diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 9837066c4e73121489509fc82a07ff6e5c72eb59..5109bb5083a03c8c118f84718daf5026ff1923d7 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -17,7 +17,7 @@ stages:
variables:
ACCOUNT_CPU: "ka1125"
ACCOUNT_GPU: "bk1341"
- SLURM_OPTIONS_CPU: "--account=$ACCOUNT_CPU --partition=shared"
+ SLURM_OPTIONS_CPU: "--account=$ACCOUNT_CPU --partition=shared --time=00:10:00"
SLURM_OPTIONS_GPU: "--account=$ACCOUNT_GPU --partition=gpu --gpus=1"
SLURM_NTASKS: "--ntasks=1"
GIT_CONFIG_COUNT: 1
@@ -94,7 +94,8 @@ nvhpc_gpu:
- module load git gcc/11.2.0-gcc-11.2.0 nvhpc/24.7-gcc-11.2.0
- mkdir nvhpc_gpu
- cd nvhpc_gpu
- - /sw/spack-levante/cmake-3.23.1-q5kzz6/bin/cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_C_COMPILER=nvc -DCMAKE_Fortran_COMPILER=nvfortran -DIM_ENABLE_OPENACC=ON
+ - export LD_LIBRARY_PATH=/sw/spack-levante/gcc-11.2.0-bcn7mb/lib64:$LD_LIBRARY_PATH
+ - /sw/spack-levante/cmake-3.23.1-q5kzz6/bin/cmake .. -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_C_COMPILER=nvc -DCMAKE_Fortran_COMPILER=nvfortran -DIM_ENABLE_GPU=nvidia-sm80 -DIM_ENABLE_OPENACC=ON
- make VERBOSE=1
- make test
tags:
diff --git a/src/horizontal/mo_lib_divrot.cpp b/src/horizontal/mo_lib_divrot.cpp
index d086e8ba4d7910986117e912fa32cdc4f4426ac5..0138cc0408c6ee068a73bafed2d7001bb51eaae9 100644
--- a/src/horizontal/mo_lib_divrot.cpp
+++ b/src/horizontal/mo_lib_divrot.cpp
@@ -36,9 +36,6 @@ void recon_lsq_cell_l(const T *p_cc, const int *cell_neighbor_idx,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T *> z_d("z_d", lsq_dim_c);
- Kokkos::View<T *> z_qt_times_d("z_qt_times_d", lsq_dim_unk);
-
UnmanagedConstInt3D iidx(cell_neighbor_idx, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(cell_neighbor_blk, nproma, nblks_c, lsq_dim_c);
@@ -54,35 +51,38 @@ void recon_lsq_cell_l(const T *p_cc, const int *cell_neighbor_idx,
nblks_c);
UnmanagedConstT3D lsq_moments_view(lsq_moments, nproma, nblks_c, lsq_dim_unk);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_l_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_d(0) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ T z_d[3]; // Local array instead of shared View
+ T z_qt_times_d[2];
+
+ z_d[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
p_cc_view(jc, jk, jb);
- z_d(1) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ z_d[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
p_cc_view(jc, jk, jb);
- z_d(2) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ z_d[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
p_cc_view(jc, jk, jb);
// matrix multiplication Q^T d (partitioned into 2 dot products)
- z_qt_times_d(0) = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d(0) +
- lsq_qtmat_c_view(jc, 0, 1, jb) * z_d(1) +
- lsq_qtmat_c_view(jc, 0, 2, jb) * z_d(2);
- z_qt_times_d(1) = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d(0) +
- lsq_qtmat_c_view(jc, 1, 1, jb) * z_d(1) +
- lsq_qtmat_c_view(jc, 1, 2, jb) * z_d(2);
+ z_qt_times_d[0] = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 0, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 0, 2, jb) * z_d[2];
+ z_qt_times_d[1] = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 1, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 1, 2, jb) * z_d[2];
p_coeff_view(2, jc, jk, jb) =
- lsq_rmat_rdiag_c_view(jc, 1, jb) * z_qt_times_d(1);
+ lsq_rmat_rdiag_c_view(jc, 1, jb) * z_qt_times_d[1];
p_coeff_view(1, jc, jk, jb) =
lsq_rmat_rdiag_c_view(jc, 0, jb) *
- (z_qt_times_d(0) -
+ (z_qt_times_d[0] -
lsq_rmat_utri_c_view(jc, 0, jb) * p_coeff_view(2, jc, jk, jb));
p_coeff_view(0, jc, jk, jb) = p_cc_view(jc, jk, jb);
});
@@ -124,8 +124,6 @@ void recon_lsq_cell_l_svd(const T *p_cc, const int *cell_neighbor_idx,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T *> z_b("z_b", lsq_dim_c);
-
UnmanagedConstInt3D iidx(cell_neighbor_idx, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(cell_neighbor_blk, nproma, nblks_c, lsq_dim_c);
@@ -136,31 +134,32 @@ void recon_lsq_cell_l_svd(const T *p_cc, const int *cell_neighbor_idx,
lsq_dim_c, nblks_c);
UnmanagedConstT3D lsq_moments_view(lsq_moments, nproma, nblks_c, lsq_dim_unk);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_l_svd_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_b(0) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ T z_b[3]; // Local array instead of shared View
+ z_b[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
p_cc_view(jc, jk, jb);
- z_b(1) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ z_b[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
p_cc_view(jc, jk, jb);
- z_b(2) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ z_b[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
p_cc_view(jc, jk, jb);
p_coeff_view(2, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 1, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 1, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 1, 2, jb) * z_b(2);
+ lsq_pseudoinv_view(jc, 1, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 1, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 1, 2, jb) * z_b[2];
p_coeff_view(1, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 0, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 0, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 0, 2, jb) * z_b(2);
+ lsq_pseudoinv_view(jc, 0, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 0, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 0, 2, jb) * z_b[2];
p_coeff_view(0, jc, jk, jb) = p_cc_view(jc, jk, jb);
});
if (l_consv) {
@@ -201,9 +200,6 @@ void recon_lsq_cell_q(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T ***> z_d("z_d", lsq_dim_c, nproma, nlev);
- Kokkos::View<T *> z_qt_times_d("z_qt_times_d", lsq_dim_unk);
-
UnmanagedConstInt3D iidx(lsq_idx_c, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(lsq_blk_c, nproma, nblks_c, lsq_dim_c);
@@ -221,7 +217,7 @@ void recon_lsq_cell_q(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
if (patch_id > 0 || l_limited_area) {
Kokkos::MDRangePolicy<Kokkos::Rank<4>> initPolicy(
{0, i_startidx_in, slev, i_startblk},
- {lsq_dim_unk + 1, i_endidx_in, elev, i_endblk});
+ {lsq_dim_unk + 1, i_endidx_in + 1, elev + 1, i_endblk + 1});
Kokkos::parallel_for(
"recon_lsq_cell_q_init", initPolicy,
KOKKOS_LAMBDA(const int ji, const int jc, const int jk, const int jb) {
@@ -229,103 +225,102 @@ void recon_lsq_cell_q(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
});
}
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_q_step1", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_d(0, jc, jk) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
- p_cc_view(jc, jk, jb);
- z_d(1, jc, jk) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
- p_cc_view(jc, jk, jb);
- z_d(2, jc, jk) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
- p_cc_view(jc, jk, jb);
- z_d(3, jc, jk) = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
- p_cc_view(jc, jk, jb);
- z_d(4, jc, jk) = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
- p_cc_view(jc, jk, jb);
- z_d(5, jc, jk) = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
- p_cc_view(jc, jk, jb);
- z_d(6, jc, jk) = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
- p_cc_view(jc, jk, jb);
- z_d(7, jc, jk) = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
- p_cc_view(jc, jk, jb);
- z_d(8, jc, jk) = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
+ T z_d[9]; // Local array instead of shared View
+ T z_qt_times_d[5];
+ z_d[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ p_cc_view(jc, jk, jb);
+ z_d[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ p_cc_view(jc, jk, jb);
+ z_d[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ p_cc_view(jc, jk, jb);
+ z_d[3] = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
+ p_cc_view(jc, jk, jb);
+ z_d[4] = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
+ p_cc_view(jc, jk, jb);
+ z_d[5] = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
+ p_cc_view(jc, jk, jb);
+ z_d[6] = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
+ p_cc_view(jc, jk, jb);
+ z_d[7] = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
+ p_cc_view(jc, jk, jb);
+ z_d[8] = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
p_cc_view(jc, jk, jb);
- });
- Kokkos::parallel_for(
- "recon_lsq_cell_q_step2", innerPolicy,
- KOKKOS_LAMBDA(const int jk, const int jc) {
- z_qt_times_d(0) = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(1) = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(2) = lsq_qtmat_c_view(jc, 2, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(3) = lsq_qtmat_c_view(jc, 3, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(4) = lsq_qtmat_c_view(jc, 4, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 8, jb) * z_d(8, jc, jk);
-
- p_coeff_view(5, jc, jk, jb) = ptr_rrdiag(jc, 4, jb) * z_qt_times_d(4);
+
+ z_qt_times_d[0] = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 0, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 0, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 0, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 0, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 0, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 0, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 0, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 0, 8, jb) * z_d[8];
+ z_qt_times_d[1] = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 1, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 1, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 1, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 1, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 1, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 1, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 1, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 1, 8, jb) * z_d[8];
