NI Labs Discussions
Welcome to LabVIEW High Performance Analysis Library
This is looking great, thanks! Great to have 3D FFTs (at last!) natively and fast in LabVIEW, and I'm also impressed to see SGL versions of all the routines. I've quickly benchmarked the 3D FFTs and see a similar improvement over the ASP routines as for 2D on a 2-core machine. Nice!
It does raise a few questions for me:
- There are other signal processing routines that build on top of these existing routines, prime examples being convolution and cross-correlation, both of which are often implelemented with FFTs. Are these part of the plans? It would be great to have built-in routines for these that cope with 1D/2D/3D and various representations. Also, it would be great to finally have a "correct" Normalized Cross-Correlation which is sadly missing at the moment! (see here).
- Having SGL routines is great - I am about to reimplement my 3D deconvolution using SGL/CSG and see if the accuracy is sufficient, which would effectively give twice the memory capacity. However it does point up the lack of SGL routines in the rest of the Array/ASP libraries in LabVIEW currently, even simple functions like Mean. I wonder whether there is the scope to rationalize across all of the floating-point routines in a way that enables this all to be used with the greatest efficiency.
- The only concern I have for my work is the ability to use memory efficiently. Up until now, I've obviously had to code my own 3D FFTs (as a series of 1D FFTs) and have passed into the routine an array to use for the result.
.png)
This minimises the number of arrays created, which is very important especially for large 3D Deconvolution problems. The Richardson-Lucy algorithm, for example, has 2 FFTs and 2 IFFTs per iteration, but I can ensure no extra memory is allocated along the way.
From quickly looking at the MKL documentation, it looks as though the option is there to either re-use the input array, or to provide input and output arrays. Can something like this be added?
Again, many thanks for this. It looks very useful, even at the current level of implementation.
Just doing some more benchmarking, this time of the Dot Product function. For my calculations, this High-Performance function is slightly slower than the builtin Linear Algebra function (LV 2010 and 2011), and considerably slower (3-4x) than the equivalent Multiply-Add for vectors smaller than 200. All techniques show a speed increase when used inside a parallel loop, but the order is still the same.
Have you benchmarked all of the functions against their builtin equivalents?
