[Bf-cycles] CUDA performance tests

[Brecht Van Lommel]

Thanks, that is a bit unexpected. I don't have CUDA set up on Windows
at the moment, only on Mac and Linux so can't give you a binary today
since they are platform specific.

I will try to figure out why those numbers are different before and
after __launch_bounds__ and/or do some testing on Windows soon.

On Mon, Apr 28, 2014 at 9:37 PM, Rolf Jawarsch <rolf_jawarsch at web.de> wrote:
> Here my ptxas output:
>
> pre __launch_bounds__:
>
> ptxas : info : Compiling entry function 'kernel_cuda_path_trace' for 'sm_50'
> ptxas : info : Function properties for kernel_cuda_path_trace
>     9216 bytes stack frame, 3660 bytes spill stores, 3540 bytes spill loads
>
> ptxas : info : Used 32 registers, 764 bytes cmem[0], 392 bytes cmem[2], 100
> textures
>
> post __launch_bounds__:
>
> ptxas : info : Compiling entry function 'kernel_cuda_path_trace' for 'sm_50'
> ptxas : info : Function properties for kernel_cuda_path_trace
>     9184 bytes stack frame, 2904 bytes spill stores, 3684 bytes spill loads
> ptxas : info : Used 32 registers, 764 bytes cmem[0], 428 bytes cmem[2], 100
> textures
>
> I change nothing except the parameters i post in the spreadsheet.
> Can i test the kernel you compiled?
>
> In addition, I noticed that utilization of two or more 750ti´s is not good,
> but the modified_bmw.blend has to short rendertime to reproduce his.
> But that has nothing to do with __launch_bounds__.
>
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> http://lists.blender.org/mailman/listinfo/bf-cycles
>