PVR@be.ac.rug.autoctrl (03/01/91)
Dear mailing list, I received a demonstration disk of the VECLIB libraries for the transputer and I was impressed by the performance. Has anyone done a comparison of this vector library with the Mark 2, BLAS, those from TopExpress, Parsytec, Numerical Analysis and the others there may be ? Does anyone of you know if the VECLIB version for 3L gives problems when you use it with the occam Toolset ? Finally, does anyone know if the VECLIB already contains FFT, IFFT etc ? Patrick Van Renterghem pvr@bgerug51.bitnet (EARN) Belgium pvr@autoctrl.rug.ac.be
sjb@transtech.UUCP (Manager Systems Division) (03/02/91)
> Dear mailing list, > > I received a demonstration disk of the VECLIB libraries for the transputer > and I was impressed by the performance. Has anyone done a comparison of > this vector library with the Mark 2, BLAS, those from TopExpress, Parsytec, > Numerical Analysis and the others there may be ? > > Does anyone of you know if the VECLIB version for 3L gives problems when > you use it with the occam Toolset ? > > Finally, does anyone know if the VECLIB already contains FFT, IFFT etc ? > > Patrick Van Renterghem pvr@bgerug51.bitnet (EARN) > Belgium pvr@autoctrl.rug.ac.be > Sinectonalysis has just finished porting VECLIB to GENESYS. We have not been able to establish an exact comparison with other libraries but in view the levels of performance being achieved I would be surprised if there would be anything available which is significantly faster. VECIB offers over 350 vector primitives to accelerate vector/matrix, singnal processing and neural network operations. The library is hand coded in T800 assembler. Alexander Orenshteyn (of Sinectonalysis) has got fft routines, as the results below indicate. If there are VECLIB users out there who do not have his FFTs he would more than likely be willing to sort them out. His phone number is: (617) 527 5807 The following results were obtained with his library routines being accessed from C under GENESYS. PERFORMANCE ___________ A 1024 point complex FFT is computed in 32.9 milliseconds running on a 25MHz T800. The following results are also obtained on a 25MHz T800: Function Name Desc Performance (MFLOPS) Double Single vmuladd vector scalar 1.03 1.60 multiply, add vdot dot product 1.42 2.05 vsum sum-up vector 2.11 2.29 vnorm2 norm of vector 1.67 2.33 vmulalpha scale by constant 0.61 1.06 vaddalpha add constant 0.94 1.35 vsquare square comps. 0.62 1.04 vmax find max 0.80 0.94 _/ S.J.Bradshaw sjb@transtech.com _/_/ Manager, Systems Division, ..!theory.tn.cornell.edu! _/_/_/ Transtech Parallel Systems Corp., cayuga!sjb _/_/ 120 Langmuir Laboratories, Phone (607) 257 6502 _/ 95 Brown Road, Ithaca, NY 14850, USA FAX (607) 257 3980