Jeff Terry wrote:
If I read the vDSP manual correctly, the vDSP_fft2d_zrop function can only operate on matrices up to 1024 by 1024 in size because log2nInCol nad log2nInRow must be between 3 and 10 and 2^10 is 1024.
Is my interpretation of the manual correct? Is anyone doing ffts on larger arrays? If so, what are you using?
You are reading the manual correctly. Attempting larger transforms on PowerPC definitely fails. However, the Intel implementation of vDSP uses different code, which I work on, and it does not have the same limits. Glancing at the code, I do not see a limit other than the natural address limits, but I am unprepared to make an authoritative statement.
However, even if the code works mathematically, performance may be an issue. vDSP is generally designed only for in-cache operations. Large FFTs encounter not only memory bandwidth but also cache thrashing due to how memory is mapped to cache lines. To get good performance, the code has to be designed to deal with that. If this is a feature you would like in vDSP, please file a "bug report" (also used for feature requests) at http://bugreporter.apple.com.
-- edp (Eric Postpischil)
|
