BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::Factor(Int option)
{
#ifdef BASKER_KOKKOS_TIME
Kokkos::Impl::Timer timer;
#endif
factor_notoken(option);
#ifdef BASKER_KOKKOS_TIME
stats.time_nfactor += timer.seconds();
#endif
// NDE
MALLOC_ENTRY_1DARRAY(x_view_ptr_copy, gn); //used in basker_solve_rhs - move alloc
MALLOC_ENTRY_1DARRAY(y_view_ptr_copy, gm);
MALLOC_INT_1DARRAY(perm_inv_comp_array , gm); //y
MALLOC_INT_1DARRAY(perm_comp_array, gn); //x
MALLOC_INT_1DARRAY(perm_comp_iworkspace_array, gn);
MALLOC_ENTRY_1DARRAY(perm_comp_fworkspace_array, gn);
permute_composition_for_solve();
factor_flag = BASKER_TRUE;
return 0;
}//end Factor()
BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::permute
(
ENTRY_1DARRAY vec,
INT_1DARRAY p,
Int n
)
{
ENTRY_1DARRAY temp;
MALLOC_ENTRY_1DARRAY(temp, n);
init_value(temp, n, (Entry) 0.0);
//Permute
for(Int i = 0; i < n; i++)
{
temp(i) = vec(p(i));
//temp(p(i)) = vec(i);
}
//Copy back
for(Int i = 0; i < n; i++)
{
vec(i) = temp(i);
}
FREE_ENTRY_1DARRAY(temp);
return 0;
}
int Basker<Int,Entry,Exe_Space>::Factor_Inc(Int Options)
{
factor_inc_lvl(Options);
// NDE
MALLOC_ENTRY_1DARRAY(x_view_ptr_copy, gn); //used in basker_solve_rhs - move alloc
MALLOC_ENTRY_1DARRAY(y_view_ptr_copy, gm);
MALLOC_INT_1DARRAY(perm_inv_comp_array , gm); //y
MALLOC_INT_1DARRAY(perm_comp_array, gn); //x
MALLOC_INT_1DARRAY(perm_comp_iworkspace_array, gn);
MALLOC_ENTRY_1DARRAY(perm_comp_fworkspace_array, gn);
permute_composition_for_solve();
return 0;
}
BASKER_INLINE
void BaskerMatrix<Int, Entry, Exe_Space>::init_vectors
(
Int _m, Int _n, Int _nnz
)
{
nrow = _m;
ncol = _n;
nnz = _nnz;
mnnz = _nnz;
if(ncol >= 0)
{
BASKER_ASSERT((ncol+1)>0, "matrix init_vector ncol");
MALLOC_INT_1DARRAY(col_ptr,ncol+1);
}
if(nnz > 0)
{
BASKER_ASSERT(nnz > 0, "matrix init_vector nnz");
MALLOC_INT_1DARRAY(row_idx,nnz);
MALLOC_ENTRY_1DARRAY(val,nnz);
#ifdef BASKER_INC_LVL
MALLOC_INT_1DARRAY(inc_lvl, nnz);
#endif
}
else if(nnz==0)
{
BASKER_ASSERT((nnz+1)>0, "nnz+1 init_vector ");
MALLOC_INT_1DARRAY(row_idx, nnz+1);
row_idx(0) = (Int) 0;
MALLOC_ENTRY_1DARRAY(val, nnz+1);
val(0) = (Entry) 0;
#ifdef BASKER_INC_LVL
MALLOC_INT_1DARRAY(inc_lvl, nnz+1);
#endif
}
}//end init_vectors(Int, Int, Int)
BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::permute
(
ENTRY_1DARRAY vec,
INT_1DARRAY p,
Int n, //n = size(vec) //n > m
Int m, //m = size(p)
Int start
)
{
if(m > n)
{
printf("ERROR permtue \n");
return BASKER_ERROR;
}
ENTRY_1DARRAY temp;
MALLOC_ENTRY_1DARRAY(temp, n);
init_value(temp, n, (Entry) 0.0);
//Permute
for(Int i = 0; i < n; i++)
{
temp(i) = vec(p(i));
//temp(p(i)) = vec(i);
}
//Copy back
for(Int i = 0; i < n; i++)
{
vec(i) = temp(i);
}
FREE_ENTRY_1DARRAY(temp);
return 0;
}
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::test_solve()
{
ENTRY_1DARRAY x_known;
ENTRY_1DARRAY x;
ENTRY_1DARRAY y;
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("test_solve called \n");
printf("Global pivot permuation\n");
printVec(gperm, gn);
printf("\n");
printf("Global pivot permutation inverse\n");
printVec(gpermi, gn);
printf("\n");
#endif
BASKER_ASSERT(gn > 0, "solve testsolve gn");
MALLOC_ENTRY_1DARRAY(x_known, gn);
init_value(x_known, gn , (Entry)1.0);
//temp
for(Int i = 0; i < gn; i++)
{
//x_known(i) = (Entry)(i+1);
x_known(i) = (Entry) 1.0;
}
//JDB: used for other test
//permute(x_known, order_csym_array, gn);
MALLOC_ENTRY_1DARRAY(x, gn);
init_value(x, gn, (Entry) 0.0);
BASKER_ASSERT(gm > 0, "solve testsolve gm");
MALLOC_ENTRY_1DARRAY(y, gm);
init_value(y, gm, (Entry) 0.0);
if(btf_nblks > 0)
{
sort_matrix(BTF_C);
//printMTX("C_BEFORE_SOLVE.mtx", BTF_C);
}
if(Options.btf == BASKER_TRUE)
{
//printf("btf_tabs_offset: %d ", btf_tabs_offset);
//printf("btf_nblks: %d \n", btf_nblks);
if(btf_tabs_offset != 0)
{
//printf("BTF_A spmv\n");
spmv(BTF_A, x_known,y);
if(btf_nblks> 1)
{
//printf("btf_B spmv \n");
