APOLLO - [BETA]

Writing of this tutorial is currently in progress. If you wish to follow this tutorial, please consider yourself a beta tester (and don’t hesitate to send us any improvement suggestions).

Tutorial prerequisites:

  • This tutorial has been developed and tested for Ubuntu distributions
Last update the 09.05.2019 - Tutorial created the 09.05.2019 by R. Lazcano

Project setup

In addition to the default requirements (see Requirements for Running Tutorial Generated Code), please download the following files:

APOLLO setup

To be able to apply these optimizations, first APOLLO must be installed.

APOLLO installation

  1. Download APOLLO using git: git clone https://scm.gforge.inria.fr/anonscm/git/apollo/apollo.git
  2. Go into apollo folder: cd apollo
  3. Select the latest branch: git checkout apolloPreesm
  4. Create a new directory for building APOLLO: mkdir build
  5. Read and install all the required dependencies detailed in INSTALL:
    • autoconf
    • bison
    • build-essential
    • cmake
    • flex
    • git
    • libboost-dev
    • libboost-system-dev
    • libgmp-dev
    • liblapack-dev
    • libtool
    • ninja-build
    • texi2html
    • texinfo
  6. Run: cmake .. -DAPOLLO_BUILD_JOBS=4 (replace 4 with number of available cores)
  7. Run: sudo make install -j4 (replace 4 with number of available cores)
    Note: APOLLO installation can take up to 2 hours, so setting APOLLO_BUILD_JOBS to more than 1 is recommended.
  8. Run: sudo cp ../apollo/runtime/include/apolloAPI.h /usr/local/include

How to use APOLLO with a Preesm application

As described in APOLLO website, APOLLO (Automatic speculative POLyhedral Loop Optimizer) is a compiler framework dedicated to automatic, dynamic and speculative parallelization and optimization of programs’ loop nests. APOLLO can be used to further accelerate a dataflow application by combining it with Preesm. To do so, several modifications must be done within Preesm app:

Set APOLLO as the compiler

As stated before, APOLLO is a compiler, so it has to be used to compile the application. To do so, please use the CMakeLists.txt provided in Matrix-Multiplication Preesm Project or add the following to your own CMakeLists.txt:

  1. At the beginning of CMakeLists.txt, add the following line:
    set(USEAPOLLO 1)
    

    If set to 1, this variable enables APOLLO compilation. Consequently, if you do not wish to use APOLLO, you should either comment that line or set it to 0.

  2. Just after the creation of the project, add the following lines:
project(Main)

if(${USEAPOLLO})
	find_library(apollo_LIBRARY LLVMApollo)
	find_library(apolloRuntime_LIBRARY apolloRuntime)
	if(apollo_LIBRARY AND apolloRuntime_LIBRARY)
		MESSAGE("")
		MESSAGE("WARNING: Using APOLLO compiler")
		MESSAGE("If you do not wish to use APOLLO, please deactivate USEAPOLLO in CMakeLists.txt")
		MESSAGE("")
		SET(CMAKE_C_COMPILER "/usr/local/bin/apolloc")
		add_definitions(-DAPOLLO_AVAILABLE)
	else()
		MESSAGE("WARNING: Couldn't find APOLLO libraries, so APOLLO compilation is disabled")
		set(USEAPOLLO 0)
	endif()
	
endif()

Configure the workflow of the project

  1. Open CodegenPiMM.workflow
  2. Select the ‘Code Generation’ Task
  3. Select the ‘Properties’ tab
  4. Select the ‘Basic’ tab in ‘Task Variables’
  5. Click on ‘Add…’ and write “Apollo”
  6. In the new ‘Apollo’ option of the Task Variables, write ‘true’

The resulting workflow tab should be similar to the one displayed on the following image:

Add APOLLO pragma

To use APOLLO, a specific pragma must be added surrounding the target loop nest:

#ifdef APOLLO_AVAILABLE
#pragma apollo dcop
{
#endif
	for (i = 0; i < rowsA; i++) {
		for (j = 0; j < colsB; j++) {
			mult[i*colsB + j] = 0.0;
			for (k = 0; k < colsA; k++) {
				mult[i*colsB + j] += A[i*colsA+k] * B[k*colsB+j];   
			}
		}
	}
#ifdef APOLLO_AVAILABLE
}
#endif

Generate and run the application

  1. Run the workflow selecting as Scenario the one called 1core.scenario
  2. Run the CMakeGCC.sh file –> sh CMakeGCC.sh
  3. Go to make/bin –> cd bin/make
  4. Run make
  5. Run the application ./Release/main

Please, note that you will notice that the application runs faster after 10 iterations. This happens because the first 10 iterations are considered as a training stage and, after that, the best optimization among them is chosen and applied.

If any error appears, please, create an issue on GitHub.

Updated: