BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::scotch_partition
(BASKER_MATRIX &M)
{
nd_flag = BASKER_TRUE;
if(Options.symmetric == BASKER_TRUE)
{
//printf("Scotch Symmetric\n");
part_scotch(M, part_tree);
}
else
{
//printf("Scotch Nonsymmetrix\n");
BASKER_MATRIX MMT;
AplusAT(M,MMT);
//printMTX("AAT.mtx", MMT);
part_scotch(MMT, part_tree);
FREE(MMT);
}
nd_flag == BASKER_TRUE;
//permute
//permute_col(M, part_tree.permtab);
///permute_row(M, part_tree.permtab);
permute_row(M, part_tree.permtab);
permute_col(M, part_tree.permtab);
//May need to sort row_idx
return 0;
}//end scotch_partition()
BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::permute
(
BASKER_MATRIX &M,
INT_1DARRAY row,
INT_1DARRAY col
)
{
permute_col(M,col);
permute_row(M,row);
return 0;
}//end permute(int, int)
BASKER_INLINE
int Basker<Int,Entry, Exe_Space>::find_btf(BASKER_MATRIX &M)
{
Int nblks = 0;
strong_component(M,nblks,order_btf_array,btf_tabs);
btf_flag = BASKER_TRUE;
#ifdef BASKER_DEBUG_ORDER_BTF
printf("BTF nblks returned: %d \n", nblks);
BASKER_ASSERT(nblks>1, "NOT ENOUGH BTF BLOCKS");
#endif
#ifdef BASKER_DEBUG_ORDER_BTF
if(nblks<2)
{
printf("BTF did not find enough blks\n");
}
#endif
#ifdef BASKER_DEBUG_ORDER_BTF
/*
printf("\nBTF perm: \n");
for(Int i=0; i <M.nrow; i++)
{
printf("%d, ", order_btf_array(i));
//printf("%d, ", btf_perm(i));
}
*/
printf("\n\nBTF tabs: \n");
for(Int i=0; i < nblks+1; i++)
{
printf("%d, ", btf_tabs(i));
}
printf("\n");
#endif
permute_col(M, order_btf_array);
permute_row(M, order_btf_array);
break_into_parts(M, nblks, btf_tabs);
btf_nblks = nblks;
//#ifdef BASKER_DEBUG_ORDER_BTF
printf("------------BTF CUT: %d --------------\n",
btf_tabs(btf_tabs_offset));
//#endif
return 0;
}//end find BTF
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::btf_order()
{
//1. Matching ordering on whole matrix
//currently finds matching and permutes
//found bottle-neck to work best with circuit problems
sort_matrix(A);
//printMTX("A_nonmatch.mtx", A);
match_ordering(0);
//printf("DEBUG1: done match\n");
//for debuging
sort_matrix(A);
//printMTX("A_match.mtx", A);
//2. BTF ordering on whole matrix
// Gets estimate of work on all blocks
//currently finds btf-hybrid and permutes
//A -> [BTF_A, BTF_C; 0 , BTF B]
printf("outter num_threads:%d \n", num_threads);
MALLOC_INT_1DARRAY(btf_schedule, num_threads+1);
init_value(btf_schedule, num_threads+1, 0);
find_btf(A);
if(btf_tabs_offset != 0)
{
// printf("A/B block stuff called\n");
//3. ND on BTF_A
//currently finds ND and permute BTF_A
//Would like to change so finds permuation,
//and move into 2D-Structure
//printMTX("A_BTF_FROM_A.mtx", BTF_A);
sort_matrix(BTF_A);
scotch_partition(BTF_A);
//need to do a row perm on BTF_B too
if(btf_nblks > 1)
{
permute_row(BTF_B, part_tree.permtab);
}
//needed because moving into 2D-Structure,
//assumes sorted columns
sort_matrix(BTF_A);
if(btf_nblks > 1)
{
sort_matrix(BTF_B);
sort_matrix(BTF_C);
}
//For debug
//printMTX("A_BTF_PART_AFTER.mtx", BTF_A);
//4. Init tree structure
//This reduces the ND ordering into that fits,
//thread counts
init_tree_thread();
//5. Permute BTF_A
//Constrained symamd on A
INT_1DARRAY cmember;
MALLOC_INT_1DARRAY(cmember, BTF_A.ncol+1);
init_value(cmember,BTF_A.ncol+1,(Int) 0);
for(Int i = 0; i < tree.nblks; ++i)
{
for(Int j = tree.col_tabs(i); j < tree.col_tabs(i+1); ++j)
{
cmember(j) = i;
}
}
//INT_1DARRAY csymamd_perm = order_csym_array;
MALLOC_INT_1DARRAY(order_csym_array, BTF_A.ncol+1);
//MALLOC_INT_1DARRAY(csymamd_perm, BTF_A.ncol+1);
init_value(order_csym_array, BTF_A.ncol+1,(Int) 0);
//init_value(csymamd_perm, BTF_A.ncol+1,(Int) 0);
