:X11. The X11/XCB libraries load at runtime from the launcher's library path and are not bundled. Presentation is a software copy (X11 PutImage). |
g615-v11-csf-v0.1.0-beta.3 (code commit
fc8a759e7d1b2b8de01c0e96f1fdc5e3950ba1a3, published 2026-09-19) and its
assets stay frozen by project policy. See
validation/g615-v11-csf/BETA3-PUBLICATION-ADDENDUM.md.
Two separate sets are stored per release: UPSTREAM_MATRIX_CAPABILITIES
(from docs/features.txt of the exact Mesa checkout) and
RUNTIME_DEVICE_CAPABILITIES (from on-device vulkaninfo/probe). Only
runtime-tested capabilities may be used for compatibility claims. No fake
feature bits.
Full specification and per-capability test breakdown:
docs/RUNTIME-FEATURES.md. Canonical
machine-readable evidence:
validation/g615-v11-csf/runtime-feature-matrix.json.
- Vulkan API Version:
1.4.363 - Total Extensions Exposed:
194(13 instance, 181 device); beta.2 tag:178 - Driver: Mesa
26.3.0-devel(commit5a07217f034b3e50d8c7c7794f97a2df1742613b) - GPU: Mali-G615 MC6 (
0xb8a31030) - Kernel Interface:
mali_kbase(CSF uAPI 1.21,/dev/mali0) - Patch Series ID:
sha256:c0bbdeef591b206a2f3ae36dc6191c08854399039075f8d69f103e33c0eca2f8
All 13 target groups are natively exposed and passed real Poco X6 Pro workloads. Enumeration alone is not counted as a test.
| Feature group | Status | Concise workload evidence |
|---|---|---|
| Descriptor indexing | PASS | Runtime array, partially-bound and variable-count descriptors; non-uniform sampled-image/storage-buffer indexing; checksum 123 |
| Timeline semaphore | PASS | GPU chain 1..64, CPU wait/signal, CPU-to-GPU wait, final counter 66 |
| Dynamic rendering | PASS | vkCmdBeginRendering/vkCmdEndRendering; 1,154 triangle pixels; checksum 0x9a7f1b8fec90af07 |
| Synchronization2 | PASS | vkCmdPipelineBarrier2 and vkQueueSubmit2; timeline completion 67 |
| Buffer device address | PASS | Shader dereference at index 3; readback 0x2468ace0 |
| Push descriptors | PASS | vkCmdPushDescriptorSetKHR storage buffers; readback 0xb791f3dd |
| Robust buffer access | PASS | Out-of-bounds index 64 from one bound uint returned 0x00000000 |
| Anisotropy | PASS | 4x anisotropy across 9 sampled formats; two deterministic runs |
| Wide lines | PASS | Width 5 rasterized 185 green pixels |
| Large points | PASS | Size 11 rasterized 121 yellow pixels |
| ETC2/EAC | PASS | ETC2 RGB/RGBA plus EAC R11/RG11 UNORM/SNORM sampling, filtering, and mip level 1 |
| ASTC LDR | PASS | ASTC 4x4 UNORM/SRGB sampling, filtering, mip level 1, exact checksums |
| ASTC HDR | PASS | VK_FORMAT_ASTC_4x4_SFLOAT_BLOCK_EXT; pixel [1, 2, 3, 1]; filtering and mip level 1 |
This table describes the beta.3 release. Since then textureCompressionBC
has been exposed on feature/g615-dxvk-complete through GPU compute decode
(G615 has no BC hardware); see the DXVK section below. Other unsupported
features are listed in docs/RUNTIME-FEATURES.md.