spmv(BTF_B, x_known, y);
}
}
if(btf_nblks > 1)
{
//printf("btf_c spmv \n");
spmv(BTF_C, x_known, y);
}
//return -1;
}
else
{
//printf("other\n");
//spmv(BTF_A, x_known,y);
}
//printf("\n Before Test Points \n");
//printf("i: %d x: %f y: %f \n", 0, x_known(0), y(0));
//if(gn > 24)
// {
// printf("i: %d x: %f y: %f \n", 24, x_known(24), y(24));
// }
//pivot permuation
//printVec("gperm.csc", gpermi, gn);
for(Int i = 0; i < gn; i++)
{
x(gpermi(i)) = y(i);
}
for(Int i = 0; i < gn; i++)
{
y(i) = x(i);
x(i) = 0;
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("\n\n");
//printf("Known Solution: \n");
//for(Int i = 0; i < gn; i++)
// {
// printf("%f, " , x_known(i));
// }
printf("\n\n");
printf("RHS: \n");
for(Int i =0; i < gm; i++)
{
printf("%d %f,\n ", i, y(i));
}
printf("\n\n");
#endif
if(Options.btf == BASKER_FALSE)
{
//printf("before serial solve\n");
if(btf_tabs_offset != 0)
{
serial_solve(y,x);
}
//printf("After serial solve\n");
//printf("i: %d x: %f y: %f \n", 0, x(0), y(0));
//printf("i: %d x: %f y: %f \n", 24, x(24), y(24));
}
else
{
//A\y -> y
//serial_btf_solve(y,x);
//printf("before btf serial solve\n");
serial_btf_solve(y,x);
//printf("After btf solve\n");
//printf("i: %d x: %f y: %f \n", 0, x(0), y(0));
//printf("i: %d x: %f y: %f \n", 24, x(24), y(24));
}
Entry diff =0.0;
for(Int i = 0; i < gn; i++)
{
diff += (x_known(i) - x(i));
}
diff = diff/(Entry) gn;
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("\n\n");
printf("Solve Compare: \n");
for(Int i = 0; i < gn; i++)
{
printf("%d %f %f \n",
i, x_known(i), x(i));
}
printf("\n\n");
#endif
printf("\n Test Points \n");
printf("i: %d x: %f %f \n", 0, x_known(0), x(0));
if(gn > 24)
{
printf("i: %d x: %f %f \n", 10, x_known(10), x(10));
printf("i: %d x: %f %f \n", 24, x_known(24), x(24));
}
printf("\n");
printf("TEST_SOLVE: ||x-x||/||x| = %e", diff);
printf("\n");
if((diff > -1e-2) && (diff < 1e-2))
{
printf("TEST PASSED \n");
}
return 0;
}//end test_solve
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::solve_interface
(
Entry *_x, //Solution (len = gn)
Entry *_y
)
{
//===== Move to view===========
ENTRY_1DARRAY x;
ENTRY_1DARRAY y;
MALLOC_ENTRY_1DARRAY(x, gn);
MALLOC_ENTRY_1DARRAY(y, gm);
for(Int i =0; i < gn; i++)
{
x(i) = (Entry) 0;
y(i) = (Entry) _y[i];
}
//printf("RHS: \n");
//printVec(y, gn);
//printf("\n");
//===== Permute
//printf("Permute RHS\n");
//==== Need to make this into one global perm
if(match_flag == BASKER_TRUE)
{
//printf("match order\n");
//printVec("match.txt", order_match_array, gn);
permute_inv(y,order_match_array, gn);
}
if(btf_flag == BASKER_TRUE)
{
//printf("btf order\n");
//printVec("btf.txt", order_btf_array, gn);
permute_inv(y,order_btf_array, gn);
//printVec("btf_amd.txt", order_c_csym_array, gn);
permute_inv(y,order_blk_amd_array, gn);
}
if(nd_flag == BASKER_TRUE)
{
//printf("ND order \n");
//printVec("nd.txt", part_tree.permtab, gn);
permute_inv(y,part_tree.permtab, gn);
}
if(amd_flag == BASKER_TRUE)
{
//printf("AMD order \n");
//printVec("amd.txt",order_csym_array, gn);
permute_inv(y,order_csym_array, gn);
}
//printVec("perm.txt" , gperm, gn);
permute_inv(y,gperm, gn);
solve_interface(x,y);
//Inverse perm
//Note: don't need to inverse a row only perm
if(btf_flag == BASKER_TRUE)
{
//printf("btf order\n");
//printVec(order_btf_array, gn);
permute(x,order_btf_array, gn);
}
if(nd_flag == BASKER_TRUE)
{
//printf("ND order \n");
//printVec(part_tree.permtab, gn);
permute(x,part_tree.permtab, gn);
}
if(amd_flag == BASKER_TRUE)
{
//printf("AMD order \n");
//printVec(order_csym_array, gn);
//FILE *fpamd;
//fpamd = fopen("amd.csv", "w");
//for(Int i = 0; i < gn; i++)
// {
// fprintf(fpamd, "%d \n",
// order_csym_array(i));
// }
//fclose(fpamd);
permute(x,order_csym_array, gn);
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("\n\n");
printf("X: \n");
for(Int i = 0; i < gn; i++)
{
printf("%f, " , x(i));
}
printf("\n\n");
