Hi all, I'm testing some vectorization capabilities of the icc compiler with the kernel matrix multiply code. Depending on the order of the loops the compiler vectorizes or not. Therefore I'm compiling with the option -O3 to force icc to do a loop interchange in order to be able to vectorize always the inner loop of the matrix multiply. I declare the matrices inside the main and perform the multiplication also inside the main (to avoid the need of inter procedural information). Here is the code:

int main()
{
	int dimi = 148; 
	int dimj = 148; 
	int dimk = 148; 

	int i, j, k;

	float **c = (float **)malloc(sizeof(float *)*dimi);
	c[0] = (float *)malloc(dimi * dimj * sizeof(float));
	for(i = 1; i < dimi; i++)
		c[i] = c[0] + i * dimj;

	float **b = (float **)malloc(sizeof(float *)*dimk);
	b[0] = (float *)malloc(dimk * dimj * sizeof(float));
	for(i = 1; i < dimk; i++)
		b[i] = b[0] + i * dimj;

	float **a = (float **)malloc(sizeof(float *)*dimi);
	a[0] = (float *)malloc(dimi * dimk * sizeof(float));
	for(i = 1; i < dimi; i++)
		a[i] = a[0] + i * dimk;


	inic_m(a, dimi, dimk);
	inic_m2(b, 2.0, dimk, dimj);
	inic_m3(c, 3.1, dimi, dimj);



	for (i = 0; i < dimi; i++)
	{
		for (k = 0; k < dimk; k++)
		{
			for (j = 0; j < dimj; j++)
			{
				c[i][j] += a[i][k]*b[k][j];
			}

		}
	}



	return 0;
}

matrizXmatrizTests.c(112) : (col. 2) remark: loop was not vectorized: not inner loop.
matrizXmatrizTests.c(114) : (col. 3) remark: loop was not vectorized: not inner loop.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: vector dependence: assumed FLOW dependence between reference at line 118 and reference at line 118.
matrizXmatrizTests.c(116) : (col. 4) remark: loop was not vectorized: existence of vector dependence.

For the orders KIJ and IKJ I get:

matrizXmatrizTests.c(112) : (col. 2) remark: loop was not vectorized: not inner loop.
matrizXmatrizTests.c(114) : (col. 3) remark: loop was not vectorized: not inner loop.
matrizXmatrizTests.c(116) : (col. 4) remark: LOOP WAS VECTORIZED.