Some questions on new OpenCL GPU support..

Some questions on new OpenCL GPU support..

Have same questions about new GPU support..
first seems with jointly CPU+GPU context graphics interop via all apis is disabled (gl and d3d10 interop) this isn't the case with amd ocl sdk either in fusion or cpu+gpu discrete case.. why that limitation? I'm assuming you are allowing creation of buffers and textures in same special memory region where zero copy usage of this buffers is possible either in CPU or GPU.. but seems these are is not good for direct usage in GPU by graphics APIs is that the case?.. more info please.. at least would be good to add support of graphics apis interop in cpu+gpu context but says in that case says that buffers or texture acquired via clcreatefromGLbuffer aren't zero copy if used by CPU kernels, no?second questions in graphics interop table you say that d3d10 interop is truly zero copy if using RESOURCE_MISC_SHARED but seems in OpenGL is no way of assuring that we have zero copy usage.. why that? could you clarify how to check if gl buffer or texture is efficiently mapped to gpu kernels?more info please.thanks..

2 posts / 0 new
Last post
For more complete information about compiler optimizations, see our Optimization Notice.

Interop

Right; GL/DX
interoperability on all it's flavors isn't supported at CPU+GPU context (aka
Shared Context)

This is a known issue; being evaludated for introduction at
future versions;

We are
very interested in getting an idea about what kind of usage you are seeking
here with interop at Shared Context, can u please elaborate a little bit what
usage is it targetting?

Resource Sharing

Right,
resources as buffers/textures are phsyically shared by the context devices,
CPU+GPU, in a Shared Context; Please note that u need to gauarantee alignments
requirement according to the optimization guide for assuring physical sharing
with the user hostptr (in case of CL_MEM_USE_HOST_PTR) in case its being
provided.

This is
doesn't necessarily mean that textures targetting GPU are managed less
efficient comparing to single GPU context; Please notice that in case you have
used hostptr at the texture creation which is aligned to 4K (see Intel OpenCL
Optimization Guide) at Shared Context, then you have increased performance at
clEnqueueMap/Unmap for the same texture.

Leave a Comment

Please sign in to add a comment. Not a member? Join today