Turnip Adreno 710,720,722 v4.5

[Vauzi-17]

Experimental build based on Mesa (main branch).
Not officially supported — do not report issues to Mesa upstream.

Support

• Need help? https://t.me/vauzi_17

---

Changes in v4.5

• Synced with the latest Mesa main branch

Driver Variants

• Regular / Universal (Turnip-710-720-722-v4.5.zip):
  Universal build based directly on upstream Mesa for broad compatibility across different emulators and native applications.
• Switch (Turnip-710-720-722-v4.5-sw.zip):
  Specifically tuned and optimized for Nintendo Switch emulators (Eden, etc.) on Adreno 7xx and 8xx.
• Winlator (Turnip-710-720-722-v4.5-win.zip):
  Specifically tuned and optimized for Winlator and Windows/DirectX emulation environments on Android.

Environment Configuration (default_env.conf) (Optional)

Download: default_env.conf

File Placement

Save the downloaded default_env.conf file to:

• Per-App Storage: /sdcard/Android/data/<package_name>/files/default_env.conf
  • Eden: /sdcard/Android/data/<package_name>/files/default_env.conf
  • Winlator: /sdcard/Android/data/com.winlator/files/default_env.conf
• Global Fallback: /sdcard/default_env.conf

Note: Settings already active by default in the driver are commented out with #. You only need to remove the # or adjust the value if you want to override the driver's default behavior.

Build Info

Mesa	27.0.0-devel
Mesa commit	02f13375231b (2026-10-24)
Vulkan	1.4.363
Android NDK	r29
Build date	2026-10-25

Applied Patches

Regular / Universal (Upstream)

• 01-u_gralloc-always-use-ubwc-detection.patch — u_gralloc: always use the UBWC detection path

Switch (SW)

• 01-u_gralloc-always-use-ubwc-detection.patch — u_gralloc: always use the UBWC detection path
• 02-tu-lrz-reuse-dont-care-in-sysmem.patch — tu/lrz: reuse LRZ state for DONT_CARE depth when rendering in sysmem
• 06-a8xx-cube-coord-sanitize.patch — ir3: Sanitize cube map directions on A8XX
• 07-a8xx-kgsl-ib-vbo-alias.patch — tu/kgsl: Fetch A8XX command streams through a virtual BO
• 08-tu-kgsl-cache-retired-a8xx-ib-storage.patch — tu/kgsl: Cache retired A8XX IB storage
• 09-tu-ir3-subgroup-half-warp.patch — tu,ir3: Support a required subgroup size of half a wave
• 11-wn-fix-a8xx-dev-info.py — Idempotent fixup for freedreno_dev_info.h and tu_cmd_buffer.cc
• 12-wn-apply-a8xx-gpus.py — WN-Turnip A8xx GPU support applicator for freedreno_devices.py
• 14-wn-apply-a7xx-gen2-ubwc-hint.py — WN-Turnip A7xx_gen2 UI-corruption fix for FD740 / Adreno X1-85
• 15-wn-add-aimapper-gralloc.py — Add the IMapper5 stable-C gralloc backend (SP-HAL loaded) to Mesa
• 16-wn-add-ubwc-swapchain-usage.py — Request UBWC-compressed swapchain buffers on Adreno
• 17-wn-apply-perf-variant.py — Apply aggressive autotune and variant-specific GPU clock forcing for performance
• 18-wn-a8xx-shared-mem.py — Robust equivalent of whitebelyash/mesa-unified turnip/gen8 commit
• 20-default-env.py — Built-in default environment variables for Turnip with app override support
• 30-tu-z-mutable.py — Turnip-Z port: keep MUTABLE_FORMAT images tiled (and UBWC) on all GPUs
• 31-tu-z-lrz-rescue.py — Turnip-Z port: LRZ rescue on in-render-pass depth clears
• 32-tu-z-texfast.py — Turnip-Z port: cheaper texture sampling
• 33-tu-z-bgra-swap.py — Turnip-Z port: real color swap for B8G8R8A8 on tiled images
• 36-tu-z-sysmem-only.py — Turnip-Z port: force sysmem rendering, and keep the bin-restore preamble
• 37-tu-z-bgra-companion.py — Turnip-Z port: BGRA8 companion changes with multi-mode color swap support
• 40-fd710-blob.py — FD710: consolidated hardware property and register tuning from Adreno 710 blob analysis (V@0615)
• 48-a8xx-lrz-color-mask.py — A8xx: Emit GRAS_LRZ_COLOR_COMP_MASK (0x810c) in blend dynamic state
• 50-fd720-blob.py — FD720: register tuning from Adreno 720 blob analysis (V@0676)
• 51-fd810-blob.py — FD810: unified Adreno 810 hardware tuning from Vulkan blob trace analysis (V@0800.74)
• 52-a8xx-resolve-cr-cntl.py — A8XX: Emit REG_A8XX_RB_RESOLVE_CR_CNTL=1 in preamble magic registers
• 53-a8xx-lrz-bidir-disable.py — A8XX: Set FORCE_BI_DIR_LRZ_DISABLE in sysmem and disabled LRZ render passes
• 54-a8xx-sp-hysteresis.py — A8XX: Initialize Shader Processor hysteresis thresholds in context preamble
• 55-tu-lrz-dir-change-no-invalidate.py — Preserve LRZ state across draws on depth compare-op direction change
• 56-fd722-blob.py — FD722: register tuning and hardware properties for Adreno 722 from Qualcomm blob analysis (chip_id 0x43020100)
• 60-tu-sysmem-wfi-relax.py — Relax subpass slice wait from WFI to WAIT_FOR_ME in sysmem mode
• 61-tu-lrz-fs-sideeffects-no-invalidate.py — Preserve LRZ state across draws when fragment shader has side effects
• 62-tu-draw-state-skip-identical.py — Avoid redundant dynamic draw state packet emission
• 63-ir3-sched-prefer-lower-live.py — Prioritize instructions with lower live register effect during pressure growth
• 65-tu-z-shader-cache.py — Turnip-Z port: optional on-disk shader cache for Android emulators

