Skip to content

Commit

Permalink
formatting
Browse files Browse the repository at this point in the history
  • Loading branch information
@jgfouca
jgfouca committed Jan 11, 2024
1 parent d55c19c commit f31e678
Show file tree
Hide file tree
Showing 2 changed files with 51 additions and 52 deletions.
85 changes: 42 additions & 43 deletions sparse/impl/KokkosSparse_spiluk_numeric_impl.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -58,7 +58,7 @@ struct IlukWrap {
using team_policy = typename IlukHandle::TeamPolicy;
using member_type = typename team_policy::member_type;
using range_policy = typename IlukHandle::RangePolicy;
using sview_1d = typename Kokkos::View<scalar_t *, memory_space>;
using sview_1d = typename Kokkos::View<scalar_t *, memory_space>;

static team_policy get_team_policy(const size_type nrows,
const int team_size) {
Expand Down Expand Up @@ -107,7 +107,7 @@ struct IlukWrap {
WorkViewType iw;
lno_t lev_start;

using reftype = scalar_t&;
using reftype = scalar_t &;

Common(const ARowMapType &A_row_map_, const AEntriesType &A_entries_,
const AValuesType &A_values_, const LRowMapType &L_row_map_,
Expand Down Expand Up @@ -174,7 +174,7 @@ struct IlukWrap {
// gemm, alpha is hardcoded to -1, beta hardcoded to 1
KOKKOS_INLINE_FUNCTION
void gemm(const scalar_t &A, const scalar_t &B, scalar_t &C) const {
C += -1*A*B;
C += -1 * A * B;
}

// lget
Expand Down Expand Up @@ -345,16 +345,15 @@ struct IlukWrap {

// gemm, alpha is hardcoded to -1, beta hardcoded to 1
template <typename CView>
KOKKOS_INLINE_FUNCTION
void gemm(const UValuesUnmanaged2DBlockType &A,
const LValuesUnmanaged2DBlockType &B,
CView& C) const {
KOKKOS_INLINE_FUNCTION void gemm(const UValuesUnmanaged2DBlockType &A,
const LValuesUnmanaged2DBlockType &B,
CView &C) const {
KokkosBatched::SerialGemm<KokkosBatched::Trans::NoTranspose,
KokkosBatched::Trans::NoTranspose,
KokkosBatched::Algo::Gemm::Unblocked>::
invoke<scalar_t, LValuesUnmanaged2DBlockType,
UValuesUnmanaged2DBlockType, LValuesUnmanaged2DBlockType>(
-1.0, A, B, 1.0, C);
invoke<scalar_t, LValuesUnmanaged2DBlockType,
UValuesUnmanaged2DBlockType, LValuesUnmanaged2DBlockType>(
-1.0, A, B, 1.0, C);
}

// lget
Expand Down Expand Up @@ -427,21 +426,21 @@ struct IlukWrap {
Base::lset_id(team, k2);
Kokkos::parallel_for(Kokkos::TeamThreadRange(team, k1, k2),
[&](const size_type k) {
const auto col = Base::L_entries(k);
Base::lset(k, 0.0);
Base::iw(my_team, col) = k;
});
const auto col = Base::L_entries(k);
Base::lset(k, 0.0);
Base::iw(my_team, col) = k;
});

team.team_barrier();

k1 = Base::U_row_map(rowid);
k2 = Base::U_row_map(rowid + 1);
Kokkos::parallel_for(Kokkos::TeamThreadRange(team, k1, k2),
[&](const size_type k) {
const auto col = Base::U_entries(k);
Base::uset(k, 0.0);
Base::iw(my_team, col) = k;
});
const auto col = Base::U_entries(k);
Base::uset(k, 0.0);
Base::iw(my_team, col) = k;
});

team.team_barrier();

Expand All @@ -450,14 +449,14 @@ struct IlukWrap {
k2 = Base::A_row_map(rowid + 1);
Kokkos::parallel_for(Kokkos::TeamThreadRange(team, k1, k2),
[&](const size_type k) {
const auto col = Base::A_entries(k);
const auto ipos = Base::iw(my_team, col);
if (col < rowid) {
Base::lset(ipos, Base::aget(k));
} else {
Base::uset(ipos, Base::aget(k));
}
});
const auto col = Base::A_entries(k);
const auto ipos = Base::iw(my_team, col);
if (col < rowid) {
Base::lset(ipos, Base::aget(k));
} else {
Base::uset(ipos, Base::aget(k));
}
});

team.team_barrier();

Expand All @@ -476,7 +475,8 @@ struct IlukWrap {
const auto col = Base::U_entries(kk);
const auto ipos = Base::iw(my_team, col);
if (ipos != -1) {
typename Base::reftype C = col < rowid ? Base::lget(ipos) : Base::uget(ipos);
typename Base::reftype C =
col < rowid ? Base::lget(ipos) : Base::uget(ipos);
Base::gemm(Base::uget(kk), fact, C);
}
}); // end for kk
Expand All @@ -498,17 +498,17 @@ struct IlukWrap {
k2 = Base::L_row_map(rowid + 1) - 1;
Kokkos::parallel_for(Kokkos::TeamThreadRange(team, k1, k2),
[&](const size_type k) {
const auto col = Base::L_entries(k);
Base::iw(my_team, col) = -1;
});
const auto col = Base::L_entries(k);
Base::iw(my_team, col) = -1;
});

k1 = Base::U_row_map(rowid);
k2 = Base::U_row_map(rowid + 1);
Kokkos::parallel_for(Kokkos::TeamThreadRange(team, k1, k2),
[&](const size_type k) {
const auto col = Base::U_entries(k);
Base::iw(my_team, col) = -1;
});
const auto col = Base::U_entries(k);
Base::iw(my_team, col) = -1;
});
}
};