+ z_qt_times_d[2] = lsq_qtmat_c_view(jc, 2, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 2, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 2, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 2, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 2, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 2, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 2, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 2, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 2, 8, jb) * z_d[8];
+ z_qt_times_d[3] = lsq_qtmat_c_view(jc, 3, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 3, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 3, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 3, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 3, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 3, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 3, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 3, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 3, 8, jb) * z_d[8];
+ z_qt_times_d[4] = lsq_qtmat_c_view(jc, 4, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 4, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 4, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 4, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 4, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 4, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 4, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 4, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 4, 8, jb) * z_d[8];
+
+ p_coeff_view(5, jc, jk, jb) = ptr_rrdiag(jc, 4, jb) * z_qt_times_d[4];
p_coeff_view(4, jc, jk, jb) =
ptr_rrdiag(jc, 3, jb) *
- (z_qt_times_d(3) -
+ (z_qt_times_d[3] -
ptr_rutri(jc, 0, jb) * p_coeff_view(5, jc, jk, jb));
p_coeff_view(3, jc, jk, jb) =
ptr_rrdiag(jc, 2, jb) *
- (z_qt_times_d(2) -
+ (z_qt_times_d[2] -
ptr_rutri(jc, 1, jb) * p_coeff_view(4, jc, jk, jb) -
ptr_rutri(jc, 2, jb) * p_coeff_view(5, jc, jk, jb));
p_coeff_view(2, jc, jk, jb) =
ptr_rrdiag(jc, 1, jb) *
- (z_qt_times_d(1) -
+ (z_qt_times_d[1] -
ptr_rutri(jc, 3, jb) * p_coeff_view(3, jc, jk, jb) -
ptr_rutri(jc, 4, jb) * p_coeff_view(4, jc, jk, jb) -
ptr_rutri(jc, 5, jb) * p_coeff_view(5, jc, jk, jb));
p_coeff_view(1, jc, jk, jb) =
ptr_rrdiag(jc, 0, jb) *
- (z_qt_times_d(0) -
+ (z_qt_times_d[0] -
ptr_rutri(jc, 6, jb) * p_coeff_view(2, jc, jk, jb) -
ptr_rutri(jc, 7, jb) * p_coeff_view(3, jc, jk, jb) -
ptr_rutri(jc, 8, jb) * p_coeff_view(4, jc, jk, jb) -
@@ -365,8 +360,6 @@ void recon_lsq_cell_q_svd(const T *p_cc, const int *lsq_idx_c,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T ***> z_b("z_b", lsq_dim_c, nproma, elev);
-
UnmanagedConstInt3D iidx(lsq_idx_c, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(lsq_blk_c, nproma, nblks_c, lsq_dim_c);
@@ -388,88 +381,86 @@ void recon_lsq_cell_q_svd(const T *p_cc, const int *lsq_idx_c,
});
}
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_q_svd_step1", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_b(0, jc, jk) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ T z_b[9]; // Local array instead of shared View
+ z_b[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
p_cc_view(jc, jk, jb);
- z_b(1, jc, jk) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ z_b[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
p_cc_view(jc, jk, jb);
- z_b(2, jc, jk) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ z_b[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
p_cc_view(jc, jk, jb);
- z_b(3, jc, jk) = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
+ z_b[3] = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
p_cc_view(jc, jk, jb);
- z_b(4, jc, jk) = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
+ z_b[4] = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
p_cc_view(jc, jk, jb);
- z_b(5, jc, jk) = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
+ z_b[5] = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
p_cc_view(jc, jk, jb);
- z_b(6, jc, jk) = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
+ z_b[6] = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
p_cc_view(jc, jk, jb);
- z_b(7, jc, jk) = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
+ z_b[7] = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
p_cc_view(jc, jk, jb);
- z_b(8, jc, jk) = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
+ z_b[8] = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
p_cc_view(jc, jk, jb);
- });
- Kokkos::parallel_for(
- "recon_lsq_cell_q_svd_step2", innerPolicy,
- KOKKOS_LAMBDA(const int jk, const int jc) {
+
p_coeff_view(5, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 4, 0, jb) * z_b(0, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 1, jb) * z_b(1, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 2, jb) * z_b(2, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 3, jb) * z_b(3, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 4, jb) * z_b(4, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 5, jb) * z_b(5, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 6, jb) * z_b(6, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 7, jb) * z_b(7, jc, jk) +
- lsq_pseudoinv_view(jc, 4, 8, jb) * z_b(8, jc, jk);
+ lsq_pseudoinv_view(jc, 4, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 4, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 4, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 4, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 4, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 4, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 4, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 4, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 4, 8, jb) * z_b[8];
p_coeff_view(4, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 3, 0, jb) * z_b(0, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 1, jb) * z_b(1, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 2, jb) * z_b(2, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 3, jb) * z_b(3, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 4, jb) * z_b(4, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 5, jb) * z_b(5, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 6, jb) * z_b(6, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 7, jb) * z_b(7, jc, jk) +
- lsq_pseudoinv_view(jc, 3, 8, jb) * z_b(8, jc, jk);
+ lsq_pseudoinv_view(jc, 3, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 3, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 3, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 3, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 3, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 3, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 3, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 3, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 3, 8, jb) * z_b[8];
p_coeff_view(3, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 2, 0, jb) * z_b(0, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 1, jb) * z_b(1, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 2, jb) * z_b(2, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 3, jb) * z_b(3, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 4, jb) * z_b(4, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 5, jb) * z_b(5, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 6, jb) * z_b(6, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 7, jb) * z_b(7, jc, jk) +
- lsq_pseudoinv_view(jc, 2, 8, jb) * z_b(8, jc, jk);
+ lsq_pseudoinv_view(jc, 2, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 2, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 2, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 2, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 2, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 2, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 2, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 2, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 2, 8, jb) * z_b[8];
p_coeff_view(2, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 1, 0, jb) * z_b(0, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 1, jb) * z_b(1, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 2, jb) * z_b(2, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 3, jb) * z_b(3, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 4, jb) * z_b(4, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 5, jb) * z_b(5, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 6, jb) * z_b(6, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 7, jb) * z_b(7, jc, jk) +
- lsq_pseudoinv_view(jc, 1, 8, jb) * z_b(8, jc, jk);
+ lsq_pseudoinv_view(jc, 1, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 1, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 1, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 1, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 1, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 1, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 1, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 1, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 1, 8, jb) * z_b[8];
p_coeff_view(1, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 0, 0, jb) * z_b(0, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 1, jb) * z_b(1, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 2, jb) * z_b(2, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 3, jb) * z_b(3, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 4, jb) * z_b(4, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 5, jb) * z_b(5, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 6, jb) * z_b(6, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 7, jb) * z_b(7, jc, jk) +
- lsq_pseudoinv_view(jc, 0, 8, jb) * z_b(8, jc, jk);
+ lsq_pseudoinv_view(jc, 0, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 0, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 0, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 0, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 0, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 0, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 0, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 0, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 0, 8, jb) * z_b[8];
