GTA V on Xiaomi 17T Pro (MediaTek MT6993, Mali-G1-Ultra): working configuration

[JuliusPinsker]

I have wanted to play GTA V on a phone since I was a teenager. GameNative is the app I would have tried to write myself, and this month it got me there on a Mali device. Thank you to everyone who works on it. Fork with the fixes, a release with all files and a step-by-step guide: https://github.com/JuliusPinsker/GameNative (release https://github.com/JuliusPinsker/GameNative/releases/tag/gta5-mali-g1-v1). Below is the configuration that works, with measurements, in case it helps the next Mali user. Prior art for Adreno: tarikbc/gtav-android, which saved me a lot of time.

Device and versions

• Xiaomi 17T Pro: Dimensity 9500, Mali-G1-Ultra (driver r54p1, no native BCn), HyperOS 3, Android 16, no root
• GameNative master c78519c (1.3.0), own build with the fixes linked below
• GTA V Legacy (Steam), Story Mode only

Container

Setting	Value
Container variant	Bionic, "Experimental Bionic Steam" on (starts the Rockstar stub directly)
Wine	proton-11.0-2-arm64ec-1
Emulator	FEXCore 2605, preset Intermediate
Graphics driver	Wrapper-leegao with three local patches (see below), bcnEmulation=none so DXVK passes BC formats to the wrapper
DX wrapper	DXVK-Sarek 1.14.0 (content package)
Env	DXVK_CONFIG=dxgi.maxDeviceMemory=1024;dxgi.maxSharedMemory=1024;dxgi.maxFrameRate=30
Screen	1440x720, stretch scaling
ntsync	on (default)

GTA settings.xml: borderless window at the container resolution, VSync off, texture quality normal or lower, everything else at minimum except shadows (ShadowQuality 0 freezes the game reliably).

Results

• Built-in benchmark, 1280x720: 54.1 FPS mean (passes 49.5 to 61.1). At 1440x720: 48.5 to 52.4 FPS mean, depending on temperature and battery level.
• Real driving (a small ScriptHookV autodrive script for repeatable routes): 38 to 40 FPS, CPU at 82 to 90 °C, big cores held at 1.8 to 2.7 GHz of 3.5/4.2 GHz. The frame rate follows the CPU clock, and the governor holds the big cores well below their maximum.
• City route, 15 min each: uncapped 37.4 FPS with 5.3 frames over 50 ms per 1000; capped at 30 it holds 29.5 FPS with 2.8 per 1000. A cap of 40 was worse than both (32.2 FPS, 23.7 per 1000). I play with the 30 cap.
• Stable over long sessions: two 45 min benchmark soaks without a crash, plus evenings of normal play. GPU memory about 2.5 GB, around 2 GB of RAM free.

What it took

1. DXVK memory budget. Without a working dxgi.maxDeviceMemory, GPU memory grows to 5.4 GB and the game dies after about 8 minutes. The container setting does not reach DXVK today; the env line above does. Fix: Pass DXVK_CONFIG in the format DXVK parses #2170.
2. VMA limit (vm.max_map_count 65530). On kbase every GPU allocation is its own mapping. Two wrapper fixes keep GTA under 15k mappings: Free BCn staging buffers when their command buffer is begun again leegao/bionic-vulkan-wrapper#227 and Clear prefersDedicatedAllocation in memory requirements leegao/bionic-vulkan-wrapper#228.
3. Memory headroom. Decoding BC1/BC4/BC5 to 16-bit and 8-bit targets saves about 0.75 GB at the same frame rate: Smaller targets for BCn emulation (16-bit and 8-bit formats) to cut GPU memory on Mali leegao/bionic-vulkan-wrapper#229.
4. Controller rumble that never stops (GameSir): Re-arm controller rumble from the current shared-memory value #2169.
5. wrapper-gamenative (ASTC) needs a Present fix to start at all: Present: answer QueryCapabilities and accept NotifyMSC #2171. It ran stably, but at about 20 FPS, so I stayed on wrapper-leegao.

Levers that did nothing or hurt

Measured with the benchmark; single-run spread is about ±1 FPS.

• ntsync off (PROTON_NO_NTSYNC=1): 18.7 vs 49.9 FPS. Never turn it off.
• FEX Multiblock off: 33.4 vs 48.5. TSO off: crash after about 90 s.
• Budget 512 or 2048 (vs 1024): both run into the memory wall again.
• No gain: core pinning, FEX L1/L2 cache switches, FEX code cache (about 8 FPS slower in both runs), DXVK async, prefaulting the heap, reflection/grass/population settings.
• HyperOS performance mode: no difference in sustained play; only a 90 s low-clock phase right after switching.
• Battery at 7 % cost about 5 FPS compared with 70 %.

Pitfalls

• "Activation Required": the Rockstar account has to be the one linked to the Steam copy. On the very first start after installing the app, Steam bootstrap needs one more launch.
• A VPN on the phone broke the Steam login of the game process (SteamHost: not logged on) while the app itself stayed logged in. Excluding GameNative in the VPN's split tunnelling fixed it.
• After every crash GTA rewrites settings.xml to its defaults (800x600, VSync half). Check it before the next run.
• GameNative loads a dinput8.dll in the game folder automatically (ModDllOverrides). Handy for Script Hook V, but move the files out when you want the plain game.
• DXVK_SHADER_DUMP_PATH on shared storage kills the game after a few hundred shaders; a path inside the prefix works.

One more trap: container LD_PRELOAD

Setting LD_PRELOAD in a container's environment variables replaces the preload list the bionic launcher builds (libandroid-sysvshm.so, libevshim.so, libgamenative_dns_v4mapped.so, libredirect). BionicProgramLauncherComponent builds LD_PRELOAD first and then does envVars.putAll(this.envVars) with the container env, so the user value wins outright.

Why it hurts. Anyone who preloads one extra library (I used mimalloc to cut allocator mappings) must copy the app's own entries into the container by hand. Those are absolute paths under /data/app/~~<hash>/app.gamenative-<hash>/lib/arm64/, and the hash changes with every APK install. After the next update wine fails immediately:

CANNOT LINK EXECUTABLE ... libevshim.so not found

and the game dies at launch with no hint in the UI. Without the evshim entry the controller shim is missing as well.

Suggestion. Append the container's LD_PRELOAD after the launcher's list (launcher:user), the same way the XR shim gets prepended. Replacing the whole list could stay available behind a separate switch for the rare user who needs it.

🤖 Generated with Claude Code