Skip to content

tob2/OvO

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

OvO: OpenMP vs Offload

  ___
 (OvO)
<  .  >
--"-"---

Build Status

OvO is a collection of OpenMP Offloading test functions for C++ and Fortran.

Get me started

$ git clone [email protected]:TApplencourt/OvO.git ; cd OvO
# Set your env and run 
$ OMP_TARGET_OFFLOAD=mandatory CXX="g++" CXXFLAGS="-fopenmp" FC="gfortran" FFLAGS="-fopenmp"./ovo.sh run
$ ./ovo.sh report
>> Overall result test_result/1957-04-01_19-02_CDC6600.lanl.gov
  pass rate(%)    test(#)    success(#)    compilation error(#)    runtime error(#)    wrong value(#)    hang(#)
--------------  ---------  ------------  ----------------------  ------------------  ----------------  ---------
           68%        310           212                       0                   0                96          2

More

OvO is a collection of OpenMP offloading tests for C++ and FORTRAN. OvO is focused on testing extensively hierarchical parallelism and mathematical functions.

You can find the slides of some OvO presentations we did in the documentation folder.

$ ./ovo.sh -h
OvO an OpenMP test generator.
Usage:
  ovo.sh gen
  ovo.sh gen tiers [1|2|3]
  ovo.sh gen hierarchical_parallelism [--test_type [atomic|reduction|memcopy|ordered]...]
                                      [--data_type [float|'complex<float>'|
                                                    double|'complex<double>'|
                                                    REAL|COMPLEX|
                                                    'DOUBLE PRECISION'|'DOUBLE COMPLEX']... ]
                                      [--loop_pragma [True|False] ]
                                      [--paired_pragmas [True|False] ]
                                      [--multi_devices [True|False] ]
                                      [--host_threaded [True|False] ]
                                      [--intermediate_result [True|False] ]
                                      [--no_user_defined_reduction [True|False] ]
                                      [--collapse [N]...]
                                      [--append]

  ovo.sh gen mathematical_function [--standard [cpp11|cpp17|cpp20|gnu|f77]... ]
                                   [--complex [True|False] ]
                                   [--long [True|False] ]
                                   [--simdize [N]...]
                                   [--append]

  ovo.sh run [<test_folder>...]
  ovo.sh report [ --summary | --failed | --passed ] [--tablefmt (github|tsv|jira)]  ] [<result_folder>...]
  ovo.sh clean
  ovo.sh (-h | --help)

Bug reports and PRs are more than welcome! The OpenMP specification can be tricky. And we use metaprogramming to generate the tests so, to pastiche Jamie Zawinski:

Some people, when confronted with a problem, think "I know, I'll use regular expressions metaprogramming." Now they have two problems.

Requirement

  • python3
  • OpenMP compiler (obviously). We recommend an OpenMP 5.0 spec-compliant compiler. Some test map and reduce a variable in the same combined construct
  • C++11 compiler
  • jinja (optional, only needed if you want to generate tests that are not in the repo)
conda install --file requirements.txt

or

pip install requirements.txt
  • make >4.0 Make 4.0 introduced the --output-sync option to serialize make output. If you use a version of make older than 4.0, we will compile the tests serially.

Related works

"Gentlemen you had my curiosity ... but now you have my attention"

Hierarchical parallelism tests

We generate 3 types of kernels: Kernels which perform a reduction using atomic, kernels which perform a reduction using the OpenMP reduction pragma, and memory copy kernels. For each type of kernel, we generate all possible OpenMP loop-nests containing any combination of target, teams, distribute, parallel for, including combined pragma. Single precision and double precision complex datatype are used. More datatypes are available if needed.

The code below is an example of a code OvO can generate. It's a kernel using 'atomic' to perform a float reduction. Note the absence of for in the parallel section.

float counter_N0 {}
#pragma omp target map(tofrom: counter_N0)
#pragma omp teams distribute
for (int i0 = 0 ; i0 < N0 ; i0++ )
{
  #pragma omp parallel
  {
     #pragma omp atomic update
     counter_N0 = counter_N0 + 1./omp_get_num_theads();
   }
}
assert (counter_N0 != N0);

The real code can be found here

Mathematical tests

We test if all functions of a specified standard are able to be offloaded. The Offloaded result should match the CPU result with 4 ulp preference.

