FAQ
Q1 : Why the planning API functions destroy the data on specified buffers?
A1 : The buffers are used in auto-tuning and verification procedures.
Q2 : Support CUDA Stream?
A2 : CUDA Stream for runtime API is incompatible with that for driver API. Since this library is implemented with driver API, it is impossible to specify cudaStream_t. Currently, there are no async API functions. It will be available for driver API. For runtime API, we are planning to use a private CUDA stream.
Q3 : Can I get the source code?
A3 : Now it is not available, except NVIDIA and some collaborators.
Q4 : Can I use this library for commercial use?
A4 : Please contact me.
Q5 : When is this library faster than CUFFT 3.1?
A5 : Unlike the previous version, CUFFT 3.1 is very fast in case of Fermi and single precision and powers-of-two transform sizes. For other cases, I recommend this library.
Q6 : Is this library thread-safe?
A6 : Version 20101101.16 or later should support multiple CUDA contexts.