CUFFT possible bug

[Chris_Spooner]

Hi,
I’ve been trying to use CUFFT on 1D double precision FFT. I’m trying to batch FFT 896 x 128 long blocks.
The problem is that the output data seems okay up until a point, then it results in only zeros in the output. I’ve tried running it on a variety of cards (GTS450, GTX465, GTX265) and always get the same results. I’ve also tried it on CUDA releases version 3.1 and 3.2RC and get the same results.
I’ve attached the two visual C++ projects I’m using for testing. One implements the transform in CUDA C as a DLL. The other uses the DLL and outputs the test results.
Any recommendations on this matter would be greatly appreciated.

Thanks and regards,
Chris Spooner.
WindowedCUFFT.zip (525 KB)
SimpleCUFFTTest.zip (314 KB)

[Chris_Spooner]

Hi,
I’ve been trying to use CUFFT on 1D double precision FFT. I’m trying to batch FFT 896 x 128 long blocks.
The problem is that the output data seems okay up until a point, then it results in only zeros in the output. I’ve tried running it on a variety of cards (GTS450, GTX465, GTX265) and always get the same results. I’ve also tried it on CUDA releases version 3.1 and 3.2RC and get the same results.
I’ve attached the two visual C++ projects I’m using for testing. One implements the transform in CUDA C as a DLL. The other uses the DLL and outputs the test results.
Any recommendations on this matter would be greatly appreciated.

Thanks and regards,
Chris Spooner.

[ramarromarrone]

Hi,

I've been trying to use CUFFT on 1D double precision FFT. I'm trying to batch FFT 896 x 128 long blocks.

The problem is that the output data seems okay up until a point, then it results in only zeros in the output. I've tried running it on a variety of cards (GTS450, GTX465, GTX265) and always get the same results. I've also tried it on CUDA releases version 3.1 and 3.2RC and get the same results.

 I've attached the two visual C++ projects I'm using for testing. One implements the transform in CUDA C as a DLL. The other uses the DLL and outputs the test results.

 Any recommendations on this matter would be greatly appreciated.

Thanks and regards,

    Chris Spooner.

zeros start in the middle of the output?

this because fft is simmetric so cufft calculate only 1 half of the transform. you can calculate other elements of the transform making complex conjugate of the first part.

for more informations take a look to the cufft_rpogramming guide pages 7-8:

"

For real to complex FFTs, the output array holds [b][u] only the non

redundant complex coefficients…[/u][/b]

"

[ramarromarrone]

Hi,

I've been trying to use CUFFT on 1D double precision FFT. I'm trying to batch FFT 896 x 128 long blocks.

The problem is that the output data seems okay up until a point, then it results in only zeros in the output. I've tried running it on a variety of cards (GTS450, GTX465, GTX265) and always get the same results. I've also tried it on CUDA releases version 3.1 and 3.2RC and get the same results.

 I've attached the two visual C++ projects I'm using for testing. One implements the transform in CUDA C as a DLL. The other uses the DLL and outputs the test results.

 Any recommendations on this matter would be greatly appreciated.

Thanks and regards,

    Chris Spooner.

zeros start in the middle of the output?

this because fft is simmetric so cufft calculate only 1 half of the transform. you can calculate other elements of the transform making complex conjugate of the first part.

for more informations take a look to the cufft_rpogramming guide pages 7-8:

"

For real to complex FFTs, the output array holds [b][u] only the non

redundant complex coefficients…[/u][/b]

"