int main(){
TYPE nzval[NNZ];
TYPE x[N];
TYPE y[N];
int colind[NNZ];
int rowptr[N+1];
int i;
srand48(8650341L);
fillVal(nzval);
fill(x);
initMat(colind, rowptr);
initOut(y);
spmv(nzval, colind, rowptr, x, y);
printf("\n");
for(i = 0; i < N; i++){
printf("%d ", y[i]);
}
printf("\n");
return 0;
}
void
spmv(const char mode[],
const AlphaType& alpha,
const AMatrix& A,
const XVector& x,
const BetaType& beta,
const YVector& y) {
typedef typename Kokkos::Impl::if_c<XVector::rank==2,RANK_TWO,RANK_ONE>::type RANK_SPECIALISE;
spmv(mode,alpha,A,x,beta,y,RANK_SPECIALISE());
}
virtual void solve(int steps = 100000, T tolerance = (T)1e-6){
cgfassert(this->mat->getWidth() == this->b->getSize());
cgfassert(this->mat->getWidth() == this->mat->getHeight());
cgfassert(this->mat->getWidth() == this->x->getSize());
this->iterations = 0;
for(int i=0;i<this->mat->getHeight();i++){
(*C)[i] = (T)1.0/Sqrt(Fabs((*this->mat)[i][i]));
}
/*r = Ax*/
spmv(*r, *(this->mat), *(this->x));
/*r = b - r*/;
Vector<T>::sub(*r, *(this->b), *r);
/*w = C * r*/
Vector<T>::mul(*w, *C, *r);
/*v = C * w*/
Vector<T>::mul(*v, *C, *w);
/*s1 = w * w*/
Vector<T>::mul(*scratch1, *w, *w);
T alpha;
/*alpha = sum(s1)*/
alpha = scratch1->sum();
int k=0;
while(k<steps){
/*s1 = v * v*/
this->iterations++;
Vector<T>::mul(*scratch1, *v, *v);
T residual;
residual = scratch1->sum();
if(Sqrt(Fabs(residual)) < (tolerance*bnorm /*+ tolerance*/)){
warning("CG::Success in %d iterations, %10.10e, %10.10e", k, residual, Sqrt(residual));
return;
}
/*u = A*v*/
spmv(*u, *(this->mat), *v);
/*s1 = v * u*/
Vector<T>::mul(*scratch1, *v, *u);
T divider;
/*divider = sum(s1)*/
divider = scratch1->sum();
T t = alpha/divider;
/*x = x + t*v*/
/*r = r - t*u*/
/*w = C * r*/
Vector<T>::mfadd(*(this->x), t, *v, *(this->x));
Vector<T>::mfadd(*r, -t, *u, *r);
Vector<T>::mul (*w, *C, *r);
/*s1 = w*w*/
Vector<T>::mul(*scratch1, *w, *w);
/*beta = sum(s1)*/
T beta = scratch1->sum();
#if 1
if(beta < (tolerance*bnorm + tolerance)){
T rl = r->length2();
if(Sqrt(rl)<(tolerance*bnorm + tolerance)){
warning("CG::Success in %d iterations, %10.10e, %10.10e", k, rl,
Sqrt(Fabs(rl)));
return;
}
}
#endif
T s = beta/alpha;
/*s1 = C * w*/
Vector<T>::mul(*scratch1, *C, *w);
/*v = s1 + s * v*/
Vector<T>::mfadd(*v, s, *v, *scratch1);
alpha = beta;
k++;
}
message("Unsuccesfull");
throw new SolutionNotFoundException(__LINE__, __FILE__,
"Number of iterations exceeded.");
}
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
void run_benchmark( void *vargs ) {
struct bench_args_t *args = (struct bench_args_t *)vargs;
spmv( args->val, args->cols, args->rowDelimiters, args->vec, args->out );
}
void
spmv(const char mode[],
const AlphaType& alpha_in,
const AMatrix& A,
const XVector& x,
const BetaType& beta_in,
const YVector& y,
const RANK_TWO) {
typedef typename Impl::GetCoeffView<AlphaType, typename XVector::device_type>::view_type alpha_view_type;
typedef typename Impl::GetCoeffView<BetaType, typename XVector::device_type>::view_type beta_view_type;
//alpha_view_type alpha = Impl::GetCoeffView<AlphaType, typename XVector::device_type>::get_view(alpha_in,x.dimension_1());
//beta_view_type beta = Impl::GetCoeffView<AlphaType, typename XVector::device_type>::get_view(beta_in, x.dimension_1());
#ifdef KOKKOS_HAVE_CXX11
// Make sure that both x and y have the same rank.
static_assert (XVector::rank == YVector::rank, "KokkosBlas::spmv: Vector ranks do not match.");
// Make sure that y is non-const.
static_assert (Kokkos::Impl::is_same<typename YVector::value_type,
typename YVector::non_const_value_type>::value,
"KokkosBlas::spmv: Output Vector must be non-const.");
#else
// We prefer to use C++11 static_assert, because it doesn't give
// "unused typedef" warnings, like the constructs below do.
