[Android] Video color-corrupted on TCL/PowerVR with every fvp decoder, but media_kit (also libmpv) renders fine on the same device

[abdelaziz-mahdy]

On a TCL Google TV (Realtek RTD2875P SoC, PowerVR GPU, Android 12),
fvp plays streams (audio fine, decoding runs at full frame rate) but the
video picture is heavily color-corrupted — chromatic noise + block artifacts
across the whole frame — and stutters.

The decisive finding: media_kit (also libmpv-based) renders the exact same
streams perfectly on the exact same device. Since both fvp/MDK and media_kit
use libmpv to demux/decode, the difference is the renderer — which points at
MDK's GL renderer / texture path on this PowerVR driver, not the GPU, libmpv,
or decoding.

Screenshot of the corruption:

Decisive comparison

Engine (Flutter)	Demux/decode	Renders correctly on this device?
fvp 0.36.x (MDK GL renderer)	libmpv	❌ corrupted
media_kit 1.2.6 (mpv vo_gpu)	libmpv	✅ perfect (correct colors, smooth)

Same decoding stack, same streams, same GPU — only the renderer differs. fvp/MDK
corrupts; mpv's vo_gpu does not.

Environment

fvp	0.36.1 (MDK 0.36.0 (git d3ec2cc))
Flutter	3.44.2 (stable), engine 77e2e94772, Dart 3.12.2
video_player	2.11.1
ABI in use	armeabi-v7a (32-bit) — MDK banner: Build for: Android21/31(12.0) ... 4KB ARMv7

Device

prop	value
model	TCL Smart TV G10_4K_US_NF (device G10)
board / platform	rtd6748
SoC	Realtek RTD2875P
GPU	PowerVR (ro.hardware.vulkan = powervr)
OS	Android 12 (API 31)
display	3840×2160 @ 60
EGL	1.5 Android META-EGL

Observations

• Audio is correct.
• Decoding is healthy: MediaCodec logs show steady queue/release at the
  content's native rate (~24–25 fps): [QIB]V, QIB:24/s / [ROB]V, ROB:25/s.
• Video is color-garbled (looks like a YUV plane / stride / color-format
  misinterpretation) and stutters despite the decoder running at rate.
• Non-video Flutter UI renders perfectly on the same GPU (Skia) — only the
  fvp video texture is wrong.

What I tried (every config shows the same corruption)

registerWith config	Result
default (AMediaCodec, image=1)	corrupted
video.decoders: ['AMediaCodec:image=0'] (SurfaceTexture/OES)	corrupted
video.decoders: ['AMediaCodec:copy=1']	corrupted
video.decoders: ['FFmpeg'] (software)	corrupted ← rules out the decoder
tunnel: true	renders nothing (black) + crash on player dispose

Reproduces under both Flutter Skia (EnableImpeller=false) and Impeller
(Vulkan).

Questions

1. Given media_kit's mpv vo_gpu renders correctly on this PowerVR driver, is
   there an MDK renderer / global option (color format, texture format, GL
   workaround) that would fix MDK's output here?
2. Is there a supported way to render fvp to a native Android SurfaceView
   (system compositor) for video_player-based integrations?

Logs

MDK init banner:

MDK: 0.36.0 (git d3ec2cc) - Multimedia Development Kit.
MDK: Build for: Android21/31(12.0) 5.4.284+ 4KB ARMv7; libc++180000; Clang18.0.4 ... NDK27.3.13750724. Smart TV TCL TCL

app-full-log.txt

mdk-extract.txt

Full MDK debug log (SetGlobalOption("logLevel","all")) and the corrupted-frame
screenshot available on request — happy to capture whatever traces help.

[wang-bin]

the log is incomplete

[abdelaziz-mahdy]

sorry yeah it was incomplete
Added here

fvp-capture-full.log

fvp-capture-mdk.log

[wang-bin]

can you show me the full log for FFmpeg software decoder?

[abdelaziz-mahdy]

Sorry for the delay getting back to the TV. I updated fvp to the latest and rebuilt with the latest MDK SDK (FVP_DEPS_LATEST=1), then re-tested.