Expand Down Expand Up @@ -556,7 +556,7 @@ struct IlukWrap {

level_nchunks_h = thandle.get_level_nchunks();
level_nrowsperchunk_h = thandle.get_level_nrowsperchunk();
iw = thandle.get_iw();
iw = thandle.get_iw();

// Main loop must be performed sequential. Question: Try out Cuda's graph
// stuff to reduce kernel launch overhead
Expand All @@ -574,11 +574,10 @@ struct IlukWrap {
else
lvl_nrows_chunk = level_nrowsperchunk_h(lvl);

team_policy tpolicy =
get_team_policy(lvl_nrows_chunk, team_size);
team_policy tpolicy = get_team_policy(lvl_nrows_chunk, team_size);
KernelLaunchMacro(A_row_map, A_entries, A_values, L_row_map,
L_entries, L_values, U_row_map, U_entries,
U_values, tpolicy, "parfor_tp1", level_idx, iw,
L_entries, L_values, U_row_map, U_entries, U_values,
tpolicy, "parfor_tp1", level_idx, iw,
lev_start + lvl_rowid_start, TPF, TPB,
thandle.is_block_enabled(), block_size);
Kokkos::fence();
Expand Down Expand Up @@ -720,27 +719,27 @@ struct IlukWrap {
if ((lvl_rowid_start_v[i] + lvl_nrowsperchunk_h_v[i](lvl)) >
(lvl_end_v[i] - lvl_start_v[i]))
lvl_nrows_chunk =
(lvl_end_v[i] - lvl_start_v[i]) - lvl_rowid_start_v[i];
(lvl_end_v[i] - lvl_start_v[i]) - lvl_rowid_start_v[i];
else
lvl_nrows_chunk = lvl_nrowsperchunk_h_v[i](lvl);

// 1.b. Create functor for stream i-th and launch
team_policy tpolicy = get_team_policy(
execspace_v[i], lvl_nrows_chunk, team_size_v[i]);
execspace_v[i], lvl_nrows_chunk, team_size_v[i]);
KernelLaunchMacro(A_row_map_v[i], A_entries_v[i], A_values_v[i],
L_row_map_v[i], L_entries_v[i], L_values_v[i],
U_row_map_v[i], U_entries_v[i], U_values_v[i],
tpolicy, "parfor_tp1", lvl_idx_v[i], iw_v[i],
lvl_start_v[i] + lvl_rowid_start_v[i], TPF,
TPB, is_block_enabled_v[i], block_size_v[i]);
lvl_start_v[i] + lvl_rowid_start_v[i], TPF, TPB,
is_block_enabled_v[i], block_size_v[i]);
// 1.c. Ready to move to next chunk
lvl_rowid_start_v[i] += lvl_nrows_chunk;
} // end if (chunkid < lvl_nchunks_h_v[i](lvl))
} // end if (stream_have_level_v[i])
} // end for streams
} // end for chunkid
} // end for lvl
} // end iluk_numeric_streams
} // end iluk_numeric_streams

}; // IlukWrap

Expand Down
18 changes: 9 additions & 9 deletions sparse/unit_test/Test_Sparse_spiluk.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -120,8 +120,7 @@ struct SpilukTest {
}

static void check_result(const RowMapType& row_map,
const EntriesType& entries,
const ValuesType& values,
const EntriesType& entries, const ValuesType& values,
const RowMapType& L_row_map,
const EntriesType& L_entries,
const ValuesType& L_values,
Expand Down Expand Up @@ -270,14 +269,14 @@ struct SpilukTest {
}

static void run_test_spiluk_scale() {

// Create a diagonally dominant sparse matrix to test:
constexpr auto nrows = 5000;
constexpr auto diagDominance = 2;

size_type nnz = 10 * nrows;
auto A = KokkosSparse::Impl::kk_generate_diagonally_dominant_sparse_matrix<
Crs>(nrows, nrows, nnz, 0, lno_t(0.01 * nrows), diagDominance);
auto A =
KokkosSparse::Impl::kk_generate_diagonally_dominant_sparse_matrix<Crs>(
nrows, nrows, nnz, 0, lno_t(0.01 * nrows), diagDominance);

// Pull out views from CRS
RowMapType row_map("row_map", A.graph.row_map.extent(0));
Expand All @@ -296,7 +295,6 @@ struct SpilukTest {
}

static void run_test_spiluk_scale_blocks() {

// Create a diagonally dominant sparse matrix to test:
constexpr auto nrows = 5000;
constexpr auto diagDominance = 2;
Expand All @@ -308,8 +306,9 @@ struct SpilukTest {
const size_type block_size = 10;

size_type nnz = 10 * nrows;
auto A = KokkosSparse::Impl::kk_generate_diagonally_dominant_sparse_matrix<
Crs>(nrows, nrows, nnz, 0, lno_t(0.01 * nrows), diagDominance);
auto A =
KokkosSparse::Impl::kk_generate_diagonally_dominant_sparse_matrix<Crs>(
nrows, nrows, nnz, 0, lno_t(0.01 * nrows), diagDominance);

// Pull out views from CRS
Bsr bsr(A, block_size);
Expand All @@ -327,7 +326,8 @@ struct SpilukTest {
KernelHandle kh;

run_and_check_spiluk_block(kh, brow_map, bentries, bvalues,
SPILUKAlgorithm::SEQLVLSCHD_TP1, fill_lev, block_size);
SPILUKAlgorithm::SEQLVLSCHD_TP1, fill_lev,
block_size);
}

static void run_test_spiluk_streams(SPILUKAlgorithm test_algo, int nstreams) {
Expand Down

0 comments on commit f31e678

Please sign in to comment.