printf("RHS: \n");
for(Int i =0; i < gm; i++)
{
printf("%f, ", y(i));
}
printf("\n\n");
#endif
for(Int i = 0; i < gn; i++)
{
_x[i] = x(i);
}
#ifndef BASKER_KOKKOS
FREE_ENTRY_1DARRAY(x);
FREE_ENTRY_1DARRAY(y);
#endif
return 0;
}
BASKER_INLINE
int Basker<Int, Entry, Exe_Space>::permute_col
(
BASKER_MATRIX &M,
INT_1DARRAY col
)
{
if((M.ncol == 0)||(M.nnz == 0))
return 0;
Int n = M.ncol;
Int nnz = M.nnz;
//printf("Using n: %d nnz: %d \n", n, nnz);
INT_1DARRAY temp_p;
MALLOC_INT_1DARRAY(temp_p, n+1);
init_value(temp_p, n+1, (Int)0);
INT_1DARRAY temp_i;
MALLOC_INT_1DARRAY(temp_i, nnz);
init_value(temp_i, nnz, (Int)0);
ENTRY_1DARRAY temp_v;
MALLOC_ENTRY_1DARRAY(temp_v, nnz);
init_value(temp_v, nnz, (Entry)0.0);
//printf("done with init \n");
//Determine column ptr of output matrix
for(Int j = 0; j < n; j++)
{
Int i = col (j);
temp_p (i+1) = M.col_ptr (j+1) - M.col_ptr (j);
}
//Get ptrs from lengths
temp_p (0) = 0;
for(Int j = 0; j < n; j++)
{
temp_p (j+1) = temp_p (j+1) + temp_p (j);
}
//copy idxs
for(Int ii = 0; ii < n; ii++)
{
Int ko = temp_p (col (ii) );
for(Int k = M.col_ptr (ii); k < M.col_ptr (ii+1); k++)
{
temp_i (ko) = M.row_idx (k);
temp_v (ko) = M.val (k);
ko++;
}
}
//copy back int A
for(Int ii=0; ii < n+1; ii++)
{
M.col_ptr (ii) = temp_p (ii);
}
for(Int ii=0; ii < nnz; ii++)
{
M.row_idx (ii) = temp_i (ii);
M.val (ii) = temp_v (ii);
}
FREE_INT_1DARRAY(temp_p);
FREE_INT_1DARRAY(temp_i);
FREE_ENTRY_1DARRAY(temp_v);
return 0;
}//end permute_col(int)
BASKER_INLINE
void BaskerMatrix<Int,Entry,Exe_Space>::convert2D
(
BASKER_MATRIX &M,
BASKER_BOOL alloc,
Int kid
)
{
if(nnz == 0)
{
for(Int i = 0; i < ncol+1; i++)
{
col_ptr(i) = 0;
}
MALLOC_INT_1DARRAY(row_idx, 1);
row_idx(0) = (Int) 0;
MALLOC_ENTRY_1DARRAY(val, 1);
val(0) = (Entry) 0;
return;
}
//info();
//We could check some flag ??
//We assume a pre-scan has already happened
if(alloc == BASKER_TRUE)
{
//printf("ALLOC\n");
if(nnz > 0)
{
BASKER_ASSERT(nnz > 0, "matrix row nnz 2");
MALLOC_INT_1DARRAY(row_idx, nnz);
}
else if(nnz ==0)
{
BASKER_ASSERT((nnz+1)>0, "matrix row nnz 3");
MALLOC_INT_1DARRAY(row_idx, nnz+1);
}
}
//init_value(row_idx, nnz, (Int) 0);
//printf("clear row: %d \n", nnz);
for(Int i = 0; i < nnz; ++i)
{
//printf("clear row_idx(%d) \n", i);
row_idx(i) = 0;
}
if(alloc == BASKER_TRUE)
{
if(nnz > 0)
{
BASKER_ASSERT(nnz > 0, "matrix nnz 4");
MALLOC_ENTRY_1DARRAY(val, nnz);
}
else if(nnz == 0)
{
BASKER_ASSERT((nnz+1) > 0, "matrix nnz 5");
MALLOC_ENTRY_1DARRAY(val, nnz+1);
}
}
//init_value(val, nnz, (Entry) 0);
for(Int i = 0; i < nnz; ++i)
{
val(i) = 0;
}
Int temp_count = 0;
for(Int k = scol; k < scol+ncol; ++k)
{
//note col_ptr[k-scol] contains the starting index
if(col_ptr(k-scol) == BASKER_MAX_IDX)
{
col_ptr(k-scol) = temp_count;
//printf("continue called, k: %d \n", k);
continue;
}
for(Int i = col_ptr(k-scol); i < M.col_ptr(k+1); i++)
{
Int j = M.row_idx(i);
//printf("i: %d j:%d \n", i,j);
if(j >= srow+nrow)
{
break;
}
//printf("writing row_dix: %d i: %d val: %d nnz: %d srow: %d nrow: %d \n",
// temp_count, i, j, nnz,
// srow, nrow);
//BASKER_ASSERT(temp_count < nnz, "2DConvert, too many values");
//row_idx[temp_count] = j;
if(j-srow <0)
{
std::cout << "kid: " << kid
<< " j: " << j
<< " srow: " << srow
<< " k: " << k
<< " idx: " << i
<< std::endl;
BASKER_ASSERT(0==1, "j-srow NO");
}
row_idx(temp_count) = j-srow;
val(temp_count) = M.val(i);
temp_count++;
}
col_ptr(k-scol) = temp_count;
}
//col_ptr[0] = 0;
//NO!!1
//Slide over the col counts
for(Int i = ncol; i > 0; i--)
{
col_ptr(i) = col_ptr(i-1);
}
col_ptr(0) = (Int) 0;
//info();
//print();
}//end convert2d(Matrix)
BASKER_INLINE
int Basker<Int, Entry,Exe_Space>::break_into_parts2
(
BASKER_MATRIX &M,
Int nblks,
INT_1DARRAY btf_tabs
)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("break_into_parts2 called \n");
printf("nblks: %d \n", nblks);
#endif
Options.btf = BASKER_TRUE;
//Alg.