Versions

• fvp 0.37.3
• MDK 0.37.0 (git f24275a) — banner: Build for: Android21/31(12.0) … NDK27.3.13750724; Jul 1 2026
• FFmpeg runtime git-2026-06-21-ff1e1b5-avbuild

Result: still corrupted with both decoders (screenshots unchanged from before). Since software decode corrupts too, this points at the MDK GL renderer / EGLImage texture path on PowerVR, not the decoder. Relevant driver errors in the logs:

• E IMGSRV: IsTextureConsistent: IMGEGLImage is not consistent
• E IMGSRV: DevmemAcquireCpuVirtAddr() failed (PVRSRV_ERROR_DEVICEMEM_MAP_FAILED) / Failed to lock buffer

First log config — software FFmpeg decoder
fvp.registerWith(options: {'video.decoders': ['FFmpeg']});
→ log confirms opening ffmpeg video decoder: h264 / yuv420p using avcodec software decoder
fvp-ffmpeg.log

Second log config — default / hardware decoder
fvp.registerWith(); (Android default priority ['AMediaCodec', 'FFmpeg', 'dav1d'])
→ log confirms creating video decoder: AMediaCodec
fvp-normal.log

[wang-bin]

can you see the same driver error log when using libmpv?

[abdelaziz-mahdy]

Yes — but it splits into two different errors, and the split is the interesting part.

I ran the same stream through media_kit (libmpv, mpv vo_gpu renderer) on the exact same device/GPU, where it renders correctly, and compared the logs against the fvp/MDK runs from my previous comment:

DevmemAcquireCpuVirtAddr() failed (PVRSRV_ERROR_DEVICEMEM_MAP_FAILED) / lock: Failed to lock buffer
→ Same error appears with libmpv too (hundreds of times). It's logged by the PowerVR driver service process (IMGSRV, pid ~401), not the app process, and it shows up even though libmpv renders perfectly. So this one looks like benign/background driver noise, unrelated to the corruption.

IsTextureConsistent: IMGEGLImage is not consistent
→ Does NOT appear with libmpv (0 occurrences). It only shows up with fvp/MDK, and it's logged inside the app process (same pid as the player) during MDK's EGLImage texture path. This is the one that tracks with the corrupted output.

Error	fvp/MDK (corrupt)	libmpv/media_kit (correct)	logged by
DevmemAcquireCpuVirtAddr … MAP_FAILED / Failed to lock buffer	yes	yes	driver service (IMGSRV pid ~401)
IsTextureConsistent: IMGEGLImage is not consistent	yes (app pid)	no	app process (MDK)

So the driver-service errors are shared, but the IMGEGLImage is not consistent one is unique to MDK's texture path.

(Full libmpv logcat available if you want it.)

[wang-bin]

the result from fable 5:
fvp-374-analysis-zh.html
fvp-374-analysis-zh.md

seems reasonable, you can try to add putenv("EGL_SDR_DEPTH=8"); or putenv("GL_BPC=8"); in JNI_OnLoad() in fvp_plugin.cpp, if it works, i will try to find a better solution to detect 10bit surface support. Can you enable mpv log? I have to confirm the output format used by mpv is 8bit.

[abdelaziz-mahdy]

EGL_SDR_DEPTH=8 fixes it — added via putenv() in JNI_OnLoad(), rebuilt, same device and streams (default AMediaCodec zero-copy path):

I/MDK: 0.37.0 (git b345e19) - Multimedia Development Kit.
[mdk] FINE: select a 8bits EGLConfig
[mdk] FINE: EGL_NATIVE_VISUAL_ID: 1        // previously: 10bits → visual id 43 (RGBA_1010102)
[mdk] FINE: player 1st video frame rendered @194.036000ms

• IsTextureConsistent: IMGEGLImage is not consistent — 0 occurrences over full playback (previously fired at first bind)
• IMGSRV now allocates the window buffers as fmt:0xc101 (8888 + vendor bits); no 0xc12b (1010102) allocations anywhere
• Picture is completely clean on all decoders, no stutter, seek/resume normal

Re mpv's output format — verbose log (msg-level=all=v) on the same device/stream:

[vo/gpu/opengl] v: Initializing GPU context 'android'
[vo/gpu/opengl] v: GL_RENDERER='PowerVR B-Series BXE-4-32'
[vo/gpu] v: Reported display depth: 8

mpv takes the driver's first ≥8-bit config (RGBA_8888) and never enters the 1010102 path — its buffers also allocate as fmt:0xc101/0x1 with zero consistency errors.

