static void getri(void *descr[], int type) {
// Computes the Echelon form of A, stores the corresponding inverted multiples
// in the lower part
// ------------------------------
// | A | | | | | | |
// |----------------------------|
// | | | | | | | |
// |----------------------------|
// | | | | | | | |
// |----------------------------|
// | | | | | | | |
// ------------------------------
unsigned int i, j, k;
unsigned int *sub_a = (unsigned int *)STARPU_MATRIX_GET_PTR(descr[0]);
unsigned int x_dim = STARPU_MATRIX_GET_NX(descr[0]);
unsigned int y_dim = STARPU_MATRIX_GET_NY(descr[0]);
unsigned int offset_a = STARPU_MATRIX_GET_OFFSET(descr[0]);
unsigned int ld_a = STARPU_MATRIX_GET_LD(descr[0]);
unsigned int mult = 0;
#if DEBUG0
printf("\n --- GETRI ---\n");
#endif
#if DEBUG
printf("x_dim = %u\n", x_dim);
printf("y_dim = %u\n", y_dim);
printf("ld_a = %u\n", ld_a);
#endif
for (i = 0; i < y_dim; ++i) {
// compute inverse
neg_inv_piv[i+offset_a] = negInverseModP(sub_a[i+i*ld_a], prime);
#if DEBUG
printf("sub_a[%u] = %u\n", i+i*ld_a,sub_a[i+i*ld_a]);
printf("inv = %u\n", neg_inv_piv[i+offset_a]);
#endif
for (j = i+1; j < x_dim; ++j) {
// multiply by corresponding coeff
mult = (neg_inv_piv[i+offset_a] * sub_a[i+j*ld_a]); //% prime;
#if DEBUG
printf("sub_a[%u] = %u\n", i+j*ld_a,sub_a[i+j*ld_a]);
printf("mult = %u\n", mult);
#endif
sub_a[i+j*ld_a] = mult;
for (k = i+1; k < y_dim; ++k) {
sub_a[k+j*ld_a] += (sub_a[k+i*ld_a] * mult);
//sub_a[k+j*ld_a] %= prime;
}
}
}
#if DEBUG0
printf("\n --- GETRI DONE ---\n");
printf("TASKS READY %d\n", starpu_task_nready());
printf("TASKS SUBMITTED %d\n", starpu_task_nsubmitted());
#endif
}
{
float *subB;
uint32_t nxB, nyB;
uint32_t ldB;
float east,west,north,south;
uint32_t offset,dimension,offset_row_shift;
subB = (float *)STARPU_MATRIX_GET_PTR(descr[0]);
nxB = STARPU_MATRIX_GET_NX(descr[0]);
nyB = STARPU_MATRIX_GET_NY(descr[0]);
ldB = STARPU_MATRIX_GET_LD(descr[0]);
unsigned i,j;
unsigned int x,y;
offset=STARPU_MATRIX_GET_OFFSET(descr[0]);
//Based on offset we need to calculate the location of north,south,east and west in matrix A.
offset_row_shift=offset;
for(i=0;i<nxB;i++)
{
for(j=0;j<nyB;j++)
{
dimension=offset/4;
x=floor(dimension/ydim);
y=dimension%ydim;
if(x==0 && y!=0)
{
north=0;
west=A[x*ydim+(y-1)];
south=A[(x+1)*ydim+y];
if(y==ydim-1)