//
// Make sure that both x and y have the same rank.
typedef typename
Kokkos::Impl::StaticAssert<XVector::rank == YVector::rank>::type Blas1_spmv_vector_ranks_do_not_match;
#endif // KOKKOS_HAVE_CXX11
// Check compatibility of dimensions at run time.
if((mode[0]==NoTranspose[0])||(mode[0]==Conjugate[0])) {
if ((x.dimension_1 () != y.dimension_1 ()) ||
(static_cast<size_t> (A.numCols ()) > static_cast<size_t> (x.dimension_0 ())) ||
(static_cast<size_t> (A.numRows ()) > static_cast<size_t> (y.dimension_0 ()))) {
std::ostringstream os;
os << "KokkosBlas::spmv: Dimensions do not match: "
<< ", A: " << A.numRows () << " x " << A.numCols()
<< ", x: " << x.dimension_0 () << " x " << x.dimension_1 ()
<< ", y: " << y.dimension_0 () << " x " << y.dimension_1 ();
Kokkos::Impl::throw_runtime_exception (os.str ());
}
} else {
if ((x.dimension_1 () != y.dimension_1 ()) ||
(static_cast<size_t> (A.numCols ()) > static_cast<size_t> (y.dimension_0 ())) ||
(static_cast<size_t> (A.numRows ()) > static_cast<size_t> (x.dimension_0 ()))) {
std::ostringstream os;
os << "KokkosBlas::spmv: Dimensions do not match (transpose): "
<< ", A: " << A.numRows () << " x " << A.numCols()
<< ", x: " << x.dimension_0 () << " x " << x.dimension_1 ()
<< ", y: " << y.dimension_0 () << " x " << y.dimension_1 ();
Kokkos::Impl::throw_runtime_exception (os.str ());
}
}
typedef KokkosSparse::CrsMatrix<typename AMatrix::const_value_type,
typename AMatrix::const_ordinal_type,
typename AMatrix::device_type,
typename AMatrix::memory_traits,
typename AMatrix::const_size_type> AMatrix_Internal;
AMatrix_Internal A_i = A;
// Call single vector version if appropriate
if( x.dimension_1() == 1) {
typedef Kokkos::View<typename XVector::const_value_type*,
typename Kokkos::Impl::if_c<Kokkos::Impl::is_same<typename YVector::array_layout,Kokkos::LayoutLeft>::value,
Kokkos::LayoutLeft,Kokkos::LayoutStride>::type,
typename XVector::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged|Kokkos::RandomAccess> > XVector_SubInternal;
typedef Kokkos::View<typename YVector::non_const_value_type*,
typename Kokkos::Impl::if_c<Kokkos::Impl::is_same<typename YVector::array_layout,Kokkos::LayoutLeft>::value,
Kokkos::LayoutLeft,Kokkos::LayoutStride>::type,
typename YVector::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > YVector_SubInternal;
XVector_SubInternal x_i = Kokkos::subview(x,Kokkos::ALL(),0);
YVector_SubInternal y_i = Kokkos::subview(y,Kokkos::ALL(),0);
alpha_view_type alpha = Impl::GetCoeffView<AlphaType, typename XVector::device_type>::get_view(alpha_in,x.dimension_1());
beta_view_type beta = Impl::GetCoeffView<BetaType, typename XVector::device_type>::get_view(beta_in, x.dimension_1());
typename alpha_view_type::non_const_type::HostMirror h_alpha =
Kokkos::create_mirror_view(alpha);
Kokkos::deep_copy(h_alpha,alpha);
typename beta_view_type::non_const_type::HostMirror h_beta =
Kokkos::create_mirror_view(beta);
Kokkos::deep_copy(h_beta,beta);
spmv(mode,h_alpha(0),A,x_i,h_beta(0),y_i);
return;
} else {
typedef Kokkos::View<typename XVector::const_value_type**,
typename XVector::array_layout,
typename XVector::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged|Kokkos::RandomAccess> > XVector_Internal;
typedef Kokkos::View<typename YVector::non_const_value_type**,
typename YVector::array_layout,
typename YVector::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > YVector_Internal;
XVector_Internal x_i = x;
YVector_Internal y_i = y;
typedef Kokkos::View<typename alpha_view_type::const_value_type*,
typename alpha_view_type::array_layout,
typename alpha_view_type::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > alpha_view_type_Internal;
typedef Kokkos::View<typename beta_view_type::const_value_type*,
typename beta_view_type::array_layout,
typename beta_view_type::device_type,
Kokkos::MemoryTraits<Kokkos::Unmanaged> > beta_view_type_Internal;
//alpha_view_type_Internal alpha_c = alpha;
//beta_view_type_Internal beta_c = beta;
return Impl::SPMV_MV<typename alpha_view_type_Internal::value_type*,
typename alpha_view_type_Internal::array_layout,
typename alpha_view_type_Internal::device_type,
typename alpha_view_type_Internal::memory_traits,
typename AMatrix_Internal::value_type,
typename AMatrix_Internal::ordinal_type,
typename AMatrix_Internal::device_type,
typename AMatrix_Internal::memory_traits,
typename AMatrix_Internal::size_type,
typename XVector_Internal::value_type**,
typename XVector_Internal::array_layout,
typename XVector_Internal::device_type,
typename XVector_Internal::memory_traits,
typename beta_view_type_Internal::value_type*,
typename beta_view_type_Internal::array_layout,
typename beta_view_type_Internal::device_type,
typename beta_view_type_Internal::memory_traits,
typename YVector_Internal::value_type**,
typename YVector_Internal::array_layout,
typename YVector_Internal::device_type,
typename YVector_Internal::memory_traits>::spmv_mv(mode,alpha_in,A,x,beta_in,y);
}
}