Hi,
any news on the next version of AMD's OpenCL SDK? Will there be a beta version in the near future which let's developers test new features, e.g. image support on GPUs? I am mostly interested in evaluating OpenCL on AMD GPUs for medical image processing and visualization tasks. However, currently no image support is a big problem for many algorithms. Could you please if possible comment on upcoming releases and features?
Thx
Kutter,
The new SDK is going to be released soon. You can expect image support to be present in the new SDK.
Originally posted by: omkaranathan Kutter,
The new SDK is going to be released soon.
Can I ask you if CrossFire and 5970 are going to be supported too ?
I hope the new FirePros like the V7800 could be supported!
Also VS 2010, pls!
thx
Originally posted by: davibu
Can I ask you if CrossFire and 5970 are going to be supported too ?
The 5970 is already supported, at least accoring to http://developer.amd.com/gpu/ATIStreamSDK/Pages/default.aspx
And mine is working fine with the current SDK.
Why would CrossFire be useful for OpenCL? Could it be used to copy data between GPUs directly rather than via host memory?
Originally posted by: dominik_g Originally posted by: davibu
Can I ask you if CrossFire and 5970 are going to be supported too ?
The 5970 is already supported, at least accoring to http://developer.amd.com/gpu/ATIStreamSDK/Pages/default.aspx
And mine is working fine with the current SDK.
Why would CrossFire be useful for OpenCL? Could it be used to copy data between GPUs directly rather than via host memory?
There have been many post in this forum about the second GPU of 5970 returning incorrect values. Are you using both GPUs at the same time on your ?
It is the same for 2 cards connected in CrossFire (in the case of a 5870+5850, the 5850 is incorrectly identified as another 5870 too).
Originally posted by: davibu
There have been many post in this forum about the second GPU of 5970 returning incorrect values. Are you using both GPUs at the same time on your ?
I've got a program that uses the CPUs as well as the four GPUs in 2 5970s (not using CrossFire). I'm using a thread per device and I synchronize the threads to make sure they're executing simultaneously (I checked that using clGetEventProfilingInfo).
All 5 devices multiply two different random matrices and I compared the output to a sequentially computed result. In all cases the result was correct.
It is the same for 2 cards connected in CrossFire (in the case of a 5870+5850, the 5850 is incorrectly identified as another 5870 too).
I see. But is there any advantage of using CrossFire in OpenCL or is it just for compatibility?
Originally posted by: dominik_g
I see. But is there any advantage of using CrossFire in OpenCL or is it just for compatibility?
Mostly compatibility and a simple way to not have to attach a monitor to each card. May be the CorssFire connection could be used also for optimizing some GPU-to-GPU memory transfer but only ATI people know if that is even possible.
Originally posted by: davibu Originally posted by: dominik_g
I see. But is there any advantage of using CrossFire in OpenCL or is it just for compatibility?
Mostly compatibility and a simple way to not have to attach a monitor to each card. May be the CorssFire connection could be used also for optimizing some GPU-to-GPU memory transfer but only ATI people know if that is even possible.
I'm running Linux and I currently have no monitor connected to my computer at all. I had to run "aticonfig --adapter=all --initial" and then set the environment variable DISPLAY to ":0". After that all cards were identified correctly.
Can we possibly expect the "byte addressable store" extension for GPUs in the next SDK ?? or is that something that the hardware cannot provide ?
For applications like a 256 bin Histogram, it would be very useful,
Thank You
Perhaad
Originally posted by: perhaad Can we possibly expect the "byte addressable store" extension for GPUs in the next SDK ?? or is that something that the hardware cannot provide ?
For applications like a 256 bin Histogram, it would be very useful,
In coming release, byte addressable store is supported for Evergreen(8
XX) cards.
Looking at IL docs it isn't possible to address memory on byte boundary, so all these accesses will be simply replaced by
AND dword ptr [aligned by DWORD mem],mask
OR dword ptr [aligned by DWORD mem],value
sequence, thus doubling memory bandwidth. So implementing 'byte addressable store' won't help much to increase performance, isn't it?
Estimate time... ZERO ! v2.1 is rdy!