How to set parameters to simulate GPU480

[정광현]

Hello,

Currently, I am studying hardware simulation. And I am learning about Macsim to see if I can use macsim in my research.

My target architecture to simulate using macsim is NVIDIA GTX480, so I have to control some configuration files to specify GTX480.

Below are the settings I tried.

• kernel_info (one of steps to generate GPU trace)
  
• There are two kernels in convolutionSeparable that I've used as a benchmark. To confirm register usage and shared memory usage of each kernel, I used nvcc by adding option '-cubin -gencode arch=compute_20,code=compute_20 -arch=sm_20 -Xptxas=-v'.
• Each parameter that nvcc shows is
• _Z21convolutionRowsKernelPfS_iii(kernel_name)   29(register_usage)   5184(shared_memory_usage)
  
• _Z24convolutionColumnsKernelPfS_iii 19 2569
  

• params.in (configuration file to specify architecture)
• I modified it based on GTX465 params which is included in git as a sample.
• The followings are the modified parts
• num_sim_cores 15
• num_sim_small_cores 15
• clock_gpu 1.4
• clock_mc 1.8
  

But, since I am not familiar with macsim, I fount out that a simulation result about cycles is different with real HW cycles measured by NVPROF when I use this configurations.

So, I have some questions.

1. Are there any problems the settings I've used?
2. If you have configulation files to simulate GTX480, could I get it? It doesn't have to be perfect.
3. Could you give me advises when measuring cycles using macsim? I want to compare with NVPROF. Any advises would be great help to me.

I think Macsim is the most well designed GPU simulator and I really want to use it for my research.

I'm looking forward to hearing from you.

Thanks in advance.

Best regard,
Kwang-hyun Chung