Posts: 110
Threads: 4
Joined: Jul 2015
Reputation:
1
07-02-2015, 05:13 PM
(This post was last modified: 07-02-2015, 05:18 PM by sirdaniel.)
Hi all.
Some sort of avoiding transfers between CPU<>GPU was creating an APU, which has same physical memory for CPU and GPU. BUt thats not all, because CPU cant normally grab data that is assigned to GPU. To help that, they created a zero_copy in OpenCL API. So a small OpenCL program can be written with zero copy so the data can be copied from GPU to CPU pool just instantly in memory. The effect would be like no copy at all or instant copy (with the speed of RAM). This is somehow "software" solution. The "hardware" one is in Carrizo which should be compatible with HSA 1 and that would be accessing the same data in memory directly by GPU or CPU.
I dont know if such big project like pcsx2/gsdx can be programmed that way, but there is actually an Opencl version. I am just a user, like read technical stuff, but dont know programming. Any dev can look on datasheets and try this? Unfortunately, AMD's hype got lowered on this matter. So, maybe that small chip Carrizo will make progress on speed?
Posts: 6.069
Threads: 68
Joined: May 2010
Reputation:
167
Location: Grenoble, France
Yes we know about it and GSdx already has been ported to openCL but it was never tested on APU (I don't think the port uses zero copy anyway).
APU win in memory transfer but loose on GPU/CPU side. In my opinion to really exploit APU, GSdx will need a dedicated texture cache/vertex buffer with zero copy feature. However all PCSX2's dev have a dedicated GPU so it won't be for today.
Posts: 8.597
Threads: 105
Joined: May 2014
Reputation:
168
Location: 127.0.0.1
on software mode, only a rendering of a texture quad is done. CPU is the important factor influencing speed on software mode.
We're supposed to be working as a team, if we aren't helping and suggesting things to each other, we aren't working as a team.
- Refraction