Skip to content

Latest commit

 

History

History
82 lines (64 loc) · 5.27 KB

README.md

File metadata and controls

82 lines (64 loc) · 5.27 KB
Raw

DiscoPoP graph analyzer

DiscoPoP profiler is accompanied by a Python framework, specifically designed to analyze the profiler output files, generate a CU graph, detect potential parallel patterns, and suggest OpenMP parallelizations. Currently, the following five patterns can be detected:

  • Reduction
  • Do-All
  • Pipeline
  • Geometric Decomposition
  • Task Parallelism

Getting started

We assume that you have already run the DiscoPoP profiler on the target sequential application, and the following files are created in the current working directory:

  • Data.xml (CU information in XML format created by CUGeneration pass)
  • <app_name>_dep.txt (Data dependences created by DPInstrumentation pass)
  • reduction.txt and loop_counter_output.txt (Reduction operations and loop iteration data identified by DPReduction pass)

In case any of the files mentioned above are missing, please follow the DiscoPoP manual to generate them.

In addition to the already mentioned files, a file named <app_name>_CUInstResult.txt is required for the task parallelism detection. In order to generate it, the following sequence of commands can be used:

python3 -m discopop_explorer --path=<path> --cu-xml=<cuxml> --dep-file=<depfile> --loop-counter=<loopcount> --reduction=<reduction> --generate-data-cu-inst=<outputdir>
clang++ -S -emit-llvm -c -std=c++11 -g <DISCOPOP_PATH>/CUInstantiation/RT/CUInstantiation_iFunctions.cpp -o iFunctions_CUInst.ll
clang++ -g -O0 -emit-llvm -fno-discard-value-names -c <C_File> -o tmp_target_app.ll
<CLANG_BIN_DIR>/opt -S -load=<PATH_TO_DISCOPOP_BUILD_DIR>/libi/LLVMCUInstantiation.so -CUInstantiation -input=Data_CUInst.txt tmp_target_app.ll -fm-path=FileMapping.txt -o tmp_target_app_instrumented.ll
clang++ tmp_target_app_instrumented.ll iFunctions_CUInst.ll -o <app_name>_cui -L$PATH_TO_DISCOPOP_BUILD_DIR/rtlib -lDiscoPoP_RT -lpthread -o <app_name>_cui
rm tmp_target_app.ll tmp_target_app_instrumented.ll iFunctions_CUInst.ll
./<app_name>_cui

Pre-requisites

To use the graph analyzer tool, you need to have Python 3.6+ installed on your system. Further Python dependencies can be installed using the following command: pip install -r requirements.txt

Usage

To run the graph analyzer, you can use the following command:

python3 -m discopop_explorer --path <path-to-your-output>

You can specify the path to DiscoPoP output files. Then, the Python script searches within this path to find the required files. Nevertheless, if you are interested in passing a specific location to each file, here is the detailed usage:

`discopop_explorer [--path <path>] [--cu-xml <cuxml>] [--dep-file <depfile>] [--plugins <plugs>] [--loop-counter <loopcount>] [--reduction <reduction>] [--json <json>] [--fmap <fmap>] [--cu-inst-res <cuinstres>] [--llvm-cxxfilt-path <cxxfp>] [--generate-data-cu-inst <outputdir>]`

Options:

    --path=<path>               Directory with input data [default: ./]
    --cu-xml=<cuxml>            CU node xml file [default: Data.xml].
    --dep-file=<depfile>        Dependencies text file [default: dep.txt].
    --loop-counter=<loopcount>  Loop counter data [default: loop_counter_output.txt].
    --reduction=<reduction>     Reduction variables file [default: reduction.txt].
    --cu-inst-res=<cuinstres>   CU instantiation result file. Task Pattern Detector is executed if this option is set.
    --llvm-cxxfilt-path=<cxxfp> Path to llvm-cxxfilt executable. Required for Task Pattern Detector
                                if non-standard path should be used.
    --plugins=<plugs>           Plugins to execute
    --fmap=<fmap>               File mapping [default: FileMapping.txt]
    --json                      Output result as a json file to specified path
    --generate-data-cu-inst=<outputdir>     Generates Data_CUInst.txt file and stores it in the given directory.
                                            Stops the regular execution of the discopop_explorer.
                                            Requires --cu-xml, --dep-file, --loop-counter, --reduction.
    -h --help                   Show this screen.
    --version                   Show version.

By default, running the graph analyzer will print out the list of patterns along with OpenMP parallelization suggestions to the standard output. You can also obtain the results in JSON format by passing --json argument to the Python script.

Walkthrough example

The test/ folder contains a number of precomputed inputs for testing the tool, e.g., atax from Polybench benchmark suite. You can try out this example workflow.

test/reduction/ contains source code and precomputed DiscoPoP output for a simple reduction loop. The loop itself sums up all numbers from 1 to n.

You can run DiscoPoP on main.c or just use included output.

After that, you can run discopop_explorer. The --path argument should point to the output of the DiscoPoP.

In this example, the output for reduction will point to the lines 6-9, and it will suggest pragma omp parallel for OpenMP directive for parallelizing the loop. You will also find i classified as a private variable and sum as a reduction variable. Thus, the parallelization directive would be suggested as follows:

#pragma omp parallel for private(i) reduction(+:sum)

The suggested pattern is demonstrated in mainp.c