// A -> [BTF_A BTF_B]
// [0 BTF_C]
//1. Run backward through the btf_tabs to find size C
//2. Form A,B,C based on size in 1.
//Short circuit,
//If nblks == 1, than only BTF_A exists
if(nblks == 1)
{
//#ifdef BASKER_DEBUG_ORDER_BTF
printf("Short Circuit part_call \n");
//#endif
BTF_A = A;
//Options.btf = BASKER_FALSE;
btf_tabs_offset = 1;
return 0;
}
//Step 1.
//Find total work estimate
Int total_work_estimate = 0;
for(Int b = 0; b < nblks; b++)
{
total_work_estimate += btf_blk_work(b);
}
//Set a class variable to use later
btf_total_work = total_work_estimate;
//printf("Total work estimate: %d \n",
// total_work_estimate);
//printf("num_threads: %d epsilon: %f \n",
// num_threads,
// ((double)1/num_threads) +
// ((double)BASKER_BTF_IMBALANCE));
Int break_size = ceil((double)total_work_estimate*(
((double)1/num_threads) +
((double)BASKER_BTF_IMBALANCE)));
printf("Break size: %d \n", break_size);
Int t_size = 0;
Int scol = M.ncol;
Int blk_idx = nblks;
BASKER_BOOL move_fwd = BASKER_TRUE;
while(move_fwd==BASKER_TRUE)
{
//printf("------TEST blk_idx: %d \n",
// blk_idx);
Int blk_work = btf_blk_work(blk_idx-1);
Int blk_size = btf_tabs(blk_idx) -
btf_tabs(blk_idx-1);
#ifdef BASKER_DEBUG_ORDER_BTF
printf(" \n move_fwd loop \n");
BASKER_ASSERT(blk_idx>=0, "btf blk idx off");
BASKER_ASSERT(blk_work>=0, "btk_work wrong");
BASKER_ASSERT(blk_size>0, "btf blk size wrong");
printf("blk_idx: %d blk_work: %d break_size: %d \n",
blk_idx, blk_work, break_size);
#endif
//Should be end
//if(((blk_work < break_size) ||
// (blk_size < BASKER_BTF_SMALL)) &&
// (blk_idx > 1))
//Continue to be in btf
if(((blk_work < break_size) &&
(blk_idx > 1)))
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("first choice \n");
#endif
t_size = t_size+blk_size;
blk_idx = blk_idx-1;
scol = btf_tabs[blk_idx];
}
//break due to size
else if(blk_work >= break_size)
{
printf("break due to size\n");
move_fwd = BASKER_FALSE;
}
//break due to end
else if(blk_idx == 1)
{
printf("break last blk\n");
blk_idx = 0;
t_size = t_size + blk_size;
scol = btf_tabs[blk_idx];
move_fwd = BASKER_FALSE;
}
//should not be called
else
{
BASKER_ASSERT(1==0, "btf order break");
move_fwd = BASKER_FALSE;
}
}//end while(move_fwd)
//#ifdef BASKER_DEBUG_ORDER_BTF
printf("Done finding BTF2 cut. Cut size: %d scol: %d \n",
t_size, scol);
printf("Done finding BTF2 cut. blk_idx: %d \n",
blk_idx);
//BASKER_ASSERT(t_size > 0, "BTF CUT SIZE NOT BIG ENOUGH\n");
BASKER_ASSERT((scol >= 0) && (scol < M.ncol), "SCOL\n");
//#endif
//Comeback and change
btf_tabs_offset = blk_idx;
//2. Move into Blocks
if(btf_tabs_offset != 0)
{
//--Move A into BTF_A;
BTF_A.set_shape(0, scol, 0, scol);
BTF_A.nnz = M.col_ptr(scol);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Init BTF_A. ncol: %d nnz: %d \n",
scol, BTF_A.nnz);
#endif
if(BTF_A.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_A.ncol >= 0, "BTF_A, col_ptr");
MALLOC_INT_1DARRAY(BTF_A.col_ptr, BTF_A.ncol+1);
BASKER_ASSERT(BTF_A.nnz > 0, "BTF_A, nnz");
MALLOC_INT_1DARRAY(BTF_A.row_idx, BTF_A.nnz);
MALLOC_ENTRY_1DARRAY(BTF_A.val, BTF_A.nnz);
BTF_A.fill();
}
Int annz = 0;
for(Int k = 0; k < scol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("copy column: %d into A_BTF, [%d %d] \n",
k, M.col_ptr(k), M.col_ptr(k+1));
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
//printf("annz: %d i: %d \n", annz, i);
