Acronimul Proiectului: pmake3d

• Nume Scurt: pmake3d
• SVN: ​https://svn-batch.grid.pub.ro/svn/PP2009/proiecte/pmake3d

• Membrii echipei: Silviu Baranga - SCPD, Adriana Szekeres - SCPD;
• Descriere proiect: Make3d is a program developed at Stanford. It takes a single 2-D image and generates a 3-D model, completely automatically. This project aims at parallelizing the make3d code, using CUDA. (More details here: ​http://make3d.stanford.edu/)

Activitate proiect

1. Analizarea codului make3d (acesta se afla in folderul make3d_original de pe svn) in vederea paralelizarii:

• pentru rularea codului, se dau urmatoarele comenzi in matlab (cu mici schimbari ale cailor):

  >> cd PP/Make3dSingleImageStanford_version0.1/LearningCode/
  >> InitialPath(true)
  >> profile on -history 
  >> OneShot3dEfficient('/home/szekeres/Desktop/PP/Make3dSingleImageStanford_version0.1/park0067_.jpg','/home/szekeres/Desktop/PP/Make3dSingleImageStanford_version0.1/out','1','/home/szekeres/Desktop/PP/MakedLearnedParameters_v0.1/', '/home/szekeres/Desktop/PP/MakedLearnedParameters_v0.1/', [])
  >> profile viewer
  

2. Incercarea paralelizarii functiei matlab sigmoidLogBarrierSolver cu Jacket ( ​http://www.accelereyes.com/products).

• initiativa esuata datorita faptului ca Jacket nu suporta inca adunarea/scaderea matricelor sparse (​http://www.accelereyes.com/forums/viewtopic.php?f=7&t=1167).

3. Paralelizarea functiei C++ segmentImgOpt folosind CUDA. Functia este de fapt implementarea unui algoritm de segmentare a unei imagini ppm.

• algoritmul implementat in C++, precum si paperul aferent se gasesc la adresa: ​http://people.cs.uchicago.edu/~pff/segment/
• codul paralelizat (segment.cu) se gaseste in folderul segment de pe svn
• pentru compilare si rulare se poate folosi Makefile-ul aflat in folderul segment de pe svn, in care se vor inlocui caile spre headerele CUDA, biblioteca CUDPP, si alte headere si biblioteci necesare.

INCDIR = -I. -I/usr/local/cuda/sdk/C/common/inc -I/home/szekeres/Desktop/cudpp_1.1/cudpp/include 
LIBS = -lcudpp
LIBDIR = -L/home/szekeres/Desktop/cudpp_1.1/lib
DBG    = -g
CPP    = g++
CFLAGS = $(DBG) $(INCDIR)
LINK   = -lm 

.cpp.o:
	$(CPP) $(CFLAGS) -c $< -o $@

all: parallel sequential

parallel:
	nvcc segment.cu -o parallel ${INCDIR} ${LIBDIR} ${LIBS}

sequential: segment.cpp segment-image.h segment-graph.h disjoint-set.h
	$(CPP) $(CFLAGS) -o sequential segment.cpp $(LINK)

clean:
	/bin/rm -f parallel  sequential *.o

clean-all: clean
	/bin/rm -f *~ 

run_parallel:
	./parallel 0.8 500 20 lanscape_vectors.ppm out.ppm

run_sequential:
	./sequential 0.8 500 20 lanscape_vectors.ppm out.ppm

4. Evaluarea implementarii:

   818x535 image:
       original code: 648 ms
       parallelized version: 336 ms
       speedup: 1.92x
   1389x822 image:
       original code: 1441 ms
       parallelized version: 662 ms 
       speedup: 2.17x