Is there any info on how to implement fft and ifft algorithms that
aren't recursive? I'd like to possibly calculate a really long on
inside a signal chain so attempting to compute a 2 second fft in one
function call would probably be disastrous.
Gosh, Steve... apologies for my naivety. The only implementation I've
seen has been recursive. I'm a bit of a noob in this regard.
Gosh, Steve... apologies for my naivety. The only implementation I've seen has been recursive. I'm a bit of a noob in this regard.
The point of the original question is, how can I spread the computation of an FFT out over multiple vector-based performance routine calls. (not sure the exact term for that.)
Ouch!
Gosh, Steve... apologies for my naivety. The only implementation I've seen has been recursive. I'm a bit of a noob in this regard. I'm definitely looking to compute these frames with a high degree of latency. The idea here is that we trade immediacy for spectral accuracy because, in my use case, latency isn't a huge problem as long as it stays under a second or two. In my current implementation I use a low priority thread and do it on the graphics processor since it could be up to a few seconds of 44.1k audio.
The point of the original question is, how can I spread the computation of an FFT out over multiple vector-based performance routine calls. (not sure the exact term for that.)
Gosh I had no clue I could be interpreted so many ways. I am very sorry. Let me put it in pseudocode:
cnt = 0
maxCnt= 2000
audioPerform(double *in, double *out){
//this will interrupt audio!!!!
if(cnt==0)
Compute an entire 16777216 bin fft
cnt =(cnt+1)%maxCnt
}
audioPerform(double *in, double *out){
//this is presumable will not stop audio!!!!
Compute (16777216/maxCnt) bins of a 16777216 Bin fft.
}
Am 22.11.19 um 19:04 schrieb MatthewA:
The point of the original question is, how can I spread the
computation of an FFT out over multiple vector-based performance
routine calls.-a (not sure the exact term for that.)
Ah, so the "real question" is: How can I efficiently compute an FFT on a parallel vector processor like a GPU?
Unfortunately, that question is nontrivial, because, as you have
discovered, there is a lot of dopendency between the input and output values. Each input value in a FFT has an influence onto each output
value, which makes parallel processing hard.
Is there any info on how to implement fft and ifft algorithms that aren't recursive? I'd like to possibly calculate a really long on inside a signal chain so attempting to compute a 2 second fft in one function call would probably be disastrous.
-Matt
The point of the original question is, how can I spread the computation of an FFT out over multiple vector-based performance routine calls. (not sure the exact term for that.)If I understand your question correctly, you want to spread the computation of an N-point DFT across P processors, with each processor computing a unique portion of the transform, such that they can all be combined afterward.
Is there any info on how to implement fft and ifft algorithms
that aren't recursive? I'd like to possibly calculate a really
long on inside a signal chain so attempting to compute
a 2 second fft in one function call would probably be disastrous.
| Sysop: | Amessyroom |
|---|---|
| Location: | Fayetteville, NC |
| Users: | 76 |
| Nodes: | 6 (0 / 6) |
| Uptime: | 53:56:57 |
| Calls: | 1,265 |
| Calls today: | 1 |
| Files: | 1,362 |
| D/L today: |
3 files (2,684K bytes) |
| Messages: | 285,448 |