BTF_A.row_idx(annz) = M.row_idx(i);
BTF_A.val(annz) = M.val(i);
annz++;
}
BTF_A.col_ptr(k+1) = annz;
}
}//no A
//Fill in B and C at the same time
INT_1DARRAY cws;
BASKER_ASSERT((M.ncol-scol+1) > 0, "BTF_SIZE MALLOC");
MALLOC_INT_1DARRAY(cws, M.ncol-scol+1);
init_value(cws, M.ncol-scol+1, (Int)M.ncol);
BTF_B.set_shape(0 , scol,
scol, M.ncol-scol);
BTF_C.set_shape(scol, M.ncol-scol,
scol, M.ncol-scol);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Set Shape BTF_B: %d %d %d %d \n",
BTF_B.srow, BTF_B.nrow,
BTF_B.scol, BTF_B.ncol);
printf("Set Shape BTF_C: %d %d %d %d \n",
BTF_C.srow, BTF_C.nrow,
BTF_C.scol, BTF_C.nrow);
#endif
//Scan and find nnz
//We can do this much better!!!!
Int bnnz = 0;
Int cnnz = 0;
for(Int k = scol; k < M.ncol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Scanning nnz, k: %d \n", k);
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
if(M.row_idx(i) < scol)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Upper, %d %d \n",
scol, M.row_idx(i));
#endif
bnnz++;
}
else
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Lower, %d %d \n",
scol, M.row_idx(i));
#endif
cnnz++;
}
}//over all nnz in k
}//over all k
#ifdef BASKER_DEBUG_ORDER_BTF
printf("BTF_B nnz: %d \n", bnnz);
printf("BTF_C nnz: %d \n", cnnz);
#endif
BTF_B.nnz = bnnz;
BTF_C.nnz = cnnz;
//Malloc need space
if((BTF_B.v_fill == BASKER_FALSE) &&
(BTF_B.nnz > 0))
//if(BTF_B.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_B.ncol >= 0, "BTF_B ncol");
MALLOC_INT_1DARRAY(BTF_B.col_ptr, BTF_B.ncol+1);
BASKER_ASSERT(BTF_B.nnz > 0, "BTF_B.nnz");
MALLOC_INT_1DARRAY(BTF_B.row_idx, BTF_B.nnz);
MALLOC_ENTRY_1DARRAY(BTF_B.val, BTF_B.nnz);
BTF_B.fill();
}
if(BTF_C.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_C.ncol >= 0, "BTF_C.ncol");
MALLOC_INT_1DARRAY(BTF_C.col_ptr, BTF_C.ncol+1);
BASKER_ASSERT(BTF_C.nnz > 0, "BTF_C.nnz");
MALLOC_INT_1DARRAY(BTF_C.row_idx, BTF_C.nnz);
MALLOC_ENTRY_1DARRAY(BTF_C.val, BTF_C.nnz);
BTF_C.fill();
}
//scan again (Very bad!!!)
bnnz = 0;
cnnz = 0;
for(Int k = scol; k < M.ncol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Scanning nnz, k: %d \n", k);
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
if(M.row_idx(i) < scol)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Upper, %d %d \n",
scol, M.row_idx[i]);
#endif
BASKER_ASSERT(BTF_B.nnz > 0, "BTF B uninit");
//BTF_B.row_idx[bnnz] = M.row_idx[i];
//Note: do not offset because B srow = 0
BTF_B.row_idx(bnnz) = M.row_idx(i);
BTF_B.val(bnnz) = M.val(i);
bnnz++;
}
else
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz Lower,k: %d %d %d %f \n",
k, scol, M.row_idx[i],
M.val(i));
#endif
//BTF_C.row_idx[cnnz] = M.row_idx[i];
BTF_C.row_idx(cnnz) = M.row_idx(i)-scol;
BTF_C.val(cnnz) = M.val(i);
cnnz++;
}
}//over all nnz in k
if(BTF_B.nnz > 0)
{
BTF_B.col_ptr(k-scol+1) = bnnz;
}
BTF_C.col_ptr(k-scol+1) = cnnz;
}//over all k
#ifdef BASKER_DEBUG_ORDER_BTF
printf("After BTF_B nnz: %d \n", bnnz);
printf("After BTF_C nnz: %d \n", cnnz);
#endif
//printf("\n\n");
//printf("DEBUG\n");
//BTF_C.print();
//printf("\n\n");
return 0;
}//end break_into_parts2 (based on imbalance)
BASKER_INLINE
int Basker<Int, Entry,Exe_Space>::break_into_parts
(
BASKER_MATRIX &M,
Int nblks,
INT_1DARRAY btf_tabs
)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("break_into_parts called \n");
printf("nblks: %d \n", nblks);
#endif
Options.btf = BASKER_TRUE;
//Alg.