DXVK Native v3.1.1 on the Poco X6 Pro creates a D3D11 device at
feature level 11_0 (D3D11 HRESULT=0x00000000 feature_level=0xb000);
D3D11 and D3D9 draw workloads pass. Every feature below runs on the GPU,
with no CPU fallback or simulation, and is exposed only after device proof.
| Feature | Implementation | Device proof (CTS Pass / Fail) |
|---|---|---|
geometryShader |
VS/GS run as compute on the GPU pre-raster path, 64 invocations (071) | geometry 193 / 0; instanced 20 / 0 |
tessellationShader |
VS, TCS, tessellator and TES as compute, GPU chunking | tessellation 526 / 0 |
VK_EXT_transform_feedback |
GPU capture kernel, 4 streams, counters, queries | transform_feedback 15793 / 0 (2 intermittent DeviceLost) |
textureCompressionBC |
GPU compute decode of BC1-7 | BC subset 1863 / 0; copy_and_blit 9620 / 0 |
shaderClipDistance, shaderCullDistance |
NIR lowering | device matrix 0 fail |
multiViewport |
16 viewports, GS viewport index honoured (076) | device matrix 0 fail, scissor 88 / 88 |
fillModeNonSolid |
GPU kernel builds line/point primitives | 17/17 pixel-exact |
pipelineStatisticsQuery |
049-054 | statistics_query 15374 / 0 |
VK_KHR_incremental_present, VK_EXT_swapchain_colorspace, VK_EXT_image_compression_control |
upstream backports | device probes 0 fail |
VK_EXT_multi_draw |
072 | multi_draw 12704 / 0 |
VK_EXT_primitives_generated_query |
073 | primitives_generated_query 75206 / 0 |
VK_EXT_memory_priority, VK_EXT_pageable_device_local_memory |
069 | 224 / 0, 202 / 0; api.info 7799 / 0 |
alphaToOne |
070 | alphaToOne 123 / 0 |
variableMultisampleRate |
074 | variable_rate 504 / 0 standalone (tmp/cts/p5-vmsr/summary.txt); combined regression run (geometry + tessellation + transform_feedback.simple + variable_rate: 15955 cases, 6210 pass, 9745 NotSupported, 0 fail; source tmp/cts/r6-geo/status.txt) |
sync_fd export |
075 (kbase KCPU queue) | sync_fd 1996 / 0 |
vertexPipelineStoresAndAtomics (v10-v12) |
078-082, vertex stage on the compute pre-raster path | atomic_operations *_vertex* 66 / 0; 12132-case tess/geometry/xfb/draw list 7940 / 0, 0 DeviceLost |
| per-viewport depth clamp/clip (beta.9) | 089, GS-selected viewports drawn as ordered runs | panvk-test gs_viewport_depth exact (no CTS) |
shaderOutputViewportIndex from VS/TES (beta.9, v10/v11) |
090 | panvk-test vs_viewport_index 6/6 (no CTS) |
depthBounds (beta.9, v10/v11) |
092, fragment-shader emulation | 2059 depth-bounds CTS cases: 1774 pass / 0 fail / 285 NotSupported; panvk-test depth_bounds 6/6 |
DXVK feature level 11_1 has not been re-checked on the beta.7 to beta.9
builds yet. The Android driver has X11 surfaces (beta.8), but under Proton 11
(i686 through wow64) Wine fails to create the Vulkan surface before the
driver is called, so DXVK presentation there is still blocked. depthBounds
lowered pipelines lose FPK, and EarlyFragmentTests shaders with depth writes
compare against the fragment's new depth (see the beta.9 release notes).
Still missing: robustImageAccess2 (the vkd3d-proton device-create
blocker) and
sparse (FL12_0, NO-GO on Kbase). The X11 present teardown hang was seen once
under Xvfb only and is unverified on Android. Progress and TODOs:
worklogs/g615-dxvk/PROGRESS.md. Roadmap:
docs/plans/PANVK_MASTER_ROADMAP.md.
Machine-evaluated against stock tagged profiles. No fake feature bits. Overall result: FAIL. DXVK 2.7.1/3.1.1 COMMON and vkd3d README hard gates PASS; D3D9/D3D10/D3D11 profiles and vkd3d 2.14.1/3.0.1 device/profile baseline FAIL. Stock DXVK/vkd3d smoke is BLOCKED (Android ICD is not a legal host for those Windows binaries). CTS is BLOCKED.