Could fvp expose this on Android — setRenderAPI() with GLRenderAPI.depth = 8 like the Windows D3D11 path, ideally as a registerWith option? Happy to test any patch on this device.

[wang-bin]

I see mpv default is 8bit, flutter default is 8bit too, it's fine here. You can try to add these lines before player->updateNativeSurface(surface, w, h);

mdk::GLRenderAPI ra{};
ra.depth = 8;
player->setRenderAPI(&ra, surface);

An ideal approach is detecting whether rgb10a2 is well supported by driver, but egl accepts this format w/o error.

[abdelaziz-mahdy]

We shipped a slightly different workaround on the app side and it works well: setting the env var before the Flutter engine loads any native libs, in MainActivity.onCreate():

Os.setenv("EGL_SDR_DEPTH", "8", true)

Since then: select a 8bits EGLConfig, zero IsTextureConsistent errors, clean playback on all decoders.

Happy to turn your setRenderAPI(depth = 8) snippet into a PR — though unconditional 8-bit would cost 10-bit/HDR output on devices where rgb10a2 works fine, so it probably wants to be opt-in or detected.

On detection — since the failure only shows when a second GL context samples the buffer (EGL reports no error at config/alloc/render time), a query can't catch it, but a one-time probe can: render a known gradient into an RGBA_1010102 ImageReader surface from one EGL context, import the buffer as an EGLImage in a second context, read back and compare, and pick depth = 8 only on mismatch. That can live entirely in fvp_plugin.cpp before updateNativeSurface(), cached after first run. I'm happy to implement and test it (this device reliably reproduces the failure) if you'd take it as a follow-up PR.

[wang-bin]

Can you try eglChooseConfig with the following attributes

#define EGL_WARN(x, ...) EGL_RUN_CHECK(x)
#define EGL_RUN_CHECK(x, ...) do { \
        while (eglGetError() != EGL_SUCCESS) {} \
        (x); \
        const EGLint err = eglGetError(); \
        if (err != EGL_SUCCESS) { \
            std::clog << #x " EGL ERROR (" << std::hex << err << std::dec << ") @" << __LINE__ << __FUNCTION__ << std::endl; \
            __VA_ARGS__; \
        } \
    } while(false)

EGLint ca[] = { 
  EGL_CONFIG_CAVEAT, EGL_NONE,  // default is EGL_DONT_CARE and may select slow config
  EGL_BUFFER_SIZE, 32, 
  EGL_RED_SIZE, 10, 
  EGL_GREEN_SIZE, 10,
  EGL_BLUE_SIZE, 10,
  EGL_ALPHA_SIZE, 2,
  EGL_RENDERABLE_TYPE, EGL_OPENGL_ES2_BIT | EGL_OPENGL_ES3_BIT,
  EGL_SURFACE_TYPE, EGL_WINDOW_BIT,
  EGL_NONE
};

    EGLint nb_cfgs; // eglGetConfigs
    vector<EGLConfig> cfgs; // sort by best match(and highest depth?)
    EGLBoolean ret = EGL_FALSE;
    EGL_WARN( ret = eglChooseConfig(disp, ca, nullptr, 0, &nb_cfgs));
// log ret and nb_cfgs
    if (ret == EGL_FALSE || nb_cfgs < 1) {
      ....
    }

in libmdk, EGL_CONFIG_CAVEAT is not set, so may get a bad config. if set to EGL_NONE and see error log from EGL_WARN(), then we can use this solution. You can also try EGL_CONFORMANT attribute

[wang-bin]

my phone only accepts EGL_DONT_CARE and EGL_NON_CONFORMANT_CONFIG for EGL_CONFIG_CAVEAT. can you test EGL_NONE, EGL_NON_CONFORMANT_CONFIG and EGL_SLOW_CONFIG?