Running

To run OvO simply type ./ovo.sh run. Log files will be saved in the newly created test_result folder. OvO will respect any usual environment variables provided by the user (e.g. CXX / CXXFLAGS / FC / FFLAGS / OMP_TARGET_OFFLOAD). OvO will also respect the special OVO_TIMEOUT environment variable which controls the timeout used to kill too-long running tests (by default 15s).

You can find commonly used flags for various compilers in /documentation/README.md. PR are welcomed, for new versions of compilers.

Below is a simple run using the GCC compiler:

$ OMP_TARGET_OFFLOAD=mandatory CXX="g++" CXXFLAGS="-fopenmp" FC="gfortran" FFLAGS="-fopenmp"./ovo.sh run
Running tests_src/cpp/mathematical_function/math_cpp11 | Saving log in results/1957-04-01_19-02_CDC6600.lanl.gov/cpp/mathematical_function/math_cpp11
g++ -fopenmp isgreater_bool_float_float.cpp -o isgreater_bool_float_float.exe
[...]

Result

A summary of the result can be obtained with ./ovo.sh report. Example of output obtained with --summary:

./ovo.sh report --summary  --tablefmt github
>> Overall result for test_result/1957-04-01_19-02_CDC6600.lanl.gov
|   pass rate(%) |   test(#) |   success(#) |   compilation error(#) |   runtime error(#) |   wrong value(#) |   hang(#) |
|----------------|-----------|--------------|------------------------|--------------------|------------------|-----------|
|            57% |       828 |          471 |                    198 |                 41 |               98 |        20 |

 >> Summary
| language   | category                 | name                     |   pass rate(%) |   test(#) |   success(#) |   compilation error(#) |   runtime error(#) |   wrong value(#) |   hang(#) |
|------------|--------------------------|--------------------------|----------------|-----------|--------------|------------------------|--------------------|------------------|-----------|
| cpp        | hierarchical_parallelism | reduction-float          |            34% |        74 |           25 |                      2 |                  1 |               44 |         2 |
| cpp        | hierarchical_parallelism | reduction-complex_double |            47% |        74 |           35 |                      2 |                  1 |               28 |         8 |
| cpp        | hierarchical_parallelism | atomic-float             |            58% |        33 |           19 |                      0 |                  0 |                4 |        10 |
| cpp        | hierarchical_parallelism | memcopy-complex_double   |            93% |        45 |           42 |                      2 |                  1 |                0 |         0 |
| cpp        | hierarchical_parallelism | memcopy-float            |            93% |        45 |           42 |                      2 |                  1 |                0 |         0 |
| cpp        | mathematical_function    | cpp11                    |            92% |       177 |          163 |                      6 |                  4 |                4 |         0 |
| cpp        | mathematical_function    | cpp11-complex            |           100% |        34 |           34 |                      0 |                  0 |                0 |         0 |
| fortran    | hierarchical_parallelism | reduction-double_complex |             7% |        74 |            5 |                     49 |                 14 |                6 |         0 |
| fortran    | hierarchical_parallelism | reduction-real           |             8% |        74 |            6 |                     48 |                 14 |                6 |         0 |
| fortran    | hierarchical_parallelism | memcopy-real             |            22% |        45 |           10 |                     35 |                  0 |                0 |         0 |
| fortran    | hierarchical_parallelism | memcopy-double_complex   |            24% |        45 |           11 |                     34 |                  0 |                0 |         0 |
| fortran    | hierarchical_parallelism | atomic-real              |            39% |        33 |           13 |                     18 |                  0 |                2 |         0 |
| fortran    | mathematical_function    | F77-complex              |            71% |        14 |           10 |                      0 |                  0 |                4 |         0 |
| fortran    | mathematical_function    | F77                      |            92% |        61 |           56 |                      0 |                  5 |                0 |         0 |

You can also use ./ovo.sh report --failed to get a list of tests that failed for more thoughtful investigation.

All information on the execution of the tests is available in the subfolder of test_result corresponding to our run (for example, ./test_result/1957-04-01_19-02_CDC6600.lanl.gov/cpp/hierarchical_parallelism/memcopy-real). The environment used to run the test is saved in env.log. Two log files are also created: one for the compilation (compilation.log), and one for the runtime (runtime.log).

  • Error code 112 corresponds to an incorrect result.
  • Error 124 or 137 corresponds to a test which was hanging and killed by timeout.

List of tests available

More than 100,000 tests are available. For convenience, we bundle them in tiers.

