No performance improvement in PIM as compared to cpu

[jasmiine28k]

./build/X86/gem5.opt --debug-flags=PIM configs/example/se.py --cpu-type=TimingSimpleCPU --cpu-clock=2GHz --caches --l2cache --l1d_size=32kB --l1i_size=32kB --l2_size=2MB --l1d_assoc=4 --l1i_assoc=4 --l2_assoc=8 --enable-pim --pim-type=cpu --kernel-type=adder --num-pim-kernels=16 --mem-size=512MB --coherence-granularity=64B -c ./tests/test-progs/example_pim/test_cpu
gives no. of ticks=4356369500

./build/X86/gem5.opt --debug-flags=PIM configs/example/se.py --cpu-type=TimingSimpleCPU --cpu-clock=2GHz --caches --l2cache --l1d_size=32kB --l1i_size=32kB --l2_size=2MB --l1d_assoc=4 --l1i_assoc=4 --l2_assoc=8 --enable-pim --pim-type=kernel --kernel-type=adder --num-pim-kernels=16 --mem-size=512MB --coherence-granularity=64B -c ./tests/test-progs/example_pim/test_pim
gives no. of ticks=4890508500
Why is there more no. of ticks in case of PIM as compared to CPU?

[dxlet]

But the paper argue that it supports dynamic feedback to decide whether a PIM instruction should be executed in memory by using IPC.If it works,my prior guess may be wrong.

Do you find the corresponding dynamic feedback source code? I can not find it in source code, it may not be implemented.

[Sayan3116]

./build/X86/gem5.opt –debug-flags=PIM configs/example/se.py –cpu-type=TimingSimpleCPU –cpu-clock=2GHz –caches –l2cache –l1d_size=32kB –l1i_size=32kB –l2_size=2MB –l1d_assoc=4 –l1i_assoc=4 –l2_assoc=8 –enable-pim –pim-type=hybrid –kernel-type=adder –num-pim-kernels=16 –mem-size=512MB –coherence granularity=64B -c tests/test-progs/pim-hello/hellopim

I am trying this code, but it gives me an error that I'm trying to implement Hybrid without first configuring it. Can anyone tell me how to configure it?

[m1chaelyeung]

Guess that it's because of the cache hit rate. The memory-side accesses DRAM everytime although the access process is shorter.If the data locality of the program is high,CPU may get more performance improvement.So it may according to your program.But the paper argue that it supports dynamic feedback to decide whether a PIM instruction should be executed in memory by using IPC.If it works,my prior guess may be wrong.
By the way, should "pim-type=cpu" match "num-pim-processors"?

I want to know why you say that pimkernels access DRAM with shorter time, I think pimkernels are also put outside the memory and they access DRAM through membus, am i right?

[ChenfengZhao]

./tests/test-progs/example_pim/test_cpu

May I ask how do you use PIM CPUs and what is the source code of your "example_pim/test_cpu" program?