csymamd_order(BTF_A, order_csym_array, cmember);
//csymamd_order(BTF_A, csymamd_perm, cmember);
//permute(BTF_A, csymamd_perm, csymamd_perm);
permute_col(BTF_A, order_csym_array);
sort_matrix(BTF_A);
permute_row(BTF_A, order_csym_array);
sort_matrix(BTF_A);
//printMTX("A_BTF_AMD.mtx", BTF_A);
if(btf_nblks > 1)
{
permute_row(BTF_B, order_csym_array);
sort_matrix(BTF_B);
//printMTX("B_BTF_AMD.mtx", BTF_B);
sort_matrix(BTF_C);
//printMTX("C_BTF_AMD.mtx", BTF_C);
}
//6. Move to 2D Structure
//finds the shapes for both view and submatrices,
//need to be changed over to just submatrices
matrix_to_views_2D(BTF_A);
//finds the starting point of A for submatrices
find_2D_convert(BTF_A);
//now we can fill submatrices
#ifdef BASKER_KOKKOS
kokkos_order_init_2D<Int,Entry,Exe_Space> iO(this);
Kokkos::parallel_for(TeamPolicy(num_threads,1), iO);
Kokkos::fence();
#else
//Comeback
#endif
//printMTX("BTF_A.mtx", BTF_A);
}//if btf_tab_offset == 0
if(btf_nblks > 1)
{
sort_matrix(BTF_C);
//printMTX("C_TEST.mtx", BTF_C);
//Permute C
MALLOC_INT_1DARRAY(order_c_csym_array, BTF_C.ncol+1);
init_value(order_c_csym_array, BTF_C.ncol+1,(Int) 0);
printf("BEFORE \n");
//csymamd_order(BTF_C, order_c_csym_array, cmember);
blk_amd(BTF_C, order_c_csym_array);
printf("After perm\n");
permute_col(BTF_C, order_c_csym_array);
sort_matrix(BTF_C);
permute_row(BTF_C, order_c_csym_array);
sort_matrix(BTF_C);
if(btf_tabs_offset != 0)
{
permute_col(BTF_B, order_c_csym_array);
sort_matrix(BTF_B);
//printMTX("BTF_B.mtx", BTF_B);
}
//printMTX("BTF_C.mtx", BTF_C);
}
//printf("Done with ordering\n");
return 0;
}//end btf_order
BASKER_INLINE
int Basker<Int,Entry, Exe_Space>::find_btf2
(
BASKER_MATRIX &M
)
{
Int nblks = 0;
strong_component(M,nblks,order_btf_array,btf_tabs);
btf_nblks = nblks;
btf_flag = BASKER_TRUE;
//#ifdef BASKER_DEBUG_ORDER_BTF
printf("BTF nblks returned: %d \n", nblks);
//BASKER_ASSERT(nblks>1, "NOT ENOUGH BTF BLOCKS");
//#endif
#ifdef BASKER_DEBUG_ORDER_BTF
if(nblks<2)
{
printf("BTF did not find enough blks\n");
}
#endif
//#ifdef BASKER_DEBUG_ORDER_BTF
/*
printf("\nBTF perm: \n");
for(Int i=0; i <M.nrow; i++)
{
printf("%d, ", order_btf_array(i));
//printf("%d, ", btf_perm(i));
}
*/
printf("num_threads: %d \n", num_threads);
printf("\n\nBTF tabs: \n");
for(Int i=0; i < nblks+1; i++)
{
printf("%d, ", btf_tabs(i));
}
printf("\n");
// #endif
permute_col(M, order_btf_array);
permute_row(M, order_btf_array);
MALLOC_INT_1DARRAY(order_blk_amd_array, M.ncol);
init_value(order_blk_amd_array, M.ncol, (Int)0);
MALLOC_INT_1DARRAY(btf_blk_nnz, nblks+1);
init_value(btf_blk_nnz, nblks+1, (Int) 0);
MALLOC_INT_1DARRAY(btf_blk_work, nblks+1);
init_value(btf_blk_work, nblks+1, (Int) 0);
//Find AMD blk ordering, get nnz, and get work
btf_blk_amd( M, order_blk_amd_array,
btf_blk_nnz, btf_blk_work);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("blk_perm:\n");
for(Int i = 0; i < M.ncol; i++)
{
printf("(%d,%d) ", i, order_blk_amd_array(i));
}
printf("\n");
printf("id/blk_size/blk_nnz/work: \n");
for(Int i = 0; i < nblks; i++)
{
printf("(%d, %d, %d, %d) ", i,
btf_tabs(i+1)-btf_tabs(i),
btf_blk_nnz(i), btf_blk_work(i));
}
printf("\n");
#endif
//printMTX("A_BEFORE.mtx", M);
//printVec("AMD.txt", order_blk_amd_array, M.ncol);
permute_col(M, order_blk_amd_array);
permute_row(M, order_blk_amd_array);
sort_matrix(M);
//changed col to row, error.
//print to see issue
//printMTX("A_AMD.mtx", M);
break_into_parts2(M, nblks, btf_tabs);
//find schedule
find_btf_schedule(M, nblks, btf_tabs);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("------------BTF CUT: %d --------------\n",
btf_tabs(btf_tabs_offset));
#endif
return 0;
}//end find BTF(nnz)