Canonical report:
validation/g615-v11-csf/P23-DXVK-VKD3D-COMPLIANCE-MATRIX.md.
JSON:
validation/g615-v11-csf/p23-dxvk-vkd3d-compliance-matrix.json.
Gap report:
validation/g615-v11-csf/dxvk-vkd3d-gap-report.md.
Capability dump:
validation/g615-v11-csf/consumer-capabilities.json.
| Target | Result |
|---|---|
| DXVK 1.10.3 D3D9 / D3D10 / FL10.1 / FL11.0 | FAIL |
| DXVK 2.7.1 COMMON | PASS |
| DXVK 2.7.1 D3D9 / D3D10_10_1 / D3D11_11_0 / D3D11_11_1 | FAIL |
| DXVK 3.1.1 COMMON | PASS |
| DXVK 3.1.1 D3D9 / D3D10_10_1 / D3D11_11_0 / D3D11_11_1 | FAIL |
| vkd3d-proton 2.0 HARD / DEVICE_CREATE | PASS |
| vkd3d-proton 2.14.1 HARD | PASS |
| vkd3d-proton 2.14.1 PROFILE_BASELINE / DEVICE_CREATE | FAIL |
| vkd3d-proton 3.0.1 HARD | PASS |
| vkd3d-proton 3.0.1 PROFILE_BASELINE / DEVICE_CREATE | FAIL |
| MAX_FEATURE_LEVEL / D3D_FEATURE_LEVEL | NOT_AVAILABLE |
At beta.3, still false (no spoofing): geometryShader, tessellationShader,
fillModeNonSolid, multiViewport, shaderClipDistance,
shaderCullDistance, textureCompressionBC, transform feedback,
pipelineStatisticsQuery, robustImageAccess2, sparse.
Proven on G615: VK_EXT_robustness2 buffer + nullDescriptor,
maxPushConstantsSize=256, VK_KHR_push_descriptor / maxPushDescriptors=32,
DXVK/vkd3d common easy gates. Advertised UAB limits are 1,048,576; 1M
descriptor stress is implemented but not yet run to completion on device.
Phase evidence:
P6,
P7,
P8,
P9,
P10,
P11,
P12,
P13,
P14,
P15,
P16,
P17,
P18,
P19,
P20,
P21,
P22.
Sparse: docs/KBASE-SPARSE-FEASIBILITY.md
(NO-GO on Kbase UAPI 1.21). Profiles:
validation/requirements/.
Evaluators: scripts/evaluate-vulkan-profile.py,
scripts/evaluate-dxvk-vkd3d-compliance-matrix.py.
Workloads: tests/dxvk-vkd3d/.
At beta.3 the next blocker was robustImageAccess2 (vkd3d 2.14.1/3.0.1
DEVICE_CREATE). It still is; geometry, fill, clip/cull and BC have since
landed (see Current state above).
These 16 beta.3 additions are exposed and workload-tested:
VK_KHR_compute_shader_derivatives, VK_KHR_copy_memory_indirect,
VK_KHR_internally_synchronized_queues, VK_KHR_maintenance7,
VK_KHR_maintenance8, VK_KHR_maintenance9, VK_KHR_present_id2,
VK_KHR_present_wait2, VK_KHR_shader_constant_data, VK_KHR_shader_fma,
VK_KHR_shader_relaxed_extended_instruction,
VK_KHR_shader_untyped_pointers, VK_KHR_surface_maintenance1,
VK_KHR_swapchain_maintenance1, VK_KHR_unified_image_layouts, and
VK_GOOGLE_user_type.
VK_KHR_depth_clamp_zero_one, VK_KHR_pipeline_binary, and
VK_KHR_robustness2 are disabled and not exposed because required workload
proof is absent. VK_GOOGLE_display_timing is not exposed without a proven
Android timing implementation.
The immutable beta.2 tag exposed 12 instance and 166 device extensions (178 total). Its post-tag documentation commit is distinct from the tagged code as recorded under Current status. Everything below is the published beta.3 inventory generated from the canonical matrix.
