BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::nfactor_diag_error
(
INT_1DARRAY threads_start
)
{
Int nthread_remalloc = 0;
for(Int ti = 0; ti < num_threads; ti++)
{
//Note: jdb we can make this into a switch
if(thread_array(ti).error_type ==
BASKER_ERROR_NOERROR)
{
threads_start(ti) = BASKER_MAX_IDX;
continue;
}//end if NOERROR
if(thread_array(ti).error_type ==
BASKER_ERROR_SINGULAR)
{
printf("ERROR THREAD: %d DIAGBLK SINGULAR: %d\n",
ti,
thread_array(ti).error_blk);
return BASKER_ERROR;
}//end if SINGULAR
if(thread_array(ti).error_type ==
BASKER_ERROR_NOMALLOC)
{
printf("ERROR THREADS: %d DIAGBLK NOMALLOC: %d\n",
ti,
thread_array(ti).error_blk);
return BASKER_ERROR;
}//end if NOMALLOC
if(thread_array(ti).error_type ==
BASKER_ERROR_REMALLOC)
{
BASKER_ASSERT(thread_array(ti).error_blk > 0,
"nfactor_diag_error error_blk");
printf("ERROR THREADS: %d DIAGBLK MALLOC: %d \n",
ti,
thread_array(ti).error_blk);
//Clean the workspace
printf("test: %d %d \n",
thread_array(ti).iws_size*thread_array(ti).iws_mult,
thread_array(ti).ews_size*thread_array(ti).ews_mult);
for(Int i = 0;
i < thread_array(ti).iws_size*thread_array(ti).iws_mult;
i++)
{
thread_array(ti).iws(i) = (Int) 0;
}
for(Int i = 0;
i < thread_array(ti).ews_size*thread_array(ti).ews_mult;
i++)
{
thread_array(ti).ews(i) = (Entry) 0;
}
//Resize L
BASKER_MATRIX &L = LBTF(thread_array(ti).error_blk);
REALLOC_INT_1DARRAY(L.row_idx,
L.nnz,
thread_array(ti).error_info);
REALLOC_ENTRY_1DARRAY(L.val,
L.nnz,
thread_array(ti).error_info);
L.nnz = thread_array(ti).error_info;
for(Int i = 0; i < L.ncol; i++)
{
L.col_ptr(i) = 0;
}
for(Int i = L.srow; i < (L.srow+L.nrow); i++)
{
gperm(i) = BASKER_MAX_IDX;
}
//Resize U
BASKER_MATRIX &U = UBTF(thread_array(ti).error_blk);
REALLOC_INT_1DARRAY(U.row_idx,
U.nnz,
thread_array(ti).error_info);
REALLOC_ENTRY_1DARRAY(U.val,
U.nnz,
thread_array(ti).error_info);
U.nnz = thread_array(ti).error_info;
for(Int i = 0; i < U.ncol; i++)
{
U.col_ptr(i) = 0;
}
printf("Setting thread start(%d) %d \n",
ti, thread_array(ti).error_blk);
threads_start(ti) = thread_array(ti).error_blk;
//Reset
thread_array(ti).error_type = BASKER_ERROR_NOERROR;
thread_array(ti).error_blk = BASKER_MAX_IDX;
thread_array(ti).error_info = BASKER_MAX_IDX;
nthread_remalloc++;
}//if REMALLOC
}//for all threads
if(nthread_remalloc == 0)
{
return BASKER_SUCCESS;
}
else
{
return nthread_remalloc;
}
//Should never be here
BASKER_ASSERT(0==1, "nfactor_diag_error, should never");
return BASKER_SUCCESS;
}//end nfactor_diag_error
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::serial_btf_solve
(
ENTRY_1DARRAY y,
ENTRY_1DARRAY x
)
{
for(Int i = 0; i < gn; ++i)
{
x(i) = y(i);
y(i) = (Entry) 0.0;
}
//printf("Test \n");
//Start in C and go backwards
//In first level, only due U\L\x->y
for(Int b = (btf_nblks-btf_tabs_offset)-1;
b>= 0; b--)
{
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("\n\n btf b: %d \n", b);
#endif
//---Lower solve
BASKER_MATRIX &LC = LBTF(b);
//L\x -> y
lower_tri_solve(LC,x,y);
BASKER_MATRIX &UC = UBTF(b);
//U\x -> y
upper_tri_solve(UC,x,y);
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("Before spmv\n");
printf("Inner Vector y print\n");
printVec(y, gn);
printf("Inner Vector x print\n");
printVec(x, gn);
printf("\n");
#endif
//-----Update
//if(b > btf_tabs_offset)
{
//x = BTF_C*y;
//printf("spmv tab: %d \n", b+btf_tabs_offset);
spmv_BTF(b+btf_tabs_offset,
BTF_C, y, x);
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("After spmv\n");
printf("Inner Vector y print\n");
printVec(y, gn);
printf("Inner Vector x print\n");
printVec(x, gn);
#endif
//BASKER_MATRIX &UC = UBTF[b];
//U\x -> y
//upper_tri_solve(UC,x,y);
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("Done, BTF-C Solve \n");
printf("\n x \n");
printVec(x, gn);
printf("\n y \n");
printVec(y, gn);
printf("\n\n");
#endif
//Update B
//BTF_B*y -> x
if(btf_tabs_offset != 0)
{
neg_spmv(BTF_B,y,x);
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("Done, SPMV BTF_B UPDATE \n");
printf("\n x \n");
printVec(x, gn);
printf("\n y \n");
printVec(y, gn);
printf("\n\n");
#endif
//now do the forward backwared solve
//L\x ->y
serial_forward_solve(x,y);
//U\y->x
serial_backward_solve(y,x);
//copy lower part down
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("copying lower starting: %d \n",
btf_tabs[btf_tabs_offset]);
#endif
for(Int i = btf_tabs(btf_tabs_offset); i < gn; ++i)
{
//x[i] = y[i];
x(i) = y(i);
}
//Comeback and fix
return 0;
}//end serial_btf_solve