Progetto Sistemi Digitali M

-- Scroll down for English version --

Il progetto consiste nell'implementazione SIMD e CUDA dell’algoritmo Fixed Window per il calcolo della Disparity Map da un’immagine Stereo.

Struttura del progetto

• dataset: contiene i file ppm dei dataset.
• include: cartella con i file di include.
• nc-profiling: file di profiling di Nsight Compute.
• src: cartella con i file di implementazione delle librerie sviluppate.
  • cuda_main.cu: implementazioni dell'algoritmo FW CUDA.
  • stereo_fw.cpp: implementazioni dell'algoritmo FW SISD e SIMD.
  • utils.cpp: funzioni di utilità per trattare i file ppm e convertire array di tipi diversi.
• CMakeLists.txt: file CMake.
• main.cpp: file main C++.
• README.md: file README attuale con struttura del progetto e istruzioni per l'esecuzione.

Istruzioni

• Compilazione ed esecuzione con CMake:
  • creare la cartella "build" e navigarci all'interno: mkdir build && cd build
  • eseguire cmake: cmake ..
  • compilazione: make
  • esecuzione: ./main 0 16 5

Per cambiare le impostazioni di ottimizzazione (autovettorizzazione) è necessario modificare il file "CMakeLists.txt".

• Get CPU info (Ubuntu): lscpu, in the flags section are listed the available instruction sets (like AVX2).

• Compilazione senza CMake (NON PIÙ SUPPORTATO):

  • Compile: g++ hello.cpp -o hello
  • Compile with optimization: g++ hello.cpp -Ox -o hello where x in [0,3]. with -mavx2 it is possible to specify the AVX2 family
  • Compile with maximum optimization that targets the native architecture: g++ hello.cpp -O3 -march=native -o hello
  • Compile with OpenMP: g++ -fopenmp hello.cpp -o hello
  • Execution: ./hello
  • CUDA Compile: nvcc -Wno-deprecated-gpu-targets hello.cu -o hello
  • CUDA Execution: ./hello

Risultati

I risultati sono presentati in: resources/Project_Presentation_IT.pdf

---

Accelerated Processing Systems Project

The project consists of the SIMD and CUDA implementation of the Fixed Window algorithm for computing the Disparity Map from a Stereo image.

Project Structure

• dataset: contains the dataset ppm files.
• include: folder with include files.
• nc-profiling: Nsight Compute profiling files.
• src: folder with implementation files of the developed libraries.
  • cuda_main.cu: CUDA FW algorithm implementations.
  • stereo_fw.cpp: SISD and SIMD FW algorithm implementations.
  • utils.cpp: utility functions for handling ppm files and converting arrays of different types.
• CMakeLists.txt: CMake file.
• main.cpp: C++ main file.
• README.md: current README file with project structure and execution instructions.

Instructions

Compilation and execution with CMake:

• create the "build" folder and navigate into it: mkdir build && cd build
• run cmake: cmake ..
• compile: make
• run: ./main 0 16 5

To change optimization settings (autovectorization), edit the "CMakeLists.txt" file.

Get CPU info (Ubuntu): lscpu, the available instruction sets (like AVX2) are listed in the flags section.

Compilation without CMake (NO LONGER SUPPORTED):

• Compile: g++ hello.cpp -o hello
• Compile with optimization: g++ hello.cpp -Ox -o hello where x ∈ [0,3]. With -mavx2 it is possible to specify the AVX2 family.
• Compile with maximum optimization targeting the native architecture: g++ hello.cpp -O3 -march=native -o hello
• Compile with OpenMP: g++ -fopenmp hello.cpp -o hello
• Execution: ./hello
• CUDA Compile: nvcc -Wno-deprecated-gpu-targets hello.cu -o hello
• CUDA Execution: ./hello

Results

Results can be found in: resources/Project_Presentation_IT.pdf