OpenCL Port for Thea Presto (wip)

Fri Sep 26, 2014 10:45 am

eph wrote: How would you say the current OpenCL port compares to CUDA speed? Does it have the potential to equal / surpass it in terms of performance?

Theoretically speaking, the expectation is that going with a more direct way (Presto CPU or proprietary CUDA) will give better performance compared to OpenCL. From there on, unless we have a stable build and tune it (i.e. optimizing it as much as we did with the other APIs) the comparison will not be fair, so it is only a qualitative index now.

Best wishes

Fri Sep 26, 2014 2:51 pm

That's amazing news.

Sun Sep 28, 2014 11:34 pm

great news!

Mon Sep 29, 2014 1:59 am

Wow thats one mean feat !! nice work !

I wonder how the new nvidia 980's will go with OpenCL ? From what I have scene the luxmark bench marks look impressive.

cheers

Mon Sep 29, 2014 6:49 am

I can say already (after a few tests) that going with CUDA for NVidia produces faster results than going with OpenCL. Although, the OpenCL port is not as optimized yet, this is to be expected because OpenCL is one layer more abstract than CUDA (compilation goes through CUDA as far as I know).

Best wishes

Mon Sep 29, 2014 12:36 pm

Great work! From what I heard about Open CL I thought it would be almost impossible to do. I would have thought it to be simpler to get Thea working with intel phi than Open CL from what I read.

It's exciting to know that when done we can compare Nvidia to AMD. Do you know yet if there will be a minimum AMD GPU we will need for it to work or will something old like a Radeon 8500 or 4770 work?

Mon Sep 29, 2014 5:28 pm

giannis wrote:I can say already (after a few tests) that going with CUDA for NVidia produces faster results than going with OpenCL. Although, the OpenCL port is not as optimized yet, this is to be expected because OpenCL is one layer more abstract than CUDA (compilation goes through CUDA as far as I know).

Best wishes

CUDA is faster (than OpenCL) when using VRay RT in GPU mode. CUDA is Nvidia's dedicated tech, so why wouldn't it be more efficient than an Open Standard.

Mon Sep 29, 2014 5:49 pm

giannis wrote:I can say already (after a few tests) that going with CUDA for NVidia produces faster results than going with OpenCL. Although, the OpenCL port is not as optimized yet, this is to be expected because OpenCL is one layer more abstract than CUDA (compilation goes through CUDA as far as I know).

Very understandable. Plus they probably prefer optimising CUDA than OpenCL

What about technical possibilities, would it be easier / more difficult to develop i.e. Displacement in OpenCL..?

Tue Sep 30, 2014 8:22 am

PixelMango wrote:
giannis wrote:I can say already (after a few tests) that going with CUDA for NVidia produces faster results than going with OpenCL. Although, the OpenCL port is not as optimized yet, this is to be expected because OpenCL is one layer more abstract than CUDA (compilation goes through CUDA as far as I know).

Best wishes
CUDA is faster (than OpenCL) when using VRay RT in GPU mode. CUDA is Nvidia's dedicated tech, so why wouldn't it be more efficient than an Open Standard.

Exactly!

Tue Sep 30, 2014 8:24 am

eph wrote:
giannis wrote:I can say already (after a few tests) that going with CUDA for NVidia produces faster results than going with OpenCL. Although, the OpenCL port is not as optimized yet, this is to be expected because OpenCL is one layer more abstract than CUDA (compilation goes through CUDA as far as I know).
Very understandable. Plus they probably prefer optimising CUDA than OpenCL

What about technical possibilities, would it be easier / more difficult to develop i.e. Displacement in OpenCL..?

In Presto framework, it poses the same level of difficulty from the pure development side but if you add more things to the mix (like optimization, debugging tools and others), again a dedicated tech will make the whole job easier.

Best wishes