This file will introduce the step to get the roofline date through mperf and plot the roofline graph.

get the data for roofline

get the data for the platform

arm cpu

• build mperf for arm64 cpu

  ./android_build.sh -m arm64-v8a
  

  get peak memory bandwidth of the cpu

  • copy the executable file apps/cpu_spec_dram_bw to your android phone and run it, you will get the GBPs for the cpu

  get peak compute bandwidth of the cpu

  • copy the executable file apps/cpu_inst_gflops_latency to your android phone and run it, you will get the GFLOPs for the cpu

mobile gpu

• build mperf for the adreno or mali

  ./android_build.sh -m arm64-v8a -g [adreno | mali]
  

  get peak memory bandwidth of the gpu

  • copy the executable file apps/gpu_spec_dram_bw to your android phone and run it, you will get the GBPs for the gpu

  get peak compute bandwidth of the gpu

  • copy the executable file apps/gpu_inst_gflops_latency to your android phone and run it, you will get the GFLOPs for the gpu

get the data for your function call

arm cpu

• enable the following metrics for mperf tma in cpu_tma_transpose

  mpf_tma.init(
          {"Metric_GFLOPs_Use", "Metric_DRAM_BW_Use", "Metric_L3_BW_Use", "Metric_L2_BW_Use"});
  

  recompile, copy to your phone, and run it, you will get the memory and compute bandwidth for your function call. And in this case, the function is an neon transpose implement.

mobile gpu

mali gpu

• recompile mperf with mali gpu enabled, and copy the executable file apps/gpu_mali_pmu_test to your android phone and run it, you will get the GBPs and GFLOPs for your custom gpu kernel call on mali gpu.

adreno gpu

• recompile mperf with adreno gpu enabled, and copy the executable file apps/gpu_adreno_pmu_test to your android phone and run it, you will get the GBPs and GFLOPs for your custom gpu kernel call on adreno gpu.

plot roofline

• now, you get all the data need to plot roofline, please edit these values into a roofline_data.txt file like roofline_data.txt, and you may need edit the params about axes coordinate range at plot_roofline.py according to values of the peak memory and compute bandwidth on your platform, and finally, plot it

  python3 plot_roofline.py ./roofline_data.txt
  

  see the roofline_readme for more details.