I mapped an algorithm which uses R2C and C2R FFTs from fftw3 to clFFT. When comparing the outputs, I encountered a non negligible offset (in the 10e-3 range with MSE) in the backward FFT (hermitian-interleaved -> real, out-of-place). I then compared the intermediate results in the algorithm which all came close to the correct fftw3 version (offsets in range 10e-21). I also triple-checked all strides and transforming dimensions. Could there be any other reason for this huge discrepancy or did anyone else face similar accuracy drops during backward FFT?
I mapped an algorithm which uses R2C and C2R FFTs from fftw3 to clFFT. When comparing the outputs, I encountered a non negligible offset (in the 10e-3 range with MSE) in the backward FFT (hermitian-interleaved -> real, out-of-place). I then compared the intermediate results in the algorithm which all came close to the correct fftw3 version (offsets in range 10e-21). I also triple-checked all strides and transforming dimensions. Could there be any other reason for this huge discrepancy or did anyone else face similar accuracy drops during backward FFT?