p_coeff_view(0, jc, jk, jb) =
p_cc_view(jc, jk, jb) -
p_coeff_view(1, jc, jk, jb) * lsq_moments_view(jc, jb, 0) -
@@ -505,9 +496,6 @@ void recon_lsq_cell_c(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T ***> z_d("z_d", lsq_dim_c, nproma, elev);
- Kokkos::View<T *> z_qt_times_d("z_qt_times_d", 9);
-
UnmanagedConstInt3D iidx(lsq_idx_c, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(lsq_blk_c, nproma, nblks_c, lsq_dim_c);
@@ -533,146 +521,146 @@ void recon_lsq_cell_c(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
});
}
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_c_step1", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_d(0, jc, jk) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ T z_d[9]; // Local array instead of shared View
+ T z_qt_times_d[9]; // Local array instead of shared View
+
+ z_d[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
p_cc_view(jc, jk, jb);
- z_d(1, jc, jk) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ z_d[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
p_cc_view(jc, jk, jb);
- z_d(2, jc, jk) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ z_d[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
p_cc_view(jc, jk, jb);
- z_d(3, jc, jk) = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
+ z_d[3] = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
p_cc_view(jc, jk, jb);
- z_d(4, jc, jk) = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
+ z_d[4] = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
p_cc_view(jc, jk, jb);
- z_d(5, jc, jk) = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
+ z_d[5] = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
p_cc_view(jc, jk, jb);
- z_d(6, jc, jk) = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
+ z_d[6] = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
p_cc_view(jc, jk, jb);
- z_d(7, jc, jk) = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
+ z_d[7] = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
p_cc_view(jc, jk, jb);
- z_d(8, jc, jk) = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
+ z_d[8] = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
p_cc_view(jc, jk, jb);
- });
- Kokkos::parallel_for(
- "recon_lsq_cell_c_step2", innerPolicy,
- KOKKOS_LAMBDA(const int jk, const int jc) {
- z_qt_times_d(0) = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 0, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(1) = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 1, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(2) = lsq_qtmat_c_view(jc, 2, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 2, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(3) = lsq_qtmat_c_view(jc, 3, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 3, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(4) = lsq_qtmat_c_view(jc, 4, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 4, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(5) = lsq_qtmat_c_view(jc, 5, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 5, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(6) = lsq_qtmat_c_view(jc, 6, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 6, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(7) = lsq_qtmat_c_view(jc, 7, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 7, 8, jb) * z_d(8, jc, jk);
- z_qt_times_d(8) = lsq_qtmat_c_view(jc, 8, 0, jb) * z_d(0, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 1, jb) * z_d(1, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 2, jb) * z_d(2, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 3, jb) * z_d(3, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 4, jb) * z_d(4, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 5, jb) * z_d(5, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 6, jb) * z_d(6, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 7, jb) * z_d(7, jc, jk) +
- lsq_qtmat_c_view(jc, 8, 8, jb) * z_d(8, jc, jk);
-
- p_coeff_view(9, jc, jk, jb) = ptr_rrdiag(jc, 8, jb) * z_qt_times_d(8);
+
+ z_qt_times_d[0] = lsq_qtmat_c_view(jc, 0, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 0, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 0, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 0, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 0, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 0, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 0, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 0, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 0, 8, jb) * z_d[8];
+ z_qt_times_d[1] = lsq_qtmat_c_view(jc, 1, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 1, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 1, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 1, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 1, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 1, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 1, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 1, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 1, 8, jb) * z_d[8];
+ z_qt_times_d[2] = lsq_qtmat_c_view(jc, 2, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 2, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 2, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 2, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 2, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 2, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 2, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 2, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 2, 8, jb) * z_d[8];
+ z_qt_times_d[3] = lsq_qtmat_c_view(jc, 3, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 3, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 3, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 3, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 3, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 3, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 3, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 3, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 3, 8, jb) * z_d[8];
+ z_qt_times_d[4] = lsq_qtmat_c_view(jc, 4, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 4, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 4, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 4, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 4, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 4, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 4, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 4, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 4, 8, jb) * z_d[8];
+ z_qt_times_d[5] = lsq_qtmat_c_view(jc, 5, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 5, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 5, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 5, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 5, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 5, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 5, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 5, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 5, 8, jb) * z_d[8];
+ z_qt_times_d[6] = lsq_qtmat_c_view(jc, 6, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 6, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 6, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 6, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 6, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 6, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 6, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 6, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 6, 8, jb) * z_d[8];
+ z_qt_times_d[7] = lsq_qtmat_c_view(jc, 7, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 7, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 7, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 7, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 7, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 7, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 7, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 7, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 7, 8, jb) * z_d[8];
+ z_qt_times_d[8] = lsq_qtmat_c_view(jc, 8, 0, jb) * z_d[0] +
+ lsq_qtmat_c_view(jc, 8, 1, jb) * z_d[1] +
+ lsq_qtmat_c_view(jc, 8, 2, jb) * z_d[2] +
+ lsq_qtmat_c_view(jc, 8, 3, jb) * z_d[3] +
+ lsq_qtmat_c_view(jc, 8, 4, jb) * z_d[4] +
+ lsq_qtmat_c_view(jc, 8, 5, jb) * z_d[5] +
+ lsq_qtmat_c_view(jc, 8, 6, jb) * z_d[6] +
+ lsq_qtmat_c_view(jc, 8, 7, jb) * z_d[7] +
+ lsq_qtmat_c_view(jc, 8, 8, jb) * z_d[8];
+
+ p_coeff_view(9, jc, jk, jb) = ptr_rrdiag(jc, 8, jb) * z_qt_times_d[8];
p_coeff_view(8, jc, jk, jb) =
ptr_rrdiag(jc, 7, jb) *
- (z_qt_times_d(7) -
+ (z_qt_times_d[7] -
ptr_rutri(jc, 0, jb) * p_coeff_view(9, jc, jk, jb));
p_coeff_view(7, jc, jk, jb) =
ptr_rrdiag(jc, 6, jb) *
- (z_qt_times_d(6) -
+ (z_qt_times_d[6] -
(ptr_rutri(jc, 1, jb) * p_coeff_view(8, jc, jk, jb) +
ptr_rutri(jc, 2, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(6, jc, jk, jb) =
ptr_rrdiag(jc, 5, jb) *
- (z_qt_times_d(5) -
+ (z_qt_times_d[5] -
(ptr_rutri(jc, 3, jb) * p_coeff_view(7, jc, jk, jb) +
ptr_rutri(jc, 4, jb) * p_coeff_view(8, jc, jk, jb) +
ptr_rutri(jc, 5, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(5, jc, jk, jb) =
ptr_rrdiag(jc, 4, jb) *
- (z_qt_times_d(4) -
+ (z_qt_times_d[4] -
(ptr_rutri(jc, 6, jb) * p_coeff_view(6, jc, jk, jb) +
ptr_rutri(jc, 7, jb) * p_coeff_view(7, jc, jk, jb) +
ptr_rutri(jc, 8, jb) * p_coeff_view(8, jc, jk, jb) +
ptr_rutri(jc, 9, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(4, jc, jk, jb) =
ptr_rrdiag(jc, 3, jb) *
- (z_qt_times_d(3) -
+ (z_qt_times_d[3] -
(ptr_rutri(jc, 10, jb) * p_coeff_view(5, jc, jk, jb) +
ptr_rutri(jc, 11, jb) * p_coeff_view(6, jc, jk, jb) +
