cblas_dasum is not adding up negative numbers

[Auvi_R_]

Hello,

I am trying to get the sum of a vector using MKL's cblas_dasum function in C using Intel C++ compiler but the negative numbers are not subtracted. I would expect 40.0 + 2.0 - 42.0 == 0.0 but I am getting 40.0 + 2.0 - 42.0 == 84.0. Here is the code:

/* compiled with  icc -mkl blastest.c -o blastest */

#include <stdio.h>
#include <stdlib.h>
#include "mkl.h"

#define N 3

int main() {
  double *v;
  v = (double *) mkl_malloc(N * sizeof( double ), 64);
  v[0] = (double) 40.0;
  v[1] = (double)  2.0;
  v[2] = (double) -42.0;
  printf("Sum  = %lf\n", cblas_dasum(N, v, 1));
  return 0;
}

Can anybody tell me what's wrong in the code as I am getting 84.0 as the answer.

Thanks,

Auvi

[mecej4]

Please read the manual page containing the specification of ?asum. The routines sasum, dasum return the sum of the absolute values of the input vector array elements. The casum and zasum routines return the sum of the absolute values of the real plus the absolute values of the imaginary parts of the input vector array elements.

[Auvi_R_]

Thanks for opening my eyes! This was an oversight on my part as I did download the documentation and tried to quickly locate a "sum" function. I don't see just a "dsum" function though that will work for my purpose. What is the best way to achieve this? A dot product of two vectors where one of them is just "1"s? summing up vectors is something I do all the time in other environments (numpy, MATLAB) and I'm looking for a straightforward way to achieve that in MKL.

Thanks in advance,

[TimP]

Intel c++ offers __sec_reduce_add as well as accumulate and so on. ? Sum is a common blas extension not offered by netlib or mkl.

[Gennady_F_Intel]

Auvi, 

mkl doesn't provide such functionality. In the case if this is "hot spot" into your application - you can use Tim's suggestion to use Cilk Plus option or  try to use another performance libraries 0 Intel(R) IPP which provides computing the sum of the elements of vectors:  ippsSum_32f and etc...  . In the case if this is not "hot spot" -- the plain code compiled by Intel Compiler would give you good results from performance point of view. 

--Gennady

[TimP]

My tests show no performance difference among C++ STL, Cilk(tm) Plus, and plain C implementations. A threaded MKL or IPP function might show an advantage for large enough arrays (10000?), while in-line code should be faster for small arrays (<500).

[Auvi_R_]

Thanks everyone for the comments. For the time being I'm keeping a vector called ones with sufficiently large number of 1.0's and using it with cblas_ddot whenever I need to sum. As the ones vector is created only once, I am okay with it.