I am not sure it is correct or not, or caused by some other reasons.
I filtered some real signals by FFT. and plus them. (0.9 support real FFT)
I did the same thing with the intel mkl FFT.
It seems that the result from cudaFFT contains some low-frequency artifacts.
I am not sure why, I guess that the cudaFFT C2R part does not consider the “Hermitian” redundancy, so the minus frequency part is ignored.
Any comment, please.
CUFFT returns only the non-redundant coefficients.
Can you be more specific about the transform you are doing? Is it 1D, 2D? What is the transform size? What is the dynamic range of your input signal?
What I did is filtering many 1D 2048 signals. I list a sample code
cufftHandle cudaFFT_Plan_R2C_Once = 0; // for filter kernel
cufftHandle cudaFFT_Plan_R2C = 0; // for forward transform
cufftHandle cudaFFT_Plan_C2R = 0; // for backward transform
int filterWidth = 2048;
int NumOfTransform = 480;
cufftComplex * filter; // the filter kernel on the device
cufftComplex * data; // data on the device
int width = 640 // the signal length
float * realSignal //real signal on host (= new float[width*NumOfTransform])
//prepare the device memory
cudaMalloc((void **) &filter, sizeof(cufftComplex)*(filterWidth/2 + 1) );
cudaMalloc((void **) &data, sizeof(cufftComplex)*(filterWidth/2 + 1)*NumOfTransform);
cudaMemset(data, 0, sizeof(cufftComplex)*(filterWidth/2 + 1)*NumOfTransform);
//prepare the filter kernel
... ...
//Transform the filter kernel
cufftPlan1d(&cudaFFT_Plan_R2C_Once, filterWidth, CUFFT_R2C, 1);
cufftExecR2C(cudaFFT_Plan_R2C_Once, (cufftReal *)filter, filter);
//normalize the filter FFT coefs here
... ...
//prepare the cuFFT Handle
cufftPlan1d(&cudaFFT_Plan_R2C, filterWidth, CUFFT_R2C, NumOfTransform);
cufftPlan1d(&cudaFFT_Plan_C2R, filterWidth, CUFFT_C2R, NumOfTransform);
//prepare the real signal
cudaMemcpy2D( ((float *) data) + (filterWidth / 2) - (width / 2), sizeof(cufftComplex)*(filterWidth /2 + 1), realSignal, width*sizeof(float), width*sizeof(float), NumOfTransform, cudaMemcpyHostToDevice);
//forward transform
cufftExecR2C(cudaFFT_Plan_R2C, (cufftReal *)data, data);
//filtration by complex multify
Filtration_Complex_Point_Wise_Mul_Cuda_Kernel<<< __dim_Grid__, __dimBlock__ >>>(data, filter, (filterWidth /2 + 1), NumOfTransform);
//backward transform
cufftExecC2R(cudaFFT_Plan_C2R, data, (cufftReal *)data);
What I did were just some simple filtration of real signal by a real kernel.
My question is about
cufftExecC2R
I wonder that the minus frequency part is ignored by this command.
Thank you.