ptr_rutri(jc, 12, jb) * p_coeff_view(7, jc, jk, jb) +
@@ -680,7 +668,7 @@ void recon_lsq_cell_c(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
ptr_rutri(jc, 14, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(3, jc, jk, jb) =
ptr_rrdiag(jc, 2, jb) *
- (z_qt_times_d(2) -
+ (z_qt_times_d[2] -
(ptr_rutri(jc, 15, jb) * p_coeff_view(4, jc, jk, jb) +
ptr_rutri(jc, 16, jb) * p_coeff_view(5, jc, jk, jb) +
ptr_rutri(jc, 17, jb) * p_coeff_view(6, jc, jk, jb) +
@@ -689,7 +677,7 @@ void recon_lsq_cell_c(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
ptr_rutri(jc, 20, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(2, jc, jk, jb) =
ptr_rrdiag(jc, 1, jb) *
- (z_qt_times_d(1) -
+ (z_qt_times_d[1] -
(ptr_rutri(jc, 21, jb) * p_coeff_view(3, jc, jk, jb) +
ptr_rutri(jc, 22, jb) * p_coeff_view(4, jc, jk, jb) +
ptr_rutri(jc, 23, jb) * p_coeff_view(5, jc, jk, jb) +
@@ -699,7 +687,7 @@ void recon_lsq_cell_c(const T *p_cc, const int *lsq_idx_c, const int *lsq_blk_c,
ptr_rutri(jc, 27, jb) * p_coeff_view(9, jc, jk, jb)));
p_coeff_view(1, jc, jk, jb) =
ptr_rrdiag(jc, 0, jb) *
- (z_qt_times_d(0) -
+ (z_qt_times_d[0] -
(ptr_rutri(jc, 28, jb) * p_coeff_view(2, jc, jk, jb) +
ptr_rutri(jc, 29, jb) * p_coeff_view(3, jc, jk, jb) +
ptr_rutri(jc, 30, jb) * p_coeff_view(4, jc, jk, jb) +
@@ -748,8 +736,6 @@ void recon_lsq_cell_c_svd(const T *p_cc, const int *lsq_idx_c,
Kokkos::MemoryUnmanaged>
UnmanagedConstInt3D;
- Kokkos::View<T *> z_b("z_b", 9);
-
UnmanagedConstInt3D iidx(lsq_idx_c, nproma, nblks_c, lsq_dim_c);
UnmanagedConstInt3D iblk(lsq_blk_c, nproma, nblks_c, lsq_dim_c);
@@ -761,13 +747,13 @@ void recon_lsq_cell_c_svd(const T *p_cc, const int *lsq_idx_c,
UnmanagedConstT3D lsq_moments_view(lsq_moments, nproma, nblks_c, lsq_dim_unk);
if (patch_id > 0 || l_limited_area) {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<3>> initPolicy(
- {slev, i_startidx, 0}, {elev, i_endidx, lsq_dim_unk + 1});
+ {slev, i_startidx, 0}, {elev + 1, i_endidx + 1, lsq_dim_unk + 1});
Kokkos::parallel_for(
"recon_lsq_cell_c_svd_init", initPolicy,
KOKKOS_LAMBDA(const int jk, const int jc, const int ji) {
@@ -776,125 +762,126 @@ void recon_lsq_cell_c_svd(const T *p_cc, const int *lsq_idx_c,
}
}
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"recon_lsq_cell_c_svd_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
- z_b(0) = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
+ T z_b[9]; // Local array instead of shared View
+ z_b[0] = p_cc_view(iidx(jc, jb, 0), jk, iblk(jc, jb, 0)) -
p_cc_view(jc, jk, jb);
- z_b(1) = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
+ z_b[1] = p_cc_view(iidx(jc, jb, 1), jk, iblk(jc, jb, 1)) -
p_cc_view(jc, jk, jb);
- z_b(2) = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
+ z_b[2] = p_cc_view(iidx(jc, jb, 2), jk, iblk(jc, jb, 2)) -
p_cc_view(jc, jk, jb);
- z_b(3) = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
+ z_b[3] = p_cc_view(iidx(jc, jb, 3), jk, iblk(jc, jb, 3)) -
p_cc_view(jc, jk, jb);
- z_b(4) = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
+ z_b[4] = p_cc_view(iidx(jc, jb, 4), jk, iblk(jc, jb, 4)) -
p_cc_view(jc, jk, jb);
- z_b(5) = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
+ z_b[5] = p_cc_view(iidx(jc, jb, 5), jk, iblk(jc, jb, 5)) -
p_cc_view(jc, jk, jb);
- z_b(6) = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
+ z_b[6] = p_cc_view(iidx(jc, jb, 6), jk, iblk(jc, jb, 6)) -
p_cc_view(jc, jk, jb);
- z_b(7) = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
+ z_b[7] = p_cc_view(iidx(jc, jb, 7), jk, iblk(jc, jb, 7)) -
p_cc_view(jc, jk, jb);
- z_b(8) = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
+ z_b[8] = p_cc_view(iidx(jc, jb, 8), jk, iblk(jc, jb, 8)) -
p_cc_view(jc, jk, jb);
p_coeff_view(9, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 8, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 8, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 8, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 8, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 8, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 8, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 8, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 8, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 8, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 8, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 8, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 8, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 8, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 8, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 8, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 8, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 8, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 8, 8, jb) * z_b[8];
p_coeff_view(8, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 7, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 7, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 7, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 7, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 7, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 7, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 7, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 7, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 7, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 7, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 7, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 7, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 7, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 7, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 7, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 7, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 7, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 7, 8, jb) * z_b[8];
p_coeff_view(7, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 6, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 6, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 6, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 6, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 6, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 6, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 6, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 6, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 6, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 6, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 6, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 6, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 6, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 6, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 6, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 6, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 6, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 6, 8, jb) * z_b[8];
p_coeff_view(6, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 5, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 5, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 5, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 5, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 5, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 5, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 5, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 5, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 5, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 5, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 5, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 5, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 5, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 5, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 5, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 5, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 5, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 5, 8, jb) * z_b[8];
p_coeff_view(5, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 4, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 4, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 4, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 4, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 4, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 4, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 4, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 4, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 4, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 4, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 4, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 4, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 4, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 4, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 4, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 4, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 4, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 4, 8, jb) * z_b[8];
p_coeff_view(4, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 3, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 3, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 3, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 3, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 3, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 3, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 3, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 3, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 3, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 3, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 3, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 3, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 3, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 3, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 3, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 3, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 3, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 3, 8, jb) * z_b[8];
