Skip to content

Latest commit

 

History

History
76 lines (67 loc) · 3.18 KB

GTC2018-step0.md

File metadata and controls

76 lines (67 loc) · 3.18 KB
Raw

The initial step of every paralelization effort is to check whether it makes sense to apply optimizations or not. In the case of accelerator programming a certain number of questions have to be answered before the first line of code is written. This includes:

  • Understanding the program structure and how data is passed through the call tree
  • Profiling the CPU-only version of the application and identifying computationally-intense "hot spots"
  • Identify which loop nests dominate the runtime
  • Are the loop nests suitable for an accelerator?
  • Insuring that the algorithms you are considering for acceleration are safely parallel

Firt. Compile & execute the sequential version of FWI:

$ mkdir build
$ cd build
$ cmake -DCMAKE_C_COMPILER=pgcc ..
$ make

Then you execute FWI either using make irun or bin/fwi fwi_schedule.txt. You should see something like this:

Number of frequencies 1
Number of shots 1
Number of gradient iterations 1
Number of test iterations 1
Output directory path: results
FWI Program finished in 1059.629833 seconds

Then we are going to profile FWI to search for hot spots using make iprofile which will call nvprof --cpu-profiling on (or call nvprof directly from console: nvprof --cpu-profiling on bin/fwi fwi_schedule.txt).

$ cmake -DCMAKE_C_COMPILER=pgcc ..
$ make
$ nvprof --cpu-profiling on --cpu-profiling-percentage-threshold 1 bin/fwi fwi_schedule.txt
Number of frequencies 1
Number of shots 1
Number of gradient iterations 1
Number of test iterations 1
Output directory path: results
FWI Program finished in 1059.575686 seconds

======== CPU profiling result (bottom up):
Time(%)      Time  Name
 19.99%  211.766s  IDX
 16.53%  175.115s  compute_component_scell_TR
 16.28%  172.505s  compute_component_scell_BL
 15.63%  165.634s  compute_component_scell_BR
  9.47%  100.343s  compute_component_scell_TL
  5.87%  62.2217s  compute_component_vcell_TR
  5.83%  61.7416s  compute_component_vcell_BR
  5.17%  54.7615s  compute_component_vcell_BL
  5.15%  54.5714s  compute_component_vcell_TL

======== Data collected at 100Hz frequency
======== Percentage threshold: 1%

We can see that scell and vcell functions dominate the execution time. IDX is the function that linearizes the (i,j,k) triplet into the linear index. Usually the compiler is smart enough to inline it, but in this execution it didn't. Since we know that IDX is only called inside scell and vcell functions, we can safely split the IDX execution time among scell and vcell functions. Therefore we can safely say that scell and vcell accounts for the 99% of the execution time of the application.

If we take a look at those functions in src/fwi_propagator.c we will arrive to this conclusions:

  1. They are embarrassingly parallel
  2. All TR/TR/BL/BR are very similar
  3. We have to apply the same parallelization strategy for all scell and vcell functions.

Now we will recompile enabling OpenMP and execute the application to measure the performance of OpenMP vs the serial implementation:

Number of frequencies 1
Number of shots 1
Number of gradient iterations 1
Number of test iterations 1
Output directory path: results
FWI Program finished in 120.587904 seconds