Full beta.3 list: 194 extensions (13 instance + 181 device)
VK_KHR_android_surface, VK_KHR_device_group_creation,
VK_KHR_external_fence_capabilities, VK_KHR_external_memory_capabilities,
VK_KHR_external_semaphore_capabilities,
VK_KHR_get_physical_device_properties2,
VK_KHR_get_surface_capabilities2, VK_KHR_surface,
VK_KHR_surface_maintenance1, VK_EXT_debug_report, VK_EXT_debug_utils,
VK_EXT_headless_surface, VK_EXT_surface_maintenance1.
VK_KHR_8bit_storage, VK_KHR_16bit_storage, VK_KHR_bind_memory2,
VK_KHR_buffer_device_address, VK_KHR_calibrated_timestamps,
VK_KHR_compute_shader_derivatives, VK_KHR_cooperative_matrix,
VK_KHR_copy_commands2, VK_KHR_copy_memory_indirect,
VK_KHR_create_renderpass2, VK_KHR_dedicated_allocation,
VK_KHR_depth_stencil_resolve, VK_KHR_descriptor_update_template,
VK_KHR_device_group, VK_KHR_draw_indirect_count,
VK_KHR_driver_properties, VK_KHR_dynamic_rendering,
VK_KHR_dynamic_rendering_local_read, VK_KHR_external_fence,
VK_KHR_external_fence_fd, VK_KHR_external_memory,
VK_KHR_external_memory_fd, VK_KHR_external_semaphore,
VK_KHR_external_semaphore_fd, VK_KHR_format_feature_flags2,
VK_KHR_get_memory_requirements2, VK_KHR_global_priority,
VK_KHR_image_format_list, VK_KHR_imageless_framebuffer,
VK_KHR_index_type_uint8, VK_KHR_internally_synchronized_queues,
VK_KHR_line_rasterization, VK_KHR_load_store_op_none,
VK_KHR_maintenance1, VK_KHR_maintenance2, VK_KHR_maintenance3,
VK_KHR_maintenance4, VK_KHR_maintenance5, VK_KHR_maintenance6,
VK_KHR_maintenance7, VK_KHR_maintenance8, VK_KHR_maintenance9,
VK_KHR_map_memory2, VK_KHR_multiview,
VK_KHR_pipeline_executable_properties, VK_KHR_pipeline_library,
VK_KHR_present_id, VK_KHR_present_id2, VK_KHR_present_wait,
VK_KHR_present_wait2, VK_KHR_push_descriptor,
VK_KHR_relaxed_block_layout, VK_KHR_sampler_mirror_clamp_to_edge,
VK_KHR_sampler_ycbcr_conversion, VK_KHR_separate_depth_stencil_layouts,
VK_KHR_shader_atomic_int64, VK_KHR_shader_clock,
VK_KHR_shader_constant_data, VK_KHR_shader_draw_parameters,
VK_KHR_shader_expect_assume, VK_KHR_shader_float16_int8,
VK_KHR_shader_float_controls, VK_KHR_shader_float_controls2,
VK_KHR_shader_fma, VK_KHR_shader_integer_dot_product,
VK_KHR_shader_maximal_reconvergence, VK_KHR_shader_non_semantic_info,
VK_KHR_shader_quad_control, VK_KHR_shader_relaxed_extended_instruction,
VK_KHR_shader_subgroup_extended_types, VK_KHR_shader_subgroup_rotate,
VK_KHR_shader_subgroup_uniform_control_flow,
VK_KHR_shader_terminate_invocation, VK_KHR_shader_untyped_pointers,
VK_KHR_spirv_1_4, VK_KHR_storage_buffer_storage_class,
VK_KHR_swapchain, VK_KHR_swapchain_maintenance1,
VK_KHR_swapchain_mutable_format, VK_KHR_synchronization2,
VK_KHR_timeline_semaphore, VK_KHR_unified_image_layouts,
VK_KHR_uniform_buffer_standard_layout, VK_KHR_variable_pointers,
VK_KHR_vertex_attribute_divisor, VK_KHR_vulkan_memory_model,
VK_KHR_workgroup_memory_explicit_layout,