p_coeff_view(3, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 2, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 2, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 2, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 2, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 2, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 2, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 2, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 2, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 2, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 2, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 2, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 2, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 2, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 2, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 2, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 2, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 2, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 2, 8, jb) * z_b[8];
p_coeff_view(2, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 1, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 1, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 1, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 1, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 1, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 1, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 1, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 1, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 1, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 1, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 1, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 1, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 1, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 1, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 1, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 1, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 1, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 1, 8, jb) * z_b[8];
p_coeff_view(1, jc, jk, jb) =
- lsq_pseudoinv_view(jc, 0, 0, jb) * z_b(0) +
- lsq_pseudoinv_view(jc, 0, 1, jb) * z_b(1) +
- lsq_pseudoinv_view(jc, 0, 2, jb) * z_b(2) +
- lsq_pseudoinv_view(jc, 0, 3, jb) * z_b(3) +
- lsq_pseudoinv_view(jc, 0, 4, jb) * z_b(4) +
- lsq_pseudoinv_view(jc, 0, 5, jb) * z_b(5) +
- lsq_pseudoinv_view(jc, 0, 6, jb) * z_b(6) +
- lsq_pseudoinv_view(jc, 0, 7, jb) * z_b(7) +
- lsq_pseudoinv_view(jc, 0, 8, jb) * z_b(8);
+ lsq_pseudoinv_view(jc, 0, 0, jb) * z_b[0] +
+ lsq_pseudoinv_view(jc, 0, 1, jb) * z_b[1] +
+ lsq_pseudoinv_view(jc, 0, 2, jb) * z_b[2] +
+ lsq_pseudoinv_view(jc, 0, 3, jb) * z_b[3] +
+ lsq_pseudoinv_view(jc, 0, 4, jb) * z_b[4] +
+ lsq_pseudoinv_view(jc, 0, 5, jb) * z_b[5] +
+ lsq_pseudoinv_view(jc, 0, 6, jb) * z_b[6] +
+ lsq_pseudoinv_view(jc, 0, 7, jb) * z_b[7] +
+ lsq_pseudoinv_view(jc, 0, 8, jb) * z_b[8];
p_coeff_view(0, jc, jk, jb) =
p_cc_view(jc, jk, jb) -
p_coeff_view(1, jc, jk, jb) * lsq_moments_view(jc, jb, 0) -
@@ -936,13 +923,13 @@ void div3d(const T *vec_e, const int *cell_edge_idx, const int *cell_edge_blk,
UnmanagedConstT3D geofac_div_view(geofac_div, nproma, 3, nblks_c);
UnmanagedT3D div_vec_c_view(div_vec_c, nproma, nlev, nblks_c);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div3d_inner", innerPolicy, KOKKOS_LAMBDA(const int jk, const int jc) {
div_vec_c_view(jc, jk, jb) =
@@ -984,13 +971,13 @@ void div3d_2field(const T *vec_e, const int *cell_edge_idx,
UnmanagedConstT3D in2_view(in2, nproma, nlev, nblks_e);
UnmanagedT3D out2_view(out2, nproma, nlev, nblks_c);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div3d_2field_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1027,8 +1014,6 @@ void div4d(const int *cell_edge_idx, const int *cell_edge_blk,
typedef Kokkos::View<const T ****, Kokkos::LayoutLeft,
Kokkos::MemoryUnmanaged>
UnmanagedConstT4D;
- typedef Kokkos::View<T ***, Kokkos::LayoutLeft, Kokkos::MemoryUnmanaged>
- UnmanagedT3D;
typedef Kokkos::View<T ****, Kokkos::LayoutLeft, Kokkos::MemoryUnmanaged>
UnmanagedT4D;
typedef Kokkos::View<const int ***, Kokkos::LayoutLeft,
@@ -1043,14 +1028,14 @@ void div4d(const int *cell_edge_idx, const int *cell_edge_blk,
UnmanagedConstT4D f4din_view(f4din, nproma, nlev, nblks_e, dim4d);
UnmanagedT4D f4dout_view(f4dout, nproma, nlev, nblks_c, dim4d);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
for (int ji = 0; ji < dim4d; ++ji) {
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev[ji], i_startidx},
- {elev[ji], i_endidx});
+ {elev[ji] + 1, i_endidx + 1});
Kokkos::parallel_for(
"div4d_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1093,8 +1078,8 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
UnmanagedConstInt3D ieidx(cell_edge_idx, nproma, nblks_c, 3);
UnmanagedConstInt3D ieblk(cell_edge_blk, nproma, nblks_c, 3);
- UnmanagedConstT3D geofac_div_view(geofac_div, nproma, 4, nblks_e);
- UnmanagedConstT3D avg_coeff_view(avg_coeff, nproma, nlev, nblks_c);
+ UnmanagedConstT3D geofac_div_view(geofac_div, nproma, 3, nblks_e);
+ UnmanagedConstT3D avg_coeff_view(avg_coeff, nproma, 4, nblks_c);
UnmanagedT3D div_vec_c_view(div_vec_c, nproma, nlev, nblks_c);
@@ -1108,13 +1093,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
int i_endblk = i_endblk_in[0];
if (l2fields) {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[0], i_endidx_in[0], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step1", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1135,13 +1120,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
});
}
} else {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[0], i_endidx_in[0], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step2", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1160,13 +1145,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
i_startblk = i_startblk_in[1];
i_endblk = i_endblk_in[1];
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[1], i_endidx_in[1], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step3", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1175,13 +1160,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
}
if (l2fields) {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[1], i_endidx_in[1], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step4", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1195,13 +1180,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
i_endblk = i_endblk_in[2];
if (l2fields) {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[2], i_endidx_in[2], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step5", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1224,13 +1209,13 @@ void div_avg(const T *vec_e, const int *cell_neighbor_idx,
});
}
} else {
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[2], i_endidx_in[2], nproma, jb,
i_startblk, i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"div_avg_step6", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
@@ -1273,13 +1258,13 @@ void rot_vertex_atmos(const T *vec_e, const int *vert_edge_idx,
UnmanagedT3D rot_vec_view(rot_vec, nproma, nlev, nblks_v);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_v_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"rot_vertex_atmos_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jv) {
@@ -1326,13 +1311,13 @@ void rot_vertex_ri(const T *vec_e, const int *vert_edge_idx,
UnmanagedT3D rot_vec_view(rot_vec, nproma, nlev, nblks_v);
- for (int jb = i_startblk; jb < i_endblk; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_v_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
Kokkos::MDRangePolicy<Kokkos::Rank<2>> innerPolicy({slev, i_startidx},
- {elev, i_endidx});
+ {elev + 1, i_endidx + 1});
Kokkos::parallel_for(
"rot_vertex_atmos_inner", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jv) {
diff --git a/src/interpolation/mo_lib_interpolation_scalar.cpp b/src/interpolation/mo_lib_interpolation_scalar.cpp
index 9e4e6c5ab4a3a531cb0cad254cc46b049c5ee58f..6b761dc99ccbbeaa1849b898856c0fc6821a04ff 100644
--- a/src/interpolation/mo_lib_interpolation_scalar.cpp
+++ b/src/interpolation/mo_lib_interpolation_scalar.cpp
@@ -52,7 +52,7 @@ void verts2edges_scalar_lib(const T *p_vertex_in, const int *edge_vertex_idx,
UnmanagedConstT3D coeff_int_view(coeff_int, nproma, 2, nblks_e);
UnmanagedT3D p_edge_out_view(p_edge_out, nproma, nlev, nblks_e);
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
@@ -117,7 +117,7 @@ void cells2edges_scalar_lib(const T *p_cell_in, const int *edge_cell_idx,
i_startblk = i_startblk_in[0];
i_endblk = i_endblk_in[0];
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_e_lib(i_startidx_in[0], i_endidx_in[0], nproma, jb,
@@ -136,10 +136,7 @@ void cells2edges_scalar_lib(const T *p_cell_in, const int *edge_cell_idx,
p_edge_out_view(je, jk, jb) = p_cell_in_view(
iidx_view(je, jb, 1), jk, iblk_view(je, jb, 1));
} else {
- std::cerr << "mo_interpolation:cells2edges_scalar_lib: error in "
- "lateral boundary filling"
- << std::endl;
- std::exit(EXIT_FAILURE);
+ Kokkos::abort("mo_interpolation:cells2edges_scalar_lib: error in lateral boundary filling");
}
});
Kokkos::fence();
@@ -150,7 +147,7 @@ void cells2edges_scalar_lib(const T *p_cell_in, const int *edge_cell_idx,
i_startblk = i_startblk_in[1];
i_endblk = i_endblk_in[1];
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_e_lib(i_startidx_in[1], i_endidx_in[1], nproma, jb,
@@ -207,7 +204,7 @@ void edges2verts_scalar_lib(const T *p_edge_in, const int *vert_edge_idx,
UnmanagedConstT3D v_int_view(v_int, nproma, 6, nblks_v);
UnmanagedT3D p_vert_out_view(p_vert_out, nproma, nlev, nblks_v);
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_v_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
@@ -280,7 +277,7 @@ void edges2cells_scalar_lib(const T *p_edge_in, const int *edge_idx,
int i_startidx, i_endidx;
