Francesco Malferrari (193103), Gianluca Siligardi (185068) and Andrea Somenzi (232254)
This is the exam project for UniMORE High Performance Computing course. The goal is to minimize the execution time of the DynProg solver of the Polybench C library through:
- Profiling and bottleneck research
- Code rewriting
- Parallelization with
- OpenMP
- CUDA
/dynprog
: original version of the file assigned/fast_dynprog
: the fastest version of the written dynprog. It has been uploaded for completeness, however it has an error due to the precision of floating point type./OpenMP
/Cuda
This folder contains the sources of the first assignment:
- the rewritten DynProg
rew_dynprog.c
in/rew_dynprog
- the OpenMP parallelization of the rewritten DynProg
parallel_dynprog.c
in/parallel_dynprog
- the OpenMP parallelization over Nvidia Jetson Nano GPU of the rewritten DynProg
gpu_dynprog.c
in/gpu_dynprog
/OpenMP
folder contains also the presentation of the results of the first assignment.
This folder contains the sources of the second assignment:
Makefile
: the Makefile for .cu filescuda_dynprog.cu
: the CUDA optimization of the rewritten version ofdynprog.c
/Cuda
folder contains also the presentation of the results of the second assignment.
Into each directory run this command:
make EXT_CFLAGS="-DPOLYBENCH_TIME -DPOLYBENCH_DUMP_ARRAYS" clean all run
You can also specify the dataset size adding the flag DATASET and writing one of the listed in the header file
/dynprog/dynprog.h
, for example:
make EXT_CFLAGS="-DLARGE_DATASET -DPOLYBENCH_TIME -DPOLYBENCH_DUMP_ARRAYS" clean all run
Alternatively, if you want to use datasets not present in the header file, you can replace the flag DATASET whith the following flags setted with numbers of your interest:
-DLENGTH=10000 -DTSTEPS=20
It is also possible to specify the number of threads with the following flag:
"-DNTHREADS=4"
but in our case having a four threads machine, it was defined in the code.
Before running these commands, to compile gpu_dynprog.c
, you must run
module load clang/11.0.0 cuda/10.0
to load the cuda library.