Compute Code For AMD Cayman Lands In Mesa

Quote:

---

Originally Posted by smitty3268

You thought wrong.

In particular, one of the hardware requirements for OpenCL is a particular kind of cache - so that multiple shaders can share the same data - which wasn't implemented in Radeon hardware until the Evergreen generation.

---

Are you guys maybe thinking of the OpenCL 1.1 spec, which would be right for Evergreen? As I remember openCL 1.0 spec has been supported by AMD since 2008, supplanting "Close To Metal" for GPGPU.

IIRC DirectX10/OpenGL3 class hardware has OpenCL1.0 support, while DirectX11/OpenGL4 class hardware has OpenCL1.1 support.

Now though, an HD2600XT would be pretty slow for compute and does not support double percision as only the top end cards of each generation get that ability, even if the mid end cards of the next gen are just as, if not more powerful.

I'd just pickup an HD6670 or faster, the HD6670 would use less power, be much faster and supports OpenCL.

Quote:

---

Originally Posted by Kivada

Are you guys maybe thinking of the OpenCL 1.1 spec, which would be right for Evergreen? As I remember openCL 1.0 spec has been supported by AMD since 2008, supplanting "Close To Metal" for GPGPU.

IIRC DirectX10/OpenGL3 class hardware has OpenCL1.0 support, while DirectX11/OpenGL4 class hardware has OpenCL1.1 support.

Now though, an HD2600XT would be pretty slow for compute and does not support double percision as only the top end cards of each generation get that ability, even if the mid end cards of the next gen are just as, if not more powerful.

I'd just pickup an HD6670 or faster, the HD6670 would use less power, be much faster and supports OpenCL.

---

R6xx hw doesn't support all the necessary bits for a full OpenCL 1,0 implementation. It could support a fair amount, but not everything. R7xx hw barely made the cut. The issue is the standards weren't finalized until after the hw was designed. It's a similar situation to with DX9 cards. OpenGL 2.x was finalized after DX9 came out so no-one actually made a card that was 100% GL 2.x compliant. The larger problem with OpenCL on 6xx/7xx cards is that they don't have a dedicated compute pipeline that matches the CL pipeline. As such, you have to translate OpenCL to something more like graphics in order to run it. Someone could implement limited CL support for 6xx/7xx cards and it would probably be useful for certain tasks.