uBLAS and fast matrix multiplication
Hello, I am considering switching from in-house matrix implementation to uBLAS. Overall impression from uBLAS is quite good, but it seems that uBLAS does not take advantage of fast matrix multiplication algorithms such as Strassen or Vinograd. Is that correct? If yes, are there plans to implement such algorithms? Best regards, Kirill Lapshin.
----- Original Message -----
From: klapshin
Hello,
I am considering switching from in-house matrix implementation to uBLAS. Overall impression from uBLAS is quite good, but it seems that uBLAS does not take advantage of fast matrix multiplication algorithms such as Strassen or Vinograd. Is that correct?
Yes.
If yes, are there plans to implement such algorithms?
There was some discussion regarding these on the boost ml during the review IIRC, which convinced me that it's not necessary to provide them. Regards Joerg
If yes, are there plans to implement such algorithms?
There was some discussion regarding these on the boost ml during the review IIRC, which convinced me that it's not necessary to provide them.
I was trying to find this discussion in archives, but no luck. Does anyone has it saved by chance? Or could anyone post quick recap on what were reasons to not to provide it? The only reason I can think of is prevent unnecessary bloating of source code. But seems to me quite a lot of people actually need it. Creating a separate header file for strassen would solve the problem with compiling it every time by users who don't really need it, would not it? Regards, Kirill.
Hi Kiril, you wrote:
If yes, are there plans to implement such algorithms?
There was some discussion regarding these on the boost ml during the review IIRC, which convinced me that it's not necessary to provide them.
I was trying to find this discussion in archives, but no luck. Does anyone has it saved by chance? Or could anyone post quick recap on what were reasons to not to provide it?
Try to search http://aspn.activestate.com/ASPN/Mail/Archives/boost/ the discussion was there during uBLAS formal review.
The only reason I can think of is prevent unnecessary bloating of source code. But seems to me quite a lot of people actually need it.
uBLAS is evolving. Although I am not not using Winograd/Strassen like algorithms in my work I do not see anything bad in having those implemented for uBLAS.
Creating a separate header file for strassen would solve the problem with compiling it every time by users who don't really need it, would not it?
Regards, Kirill.
Regards, Alexei. __________________________________________________ Do you Yahoo!? Yahoo! Mail Plus - Powerful. Affordable. Sign up now. http://mailplus.yahoo.com
participants (3)
-
Alexei Novakov -
jhr.walter＠t-online.de -
klapshin <klapshin＠yahoo.com>