Archives Discussions

[zkhan]

I have downloaded and installed the latest SDK on my Windows 7 machine with an ATI Radeon HD 4350, which reportedly has a max. work group size of 128.

However, when I run the FFT sample from the SDK, I get:

Error : Out of resources!

Work-group size required by kernel : 64

Maximum size supported on GPU for this kernel : 32

I assume I should be able to execute a simple FFT with this hardware and am doing something wrong. Any help is appreciated.

[zkhan]

If I reduce the local and global thread count to 32 from 64, it does execute. Does my Radeon HD 4350 have a smaller maximum work group size than what a utility (GPU Caps Viewer) is reporting?

[MicahVillmow]

zkhan,
Because of lack of support for a hardware barrier on the 7XX series of cards, any kernel that uses a 'barrier' function runs with a max workgroup size equal to the wavefront size. The 7XX devices come with 3 different flavors of wavefront sizes, 16(710), 32(730) and 64(rest) threads per wavefront. The 4350 uses the 730 chip, so has a wavefront size equal to 32.

Also, the max workgroup size that the device supports and the max workgroup size that the kernel supports can be different based on the constraints of the kernel.

[zkhan]

Thanks a lot for that response, Micah. I'm still in the learning stages, so I'm understanding more and more of what you've said as I move along