VK_KHR_zero_initialize_workgroup_memory, VK_EXT_4444_formats,
VK_EXT_astc_decode_mode, VK_EXT_attachment_feedback_loop_dynamic_state,
VK_EXT_attachment_feedback_loop_layout, VK_EXT_border_color_swizzle,
VK_EXT_buffer_device_address, VK_EXT_calibrated_timestamps,
VK_EXT_color_write_enable, VK_EXT_conditional_rendering,
VK_EXT_conservative_rasterization, VK_EXT_custom_border_color,
VK_EXT_debug_marker, VK_EXT_depth_bias_control,
VK_EXT_depth_clamp_control, VK_EXT_depth_clamp_zero_one,
VK_EXT_depth_clip_control, VK_EXT_depth_clip_enable,
VK_EXT_descriptor_indexing, VK_EXT_device_address_binding_report,
VK_EXT_device_memory_report, VK_EXT_dynamic_rendering_unused_attachments,
VK_EXT_extended_dynamic_state, VK_EXT_extended_dynamic_state2,
VK_EXT_extended_dynamic_state3,
VK_EXT_external_memory_acquire_unmodified,
VK_EXT_external_memory_dma_buf, VK_EXT_global_priority,
VK_EXT_global_priority_query, VK_EXT_graphics_pipeline_library,
VK_EXT_hdr_metadata, VK_EXT_host_image_copy, VK_EXT_host_query_reset,
VK_EXT_image_2d_view_of_3d, VK_EXT_image_drm_format_modifier,
VK_EXT_image_robustness, VK_EXT_image_sliced_view_of_3d,
VK_EXT_image_view_min_lod, VK_EXT_index_type_uint8,
VK_EXT_inline_uniform_block, VK_EXT_legacy_dithering,
VK_EXT_line_rasterization, VK_EXT_load_store_op_none,
VK_EXT_map_memory_placed, VK_EXT_memory_budget,
VK_EXT_multisampled_render_to_single_sampled,
VK_EXT_mutable_descriptor_type, VK_EXT_nested_command_buffer,
VK_EXT_non_seamless_cube_map, VK_EXT_physical_device_drm,
VK_EXT_pipeline_creation_cache_control,
VK_EXT_pipeline_creation_feedback, VK_EXT_pipeline_robustness,
VK_EXT_present_timing, VK_EXT_primitive_topology_list_restart,
VK_EXT_private_data, VK_EXT_provoking_vertex,
VK_EXT_queue_family_foreign,
VK_EXT_rasterization_order_attachment_access, VK_EXT_rgba10x6_formats,
VK_EXT_robustness2, VK_EXT_sampler_filter_minmax,
VK_EXT_scalar_block_layout, VK_EXT_separate_stencil_usage,
VK_EXT_shader_atomic_float, VK_EXT_shader_demote_to_helper_invocation,
VK_EXT_shader_image_atomic_int64, VK_EXT_shader_module_identifier,
VK_EXT_shader_replicated_composites, VK_EXT_shader_stencil_export,
VK_EXT_shader_subgroup_ballot, VK_EXT_shader_subgroup_vote,
VK_EXT_shader_tile_image, VK_EXT_shader_uniform_buffer_unsized_array,
VK_EXT_subgroup_size_control, VK_EXT_swapchain_maintenance1,
VK_EXT_texel_buffer_alignment, VK_EXT_texture_compression_astc_hdr,
VK_EXT_tooling_info, VK_EXT_vertex_attribute_divisor,
VK_EXT_vertex_input_dynamic_state, VK_EXT_ycbcr_2plane_444_formats,
VK_EXT_ycbcr_image_arrays, VK_EXT_zero_initialize_device_memory,
VK_ANDROID_external_memory_android_hardware_buffer,
VK_ANDROID_native_buffer, VK_ARM_rasterization_order_attachment_access,
VK_ARM_scheduling_controls, VK_ARM_shader_core_builtins,
VK_ARM_shader_core_properties, VK_GOOGLE_decorate_string,
VK_GOOGLE_hlsl_functionality1, VK_GOOGLE_user_type,
VK_VALVE_mutable_descriptor_type.