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
@@ -348,7 +345,7 @@ void cells2verts_scalar_lib(const T *p_cell_in, const int *vert_cell_idx,
int i_startidx, i_endidx;
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
get_indices_v_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
@@ -429,7 +426,7 @@ void cells2verts_scalar_ri_lib(const T *p_cell_in, const int *vert_cell_idx,
int i_startidx, i_endidx;
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
get_indices_v_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
@@ -571,15 +568,15 @@ void cell_avg_lib(const T *psi_c, const int *cell_neighbor_idx,
// block indices of triangles next to each cell, dim: (nproma,nblks_c, 3)
UnmanagedConstInt3D iblk_view(cell_neighbor_blk, nproma, nblks_c,
3); // cell_neighbour_blk
- // averaging coefficients, dim: (nproma,nlev,nblks_c)
- UnmanagedConstT3D avg_coeff_view(avg_coeff, nproma, nlev, nblks_c);
+ // averaging coefficients, dim: (nproma, 4, nblks_c)
+ UnmanagedConstT3D avg_coeff_view(avg_coeff, nproma, 4, nblks_c);
// cell based variable after averaging, dim: (nproma,nlev,nblks_c)
UnmanagedT3D avg_psi_c_view(avg_psi_c, nproma, nlev, nblks_c);
int i_startidx, i_endidx;
- for (int jb = i_startblk; jb < i_endblk + 1; ++jb) {
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
get_indices_c_lib(i_startidx_in, i_endidx_in, nproma, jb, i_startblk,
i_endblk, i_startidx, i_endidx);
diff --git a/src/interpolation/mo_lib_intp_rbf.cpp b/src/interpolation/mo_lib_intp_rbf.cpp
index d1178a65397571818db4104ca546ff67eb35d01e..ce6e238f9875ba1dd1dd6621d5f8b2443f56470b 100644
--- a/src/interpolation/mo_lib_intp_rbf.cpp
+++ b/src/interpolation/mo_lib_intp_rbf.cpp
@@ -180,62 +180,62 @@ void rbf_interpol_c2grad_lib(const T *p_cell_in, const int *rbf_c2grad_idx,
"rbf_interpol_c2grad", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
grad_x_view(jc, jk, jb) =
- rbf_c2grad_coeff_view(0, 1, jc, jb) * p_cell_in_view(jc, jk, jb) +
- rbf_c2grad_coeff_view(1, 1, jc, jb) *
+ rbf_c2grad_coeff_view(0, 0, jc, jb) * p_cell_in_view(jc, jk, jb) +
+ rbf_c2grad_coeff_view(1, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(1, jc, jb), jk,
rbf_c2grad_blk_view(1, jc, jb)) +
- rbf_c2grad_coeff_view(2, 1, jc, jb) *
+ rbf_c2grad_coeff_view(2, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(2, jc, jb), jk,
rbf_c2grad_blk_view(2, jc, jb)) +
- rbf_c2grad_coeff_view(3, 1, jc, jb) *
+ rbf_c2grad_coeff_view(3, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(3, jc, jb), jk,
rbf_c2grad_blk_view(3, jc, jb)) +
- rbf_c2grad_coeff_view(4, 1, jc, jb) *
+ rbf_c2grad_coeff_view(4, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(4, jc, jb), jk,
rbf_c2grad_blk_view(4, jc, jb)) +
- rbf_c2grad_coeff_view(5, 1, jc, jb) *
+ rbf_c2grad_coeff_view(5, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(5, jc, jb), jk,
rbf_c2grad_blk_view(5, jc, jb)) +
- rbf_c2grad_coeff_view(6, 1, jc, jb) *
+ rbf_c2grad_coeff_view(6, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(6, jc, jb), jk,
rbf_c2grad_blk_view(6, jc, jb)) +
- rbf_c2grad_coeff_view(7, 1, jc, jb) *
+ rbf_c2grad_coeff_view(7, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(7, jc, jb), jk,
rbf_c2grad_blk_view(7, jc, jb)) +
- rbf_c2grad_coeff_view(8, 1, jc, jb) *
+ rbf_c2grad_coeff_view(8, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(8, jc, jb), jk,
rbf_c2grad_blk_view(8, jc, jb)) +
- rbf_c2grad_coeff_view(9, 1, jc, jb) *
+ rbf_c2grad_coeff_view(9, 0, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(9, jc, jb), jk,
rbf_c2grad_blk_view(9, jc, jb));
grad_y_view(jc, jk, jb) =
- rbf_c2grad_coeff_view(0, 2, jc, jb) * p_cell_in_view(jc, jk, jb) +
- rbf_c2grad_coeff_view(1, 2, jc, jb) *
+ rbf_c2grad_coeff_view(0, 1, jc, jb) * p_cell_in_view(jc, jk, jb) +
+ rbf_c2grad_coeff_view(1, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(1, jc, jb), jk,
rbf_c2grad_blk_view(1, jc, jb)) +
- rbf_c2grad_coeff_view(2, 2, jc, jb) *
+ rbf_c2grad_coeff_view(2, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(2, jc, jb), jk,
rbf_c2grad_blk_view(2, jc, jb)) +
- rbf_c2grad_coeff_view(3, 2, jc, jb) *
+ rbf_c2grad_coeff_view(3, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(3, jc, jb), jk,
rbf_c2grad_blk_view(3, jc, jb)) +
- rbf_c2grad_coeff_view(4, 2, jc, jb) *
+ rbf_c2grad_coeff_view(4, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(4, jc, jb), jk,
rbf_c2grad_blk_view(4, jc, jb)) +
- rbf_c2grad_coeff_view(5, 2, jc, jb) *
+ rbf_c2grad_coeff_view(5, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(5, jc, jb), jk,
rbf_c2grad_blk_view(5, jc, jb)) +
- rbf_c2grad_coeff_view(6, 2, jc, jb) *
+ rbf_c2grad_coeff_view(6, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(6, jc, jb), jk,
rbf_c2grad_blk_view(6, jc, jb)) +
- rbf_c2grad_coeff_view(7, 2, jc, jb) *
+ rbf_c2grad_coeff_view(7, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(7, jc, jb), jk,
rbf_c2grad_blk_view(7, jc, jb)) +
- rbf_c2grad_coeff_view(8, 2, jc, jb) *
+ rbf_c2grad_coeff_view(8, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(8, jc, jb), jk,
rbf_c2grad_blk_view(8, jc, jb)) +
- rbf_c2grad_coeff_view(9, 2, jc, jb) *
+ rbf_c2grad_coeff_view(9, 1, jc, jb) *
p_cell_in_view(rbf_c2grad_idx_view(9, jc, jb), jk,
rbf_c2grad_blk_view(9, jc, jb));
});
@@ -270,10 +270,10 @@ void rbf_vec_interpol_cell_lib(const T *p_vn_in, const int *rbf_vec_idx_c,
nblks_c);
UnmanagedConstInt3D rbf_vec_blk_c_view(rbf_vec_blk_c, rbf_vec_dim_c, nproma,
nblks_c);
- UnmanagedConstT4D rbf_vec_coeff_c_view(rbf_vec_coeff_c, nproma,
- nblks_c); // TODO
+ UnmanagedConstT4D rbf_vec_coeff_c_view(rbf_vec_coeff_c, rbf_vec_dim_c, 2, nproma,
+ nblks_c);
UnmanagedT3D p_u_out_view(p_u_out, nproma, nlev, nblks_c);
- UnmanagedT3D p_v_out_view(p_u_out, nproma, nlev, nblks_c);
+ UnmanagedT3D p_v_out_view(p_v_out, nproma, nlev, nblks_c);
for (int jb = i_startblk; jb <= i_endblk; ++jb) {
@@ -288,60 +288,60 @@ void rbf_vec_interpol_cell_lib(const T *p_vn_in, const int *rbf_vec_idx_c,
"rbf_vec_interpol_cell_lib", innerPolicy,
KOKKOS_LAMBDA(const int jk, const int jc) {
p_u_out_view(jc, jk, jb) =
- rbf_vec_coeff_c_view(0, 1, jc, jb) *
+ rbf_vec_coeff_c_view(0, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(0, jc, jb), jk,
rbf_vec_blk_c_view(0, jc, jb)) +
- rbf_vec_coeff_c_view(1, 1, jc, jb) *
+ rbf_vec_coeff_c_view(1, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(1, jc, jb), jk,
rbf_vec_blk_c_view(1, jc, jb)) +
- rbf_vec_coeff_c_view(2, 1, jc, jb) *
+ rbf_vec_coeff_c_view(2, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(2, jc, jb), jk,
rbf_vec_blk_c_view(2, jc, jb)) +
- rbf_vec_coeff_c_view(3, 1, jc, jb) *
+ rbf_vec_coeff_c_view(3, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(3, jc, jb), jk,
rbf_vec_blk_c_view(3, jc, jb)) +
- rbf_vec_coeff_c_view(4, 1, jc, jb) *
+ rbf_vec_coeff_c_view(4, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(4, jc, jb), jk,
rbf_vec_blk_c_view(4, jc, jb)) +
- rbf_vec_coeff_c_view(5, 1, jc, jb) *
+ rbf_vec_coeff_c_view(5, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(5, jc, jb), jk,
rbf_vec_blk_c_view(5, jc, jb)) +
- rbf_vec_coeff_c_view(6, 1, jc, jb) *
+ rbf_vec_coeff_c_view(6, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(6, jc, jb), jk,
rbf_vec_blk_c_view(6, jc, jb)) +
- rbf_vec_coeff_c_view(7, 1, jc, jb) *
+ rbf_vec_coeff_c_view(7, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(7, jc, jb), jk,
rbf_vec_blk_c_view(7, jc, jb)) +
- rbf_vec_coeff_c_view(8, 1, jc, jb) *
+ rbf_vec_coeff_c_view(8, 0, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(8, jc, jb), jk,
rbf_vec_blk_c_view(8, jc, jb));
p_v_out_view(jc, jk, jb) =
- rbf_vec_coeff_c_view(0, 2, jc, jb) *
+ rbf_vec_coeff_c_view(0, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(0, jc, jb), jk,
rbf_vec_blk_c_view(0, jc, jb)) +
- rbf_vec_coeff_c_view(1, 2, jc, jb) *
+ rbf_vec_coeff_c_view(1, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(1, jc, jb), jk,
rbf_vec_blk_c_view(1, jc, jb)) +
- rbf_vec_coeff_c_view(2, 2, jc, jb) *
+ rbf_vec_coeff_c_view(2, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(2, jc, jb), jk,
rbf_vec_blk_c_view(2, jc, jb)) +
- rbf_vec_coeff_c_view(3, 2, jc, jb) *
+ rbf_vec_coeff_c_view(3, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(3, jc, jb), jk,
rbf_vec_blk_c_view(3, jc, jb)) +
- rbf_vec_coeff_c_view(4, 2, jc, jb) *
+ rbf_vec_coeff_c_view(4, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(4, jc, jb), jk,
rbf_vec_blk_c_view(4, jc, jb)) +
- rbf_vec_coeff_c_view(5, 2, jc, jb) *
+ rbf_vec_coeff_c_view(5, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(5, jc, jb), jk,
rbf_vec_blk_c_view(5, jc, jb)) +
- rbf_vec_coeff_c_view(6, 2, jc, jb) *
+ rbf_vec_coeff_c_view(6, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(6, jc, jb), jk,
rbf_vec_blk_c_view(6, jc, jb)) +
- rbf_vec_coeff_c_view(7, 2, jc, jb) *
+ rbf_vec_coeff_c_view(7, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(7, jc, jb), jk,
rbf_vec_blk_c_view(7, jc, jb)) +
- rbf_vec_coeff_c_view(8, 2, jc, jb) *
+ rbf_vec_coeff_c_view(8, 1, jc, jb) *
p_vn_in_view(rbf_vec_idx_c_view(8, jc, jb), jk,
rbf_vec_blk_c_view(8, jc, jb));
});
diff --git a/src/support/mo_lib_loopindices.cpp b/src/support/mo_lib_loopindices.cpp
index 30c82bd2e98521f99b09abf9343ee1a5b52f6185..fcc31b6c2a6187f68c431c0f840b42f10fdd154b 100644
--- a/src/support/mo_lib_loopindices.cpp
+++ b/src/support/mo_lib_loopindices.cpp
@@ -12,21 +12,26 @@
#include <algorithm> // For std::max
// get_indices_c_lib function
-void get_indices_c_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
+void get_indices_c_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
const int i_blk, const int i_startblk, const int i_endblk,
int &i_startidx_out, int &i_endidx_out, const bool called_from_cpp=true) {
-
+
//Since code is ported incrementally from Fortran to C++, depending on where the function is called from
//(either fortran or c++), the first index should be either 0 or 1.
