Android emulator won’t boot on AMD GPU? It’s your Vulkan driver.

I lost days to this. The Android emulator would launch, show the boot animation, and just sit there. No error dialog in Android Studio, no crash popup, nothing. An infinite boot loop with zero indication of what went wrong.

If you have an AMD GPU on Linux with AMDVLK installed, this post should save you some time.

The problem

I’m working on an Android project with an emulator image (android-34, x86_64). AMD Ryzen + AMD Radeon RX (RDNA 3), Ubuntu. The emulator process starts fine. QEMU runs, ADB connects, boot animation plays. But sys.boot_completed never becomes 1. The system hangs mid-boot and never recovers.

Android Studio tells you nothing. I tried cold boots, wiping data, recreating the AVD, different API levels. None of it mattered.

Finding the root cause

I ended up working through this with Claude Code, which got me past the guessing stage and into the system logs. Took maybe 15 minutes once we stopped flailing.

Confirm the boot is stuck

adb shell getprop sys.boot_completed
# (empty)

adb shell getprop init.svc.bootanim
# running

Empty sys.boot_completed and a still-running boot animation. System is alive but stuck.

Check logcat

adb logcat -d | grep -i -E 'fatal|crash|vulkan'

This surfaced the problem immediately:

Vulkan Instance Version: 1.3.0
Device AMD Radeon RX XXXX, API Version 1.3.0, Driver Version 2.0.349
Failed to find appropriate memory type for 10 with properties 6!

The Vulkan compositor crashed during init. It couldn’t find a memory type it needed.

Check the crash buffer

adb logcat -d -b crash

Full crash trace:

signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0
Cause: null pointer dereference
Cmdline: /system/bin/surfaceflinger

backtrace:
  #00 libvulkan_compositor.so (GraphicsObject::GetVkDevice() const+0)
  #01 libvulkan_compositor.so (LayerTextureCache::Initialize()+1563)
  #02 libvulkan_compositor.so (compositor::CreateDevice()+864)
  #03 surfaceflinger (VulkanRendererDevice::init()+2284)

Followed by:

init: process 'surfaceflinger' exited 4 times before boot completed

Vulkan memory allocation fails, VkDevice comes back null, null pointer dereference, surfaceflinger crashes. Init restarts it, same crash, four times, init gives up. Boot hangs forever.

Root cause

The Android emulator uses gfxstream to translate Vulkan calls from the guest to your host GPU. It maps guest Vulkan memory types to host memory types.

AMDVLK (AMD’s official Vulkan driver) exposes AMD-specific memory property flags through the VK_AMD_device_coherent_memory extension:

# AMDVLK: 16 memory types, including non-standard ones
Type 0:  DEVICE_LOCAL                                     (0x0001)
Type 1:  HOST_VISIBLE | HOST_COHERENT                     (0x0006)
Type 2:  DEVICE_LOCAL | HOST_VISIBLE | HOST_COHERENT      (0x0007)
Type 3:  HOST_VISIBLE | HOST_COHERENT | HOST_CACHED       (0x000e)
Type 4:  DEVICE_LOCAL | DEVICE_COHERENT_AMD | DEVICE_UNCACHED_AMD  (0x00c1)  <-- non-standard
Type 5:  HOST_VISIBLE | HOST_COHERENT | DEVICE_COHERENT_AMD | ...  (0x00c6)  <-- non-standard
...

DEVICE_COHERENT_BIT_AMD and DEVICE_UNCACHED_BIT_AMD are not part of the core Vulkan spec. gfxstream doesn’t know what to do with them. The compositor asks for HOST_VISIBLE | HOST_COHERENT memory, gfxstream chokes on the extra AMD flags, the lookup fails.

I verified this by reading the gfxstream source. The bug is a missing filter at three points:

In vk_emulated_physical_device_memory.cpp, host memory properties are copied to the guest verbatim:

mHostMemoryProperties = hostMemoryProperties;
mGuestMemoryProperties = hostMemoryProperties;  // AMD flags included, no filtering

Later, when gfxstream strips HOST_COHERENT_BIT for certain configurations, it only strips the standard flag (0x4) and leaves the AMD-specific DEVICE_COHERENT_BIT_AMD (0x40) intact. The guest ends up with an inconsistent set of memory flags.

In vk_decoder_global_state.cpp, the extension filter list (kEmulatedDeviceExtensions) only blocks Android/Fuchsia/external-memory extensions. VK_AMD_device_coherent_memory is not in the list, so the guest thinks the extension is available and expects the non-standard memory types to work through gfxstream. They don’t.

And in transformToGuestMemoryRequirements(), the memoryTypeBits are remapped by index but the property flags are never masked. AMD-specific flags pass straight through to the guest compositor, which has no idea what to do with them.

gfxstream should either strip AMD-specific memory property flags from guest memory types, or filter the VK_AMD_device_coherent_memory extension from the device extension list. It does neither.

RADV (Mesa’s Vulkan driver for AMD) doesn’t have this problem because it only exposes standard Vulkan memory types:

# RADV: 11 memory types, all standard
Type 0:  DEVICE_LOCAL                                (0x0001)
Type 1:  DEVICE_LOCAL                                (0x0001)
Type 2:  HOST_VISIBLE | HOST_COHERENT                (0x0006)
Type 3:  DEVICE_LOCAL | HOST_VISIBLE | HOST_COHERENT (0x0007)
Type 4:  DEVICE_LOCAL | HOST_VISIBLE | HOST_COHERENT (0x0007)
Type 5:  HOST_VISIBLE | HOST_COHERENT | HOST_CACHED  (0x000e)
...

No vendor extensions, no confusion, gfxstream maps everything correctly.

The fix

Just remove AMDVLK:

sudo apt remove amdvlk

That’s it. RADV (which ships with Mesa and is already installed on most Linux distros) becomes the only Vulkan driver. No environment variables, no per-app overrides, no profile files to worry about. Restart Android Studio and the emulator boots.

Verify:

vulkaninfo --summary 2>&1 | grep driverName
# Should show: radv

I initially tried a softer fix, setting VK_ICD_FILENAMES to force RADV while keeping AMDVLK installed. That works too, but it’s fiddly. You have to put it in ~/.profile (not ~/.bashrc) because GUI apps like Android Studio don’t source bash configs. And you have to remember to override it back if you ever want AMDVLK for something. Removing the package is simpler and I haven’t missed it.

Quick check: are you affected?

# Is AMDVLK installed?
dpkg -l | grep amdvlk

# Which driver is active?
vulkaninfo --summary 2>&1 | grep driverName
# "AMD open-source driver" = AMDVLK (problematic)
# "radv" = RADV (works)

# Do you have non-standard memory types?
vulkaninfo 2>&1 | grep DEVICE_COHERENT
# If this returns anything, AMDVLK is active

AMDVLK vs RADV

AMD is replacing AMDVLK with RADV as their official Vulkan driver on Linux. RADV is already the default on Steam Deck, Ubuntu, Fedora, and Arch. If you still have AMDVLK installed, now is a good time to remove it.

Summary

If your Android emulator won’t boot on an AMD GPU on Linux:

adb logcat -d -b crash – look for Vulkan crashes
vulkaninfo --summary | grep driverName – check your driver
If it says “AMD open-source driver”:
```
sudo apt remove amdvlk
```
Restart Android Studio and the emulator

I wish the emulator surfaced this error somewhere visible instead of silently hanging. But at least the logcat trail is clear once you know to look.