Why is Src and Dst memory not pinned?

i completed first video filter cuda accelerated using premiere pro sdk cs6 , cc7.  if suitable cuda hardware available, use it, if not uses multithreaded software implementation.  both work extremely well.  however, cuda implementation lot faster if source , destination memory buffers "pinned" memory.  since not, must copy source , destination memory pinned buffer, , asynchronously copy cuda device memory , back.  overhead copy source/destination memory pinned memory significant.  without copy pinned memory cuda on laptop fast enought process 130 fps 1920 1080 hd video.  pinned memory 45 fps.

 

if exact same filter in directshow, source , destination buffer pools pinned , filter runs faster on premiere pro.  noticed new gpu filter example uses , ae interface, allow access pinned memory.  however, have not mastered ae interface, , reluctant giving 13 years of learning curve on premiere sdk.

 

is there reason why source , destination buffer pool in premiere pro sdk not pinned memory?

 

gene

hey gene, pinned memory applies host memory. cuda memory given through gpu suite device resident pinning not apply , can perform gpu computation directly without transfer.

More discussions in Premiere Pro SDK

adobe