int first_index;
- if (called_from_cpp)
+ int i_endidx_loc;
+ if (called_from_cpp){
first_index = 0;
- else
- first_index = 1;
-
+ i_endidx_loc = nproma - 1;
+ }
+ else {
+ first_index = 1;
+ i_endidx_loc = nproma;
+ }
+
if (i_blk == i_startblk) {
i_startidx_out = std::max(first_index, i_startidx_in);
- i_endidx_out = nproma;
+ i_endidx_out = i_endidx_loc;
if (i_blk == i_endblk) {
i_endidx_out = i_endidx_in;
}
@@ -35,43 +40,53 @@ void get_indices_c_lib(const int i_startidx_in, const int i_endidx_in, const int
i_endidx_out = i_endidx_in;
} else {
i_startidx_out = first_index;
- i_endidx_out = nproma;
+ i_endidx_out = i_endidx_loc;
}
}
// get_indices_e_lib function
-void get_indices_e_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
+void get_indices_e_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
const int i_blk, const int i_startblk, const int i_endblk,
int &i_startidx_out, int &i_endidx_out, const bool called_from_cpp=true) {
-
- //Since code is ported incrementally from Fortran to C++, depending on where the function is called from,
+
+ //Since code is ported incrementally from Fortran to C++, depending on where the function is called from,
//the first index should be either 0 or 1.
int first_index;
- if (called_from_cpp)
+ int i_endidx_loc;
+ if (called_from_cpp) {
first_index = 0;
- else
+ i_endidx_loc = nproma - 1;
+ }
+ else {
first_index = 1;
+ i_endidx_loc = nproma;
+ }
i_startidx_out = (i_blk != i_startblk) ? first_index : std::max(first_index, i_startidx_in);
- i_endidx_out = (i_blk != i_endblk) ? nproma : i_endidx_in;
+ i_endidx_out = (i_blk != i_endblk) ? i_endidx_loc : i_endidx_in;
}
// get_indices_v_lib function
-void get_indices_v_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
+void get_indices_v_lib(const int i_startidx_in, const int i_endidx_in, const int nproma,
const int i_blk, const int i_startblk, const int i_endblk,
int &i_startidx_out, int &i_endidx_out, const bool called_from_cpp=true) {
-
- //Since code is ported incrementally from Fortran to C++, depending on where the function is called from,
+
+ //Since code is ported incrementally from Fortran to C++, depending on where the function is called from,
//the first index should be either 0 or 1.
int first_index;
- if (called_from_cpp)
+ int i_endidx_loc;
+ if (called_from_cpp) {
first_index = 0;
- else
+ i_endidx_loc = nproma - 1;
+ }
+ else {
first_index = 1;
+ i_endidx_loc = nproma;
+ }
if (i_blk == i_startblk) {
i_startidx_out = i_startidx_in;
- i_endidx_out = nproma;
+ i_endidx_out = i_endidx_loc;
if (i_blk == i_endblk) {
i_endidx_out = i_endidx_in;
}
@@ -80,6 +95,6 @@ void get_indices_v_lib(const int i_startidx_in, const int i_endidx_in, const int
i_endidx_out = i_endidx_in;
} else {
i_startidx_out = first_index;
- i_endidx_out = nproma;
+ i_endidx_out = i_endidx_loc;
}
-}
\ No newline at end of file
+}
diff --git a/test/c/CMakeLists.txt b/test/c/CMakeLists.txt
index 90ab1e3d5700b8655779340d97f6cbd3650296c6..c0f7c59a86ae3afe9187dab588239e6983ae0d9b 100644
--- a/test/c/CMakeLists.txt
+++ b/test/c/CMakeLists.txt
@@ -26,10 +26,6 @@ message(CHECK_PASS "done")
# Find Kokkos (or use your existing Kokkos installation)
# find_package(Kokkos REQUIRED)
-if(IM_ENABLE_LOOP_EXCHANGE)
- target_compile_definitions(iconmath-interpolation PRIVATE __LOOP_EXCHANGE)
-endif()
-
set(SOURCES
main.cpp
test_horizontal_div.cpp
@@ -43,6 +39,10 @@ set(SOURCES
# Create the test executable from your test files, including main.cpp.
add_executable(iconmath_test_c ${SOURCES})
+if(IM_ENABLE_LOOP_EXCHANGE)
+ target_compile_definitions(iconmath_test_c PRIVATE __LOOP_EXCHANGE)
+endif()
+
target_include_directories(iconmath_test_c PRIVATE ${CMAKE_CURRENT_SOURCE_DIR})
# Link the test executable with GoogleTest and Kokkos.
diff --git a/test/c/test_horizontal_div.cpp b/test/c/test_horizontal_div.cpp
index 596d19e708d6c6e466fa2f23c579cb51f8689e19..ef95f6b72ab4f8ce8ca95dffd1e3f9f21c6f83b9 100644
--- a/test/c/test_horizontal_div.cpp
+++ b/test/c/test_horizontal_div.cpp
@@ -14,8 +14,8 @@
#include <vector>
#include <Kokkos_Core.hpp>
-#include <gtest/gtest.h>
#include <dim_helper.hpp>
+#include <gtest/gtest.h>
#include <horizontal/mo_lib_divrot.hpp>
#include <support/mo_lib_loopindices.hpp>
@@ -29,50 +29,53 @@ protected:
static constexpr int dim4d = 2; // 4th dimension size
int i_startblk = 0;
- int i_endblk = nblks_c; // Test blocks [0 .. nblks_c-1]
+ int i_endblk = nblks_c - 1; // Test blocks [0 .. nblks_c-1]
int i_startidx_in = 0;
- int i_endidx_in = nproma; // Full range: 0 .. nproma-1
+ int i_endidx_in = nproma - 1; // Full range: 0 .. nproma-1
std::vector<int> slev;
std::vector<int> elev;
bool lacc = false; // Not using ACC-specific behavior.
