Really, nobody knows ?

S.

Le 20/02/2018 à 11:57, Stéphane Mottelet a écrit :
Hello,

Continuing on this subject, Hello, I discovered that the new Scilab API allows to modify input parameters of a function (in-place assignment), e.g. I have modified the previous daxpy such that the expression

daxpy(2,X,Y)

has no output but formally does "Y+=2*X" if such an operator would exist in Scilab. In this case there is no matrix copy at all, hence no memory overhead.

Was it possible to do this with the previous API ?

S.

Le 19/02/2018 à 19:15, Stéphane Mottelet a écrit :
Hello,

After some tests, for the intended use (multiply a matrix by a scalar), dgemm is not faster that dscal, but in the C code of "iMultiRealScalarByRealMatrix", the part which takes the most of the CPU time is the call to "dcopy". For example, on my machine,  for a 10000x10000 matrix, the call to dcopy takes 540 milliseconds and the call to dscal 193 milliseconds. Continuing my explorations today, I tried to see how Scilab expressions such as

Y+2*X

are parsed and executed. To this purpose I have written an interface (daxpy.sci and daxpy.c attached) to the BLAS function "daxpy" which does "y<-y+a*x" and a script comparing the above expression to

daxpy(2,X,Y)

for two 10000x10000 matrices. Here are the results (MacBook air core i7@1,7GHz):

 daxpy(2,X,Y)
 (dcopy: 582 ms)
 (daxpy: 211 ms)

 elapsed time: 793 ms

 Y+2*X

 elapsed time: 1574 ms

Considered the above value, the explanation is that in "Y+2*X" there are *two* copies of a 10000x10000 matrix instead of only one in "daxpy(2,X,Y)". In "Y+2*X+3*Z" there will be three copies, although there could be only one if daxpy was used twice.

I am not blaming Scilab here, I am just blaming "vectorization", which can be inefficient when large objects are used. That's why explicits loops can sometimes be faster than vectorized operations in Matlab or Julia (which both use JIT compilation).

S.


Le 15/02/2018 à 17:11, Antoine ELIAS a écrit :
Hello Stéphane,

Interesting ...

In release, we don't ship the header of BLAS/LAPACK functions.
But you can define them in your C file as extern. ( and let the linker do his job )

extern int C2F(dgemm) (char *_pstTransA, char *_pstTransB, int *_piN, int *_piM, int *_piK, double *_pdblAlpha, double *_pdblA, int *_piLdA,                        double *_pdblB, int *_piLdB, double *_pdblBeta, double *_pdblC, int *_piLdC);
and

extern int C2F(dscal) (int *_iSize, double *_pdblVal, double *_pdblDest, int *_iInc);

Others BLAS/LAPACK prototypes can be found at http://cgit.scilab.org/scilab/tree/scilab/modules/elementary_functions/includes/elem_common.h?h=6.0

Regards,
Antoine
Le 15/02/2018 à 16:50, Stéphane Mottelet a écrit :
Hello all,

Following the recent discussion with fujimoto, I discovered that Scilab does not (seem to) fully use SIMD operation in BLAS as it should. Besides the bottlenecks of its code, there are also many operations of the kind

scalar*matrix

Althoug this operation is correctly delegated to the DSCAL BLAS function (can be seen in C function iMultiRealMatrixByRealMatrix in modules/ast/src/c/operations/matrix_multiplication.c) :

int iMultiRealScalarByRealMatrix(
    double _dblReal1,
    double *_pdblReal2,    int _iRows2, int _iCols2,
    double *_pdblRealOut)
{
    int iOne    = 1;
    int iSize2    = _iRows2 * _iCols2;

    C2F(dcopy)(&iSize2, _pdblReal2, &iOne, _pdblRealOut, &iOne);
    C2F(dscal)(&iSize2, &_dblReal1, _pdblRealOut, &iOne);
    return 0;
}
in the code below the product "A*1" is likely using only one processor core, as seen on the cpu usage graph and on the elapsed time,

A=rand(20000,20000);
tic; for i=1:10; A*1; end; toc

 ans  =

   25.596843

but this second piece of code is more than 8 times faster and uses 100% of the cpu,

ONE=ones(20000,1);
tic; for i=1:10; A*ONE; end; toc

 ans  =

   2.938314

with roughly the same number of multiplications. This second computation is delegated to DGEMM (C<-alpha*A*B + beta*C, here with alpha=1 and beta=0)

int iMultiRealMatrixByRealMatrix(
    double *_pdblReal1,    int _iRows1, int _iCols1,
    double *_pdblReal2,    int _iRows2, int _iCols2,
    double *_pdblRealOut)
{
    double dblOne        = 1;
    double dblZero        = 0;

    C2F(dgemm)("n", "n", &_iRows1, &_iCols2, &_iCols1, &dblOne,
               _pdblReal1 , &_iRows1 ,
               _pdblReal2, &_iRows2, &dblZero,
               _pdblRealOut , &_iRows1);
    return 0;
}
Maybe my intuition is wrong, but I have the feeling that using dgemm with alpha=0 will be faster than dscal. I plan to test this by making a quick and dirty code linked to Scilab so my question to devs is : which are the #includes to add on top of the source (C) to be able to call dgemm and dscal ?

Thanks for your help

S.



_______________________________________________
dev mailing list
dev@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/dev


--
Stéphane Mottelet
Ingénieur de recherche
EA 4297 Transformations Intégrées de la Matière Renouvelable
Département Génie des Procédés Industriels
Sorbonne Universités - Université de Technologie de Compiègne
CS 60319, 60203 Compiègne cedex
Tel : +33(0)344234688
http://www.utc.fr/~mottelet


_______________________________________________
dev mailing list
dev@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/dev


--
Stéphane Mottelet
Ingénieur de recherche
EA 4297 Transformations Intégrées de la Matière Renouvelable
Département Génie des Procédés Industriels
Sorbonne Universités - Université de Technologie de Compiègne
CS 60319, 60203 Compiègne cedex
Tel : +33(0)344234688
http://www.utc.fr/~mottelet


_______________________________________________
dev mailing list
dev@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/dev


--
Stéphane Mottelet
Ingénieur de recherche
EA 4297 Transformations Intégrées de la Matière Renouvelable
Département Génie des Procédés Industriels
Sorbonne Universités - Université de Technologie de Compiègne
CS 60319, 60203 Compiègne cedex
Tel : +33(0)344234688
http://www.utc.fr/~mottelet

_______________________________________________
dev mailing list
dev@lists.scilab.org
http://lists.scilab.org/mailman/listinfo/dev

Reply via email to