DFT/FFT on graphics hardware

Quote from lilboy716:

nitro..

SSE3, i believe has a FFT instruction.

its also available on AMD64 3500+ Venice, San Diego chips
More...
lilboy716,

Reeeeeally? Can you provide me with a link? I tried googling it, but nothing came up for me.

nitro
 
Quote from lilboy716:

ftp://download.intel.com/technology...01/art01_microarchitecture/vol8iss1_art01.pdf
More...
Ok, I see. These are not FFT routines, but routines that are used by complex FFTs that make complex FFT routines faster.

http://www.ffte.jp/

some packages are available to take advantage of it..however, i have no experience with this at all. can't help you any further
More...
Yeah, I am aware of these packages.

why do you need FFT?
More...
The very article (pdf) you gave me above gives the reason!

nitro
 
Quote from nitro:
The very article (pdf) you gave me above gives the reason!

nitro [/B]
More...
The code sequence above shows how to implement a double-precision complex multiplication using SSE2 only or with the new SSE3 instructions, where mem_X contains one complex operand and mem_Y the other; mem_Z is used to store the complex result and xmm7 is a constant used to change the sign of one data element.

Since the main speed limiter of this code is the number of execution uops (7 for SSE2, 4 for SSE3), <i><b><u>the new instructions can improve complex multiplications by up to 75%.</i></b></u>

On SPEC CPU2000, the compiler is able to use SSE3 to improve 168.wupwise by 10-15%.

Nitro...
is this 'across the board' improvement for CPU usage for ALL custom indicators that require a lot of math or am i out in left field for the reasons to implement this...

cj...

:confused:

__________________
HAVE STOP - WILL TRADE

If You Have The Vision We Have The Code
 
I doubt 95+% of standard TA indicators would benefit from any of these instructions.

nitro
Quote from EdgeHunter:

The code sequence above shows how to implement a double-precision complex multiplication using SSE2 only or with the new SSE3 instructions, where mem_X contains one complex operand and mem_Y the other; mem_Z is used to store the complex result and xmm7 is a constant used to change the sign of one data element.

Since the main speed limiter of this code is the number of execution uops (7 for SSE2, 4 for SSE3), <i><b><u>the new instructions can improve complex multiplications by up to 75%.</i></b></u>

On SPEC CPU2000, the compiler is able to use SSE3 to improve 168.wupwise by 10-15%.

Nitro...
is this 'across the board' improvement for CPU usage for ALL custom indicators that require a lot of math or am i out in left field for the reasons to implement this...

cj...

:confused:

__________________
HAVE STOP - WILL TRADE

If You Have The Vision We Have The Code
More...
 
-most traders dont have SSE3 enabled CPU

-most trading software wont be compiled with SSE3 instructions enabled. unless you're writing your own, in nitro's case.
 
You can do a really fast FFT by programming a PLD to do it in hardware. The divide is just a shift with the FFT anyhow, that is why it's called the "Fast" Fourier. I don't see how any processor could be faster than that but I never made a comparison.
 
Sorry a little OT here, but relative...

I wonder how anyone who uses Fourier deals with non-staionarity as well as impulse events that are messy (imperfect) due to the constraints of Fourier.

I found wavelet transforms much more applicable to the type of time-serie traders tend to quantify.

Any comments as to why you choose Fourier over wavelet would be greatly appreciated, since I'm always trying to improve comprehension generally.

kt
 
You must log in or register to reply here.
Back
Top