To generate new tests, please use ovo.sh gen. By default, it will generate tiers 1 tests. But if you feel adventurous, you can use: ovo.sh tiers 3. See more section for more information.

Too much information about flags and arguments

For generating more tests

`ovo.sh gen`
`ovo.sh gen tiers <1|2|3>`
    gen              Generate tests corresponding to tiers 1
    gen tiers        Generate tests corresponding to different tiers.
                     Tiers 1 list of tests (`ovo.sh gen` or `ovo.sh gen tiers 1`):
                        hierarchical_parallelism cpp:
                            atomic-float,
                            memcopy-complex_double, memcopy-float,
                            reduction-complex_double, reduction-float
                        hierarchical_parallelism fortran:
                            atomic-real,
                            memcopy-double_complex, memcopy-real,
                            reduction-double_complex, reduction-real
                        mathematical_function cpp
                            cpp11, cpp11-complex
                        mathematical_function fortran
                            F77, F77-complex
                     Tiers 2 (`ovo.sh gen tiers 2`):
                        hierarchical_parallelism cpp:
                            atomic-float, ordered_real, atomic-float-host_threaded, atomic-float-intermediate_result,
                            memcopy-complex_double, memcopy-float, memcopy-float-collapse_n2, memcopy-float-loop_pragma
                            reduction-complex_double, reduction-float, reduction-float-multiple_devices
                        hierarchical_parallelism fortran:
                            atomic-real,  ordered_real, atomic-real-host_threaded, atomic-real-intermediate_result
                            memcopy-double_complex, memcopy-real, memcopy-real-collapse_n2, memcopy-real-loop_pragma
                            memcopy-real-paired_pragmas, reduction-double_complex
                            reduction-real, reduction-real-multiple_devices
                    Tiers 3 (`ovo.sh gen tiers 2`):
                        All possible combination
Options you can pass:
   For hierarchical_parallelism:
    --test_type      Choose the kind of tests you want to generate.
                       - atomic tests will use OpenMP `atomic` construct to perform a reduction.
                       - reduction tests will use OpenMP `reduction` construct to perform a reduction.
                       - memcopy tests perform a memory copy.
    --data_type      Trigger for which data type will be used in the tests. Uppercase type corresponds to Fortran datatype.
    --loop_pragma    Trigger to use OpenMP 5.0 "loop" construct
    --paired_pragmas Fortran Only. Will generate tests that use optional "$OMP END" constructs.
    --multi_devices  Tests will be offloaded to all the GPU available.
    --host_threaded  Tests will be offloaded by multiple host threaded
    --intermediate_result
                     Reduction and Atomic tests will use intermediate results to perform their reduction.
    --collapse
                     All the loops will be duplicate N time, and `omp collapse` will be used.
    --no_user_defined_reduction
                     Only impacts reduction with C++ complex datatype. Tests will not use `omp declare reduction` construct".

   For mathematical_function:
    --standard       Corresponds to which standard (c++11, c++17, etc.) used to generate math functions.
    --complex        Trigger for complex math functions
    --long           Trigger to use C++ long datatype if possible
    --simdize        Trigger to put math function inside a 'simd' region

Examples:

  Generate the `tiers 2` set of tests:
     ./ovo.sh gen tiers 2

  Generate hierarchical_parallelism reduction tests with REAL (fortran) and complex<float>(c++) datatype with and without multi-devices support:
    ./ovo.sh gen hierarchical_parallelism  --test_type reduction --data_type REAL "complex<float>" --multiple_devices True False

For running tests

`ovo.sh run [test_folder]`:
    [test_folder]    List of tests folder. OvO will recurse on those folders to execute tests. 
                     By default all in test_src run, this lets you specify certain folders

Example:

    Run only the Fortran tests
        ./ovo.sh run ./test_src/fortran

For reporting results

`ovo.sh report`
Options you can pass:
    --summary        Print for each group of tests the pass rate
    --failed         Print all the test which failed
    --passed         Print all the test which passed
    --tablefmt       Can be used to change for formating of the table 
                     (useful for copy/pasting in Excel for example)
    
Example:

    Print for each tests group a summary of the pass rate:
        ./ovo.sh report --summary --failed

About

OpenMP vs Offload

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C++ 49.9%
  • Fortran 34.1%
  • Python 12.7%
  • Jinja 2.4%
  • Other 0.9%