Page 1 of 1

OpenCL 2.0 specs released

Posted: Wed Jul 24, 2013 3:27 pm
by Dade
Just in case someone missed the news: http://www.anandtech.com/show/7161/khro ... nnounced/3

Re: OpenCL 2.0 specs released

Posted: Sun Jul 28, 2013 9:48 am
by dbz
Shared virtual memory looks interesting. I didn't read anything on C++ support in kernels however. That is a feature I would have expeced from OpenCL 2.0.

Re: OpenCL 2.0 specs released

Posted: Sun Jul 28, 2013 1:12 pm
by Dade
dbz wrote:I didn't read anything on C++ support in kernels however. That is a feature I would have expeced from OpenCL 2.0.


It is somewhat implicit in the new SPIR 1.2 specs: you can generate SPIR code from pretty much any language (static C++, C++ AMP, etc.). SPIR code will than run on any OpenCL platform. OpenCL seems to move toward being the standard "assembler" of GPU computing. OpenCL C kernels are probably going to become something of the past, replaced by kernels written in your preferred language an than compiled in SPIR.

This is something that lately has already been shown with source-to-source compilers (for instance by Intel with their C++ AMP-to-OpenCL C translator).

I'm really hope to move away from C kernels too.

Re: OpenCL 2.0 specs released

Posted: Sun Jul 28, 2013 4:21 pm
by graphicsMan
Does "shared virtual memory" mean what I think it does? This has some pretty important implications for hardware features... How will discrete boards support this?

Re: OpenCL 2.0 specs released

Posted: Mon Jul 29, 2013 7:06 am
by Dade
graphicsMan wrote:Does "shared virtual memory" mean what I think it does? This has some pretty important implications for hardware features... How will discrete boards support this?


The shared memory concept seems heavily bound to AMD APUs, Intel new CPU/GPU, etc. (i.e. to moving the GPU on the CPU side of the PCIe bus). However some vendor may take the opposite direction and move the CPU on the GPU side of the PCIe bus (like NVIDIA is planning to do with ARM core).

I hope that someone will develop a NUMA solution for discrete GPUs otherwise their future doesn’t look good.