// A -> [BTF_A BTF_B]
// [0 BTF_C]
//1. Run backward through the btf_tabs to find size C
//2. Form A,B,C based on size in 1.
//Short circuit,
//If nblks == 1, than only BTF_A exists
if(nblks == 1)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Short Circuit part_call \n");
#endif
BTF_A = A;
//Options.btf = BASKER_FALSE;
btf_tabs_offset = 1;
return 0;
}
//Step 1.
Int t_size = 0;
Int scol = M.ncol;
Int blk_idx = nblks;
BASKER_BOOL move_fwd = BASKER_TRUE;
while(move_fwd==BASKER_TRUE)
{
Int blk_size = btf_tabs(blk_idx)-
btf_tabs(blk_idx-1);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("move_fwd loop \n");
BASKER_ASSERT(blk_idx>=0, "btf blk idx off");
BASKER_ASSERT(blk_size>0, "btf blk size wrong");
printf("blk_idx: %d blk_size: %d \n",
blk_idx, blk_size);
std::cout << blk_size << std::endl;
#endif
if((blk_size < Options.btf_large) &&
((((double)t_size+blk_size)/(double)M.ncol) < Options.btf_max_percent))
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("first choice \n");
printf("blksize test: %d %d %d \n",
blk_size, Options.btf_large,
BASKER_BTF_LARGE);
printf("blkpercent test: %f %f %f \n",
((double)t_size+blk_size)/(double)M.ncol,
Options.btf_max_percent,
(double) BASKER_BTF_MAX_PERCENT);
#endif
t_size = t_size+blk_size;
blk_idx = blk_idx-1;
scol = btf_tabs[blk_idx];
}
else
{
//printf("second choice \n");
//#ifdef BASKER_DEBUG_ORDER_BTF
printf("Cut: blk_size: %d percent: %f \n",
blk_size, ((double)t_size+blk_size)/(double)M.ncol);
if((((double)t_size+blk_size)/(double)M.ncol)
== 1.0)
{
blk_idx = 0;
t_size = t_size + blk_size;
scol = btf_tabs[blk_idx];
}
//#endif
move_fwd = BASKER_FALSE;
}
}//end while(move_fwd)
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Done finding BTF cut. Cut size: %d scol: %d \n",
t_size, scol);
//BASKER_ASSERT(t_size > 0, "BTF CUT SIZE NOT BIG ENOUGH\n");
BASKER_ASSERT((scol >= 0) && (scol < M.ncol), "SCOL\n");
#endif
//Comeback and change
btf_tabs_offset = blk_idx;
//2. Move into Blocks
if(btf_tabs_offset != 0)
{
//--Move A into BTF_A;
BTF_A.set_shape(0, scol, 0, scol);
BTF_A.nnz = M.col_ptr(scol);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Init BTF_A. ncol: %d nnz: %d \n",
scol, BTF_A.nnz);
#endif
if(BTF_A.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_A.ncol >= 0, "BTF_A, col_ptr");
MALLOC_INT_1DARRAY(BTF_A.col_ptr, BTF_A.ncol+1);
BASKER_ASSERT(BTF_A.nnz > 0, "BTF_A, nnz");
MALLOC_INT_1DARRAY(BTF_A.row_idx, BTF_A.nnz);
MALLOC_ENTRY_1DARRAY(BTF_A.val, BTF_A.nnz);
BTF_A.fill();
}
Int annz = 0;
for(Int k = 0; k < scol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("copy column: %d into A_BTF, [%d %d] \n",
k, M.col_ptr(k), M.col_ptr(k+1));
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
//printf("annz: %d i: %d \n", annz, i);
BTF_A.row_idx(annz) = M.row_idx(i);
BTF_A.val(annz) = M.val(i);
annz++;
}
BTF_A.col_ptr(k+1) = annz;
}
}//no A
//Fill in B and C at the same time
INT_1DARRAY cws;
BASKER_ASSERT((M.ncol-scol+1) > 0, "BTF_SIZE MALLOC");
MALLOC_INT_1DARRAY(cws, M.ncol-scol+1);
init_value(cws, M.ncol-scol+1, (Int)M.ncol);
BTF_B.set_shape(0 , scol,
scol, M.ncol-scol);
BTF_C.set_shape(scol, M.ncol-scol,
scol, M.ncol-scol);
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Set Shape BTF_B: %d %d %d %d \n",
BTF_B.srow, BTF_B.nrow,
BTF_B.scol, BTF_B.ncol);
printf("Set Shape BTF_C: %d %d %d %d \n",
BTF_C.srow, BTF_C.nrow,
BTF_C.scol, BTF_C.nrow);
#endif
//Scan and find nnz
//We can do this much better!!!!
