[Starpu-devel] APU optimization: is partition camping enough of a difference?

[Chris Hennick <christopherhe@trentu.ca>]

I'm running an A4-6300 (dual-core Richland) at 3811MHz CPU / 968MHz GPU, with 4 GB DDR3 and an R7 240 video card (1 GB GDDR5). For my masters thesis, I'm trying to find an application for which StarPU can be improved by being more aware of the differences between discrete and integrated GPUs.

However, when I bought the hardware, I was unaware that AMD hadn't fully implemented HSA on Richland (meaning there's no CPU<->GPU cache snooping and no shared memory address space). Since StarPU can already account for the differences in memory-transfer bandwidth and latency, both between GPUs and to the CPU, the only differences that leaves unadjusted-for are the APU's heat conduction to and from the CPU and its immunity to partition camping.

Can StarPU actually impact whether or not tradeoffs are made to avoid partition camping while running a given OpenCL codelet, or must that be done within the codelet? How can I find out how many memory controllers my R7 240 card has, and therefore how susceptible it is to partition camping anyway? And how much difference would it make if I could upgrade to an A10-7850K to get full HSA (as well as the extra cores to ensure that STARPU_OPENCL_ON_CPUS would be a realistic option)?