rodinia-nocopy performance much worse than copy version

[chen boya]

Hi, Jason

I'm using the gem5-gpu to study heterogeneous computing.

I build the gem5-gpu environment and run the benchmark rodinia and rodinia-nocopy,

for the backprop, my result shows that copy version can finish in 0.024s while nocopy version need 0.24s, 10x worse. 

the option I use is -o 16384, other options are same to the https://gem5-gpu.cs.wisc.edu/wiki/benchmarks

Gussian and hotspot also showed worse performance in nocopy version.

In my opinion, nocopy version is something like the new technology "unified memory", why this is much worse than using explicit copy? This doesn't meet my expectation.

Did you got similar results?

Best Regards

Boya

[chen boya]

from my observation, the no copy version backprop has 10x instructions of copy version

no-copy version:

build/X86_VI_hammer_GPU/gem5.opt ../gem5-gpu/configs/se_fusion.py -c /work/rodinia/bin/nocopy/gem5_fusion_backprop --access-host-pagetable -o 16384 --cpu-type DerivO3CPU --caches --l2cache

sim_seconds                                  0.246098

sim_insts                                   635075609

copy version:

build/X86_VI_hammer_GPU/gem5.opt ../gem5-gpu/configs/se_fusion.py -c /work/rodinia/bin/copy/gem5_fusion_backprop -o 16384 --cpu-type DerivO3CPU --caches --l2cache

sim_seconds                                  0.024412

sim_insts                                    63392771

The instructions are 635M vs 63M, why does this happen?

[Jason Lowe-Power]

Hi Boya,

My quick guess is that the copy version is crashing. That said, all of this code is quite old and hasn't been touched for 4-5 years. I would suggest using the amdgpu that has been upstreamed into gem5 and has current support: https://www.gem5.org/2020/05/27/modern-gpu-applications.html

Jason

--
You received this message because you are subscribed to the Google Groups "gem5-gpu Developers List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to gem5-gpu-dev...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/gem5-gpu-dev/1c2b917a-62ad-4dea-8536-04af47549ae8n%40googlegroups.com.

[chen boya]

Hi, Jason

Thank you very much. 

I will take a look on the new solution.

Best Regards

Boya