Int bnnz = 0;
Int cnnz = 0;
for(Int k = scol; k < M.ncol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Scanning nnz, k: %d \n", k);
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
if(M.row_idx(i) < scol)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Upper, %d %d \n",
scol, M.row_idx(i));
#endif
bnnz++;
}
else
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Lower, %d %d \n",
scol, M.row_idx(i));
#endif
cnnz++;
}
}//over all nnz in k
}//over all k
#ifdef BASKER_DEBUG_ORDER_BTF
printf("BTF_B nnz: %d \n", bnnz);
printf("BTF_C nnz: %d \n", cnnz);
#endif
BTF_B.nnz = bnnz;
BTF_C.nnz = cnnz;
//Malloc need space
if((BTF_B.v_fill == BASKER_FALSE) &&
(BTF_B.nnz > 0))
//if(BTF_B.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_B.ncol >= 0, "BTF_B ncol");
MALLOC_INT_1DARRAY(BTF_B.col_ptr, BTF_B.ncol+1);
BASKER_ASSERT(BTF_B.nnz > 0, "BTF_B.nnz");
MALLOC_INT_1DARRAY(BTF_B.row_idx, BTF_B.nnz);
MALLOC_ENTRY_1DARRAY(BTF_B.val, BTF_B.nnz);
BTF_B.fill();
}
if(BTF_C.v_fill == BASKER_FALSE)
{
BASKER_ASSERT(BTF_C.ncol >= 0, "BTF_C.ncol");
MALLOC_INT_1DARRAY(BTF_C.col_ptr, BTF_C.ncol+1);
BASKER_ASSERT(BTF_C.nnz > 0, "BTF_C.nnz");
MALLOC_INT_1DARRAY(BTF_C.row_idx, BTF_C.nnz);
MALLOC_ENTRY_1DARRAY(BTF_C.val, BTF_C.nnz);
BTF_C.fill();
}
//scan again (Very bad!!!)
bnnz = 0;
cnnz = 0;
for(Int k = scol; k < M.ncol; ++k)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Scanning nnz, k: %d \n", k);
#endif
for(Int i = M.col_ptr(k); i < M.col_ptr(k+1); ++i)
{
if(M.row_idx(i) < scol)
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz to Upper, %d %d \n",
scol, M.row_idx[i]);
#endif
BASKER_ASSERT(BTF_B.nnz > 0, "BTF B uninit");
//BTF_B.row_idx[bnnz] = M.row_idx[i];
//Note: do not offset because B srow = 0
BTF_B.row_idx(bnnz) = M.row_idx(i);
BTF_B.val(bnnz) = M.val(i);
bnnz++;
}
else
{
#ifdef BASKER_DEBUG_ORDER_BTF
printf("Adding nnz Lower,k: %d %d %d %f \n",
k, scol, M.row_idx[i],
M.val(i));
#endif
//BTF_C.row_idx[cnnz] = M.row_idx[i];
BTF_C.row_idx(cnnz) = M.row_idx(i)-scol;
BTF_C.val(cnnz) = M.val(i);
cnnz++;
}
}//over all nnz in k
if(BTF_B.nnz > 0)
{
BTF_B.col_ptr(k-scol+1) = bnnz;
}
BTF_C.col_ptr(k-scol+1) = cnnz;
}//over all k
#ifdef BASKER_DEBUG_ORDER_BTF
printf("After BTF_B nnz: %d \n", bnnz);
printf("After BTF_C nnz: %d \n", cnnz);
#endif
//printf("\n\n");
//printf("DEBUG\n");
//BTF_C.print();
//printf("\n\n");
return 0;
}//end break_into_parts
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::Factor(Int nrow, Int ncol,
Int nnz, Int *col_ptr, Int *row_idx, Entry *val)
{
int err = 0;
if (Options.verbose == BASKER_TRUE)
{
std::cout << "Basker Factor Called" << std::endl;
std::cout << "Matrix: " << nrow << " " << ncol << " " << nnz << std::endl;
}
/*
int err = A.copy_values(nrow, ncol, nnz, col_ptr,
row_idx, val);
*/
if((Options.same_pattern == BASKER_TRUE) && (Options.no_pivot == BASKER_FALSE))
{
printf("Warning: Same Pattern will not allow pivoting\n");
Options.no_pivot = BASKER_TRUE;
}
if(Options.transpose == BASKER_FALSE)
{
//printf("=======NO TRANS=====\n");
//A.init_matrix("Original Matrix",
// nrow, ncol, nnz, col_ptr, row_idx, val);
//A.scol = 0;
//A.srow = 0;
A.copy_values(nrow, ncol, nnz, col_ptr,
row_idx, val);
//printf("Copy done\n");
//printMTX("A_LOAD.mtx", A);
}
else
{
//printf("======TRANS=====\n");
//Will transpose and put in A using little extra
matrix_transpose(0,
nrow,
0,
ncol,
nnz,
col_ptr,
row_idx,
val,
A);
}
sort_matrix(A);
if(Options.verbose_matrix_out == BASKER_TRUE)
{
printMTX("A_Factor.mtx", A);
}
matrix_flag = BASKER_TRUE;
if(err == BASKER_ERROR)
{
return BASKER_ERROR;
}
//err = sfactor_copy();
err = sfactor_copy2();
if (Options.verbose == BASKER_TRUE)
{
printf("Basker Copy Structure Done \n");
