Hi,
I've recently noticed that not all GPU memory can be used from OpenCL on my 5870, regardless of it they are allocated in one or multiple chunks.
From browsing some older threads from 2009 I saw a comment that this is likely due to a limitation from the underying CAL which restricts us to 256MB sized chunks. And that all memory cannot be used, even if multiple buffer objects have been allocated.
Could someone in the know confirm that this is still supposed to hold in the latest drivers? What are the plans for future releases, does this feature have a priority to be fixed (ie. is it worth waiting a few months?) or is it belived to take longer before all memory can be used?
Are there any tricks of circumventing this? (i've managed to allocate slighty more memory - but not much - by using multiple buffer objects)
Many thanks for any information on this=)
Yep it's a pain not to be able to allocate all the VRAM you need...
See this:
http://forums.amd.com/devforum/messageview.cfm?catid=390&threadid=129209&enterthread=y
Thanks, this was very usefull and will hopefully save me some swapping back and forth between CPU and GPU memory as part of the computations.
I use the SDK 2.1 with Windows 7 x64 and Cat 10.5 with a Radeon 5750(512Mb) and I wasn't able to allocate two 128Mb buffers.
I placed the environment variable called GPU_MAX_HEAP_SIZE to 512 and now I can so...
About the idea to limit the buffers to 128Mb is a terrible bad idea for the motives I exposed in that other post ( PS 256Mb + 3dsmax 192Mb, both will fail, unlimiting the memory=both will run with 512Mb ).
Pls, ATI, we need that 128Mb limit removed.
Raistmer,
We are working to get this limit removed.