###July 2014 Ewan Crawford[ewan.cr@gmail.com] License: GPL

==================================================================

Tool for visualising the memory accesses of OpenCL programs. Extended to predict the cache performance of the OpenCL programs on FERMI GPUs by rearranging the memory accesses to be platform independent. Cache prediction is still under development however as the results can not yet be validated against a complete benchmark suite.

OpenCL programs are written using a wrapper created by Alberto Magni[alberto.magni86@gmail.com]. This wrapper performs instrumentation passes on the OpenCL kernels using the LLVM optimizer tool opt. These passes allow instructions to be inserted that record information about global memory accesses. Axtor is then used as a OpenCL backend to convert the instrumented LLVM IR back into an OpenCL kernel that can be executed. The wrapper allows this is process to be hidden from the user.

After execution has finished a trace of memory accesses is available dependent executed by the users hardware. In order to generalise this to a machine independent representation of GPU OpenCL execution these memory accesses are then reordered using a scheduler. Additionally this trace can be put through a NVIDIA FERMI cache simulation to predict it's performance.

However by default the original trace is plotted as a graph without any cache prediction or scheduling. This is done using R[http://www.r-project.org/]. In the graph the colour of a point represents the thread which made the accesses while the y-axis value represents the element index, which assumes that all data elements are 4 Bytes in size. However the coalescing done by GPU's is not taken into account and all plots are of individual accesses.

This project is based on my undergraduate project at Edinburgh University. The dissertation for which can be found in /support/dissertation.pdf

Example graphs from polybench benchmarks that are rearranged using the scheduler.

====================================================================

LLVM 3.3[http://llvm.org/] Axtor[bitbucket.org/gnarf/axtor/, Simon Moll]

This tool is only available for Linux, and additionally has only been tested on Ubuntu and Mint distributions.

====================================================================

$> cp SOURCE_DIR/passes/* LLVM_SOURCE_DIR/lib/Transforms/ $> cd LLVM_SOURCE_DIR/lib/Transforms $> append 'add_subdirectory(MemTrace), add_subdirectory(MemSize)' to CMakeLists.txt $> rebuild LLVM $> export VIS_PASSES as path to shared object files for passes

$> mkdir build $> cd build $> cmake [SOURCE_DIR] $> make $> export VIS_TOOLS

=====================================================================

visualise.py --Links components together so the tool can be run easily

benchmarks/ --Benchmarks implemented using wrapper

    matrix/   -- Matrix based benchmarks from NVIDIA
       mm/
       mt/
       mv/

    parboil/   -- contains selected benchmarks from parboil benchmark suite
                  [http://impact.crhc.illinois.edu/Parboil/parboil.aspx]
       stencil/
       bfs/

    polybench/ --contains complete polybench benchmark suite
                 [http://web.cse.ohio-state.edu/~pouchet/software/polybench/]

passes/ MemSize/ --LLVM pass that calculates the length of the trace so that memory can be allocated

MemTrace/       --LLVM pass that adds instructions to record memory accesses

support/

scripts/        -- Bash scripts called by OpenCL wrapper for executing LLVM passes
                   as well as R graphing

wrapper/        --Modified version of Alberto Magni's OpenCL wrapper

tools/

scheduler/      --Rearranges machine dependent memory accesses so they represent a GPU.

cacheSimulator/ --simulates cache performance of memory accesses

examples/ --Examples of graphs that can be produced using the tool.