oneMKL黑客松比赛参赛经验分享

为了完成比赛，需要下载到Intel官网下载oneMKL包。

实现方法如下：首先做出随机数

void Random(float *input,int N)
{
	VSLStreamStatePtr stream;
   	vslNewStream(&stream,VSL_BRNG_MT19937,1);
    	vsRngUniform(VSL_RNG_METHOD_UNIFORM_STD,stream,N*N,input,0.0f,1.0f);
	vslDeleteStream(&stream);
}

利用fftw3做fft

void FFTW3(float *input,fftwf_complex *output,int N)
{
	int time;
	fftw_plan flag;

	clock_t start,end;
	flag=fftwf_plan_dft_r2c_2d(N,N,input,output,FFTW_MEASURE);

	start=clock();
	for(time=0;time<1;time++) fftwf_execute(flag);
	end=clock();
	printf("when time=%-10d  , FFTW3  use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<10;time++) fftwf_execute(flag);
	end=clock();
	printf("when time=%-10d  , FFTW3  use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<100;time++) fftwf_execute(flag);
	end=clock();
	printf("when time=%-10d  , FFTW3  use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<1000;time++) fftwf_execute(flag);
	end=clock();
	printf("when time=%-10d  , FFTW3  use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

}

利用oneMKL做fftw3

void MKL(float *input,MKL_Complex8 *output,int N)
{
	int time;
	MKL_LONG sizeN[2]={N,N};
	MKL_LONG rs[3]={0,N,1};
    	MKL_LONG cs[3]={0,N/2+1,1};
	DFTI_DESCRIPTOR_HANDLE handle=NULL;

	DftiCreateDescriptor(&handle,DFTI_SINGLE,DFTI_REAL,2,sizeN);
	DftiSetValue(handle,DFTI_PLACEMENT,DFTI_NOT_INPLACE);
    	DftiSetValue(handle,DFTI_CONJUGATE_EVEN_STORAGE,DFTI_COMPLEX_COMPLEX);
    	DftiSetValue(handle,DFTI_INPUT_STRIDES,rs);
    	DftiSetValue(handle,DFTI_OUTPUT_STRIDES,cs);
	DftiCommitDescriptor(handle);
	
	clock_t start,end;
	start=clock();
	for(time=0;time<1;time++) DftiComputeForward(handle,input,output);
	end=clock();
	printf("when time=%-10d  , oneMKL use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<10;time++) DftiComputeForward(handle,input,output);
	end=clock();
	printf("when time=%-10d  , oneMKL use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<100;time++) DftiComputeForward(handle,input,output);
	end=clock();
	printf("when time=%-10d  , oneMKL use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

	start=clock();
	for(time=0;time<1000;time++) DftiComputeForward(handle,input,output);
	end=clock();
	printf("when time=%-10d  , oneMKL use %f\n",time,(double)(end-start)/CLOCKS_PER_SEC);

}

比较性能和残差

void Compare(fftwf_complex *data1,MKL_Complex8 *data2,int N)
{
	float Residual=0.00001;
	float Rr,Ri;
	int match=0;
	for(int i=0;i<(N/2+1)*N;i++)
    	{
        	Rr=fabs(data2[i].real-data1[i][0]);
        	Ri=fabs(data2[i].imag-data1[i][1]);
        	if(Rr>Residual || Ri>Residual)
		{
			match=1;
		}
   	}
	if(match==1)
	{
		printf("Result Error\n");
	}
	else
	{
		printf("Result Correct\n");
	}
}

感谢因特尔提供的比赛机会，让我认识到了做FFT的新工具，借助oneMKL强大性能，使得对矩阵运算有了更好的性能。