Hi,
I have read the post which said that this is a driver issue.
We have a scenario in which we need to do kernel processing on large amounts of data from the host memory.
How can we know if the memory transfer is done in parallel (by DMA) or by the command queue which is not parallel in that matter?
When should this issue be solved in case there is currently no DMA support?
Regards,
Tomer Gal
+1
DMA support is currently not available.
We are working towards it and you can expect it in future releases.
hopefully it will be the next release.
whether enqueueWriteBuffer does not use DMA operation? I think it should even in SDK 2.0, as we talk about transfer host memory to local video memory!
DTOP,
clEnqueueWriteBuffer is an API to write data in a buffer.But it does not use DMA and hence memory transfer and kernel execution are serialized.
With DMA it would be possible to do them parallely.
What clEnqueueWriteBuffer is using then? map/copy/unmap?
@nou...
Do you mean that this feature is already available in the hardware, just not supported by the drivers? Can we expect to see DMA enabled for hd 5870 etc in a future release?
Thank You.
Debdatta Basu.
of course it is supported in HW. but i do not work for AMD so i can't tell.
ONCE ENABLED YOU CAN EXPECT DMA SUPPORT EVEN FOR R600 DEVICES.
http://www.beyond3d.com/content/reviews/53/6
it seems that the new HD6900 cards have asynchronous kernel dispatch and dual bidirectional DMA engines. How good that may sound compared to the < HD6900 cards?