QuoteOriginally posted by: PolterQuoteOriginally posted by: renorm I wrote some LMM code (both plain and Longstaff-Scwartz) that can run almost as fast as CUDA (or even faster, depending on hardware):-)QuoteOriginally posted by: renorm but can't publish it due to lack of time and some other reasons.:-(The product is available but not shipping?

QuoteOnly small subset of GSL is needed. A nice, compact with standardised interface library to do algebra is possible. Even now, the Boost Math Toolkit has a lot of overlap with GSL. And building matrix algos on top of uBLAS is easy and transparent.What are the most important functions you use in GSL?

double.

QuoteOriginally posted by: renormNot sure about nice part, but LAPACK/BLAS and corresponding C-API (as it is found in MKL 10.3 and newer) are standardized interface for linear algebra.uBlas is no replacement for LAPACK/BLAS. As far as I know it doesn't work well with Intel/AMD/ATALS BLAS. uBlas is OK for small matrices, but too slow for large matrices. It is about 50 times slower than Intel/AMD/ATLAS BLAS and doesn't provide LAPACK functionality. It would be very bad idea to reinvent LU/QR/SVD/OLS and etc instead of using LAPACK. C++ wrappers over LAPACK/BLAS are OK, but reinventing core functionality is really really bad idea.QuoteWhat are the most important functions you use in GSL? Quadratures and some special functions.Btw, Boost math toolkit isn't designed for high performance. For example, inverse Gaussian CDF from Boost is several times slower than Acklam's method found in QuantLib. Acklam's relative error is about 1e-9, which is more than enough for MC. IIRC, MKL's implementation is about about 50 times faster than Boost. I think that is a lot, especially in LMM type of simulations. The same could be said about other distributions (Poisson, discrete and etc).We should then provide a choice for the user:1. Portable, easy to install libraries, not necessarily optimised. BTW Math Toolkit's focus was accuracy, initially.2. Specific, possibly platform-dependent libraries (that MKL costs $?) with bigger footprint and learning curve (e.g. big challenge of LAPACK on Windows3. There is a real need for lightweight libraries without having to drag in the kitchen sink in order to use them (e.g. novice C++ users).For MC and speed you need 2. But not always. More speedup <==> more effort.QuoteIt would be very bad idea to reinvent LU/QR/SVD/OLS and etc instead of using LAPACKWhy a bad idea? LU is about 8 lines of code using uBLAS. Here is a scenario; a casual C++ user creates an accurate POC app using local LU and then hands the code to C++ system integrators for acceptance testing, just by switching the interfaces.QuoteuBlas is OK for small matrices300X300? QuoteQuadratures and some special functions.Outrun has some remarks on Boost Quadrature.

Well, I programmed QR, SVD etc. before (in Fortran) and it is part of graduate numerical analysis studies. No problem, and we only need a small subset. QuoteDoes anyone is planning to run this project on anything but x86? And what do you mean by "easy to install"?For MFE/MSc students. Can we have this StyleI think x86 on Windows is popular. QuoteIf novice C++ users can learn Boost, they should have no problem with basic C-API such LAPACK/BLAS.Maybe I am missing something, *but* LAPACK install on Windows is NOT easy.QuoteSpecific, possibly platform-dependent libraries (that MKL costs $?) with bigger footprint and learning curve (e.g. big challenge of LAPACK on WindowsThis is still an issue. 'Footprint' ==> we need the full package, 'learning curve' is known to each cost centre. If it takes 2 weeks to learn something just to use QR, then you can use PLAN B.