Error using --vpp-resize mfx_ai_superres in QSVEncC 7.71 & 7.70
Closed this issue · 4 comments
Description:
While attempting to use the --vpp-resize mfx_ai_superres option in QSVEncC 7.71 and 7.70, the encoding process fails with an error. This issue appears to be related to buffer allocation and AI-based super-resolution in Intel’s Media SDK.
Environment:
QSVEncC Version: QSVEncC (x64) 7.71 (r3471) by rigaya, Oct 27 2024 12:02:14 (VC 1941/Win)
Operating System: Windows 11 x64 (build 22631), UTF-8
CPU: Intel Core i5-13600 [4.92GHz] (6P+8E, 14C/20T)
GPU: Intel Arc A770 Graphics (512EU), driver version 32.0.101.6127
Media SDK: QuickSyncVideo (hardware encoder) FF, 1st GPU(d), API v2.13
MFXVPP: Failed to get required buffer size for MFXVPP: undeveloped feature.
allocD3D11: QSVAllocator: Failed CheckRequestType: undeveloped feature.
INPUT: allocWorkSurfaces: Failed to allocate frames: undeveloped feature.
AllocFrames: Failed to allocate frames for INPUT-MFXVPP: undeveloped feature.
avout: failed to write header for output file: Invalid data found when processing input
QSVEncC.exe finished with error!
--avhw
--cqp 32:34:36
--output-res -2x1080
--codec hevc
--output-depth 10
--bframes 16
--vpp-resize mfx_ai_superres
--audio-copy
--chapter-copy
Additional Information:
Since installing the latest Intel GPU driver (32.0.101.6127), this issue has started occurring, affecting both QSVEncC versions 7.71 and 7.70.
The error only occurs when using the --output-depth 10 option; removing this option allows the encoding to proceed without issues.
The issue might be related to compatibility between 10-bit output depth, the AI super-resolution feature, and the latest Media SDK or Intel Arc drivers.
In the previous conversation, there was information that the input and output of VPP for AI super resolution is NV12.
10bit depth requires P010(10bit/16bit) format instead of NV12 (8bit). Although I actually haven't tested 10bit in older drivers, I think it can be thought as limitation of AI super-resolution.
In the previous conversation, there was information that the input and output of VPP for AI super resolution is NV12.
10bit depth requires P010(10bit/16bit) format instead of NV12 (8bit). Although I actually haven't tested 10bit in older drivers, I think it can be thought as limitation of AI super-resolution.
Thanks for the info.
Strangely, --vpp-resize mfx_ai_superres with --output-depth 10 worked fine before the latest driver update (32.0.101.6127).
I try to roll back.
I'm closing this.
I will be using another upscaler.
The driver's fallback did not help.
Thank you@rigaya for checking :-)
I'm on 32.101.6078 and this driver doesn't help either.