Get start

[Shayan_Y_]

Hi

I am newbie with Linux(I have Ubuntu 14.04) and TBB. I want to write codes with TBB library and run on computer. I have no idea how should I start .

Can anyone help me with a step by step tutorial or something that help me?

[Vladimir_P_1234567890]

Tutorial

https://www.threadingbuildingblocks.org/intel-tbb-tutorial

--Vladimir

[Shayan_Y_]

Thanks.

My first problem is how should I be sure I have TBB on my system ? I have installed parallel studio xe 2016 and in it's directory and in my terminal run this command "source iccvars.sh intel64".

My first example is 

#include <tbb/tbb.h>
#include <iostream>
#include <omp.h>

using namespace std;
using namespace tbb;
class Average {
	public:
		float* input;
		    float* output;
	void operator( )( const blocked_range<int>& range ) const {
		for( int i=range.begin(); i!=range.end( ); ++i )
      		{
		        output = (input[i-1]+input+input[i+1])/3.0f;
		}
    	}
};
void ParallelAverage( float* output, float* input, size_t n ) {
	Average avg;
	avg.input = input;
	avg.output = output;
	output[0]=(input[0]+input[1])/2.0f;
	parallel_for( blocked_range<int>( 1, n-1, 100 ), avg );
	output[n-1]=(input[n-1]+input[n-2])/2.0f;
}
void SerialAverage(float* output,float* input,int  n){
	output[0] = (input[0]+input[1])/2.0f;
	for(long int i=1; i<n-1; ++i )
      	{
		output = (input[i-1]+input+input[i+1])/3.0f;
      	}
  	output[n-1] = (input[n-1]+input[n-2])/2.0f;
}
int main()
{
	task_scheduler_init init;
	srand ( time(NULL) );
	unsigned long long int lenght=1000000;
	float input[lenght];
	float output[lenght],temp[lenght];
	double serial_timer = 0.;
	double compute_timer = 0.;
	for(size_t i=0;i<lenght;i++)
		input=rand()%100;


	compute_timer -= omp_get_wtime();
	ParallelAverage(output,input,lenght);
	compute_timer += omp_get_wtime();
	cout << "Calculation time in parallel = " << compute_timer<<endl;
	compute_timer = 0.;
	
	serial_timer -= omp_get_wtime();
	SerialAverage(temp,input,lenght);
	serial_timer += omp_get_wtime();
	cout<<"Calculation time in serial="<< serial_timer<<endl;

	return 0;
}

But calculating average of  next and previous of each item  is much faster than in serial loop than parallel one.

[Shayan_Y_]

Thanks.

My first problem is how should I be sure I have TBB on my system ? I have installed parallel studio xe 2016 and in it's directory and in my terminal run this command "source iccvars.sh intel64".

My first example is 

#include <tbb/tbb.h>
#include <iostream>
#include <omp.h>

using namespace std;
using namespace tbb;
class Average {
	public:
		float* input;
		    float* output;
	void operator( )( const blocked_range<int>& range ) const {
		for( int i=range.begin(); i!=range.end( ); ++i )
      		{
		        output = (input[i-1]+input+input[i+1])/3.0f;
		}
    	}
};
void ParallelAverage( float* output, float* input, size_t n ) {
	Average avg;
	avg.input = input;
	avg.output = output;
	output[0]=(input[0]+input[1])/2.0f;
	parallel_for( blocked_range<int>( 1, n-1, 100 ), avg );
	output[n-1]=(input[n-1]+input[n-2])/2.0f;
}
void SerialAverage(float* output,float* input,int  n){
	output[0] = (input[0]+input[1])/2.0f;
	for(long int i=1; i<n-1; ++i )
      	{
		output = (input[i-1]+input+input[i+1])/3.0f;
      	}
  	output[n-1] = (input[n-1]+input[n-2])/2.0f;
}
int main()
{
	task_scheduler_init init;
	srand ( time(NULL) );
	unsigned long long int lenght=1000000;
	float input[lenght];
	float output[lenght],temp[lenght];
	double serial_timer = 0.;
	double compute_timer = 0.;
	for(size_t i=0;i<lenght;i++)
		input=rand()%100;


	compute_timer -= omp_get_wtime();
	ParallelAverage(output,input,lenght);
	compute_timer += omp_get_wtime();
	cout << "Calculation time in parallel = " << compute_timer<<endl;
	compute_timer = 0.;
	
	serial_timer -= omp_get_wtime();
	SerialAverage(temp,input,lenght);
	serial_timer += omp_get_wtime();
	cout<<"Calculation time in serial="<< serial_timer<<endl;

	return 0;
}

But calculating average of  next and previous of each item  is much faster than in serial loop than parallel one.

[RafSchietekat]

(Please remove the second copy of your posting. Well, edit it down to "(Removed second copy)" or something.)

TBB has to get out of the starting blocks first (even task-based toolkits have to start up some worker threads), so your timings will be useless in such a small benchmark. Try a few ParallelAverage executions, and only then start the timer. Tip: decrease variance by averaging out several executions.

What's the difference in timing, if any, if you "hoist" end() out of the loop: "for( int i=range.begin(), i_end=range.end(); i!=i_end; ++i )"? This might help the compiler unroll or even auto-vectorise the loop, which can make a huge difference. Alternatively, try accepting the blocked_range by value: maybe the compiler can figure it all out by itself, then. Please give times for the original, both variants, and perhaps the combination: supposedly you should have one "slow" execution" and 2 or 3 that are close to each other and a lot faster.

Also, you should probably use a far larger grainsize for such a simple Body, especially if you have a lot of hardware parallelism on your machine (which makes the auto_partitioner pick a fairly small average effective grainsize). Please also provide timings for what you find here.

Then there's telling the compiler to vectorise, using "#pragma simd" (instead of merely hoping that it might). Same request about timings...