cuvidCreateDecoder fails with CUDA_ERROR_OUT_OF_MEMORY?

Gaming and Visualization Technologies Visualization Video Processing & Optical Flow
#1

Hi,all:
I am using Quadro M2000 card for hardware decoding.However,I fail to create the ninth decode session with params below.
CUVIDDECODECREATEINFO oVideoDecodeCreateInfo_;
memset(&oVideoDecodeCreateInfo_, 0, sizeof(CUVIDDECODECREATEINFO));

// Create video decoder
oVideoDecodeCreateInfo_.CodecType           = cudaVideoCodec_H264;
oVideoDecodeCreateInfo_.ulWidth             = 1280;
oVideoDecodeCreateInfo_.ulHeight            = 720;
oVideoDecodeCreateInfo_.ChromaFormat        = cudaVideoChromaFormat_420;
oVideoDecodeCreateInfo_.ulNumDecodeSurfaces = 16;

oVideoDecodeCreateInfo_.OutputFormat        = cudaVideoSurfaceFormat_NV12;
oVideoDecodeCreateInfo_.DeinterlaceMode     = cudaVideoDeinterlaceMode_Adaptive;

// No scaling
oVideoDecodeCreateInfo_.ulTargetWidth       = 1280;
oVideoDecodeCreateInfo_.ulTargetHeight      = 720;
oVideoDecodeCreateInfo_.ulNumOutputSurfaces  = 1;

CUresult ret;
CUvideodecoder decoder = NULL;
ret = cuvidCreateDecoder(&decoder, &oVideoDecodeCreateInfo_);

The ret is CUDA_ERROR_OUT_OF_MEMORY. But It takes no effect to reduce oVideoDecodeCreateInfo_.ulNumDecodeSurfaces from 16 to 4. So I wonder whether 8 decode sessions is the limit of Quadro M2000 card. I look to https://developer.nvidia.com/nvidia-video-codec-sdk#NVDECFeatures and https://developer.nvidia.com/video-encode-decode-gpu-support-matrix/#Decoder for help, there is no answer to my puzzle, however.
Is there any one can help?
Thanks in advance.
#2

Fix it. Too many cuda contexts lead to this problem. Just create one cuda context and push it current to each decoder thread.

GT 720 video decoder count limitation
#3

Hi,
do you push it to decoder thread via cuCtxPushCurrent or via cuCtxSetCurrent.
Do you use:
CUVIDDECODECREATEINFO.vidLock, cuvidCtxLockCreate,cuvidCtxLock/cuvidCtxUnlock to synchronize context between decoding threads or it’s enough to call cuCtxSetCurrent in each ?

My current implementation just call cuCtxSetCurrent in each decoding thread and don’t use “CUVIDDECODECREATEINFO.vidLock, cuvidCtxLockCreate,cuvidCtxLock/cuvidCtxUnlock”. Is it ok ? There is no any recommendations in NVDEC docs. I found cuCtxSetCurrent accidentally and now use it and it works but I cannot find any confirmation that it’s correct approach.

#4

maybe some contexts are not destroyed.Check you code, if so, you can call cuCtxDestroy(cuContext).