Team Failure
Team Members
- John Iannandrea, TBB Heat diffusion
- Colin Campbell, OpenMP Heat diffusion
- Mateya Lucic, Cilk Plus Heat diffusion
Assignment
Our assignment was to implement OMP, TBB, and Cilk Plus versions of a 2d diffusion algorithm.
Serial
This is the serial version of the code we have parallelized
class SerialDiffuser : public IDiffuser {
protected:
void evolveTimestep(){
for (int row = 1; row < N - 1; row++) {
for (int col = 1; col < N - 1; col++) {
float uxx = (ui[(row + 1) * N + col] - (2 * ui[row * N + col]) + ui[(row - 1) * N + col]) / delta;
float uyy = (ui[row * N + (col + 1)] - (2 * ui[row * N + col]) + ui[row * N + (col - 1)]) / delta;
u[row * N + col] = ui[row * N + col] + deltaT * diff * (uxx + uyy);
}
}
}
public:
SerialDiffuser(int _N, int _T) : IDiffuser(_N, _T) {}
void init(){
for (int row = 0; row < N; row++) {
for (int col = 0; col < N; col++) {
if ((pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) <= 0.1)
& (pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) >= 0.05))
ui[row * N + col] = 1.0;
}
}
}
void compute(){
for (int m = 1; m < timeSteps; m++) {
evolveTimestep();
std::copy(u, u + N * N, ui);
}
}
};
Omp
class OMPDiffuser : public IDiffuser {
protected:
void evolveTimestep(){
#pragma omp parallel for
for (int row = 1; row < N - 1; row++) {
for (int col = 1; col < N - 1; col++) {
float uxx = (ui[(row + 1) * N + col] - (2 * ui[row * N + col]) + ui[(row - 1) * N + col]) / delta;
float uyy = (ui[row * N + (col + 1)] - (2 * ui[row * N + col]) + ui[row * N + (col - 1)]) / delta;
u[row * N + col] = ui[row * N + col] + deltaT * diff * (uxx + uyy);
}
}
}
public:
OMPDiffuser(int _N, int _T) : IDiffuser(_N, _T) {}
void init(){
#pragma omp parallel for
for (int row = 0; row < N; row++) {
for (int col = 0; col < N; col++) {
if ((pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) <= 0.1)
& (pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) >= 0.05))
ui[row * N + col] = 1.0;
}
}
}
void compute(){
for (int m = 1; m < timeSteps; m++) {
evolveTimestep();
std::copy(u, u + N * N, ui);
}
}
};
Cilk
class CilkDiffuser : public IDiffuser {
protected:
void evolveTimestep(){
cilk_for(int row = 1; row < N - 1; row++) {
for (int col = 1; col < N - 1; col++) {
float uxx = (ui[(row + 1) * N + col] - (2 * ui[row * N + col]) + ui[(row - 1) * N + col]) / delta;
float uyy = (ui[row * N + (col + 1)] - (2 * ui[row * N + col]) + ui[row * N + (col - 1)]) / delta;
u[row * N + col] = ui[row * N + col] + deltaT * diff * (uxx + uyy);
}
}
}
public:
CilkDiffuser(int _N, int _T) : IDiffuser(_N, _T) {}
void init(){
cilk_for(int row = 0; row < N; row++) {
for (int col = 0; col < N; col++) {
if ((pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) <= 0.1)
& (pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) >= 0.05))
ui[row * N + col] = 1.0;
}
}
}
void compute(){
cilk_for(int m = 1; m < timeSteps; m++) {
evolveTimestep();
u[0:N*N] = ui[0:N*N];
}
}
};
TBB
class TBBEvolve {
float* u;
float* ui;
float delta, deltaT;
const float diff = 0.5;
int N;
public:
TBBEvolve(float* _u, float* _ui, float d, float dt, float n) : u(_u), ui(_ui), delta(d), deltaT(dt), N(n) {}
void operator()(tbb::blocked_range2d<int> r) const{
for (int row = r.rows().begin(); row < r.rows().end(); row++) {
#pragma simd
for (int col = r.cols().begin(); col < r.cols().end(); col++) {
float uxx = (ui[(row + 1) * N + col] - (2 * ui[row * N + col]) + ui[(row - 1) * N + col]) / delta;
float uyy = (ui[row * N + (col + 1)] - (2 * ui[row * N + col]) + ui[row * N + (col - 1)]) / delta;
u[row * N + col] = ui[row * N + col] + deltaT * diff * (uxx + uyy);
}
}
}
};
class TBBDiffuser : public IDiffuser {
protected:
void evolveTimestep(){
}
public:
TBBDiffuser(int _N, int _T) : IDiffuser(_N, _T) {}
void init(){
for (int row = 0; row < N; row++) {
for (int col = 0; col < N; col++) {
if ((pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) <= 0.1)
& (pow(row * dir - 0.5, 2) + pow(col * dir - 0.5, 2) >= 0.05))
ui[row * N + col] = 1.0;
}
}
}
void compute(){
for (int m = 1; m < timeSteps; m++) {
tbb::blocked_range2d<int> r(1, N - 1, 1, N - 1);
tbb::parallel_for(r, TBBEvolve(u, ui, delta, deltaT, N));
}
}
};
Results
What we found was that all the parallelization methods were all very similar. We also tested this with cuda and found cuda to be the fastest.
