openCL decompression

[KornX]

hi,
anybody ever thought about or tried openCL decompression of h264 (or maybe vc1)?
The reason I am asking is many hardware video decoders are only capable of ~1080p content...

KornX

[LoRd_MuldeR]

"Hardware" decoding doesn't happen on the actual GPU, but in a special "video engine". NVidia calls that "PureVideo HD" (VPx), AMD calls theirs "Unified Video Decoder" (UVD).

Regardless of what API you use to access the "hardware" decoder (DXVA, CUVID, VDPAU or maybe OpenCL), the limitations on what works (or doesn't work) is defined by the hardwired decoding engine...

[KornX]

I know where it happens. thats the reason why my q is to offload some off the sw decoding from the cpu to the gpu via openCL...
So there comes no api to the UVD or whatsoever, but directrly to the compute shaders...

KornX

[LoRd_MuldeR]

This most likely will run into the same problems as GPU-accelerated encoding:

Offloading only some functions of the decoding process to the GPU will have a severe delay for uploading the inputs from CPU memory to GPU memory and then downloading the results again. This delay can easily destroy any speed-benefit, even if the calculation itself runs a lot faster on the GPU. Now you may argue that the upload/download delay can be hidden using a "smart" pipelining. That may be right, but it doesn't make things easier ;)

Also, when having a fully-fledged H.264 "hardware" decoder available as a hardwired piece of silicon on pretty much any graphics card sold in the last ~7 years, motivation to write an OpenCL-based "GPU" decoder from the scratch is low...

(Once "4K" content becomes more popular, you can be pretty sure that NVidia and AMD will "upgrade" their hardware decoders accordingly - with the next GPU generation)

[Dark Shikari]

GPUs are an order of magnitude too slow to do video decoding, which is largely a linear process and can't be parallelized the way a GPU needs. That's why hardware decoders exist in the first place (besides the fact that they're nice and fast and take almost no power or silicon).