<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=utf-8">
</head>
<body text="#000000" bgcolor="#FFFFFF">
<div class="moz-cite-prefix">Hello,<br>
<br>
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<br>
<br>
Y+2*X<br>
<br>
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 <br>
<br>
daxpy(2,X,Y)<br>
<br>
for two 10000x10000 matrices. Here are the results (MacBook air
core i7@1,7GHz):<br>
<br>
daxpy(2,X,Y)<br>
(dcopy: 582 ms)<br>
(daxpy: 211 ms)<br>
<br>
elapsed time: 793 ms<br>
<br>
Y+2*X<br>
<br>
elapsed time: 1574 ms<br>
<br>
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. <br>
<br>
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).<br>
<br>
S.<br>
<br>
<br>
Le 15/02/2018 à 17:11, Antoine ELIAS a écrit :<br>
</div>
<blockquote type="cite"
cite="mid:3daebae7-eba4-a32c-5e1e-c403ece6f5fa@scilab-enterprises.com">Hello
Stéphane, <br>
<br>
Interesting ... <br>
<br>
In release, we don't ship the header of BLAS/LAPACK functions. <br>
But you can define them in your C file as extern. ( and let the
linker do his job ) <br>
<br>
extern int C2F(dgemm) (char *_pstTransA, char *_pstTransB, int
*_piN, int *_piM, int *_piK, double *_pdblAlpha, double *_pdblA,
int *_piLdA, <br>
double *_pdblB, int *_piLdB, double
*_pdblBeta, double *_pdblC, int *_piLdC); <br>
and <br>
<br>
extern int C2F(dscal) (int *_iSize, double *_pdblVal, double
*_pdblDest, int *_iInc); <br>
<br>
Others BLAS/LAPACK prototypes can be found at
<a class="moz-txt-link-freetext"
href="http://cgit.scilab.org/scilab/tree/scilab/modules/elementary_functions/includes/elem_common.h?h=6.0">http://cgit.scilab.org/scilab/tree/scilab/modules/elementary_functions/includes/elem_common.h?h=6.0</a><br>
<br>
Regards, <br>
Antoine <br>
Le 15/02/2018 à 16:50, Stéphane Mottelet a écrit : <br>
<blockquote type="cite">Hello all, <br>
<br>
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 <br>
<br>
scalar*matrix <br>
<br>
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) : <br>
<br>
<blockquote type="cite">int iMultiRealScalarByRealMatrix( <br>
double _dblReal1, <br>
double *_pdblReal2, int _iRows2, int _iCols2, <br>
double *_pdblRealOut) <br>
{ <br>
int iOne = 1; <br>
int iSize2 = _iRows2 * _iCols2; <br>
<br>
C2F(dcopy)(&iSize2, _pdblReal2, &iOne,
_pdblRealOut, &iOne); <br>
C2F(dscal)(&iSize2, &_dblReal1, _pdblRealOut,
&iOne); <br>
return 0; <br>
} <br>
</blockquote>
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, <br>
<br>
A=rand(20000,20000); <br>
tic; for i=1:10; A*1; end; toc <br>
<br>
ans = <br>
<br>
25.596843 <br>
<br>
but this second piece of code is more than 8 times faster and
uses 100% of the cpu, <br>
<br>
ONE=ones(20000,1); <br>
tic; for i=1:10; A*ONE; end; toc <br>
<br>
ans = <br>
<br>
2.938314 <br>
<br>
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) <br>
<br>
<blockquote type="cite">int iMultiRealMatrixByRealMatrix( <br>
double *_pdblReal1, int _iRows1, int _iCols1, <br>
double *_pdblReal2, int _iRows2, int _iCols2, <br>
double *_pdblRealOut) <br>
{ <br>
double dblOne = 1; <br>
double dblZero = 0; <br>
<br>
C2F(dgemm)("n", "n", &_iRows1, &_iCols2,
&_iCols1, &dblOne, <br>
_pdblReal1 , &_iRows1 , <br>
_pdblReal2, &_iRows2, &dblZero, <br>
_pdblRealOut , &_iRows1); <br>
return 0; <br>
} <br>
</blockquote>
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 ? <br>
<br>
Thanks for your help <br>
<br>
S. <br>
<br>
<br>
</blockquote>
<br>
_______________________________________________ <br>
dev mailing list <br>
<a class="moz-txt-link-abbreviated"
href="mailto:dev@lists.scilab.org">dev@lists.scilab.org</a> <br>
<a class="moz-txt-link-freetext"
href="http://lists.scilab.org/mailman/listinfo/dev">http://lists.scilab.org/mailman/listinfo/dev</a>
<br>
</blockquote>
<p><br>
</p>
<pre class="moz-signature" cols="72">--
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
<a class="moz-txt-link-freetext" href="http://www.utc.fr/%7Emottelet">http://www.utc.fr/~mottelet</a></pre>
</body>
</html>