Mesa code remains under its upstream licenses (see LICENSES/ and
NOTICE.md). Build/patch/test scaffolding in this repository is MIT unless
otherwise noted.
This project builds on the work of many people. Thank you.
- Mesa / PanVK and Panfrost (MIT): the driver this project patches. Pinned commit in
sources.lock. - Arm
mali_kbaseUAPI headers (Linux kernel, GPL-2.0 WITH Linux-syscall-note): vendored, seeNOTICE.md. - wonderkast02/panvk-g720-kbase-csf: CSF / Valhall and 5th Gen reference (kmod backend, CSF lifecycle, AHB, sync, G720 tessellation).
- LukeValen/panvk-mali-g52: JM / Bifrost reference.
- Noysz/panvk-g99-jm (MIT): JM / Valhall v9 reference. Ported or used as reference in
jm-v9003 (multiview/layered, 0042/0048), 007 (resubmit reset, 0072), 008 (async submit events, 0068), 009 (multi-draw indirect, 0044/0039), 010 (robustness2, 0045/0071), 011-013 (XFB, GS and streams, clip/cull, 0060/0063-0066), 014 (tessellation, 0062), 015 (multiViewport, 0066), 016 (queries, 0067) and 018 (per-sample shading and VMR, 0059). - FristOneRR-Panvk-Source (Mesa MIT headers): base of the JM kbase submission and v9 JM backend (
jm-v9001/002). - bmdhacks/malisx2 (MIT): v9 id table (
jm-v9021) and kbase JM fixes: atom layout probe, fault propagation, teardown wait (022) and tiler heap commit (023). - JimVulkan/mali-panvk (MIT): fill mode reference (
jm-v9019). - Upstream Mesa merge requests !43414 (G68 model row), !37969 (poly_heap on JM) and !44535 (kraid on all Valhall GPUs): v9 references.
- m8l8th814n-eng/mesa-gs101: G77/G78 model rows and the G78 transaction elimination finding (facts only, no code;
jm-v9021). - faux123: release notes reporting that cleaning the cache of imported memory before submit fixes black dashes in 32-bit DXVK games (fact only;
jm-v9024). - nangitagamer777-art/PanVK-kbase and Panvk_Kmod: implementation history and secondary reference.
- leegao/bcn_layer (MIT) with libGPUCounters, and leegao/mesa-funnymdzz: BCn compatibility and secondary references.
- Khronos Vulkan headers, registry and VK-GL-CTS: API definitions and conformance testing.
- Winlator (LGPL-2.1): built-in X server, renderer and touchpad view (vendored in
apps/panvk-launcher/third_party/winlator). - Wine (LGPL-2.1-or-later) and Valve's Proton, via the GameNative proton-wine arm64ec build. Wine 11.19
d3dcompiler_43/wined3d(with vkd3d-shader) as the HLSL compiler fix. - GameNative:
imagefs_bionicAndroid userland for Wine. - FEX-Emu (MIT): x86 / x86-64 emulation (ARM64EC and WoW64).
- DXVK (Zlib): Direct3D 8/9/10/11 to Vulkan.
- vkd3d-proton (LGPL-2.1): Direct3D 12 to Vulkan.
- cnc-ddraw (MIT): DirectDraw renderer for old games.
- Microsoft Visual C++ Redistributable (MFC140, VCOMP140): redistributed under Microsoft's redistribution terms.
- llvm-mingw: Windows cross toolchain for FEX, DXVK and vkd3d-proton.
- Apache Commons Compress (Apache-2.0), XZ for Java (0BSD) and zstd-jni (BSD-2-Clause): archive unpacking.
- Android Jetpack / AndroidX and Jetpack Compose (Apache-2.0): app UI.
- Android NDK and the Vulkan loader on Android: native test runner and Vulkan info.
Full license notices: NOTICE.md, LICENSES/ and apps/panvk-launcher/app/src/main/assets/components-NOTICE.txt.