}
//printf("Done with sfactor_copy: %d \n", err);
if(err == BASKER_ERROR)
{
return BASKER_ERROR;
}
//printf("before notoken\n");
//Kokkos::Impl::Timer timer;
if(Options.incomplete == BASKER_FALSE)
{
err = factor_notoken(0);
//printf("Notoken called\n");
}
else
{
err = factor_inc_lvl(0);
}
if(err == BASKER_ERROR)
{
return BASKER_ERROR;
}
if(Options.verbose == BASKER_TRUE)
{
printf("Basker Factor Done \n");
}
/*
std::cout << "Raw Factor Time: "
<< timer.seconds()
<< std::endl;
*/
//DEBUG_PRINT();
// NDE
MALLOC_ENTRY_1DARRAY(x_view_ptr_copy, gn); //used in basker_solve_rhs - move alloc
MALLOC_ENTRY_1DARRAY(y_view_ptr_copy, gm);
MALLOC_INT_1DARRAY(perm_inv_comp_array , gm); //y
MALLOC_INT_1DARRAY(perm_comp_array, gn); //x
MALLOC_INT_1DARRAY(perm_comp_iworkspace_array, gn);
MALLOC_ENTRY_1DARRAY(perm_comp_fworkspace_array, gn);
permute_composition_for_solve();
factor_flag = BASKER_TRUE;
return 0;
}//end Factor()
BASKER_INLINE
int Basker<Int,Entry,Exe_Space>::solve_interface
(
Entry *_x, //Solution (len = gn)
Entry *_y
)
{
//Need to modify to use global perm
INT_1DARRAY temp_array;
MALLOC_INT_1DARRAY(temp_array, gn);
//===== Move to view===========
ENTRY_1DARRAY x;
ENTRY_1DARRAY y;
MALLOC_ENTRY_1DARRAY(x, gn);
MALLOC_ENTRY_1DARRAY(y, gm);
for(Int i =0; i < gn; i++)
{
x(i) = (Entry) 0;
y(i) = (Entry) _y[i];
}
//printf("RHS: \n");
//printVec(y, gn);
//printf("\n");
//===== Permute
//printf("Permute RHS\n");
//==== Need to make this into one global perm
if(match_flag == BASKER_TRUE)
{
//printf("match order\n");
//printVec("match.txt", order_match_array, gn);
permute_inv(y,order_match_array, gn);
}
if(btf_flag == BASKER_TRUE)
{
//printf("btf order\n");
//printVec("btf.txt", order_btf_array, gn);
permute_inv(y,order_btf_array, gn);
//printVec("btf_amd.txt", order_c_csym_array, gn);
permute_inv(y,order_blk_amd_array, gn);
}
if(nd_flag == BASKER_TRUE)
{
//printf("ND order \n");
//printVec("nd.txt", part_tree.permtab, gn);
for(Int i = 0; i < BTF_A.ncol; ++i)
{
temp_array(i) = part_tree.permtab(i);
}
for(Int i = BTF_A.ncol; i < gn; ++i)
{
temp_array(i) = i;
}
//permute_inv(y,part_tree.permtab, gn);
permute_inv(y, temp_array,gn);
}
if(amd_flag == BASKER_TRUE)
{
//printf("AMD order \n");
//printVec("amd.txt",order_csym_array, gn);
for(Int i = 0; i < BTF_A.ncol; ++i)
{
temp_array(i) = order_csym_array(i);
}
for(Int i = BTF_A.ncol; i < gn; ++i)
{
temp_array(i) = i;
}
//permute_inv(y,order_csym_array, gn);
permute_inv(y,temp_array, gn);
}
//printVec("perm.txt" , gperm, gn);
permute_inv(y,gperm, gn);
solve_interface(x,y);
//Inverse perm
//Note: don't need to inverse a row only perm
if(btf_flag == BASKER_TRUE)
{
//printf("btf order\n");
//printVec(order_btf_array, gn);
permute(x,order_btf_array, gn);
}
if(nd_flag == BASKER_TRUE)
{
//printf("ND order \n");
//printVec(part_tree.permtab, gn);
for(Int i = 0; i < BTF_A.ncol; ++i)
{
temp_array(i) = part_tree.permtab(i);
}
for(Int i = BTF_A.ncol; i < gn; i++)
{
temp_array(i) = i;
}
//permute(x,part_tree.permtab, gn);
permute(x,temp_array, gn);
}
if(amd_flag == BASKER_TRUE)
{
//printf("AMD order \n");
//printVec(order_csym_array, gn);
for(Int i = 0; i < BTF_A.ncol; ++i)
{
temp_array(i) = order_csym_array(i);
}
for(Int i = BTF_A.ncol; i < gn; ++i)
{
temp_array(i) = order_csym_array(i);
}
//permute(x,order_csym_array, gn);
permute(x,temp_array,gn);
}
#ifdef BASKER_DEBUG_SOLVE_RHS
printf("\n\n");
printf("X: \n");
for(Int i = 0; i < gn; i++)
{
printf("%f, " , x(i));
}
printf("\n\n");
printf("RHS: \n");
for(Int i =0; i < gm; i++)
{
printf("%f, ", y(i));
}
printf("\n\n");
#endif
for(Int i = 0; i < gn; i++)
{
_x[i] = x(i);
}
FREE_ENTRY_1DARRAY(x);
FREE_ENTRY_1DARRAY(y);
FREE_INT_1DARRAY(temp_array);
return 0;
}
