Getting dgemm or dgemms to work in C

[echidnasunrise]

Hi there,

I cannot for the life of me get dgemm to spit out the correct answer when I call it from C. I saw your posts on a similar question and hoped you might have an answer.

I would love to have a function that that was some thing like:

void matmult(float ** A, float ** B, float ** C, int Arows, int innerdim, int Bcols)

where you could pass it three arrays float ** or float * form (I don't care, I just want something to work) the rows of the A matrix, the shared inner dimension, and the columns of the B matrix and have the answer dumped in to the C matrix.

Any thoughts? Figuring out how to correct for the column major order and the leading dimensions is difficult.

Thanks so much,
Mike

Here's what I have so far, and it only works for an NxN times and NxN matrix:

/************************************
              INCLUDES
************************************/
#include <stdio.h>
#include <stdlib.h>
#include <math.h>

/************************************
       CREATE A 1D ARRAY
************************************/
float* alloc_float_1d(int ni){
  int i;
  
  float* a = (float*) malloc(ni*sizeof(float));
  
  for(i=0;i<ni;i++)
    a[i]=0;
  
  return a;
} 


/************************************
       1D ARRAY INDEX
************************************/
int aind(int row, int col, int numcols)
{
  return (numcols*row)+col;
}

void aprint1d(float * A, int rows, int cols)
{
 for(int i=0;i<rows;i++)
    {
      for(int j=0;j<cols;j++)
	printf("%1.4f ", A[aind(i,j,cols)]);
      
      printf("\n");
    }
}


/************************************
     Multiply 2 2d(float *) ARRAYS
************************************/
void matmult(float * A, float * B, float * C, int arows, int idim, int bcols)
{

  //Perform the matrix multiplication
  sgemms(A, &arows, "N", B, &bcols, "N", C, &bcols, &arows, &bcols, &idim, 0, 0);


  float temp;
  
  //FLIP UD 
  for(int i=0;i<(int)floor(arows/2.0);i++)
    {
      for(int j=0;j<bcols;j++)
	{
	  temp=C[aind(i,j,bcols)];
	  C[aind(i,j,bcols)]=C[aind((bcols-1)-i,j,bcols)];
	  C[aind((bcols-1)-i,j,bcols)]=temp;
	}
    }
  
  //FLIP LR
  for(int i=0;i<(int)floor(arows/2.0);i++)
    {
      for(int j=0;j<bcols;j++)
	{
	  temp=C[aind(j,i,bcols)];
	  C[aind(j,i,bcols)]=C[aind(j,(bcols-1)-i,bcols)];
	  C[aind(j,(bcols-1)-i,bcols)]=temp;
	}
	}
}


//***********************************
//***********************************
//             MAIN
//***********************************
//***********************************
void main()
{
  //Square Array size
  int  i, j;
  int arows=3;  
  int idim=3;
  int bcols=3;
  
  //Allocate the arrays
  float * A=alloc_float_1d(arows*idim);
  float * B=alloc_float_1d(idim*bcols);
  float * C=alloc_float_1d(arows*bcols);
  
  int c=1; 
   for(i=0; i<arows; i++){
    for(j=0; j<idim; j++){ 
      A[aind(i,j,idim)] = c++;
    }
  }
  
  c=1;
  for(i=0; i<idim; i++){
    for(j=0; j<bcols; j++){ 
      B[aind(i,j,bcols)] = (idim*bcols+1)-c++;
    }
  }
  
  matmult(A,B,C,arows,idim,bcols);
   
  printf("A:\n");
  aprint1d(A,arows,idim);
  printf("\n");
  
  printf("B:\n");
  aprint1d(B,idim,bcols);
  printf("\n");
  
  printf("C:\n");
  aprint1d(C,arows,bcols);
  
}

 

[sedawk]

Maybe I misunderstood your question. I don't know why you had trouble in this:

"Figuring out how to correct for the column major order and the leading dimensions is difficult".

The leading dimension is the input matrix's row dimension. You have full control over them. For example,

using

sgemms(a, lda, transa, b, ldb, transb, c, ldc, l, m, n, aux, naux);

to compute C = A * B.

A is an 8x6 matrix and B is 7x4 matrix. You want C = A x B be 6x4 matrix. You can specify l=6,m=6,n=4 as a set or l=6,m=4,n=4 as another. "m" specifies how many columns in A will be multiply-and-add in B's row. The reason of having varied "m" is because A and B can have different sizes.

I guess the reason that you only had square matrix working is l,m,n arguments specification I described above. The code seems ok.