Winlator / Windows (WIN)

• 01-u_gralloc-always-use-ubwc-detection.patch — u_gralloc: always use the UBWC detection path
• 02-tu-lrz-reuse-dont-care-in-sysmem.patch — tu/lrz: reuse LRZ state for DONT_CARE depth when rendering in sysmem
• 06-a8xx-cube-coord-sanitize.patch — ir3: Sanitize cube map directions on A8XX
• 07-a8xx-kgsl-ib-vbo-alias.patch — tu/kgsl: Fetch A8XX command streams through a virtual BO
• 08-tu-kgsl-cache-retired-a8xx-ib-storage.patch — tu/kgsl: Cache retired A8XX IB storage
• 09-tu-ir3-subgroup-half-warp.patch — tu,ir3: Support a required subgroup size of half a wave
• 11-wn-fix-a8xx-dev-info.py — Idempotent fixup for freedreno_dev_info.h and tu_cmd_buffer.cc
• 12-wn-apply-a8xx-gpus.py — WN-Turnip A8xx GPU support applicator for freedreno_devices.py
• 14-wn-apply-a7xx-gen2-ubwc-hint.py — WN-Turnip A7xx_gen2 UI-corruption fix for FD740 / Adreno X1-85
• 15-wn-add-aimapper-gralloc.py — Add the IMapper5 stable-C gralloc backend (SP-HAL loaded) to Mesa
• 16-wn-add-ubwc-swapchain-usage.py — Request UBWC-compressed swapchain buffers on Adreno
• 17-wn-apply-perf-variant.py — Apply aggressive autotune and variant-specific GPU clock forcing for performance
• 18-wn-a8xx-shared-mem.py — Robust equivalent of whitebelyash/mesa-unified turnip/gen8 commit
• 20-default-env.py — Built-in default environment variables for Turnip with app override support
• 32-tu-z-texfast.py — Turnip-Z port: cheaper texture sampling
• 40-fd710-blob.py — FD710: consolidated hardware property and register tuning from Adreno 710 blob analysis (V@0615)
• 48-a8xx-lrz-color-mask.py — A8xx: Emit GRAS_LRZ_COLOR_COMP_MASK (0x810c) in blend dynamic state
• 50-fd720-blob.py — FD720: register tuning from Adreno 720 blob analysis (V@0676)
• 51-fd810-blob.py — FD810: unified Adreno 810 hardware tuning from Vulkan blob trace analysis (V@0800.74)
• 56-fd722-blob.py — FD722: register tuning and hardware properties for Adreno 722 from Qualcomm blob analysis (chip_id 0x43020100)
• 60-tu-sysmem-wfi-relax.py — Relax subpass slice wait from WFI to WAIT_FOR_ME in sysmem mode
• 61-tu-lrz-fs-sideeffects-no-invalidate.py — Preserve LRZ state across draws when fragment shader has side effects
• 62-tu-draw-state-skip-identical.py — Avoid redundant dynamic draw state packet emission
• 63-ir3-sched-prefer-lower-live.py — Prioritize instructions with lower live register effect during pressure growth
• 7c14092ffb50622f78f9c466d92f2a6c0fce71a7.patch — ir3: Only prefetch UBO descriptors
• ec5a9a6d86133c45d3e42bd3583ba3ad9572c3a2.patch — ir3: Remove texture descriptor prefetch support
• tu-sample-shading-gate.patch — tu: only force per-sample interpolation for shaders that use it

Upstream Freedreno Changes Since 4.4-exp

38 commits under src/freedreno, grouped by commit title.

