The BWs are 20mhz, 40mhz and 80Mhz respectively. The host can range from a server with 48 cores(2.6Ghz), 256GB ram to a laptop with 8 cores(3.0Ghz), 64GB ram.
Have you tried using the blocks you asked about to see what happens with them? Jeff Long recommended running top to see if the processing threads for the resampling fall behind. You can also check your memory growth to see if that happens.
You may be able to do a cheap/dirty resampling by taking the FFT at 23MHz, adding a few 0 bins to get it to 25MHz, then doing the IFFT. You'll need to do some strange FFT sizes - maybe 4600 samples at the lower rate, and pad to 5000 samples for the higher
rate. I don't think you'll need to filter, but I haven't gone through the exercise. For cores like that, tho, it should be very fast to do an FFT followed by an IFFT with some zero's shoved into a vector, especially with so many cores.
Lastly, have you run fftw wisdom to get the fastest FFT performance? That might be useful as well.
Let us know how your experimentation goes. This sounds interesting.
Brian