Skip to content

Latest commit

 

History

History
114 lines (81 loc) · 5.12 KB

readme.md

File metadata and controls

114 lines (81 loc) · 5.12 KB
Raw

Utilities

This directory contains helper files for timing (polybench-*, time_benchmark.sh, template-for-new-benchmark.c). These are obtained from PolyBench/C.

This directory also contains instructions and utilities (rose.sh, dgemvT.C, rose_dgemvT.C) for working with ROSE compiler.

About ROSE Compiler

We use ROSE compiler to automatically transform and parallelize benchmarks, for comparison. Performing any of these steps requires building the compiler from source. Build instructions are in the ROSE wiki.

After that we apply the LoopProcessor (source) tool to transform loops, and AutoPar (source) tool to parallelize them.

Note that loop and parallelization transformations are applied to the entire file1. This breaks the structure expected by the timing utilities and prevents benchmarking the transformed file, unless the original template code around benchmark kernel is restored. As the last step we extract the kernel transformed by ROSE, and substitute it back into the original benchmark template.

The rest of this guide details the ROSE compiler operations. It includes examples to help verify the expected behavior in each step.

Loop Transformation

Test program: dgemvT.C

Expected result: rose_dgemvT.C.save

Compilation

/path_to/rose/build/tutorial/loopProcessor --edg:no_warnings -w -c -fs0 -cp 0 /path_to/dgemvT.C

Expected output

array-copy dimension is 0
opt level=0

Automatic parallelization

Build the auto-parallelization tool AutoPar. Then it can be used to automatically parallelize programs, e.g.:

../autoPar --edg:no_warnings -I../headers -c /path_to/program.c

There is no expected output for this command.

Transform all original benchmarks

The following command performs these steps--transformation, parallelization and template restore--on all original benchmarks automatically, with the best effort. If it reports errors, you will need to check the output manually.

./utilities/rose.sh

Supported arguments: -r path_to/rose and -p benchmark_name.

Expected output

3mm          transformed: ✓   parallel: ✓   restored: ✕
bicg         transformed: ✓   parallel: ✓   restored: ✓
colormap     transformed: ✓   parallel: ✓   restored: ✓
conjgrad     transformed: ✓   parallel: ✓   restored: ✕
cp50         transformed: ✓   parallel: ✓   restored: ✓
deriche      transformed: ✓   parallel: ✓   restored: ✕
fdtd-2d      transformed: ✓   parallel: ✓   restored: ✕
gemm         transformed: ✓   parallel: ✓   restored: ✕
gesummv      transformed: ✓   parallel: ✓   restored: ✓
mvt          transformed: ✓   parallel: ✓   restored: ✓
remap        transformed: ✕   parallel: ✕   restored: ✓
tblshft      transformed: ✓   parallel: ✓   restored: ✓

The benchmarks with indicated error in last column are caused by removed scop/endscop pragma. This is not an error in transforming the file, but prevents restoring the templating code. It can be resolved by manually by inspecting and replacing the templating code.

Example remap fails to transform with following error. This is significant error and prevents measuring this example in the alternate case. The error message is:

array-copy dimension is 0
Error in SymbolicValGenerator::GetSymbolicVal(): unhandled type of binary operator BOP_MOD
lt-loopProcessor: /home/rose/src/src/midend/astUtil/symbolicVal/SymbolicVal.C:344: static SymbolicVal 
  SymbolicValGenerator::GetSymbolicVal(AstInterface&, const AstNodePtr&): Assertion `false' failed.
Aborted (core dumped)

There is no known fix at this time. The example is measured as equal to original, due to this transformation failure.

Footnotes

  1. Rose User Manual, 15.2.2 Partial Compilation: "At present the transformation does not properly support partial compilation. This is not a fundamental challenge to the current design, it merely reflects the incomplete state of the project. See also section 15.4." ↩