A7xx (1)

• 73c57838da freedreno/ci: lower the a702 job timeouts

Turnip (tu) (17)

• 8aa6218125 tu: Restore old predicated binning workaround for a630_sqe.fw
• e14789a70e tu: Report unsigned and mixed 4x8 dot acceleration with dp4acc
• f0bc8697b5 tu: Mark signed 4x8 integer dot products as accelerated
• fe3e563ceb tu: Set prefer_gmem for Sweet Surrender VR
• 7e29b87af3 tu: Prefer GMEM on Beat Saber
• 64245a20b7 tu: Add back float16 3D path for multisampled copies
• a1e9f88842 tu/kgsl: Fix BDA capture/replay failures due to deferred VA frees
• a799b0baca tu: Apply tu_enable_ssbo_emulation workaround to all UE5 games
• 69f1f21f02 tu: upload draw_id_offset through driver_params UBO
• 9a7372e8f7 tu: split VS driver params constants setup and emission
• fd5a6795e3 tu: rename skip_vfd to indirect in struct tu_vs_params
• 7638039f12 tu: Split large indirect count draws on A8xx
• 14656cee8c tu: Split large indirect draws on A8xx
• cadc7e42fb tu/a6xx: Invalidate shared consts before disabling them
• af60682ce5 tu: modify hull shader programming
• 4f6314dc9c tu: Drop reference taken with get_vis_stream_patchpoint_cs()
• 761d5c8a54 tu/a6xx: Set correct value for GRAS_MODE_CNTL

Compiler (ir3) (11)

• ed2b00fa28 ir3: assemble the block copy between global memory and local
• 42177b5f4f ir3: decode the block copy between global memory and local
• 8d2fe6ed62 ir3: implement global atomics on a5xx
• fe46ccc999 ir3: implement global load/store for a5xx
• ae42351705 ir3: move the ldg/stg offset handling to a header
• b8141939d2 ir3: print the half prefix on local and private load/store
• 05950e40bf ir3: reserve image dimensions for read-write images on a5xx
• 07fd8b3603 ir3: keep a5xx shaders below the descriptor window in the constant file
• 7656cf464d ir3: do not let full-register writes reset half-register delays before a6xx
• 29db4909d4 ir3/ra: fix DOUBLE_ONLY register limit unit
• 1ae1558b89 ir3: lower draw_id to use draw_id_offset

Other (9)

• 02f1337523 mesa: don't read past the name when matching a resource without its [0]
• ca442f8209 freedreno/a5xx: swap components when storing to an image
• 724b21ef16 freedreno/a5xx: let half arithmetic overflow to infinity
• 07f52057ca freedreno/a5xx: decompose strips and fans in the D3D order for the first-vertex convention
• e5c0aa8e52 freedreno/a5xx: keep hardware-binned bins below 1024 rows
• 57fea5146f ci/lava: uprev lava-job-submitter
• d1337cf9f7 freedreno/ci: reenable dEQP UBO tests
• 0e8f40efb0 freedreno: Fixup incorrectly copied A6XX_TPL1_DBG_ECO_CNTL for A650
• d1526c1885 freedreno/devices: Add the Adreno 829 msm chip id

GPU Support

Supports all GPUs currently supported by Mesa, including Adreno 6xx, 7xx, and 8xx series.
For the complete and up-to-date device list, see:
https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/freedreno/common/freedreno_devices.py

Notes

• Graphics issues may still occur depending on the game or emulator.
• Performance and compatibility vary by device, firmware, emulator version, and game.
• Recommended memory mode: sysmem (TU_DEBUG=sysmem).
• Experimental build.

---

This discussion was created from the release Turnip Adreno 710,720,722 v4.5.

[jodista]

Eden Emulator runs very smoothly on my Poco Pad, getting 40–60 FPS with the Snapdragon 7s Gen 2 and Adreno 710.

[Originalys]

Only good driver i've found for my 7s gen 4 A810, literally life and sanity saver.
Eden runs perfectly fine and without any visual bugs, unlike system drivers