- std::vector<ValueType> vec_e;
- std::vector<int> cell_edge_idx;
- std::vector<int> cell_edge_blk;
- std::vector<ValueType> geofac_div;
- std::vector<ValueType> div_vec_c;
- std::vector<ValueType> f4din;
- std::vector<ValueType> f4dout;
+ // Here we allocate Kokkos::View objects in a memory space that is directly
+ // accessible from both the host and device
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Views for the test data. All the data is assigned as one-dimensional arrays
+ Kokkos::View<ValueType *, memory_space> vec_e;
+ Kokkos::View<int *, memory_space> cell_edge_idx;
+ Kokkos::View<int *, memory_space> cell_edge_blk;
+ Kokkos::View<ValueType *, memory_space> geofac_div;
+ Kokkos::View<ValueType *, memory_space> div_vec_c;
+ Kokkos::View<ValueType *, memory_space> f4din;
+ Kokkos::View<ValueType *, memory_space> f4dout;
// Followings are needed in HorizontalDivAvgTest
- std::vector<int> cell_neighbor_idx;
- std::vector<int> cell_neighbor_blk;
- std::vector<ValueType> avg_coeff;
- std::vector<ValueType> opt_in2;
- std::vector<ValueType> opt_out2;
-
- HorizontalDivTest() {
- slev.resize(dim4d, 0);
- elev.resize(dim4d, nlev); // Full vertical range (0 .. nlev-1)
-
- vec_e.resize(dim_combine(nproma, nlev, nblks_e));
- cell_edge_idx.resize(dim_combine(nproma, nblks_c, 3));
- cell_edge_blk.resize(dim_combine(nproma, nblks_c, 3));
- geofac_div.resize(dim_combine(nproma, 3, nblks_c));
- div_vec_c.resize(dim_combine(nproma, nlev, nblks_c));
- f4din.resize(dim_combine(nproma, nlev, nblks_e, dim4d));
- f4dout.resize(dim_combine(nproma, nlev, nblks_c, dim4d));
- cell_neighbor_idx.resize(dim_combine(nproma, nblks_c, 3));
- cell_neighbor_blk.resize(dim_combine(nproma, nblks_c, 3));
- avg_coeff.resize(dim_combine(nproma, 4, nblks_c));
- opt_in2.resize(dim_combine(nproma, nlev, nblks_e));
- opt_out2.resize(dim_combine(nproma, nlev, nblks_c));
- }
+ Kokkos::View<int *, memory_space> cell_neighbor_idx;
+ Kokkos::View<int *, memory_space> cell_neighbor_blk;
+ Kokkos::View<ValueType *, memory_space> avg_coeff;
+ Kokkos::View<ValueType *, memory_space> opt_in2;
+ Kokkos::View<ValueType *, memory_space> opt_out2;
+
+ HorizontalDivTest()
+ : slev(dim4d, 0),
+ elev(dim4d, nlev - 1), // Full vertical range (0 .. nlev-1)
+ vec_e("vec_e", dim_combine(nproma, nlev, nblks_e)),
+ cell_edge_idx("cell_edge_idx", dim_combine(nproma, nblks_c, 3)),
+ cell_edge_blk("cell_edge_blk", dim_combine(nproma, nblks_c, 3)),
+ geofac_div("geofac_div", dim_combine(nproma, 3, nblks_c)),
+ div_vec_c("div_vec_c", dim_combine(nproma, nlev, nblks_c)),
+ f4din("f4din", dim_combine(nproma, nlev, nblks_e, dim4d)),
+ f4dout("f4dout", dim_combine(nproma, nlev, nblks_c, dim4d)),
+ cell_neighbor_idx("cell_neighbor_idx", dim_combine(nproma, nblks_c, 3)),
+ cell_neighbor_blk("cell_neighbor_blk", dim_combine(nproma, nblks_c, 3)),
+ avg_coeff("avg_coeff", dim_combine(nproma, 4, nblks_c)),
+ opt_in2("opt_in2", dim_combine(nproma, nlev, nblks_e)),
+ opt_out2("opt_out2", dim_combine(nproma, nlev, nblks_c)) {}
};
/// ValueTypes which the divrot tests should run with
typedef ::testing::Types<float, double> ValueTypes;
-
TYPED_TEST_SUITE(HorizontalDivTest, ValueTypes);
TYPED_TEST(HorizontalDivTest, TestDiv3DSpecific) {
@@ -86,34 +89,46 @@ TYPED_TEST(HorizontalDivTest, TestDiv3DSpecific) {
const auto &geofac_div_at = at<nproma, 3, nblks_c>;
const auto &div_vec_c_at = at<nproma, nlev, nblks_c>;
- // Initialization with specific values
+ // create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+
+ // Initialize the arrays with the same patterns as before.
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
+ vec_e_h[vec_e_at(i, k, 0)] = static_cast<TypeParam>((i + 1) * (k + 1));
}
-
// Set edge indices to point to specific cells (including self)
- this->cell_edge_idx[cell_edge_at(i, 0, 0)] = i;
- this->cell_edge_idx[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_edge_idx[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 0)] = i;
+ cell_edge_idx_h[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
// All edges are in the same block for this test
for (int j = 0; j < 3; ++j) {
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
}
- // Geometric factors
- this->geofac_div[geofac_div_at(i, 0, 0)] = 0.5;
- this->geofac_div[geofac_div_at(i, 1, 0)] = 0.3;
- this->geofac_div[geofac_div_at(i, 2, 0)] = 0.2;
+ geofac_div_h[geofac_div_at(i, 0, 0)] = static_cast<TypeParam>(0.5);
+ geofac_div_h[geofac_div_at(i, 1, 0)] = static_cast<TypeParam>(0.3);
+ geofac_div_h[geofac_div_at(i, 2, 0)] = static_cast<TypeParam>(0.2);
// Initialize div_vec_c to zero
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = 0.0;
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
}
}
-
- // Call the div3d function
+ // Copy the initialized data back to the device memory (or unified memory,
+ // which in some cases may be a no-op if already accessible on the host).
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+
+ // Call the div3d function using the device pointers from the Views.
div3d<TypeParam>(this->vec_e.data(), this->cell_edge_idx.data(),
this->cell_edge_blk.data(), this->geofac_div.data(),
this->div_vec_c.data(), this->i_startblk, this->i_endblk,
@@ -121,12 +136,20 @@ TYPED_TEST(HorizontalDivTest, TestDiv3DSpecific) {
this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 1, 0)], 3.4, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 0, 0)], 2.1, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 1, 0)], 4.2, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 0, 0)], 2.2, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 1, 0)], 4.4, 1e-6);
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(0, 0, 0)), static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(0, 1, 0)), static_cast<TypeParam>(3.4),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(1, 0, 0)), static_cast<TypeParam>(2.1),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(1, 1, 0)), static_cast<TypeParam>(4.2),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(2, 0, 0)), static_cast<TypeParam>(2.2),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h(div_vec_c_at(2, 1, 0)), static_cast<TypeParam>(4.4),
+ 1e-6);
}
TYPED_TEST(HorizontalDivTest, TestDiv3DRandom) {
@@ -140,37 +163,44 @@ TYPED_TEST(HorizontalDivTest, TestDiv3DRandom) {
const auto &geofac_div_at = at<nproma, 3, nblks_c>;
const auto &div_vec_c_at = at<nproma, nlev, nblks_c>;
- // Set up random number generators
+ // create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+
+ // Initialize the arrays with random values.
std::random_device rd;
std::mt19937 gen(rd());
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
- // Initialization with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
+ vec_e_h[vec_e_at(i, k, 0)] = real_distrib(gen);
}
- // Set random edge indices
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = int_distrib(gen);
- this->cell_edge_blk[cell_edge_at(i, 0, j)] =
- 0; // Keep in same block for simplicity
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = int_distrib(gen);
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
}
- // Random geometric factors
for (int j = 0; j < 3; ++j) {
- this->geofac_div[geofac_div_at(i, j, 0)] = real_distrib(gen);
+ geofac_div_h[geofac_div_at(i, j, 0)] = real_distrib(gen);
}
- // Initialize div_vec_c to random values
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = real_distrib(gen);
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = real_distrib(gen);
}
}
- // Call the div3d function
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+
div3d<TypeParam>(this->vec_e.data(), this->cell_edge_idx.data(),
this->cell_edge_blk.data(), this->geofac_div.data(),
this->div_vec_c.data(), this->i_startblk, this->i_endblk,
@@ -178,38 +208,40 @@ TYPED_TEST(HorizontalDivTest, TestDiv3DRandom) {
this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+
// Calculate reference values separately and verify results
std::vector<TypeParam> ref_div_vec_c(nproma * nlev * nblks_c, 0.0);
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)])] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)])] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)])] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)])] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)])] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)])] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(i, k, 0)],
- ref_div_vec_c[div_vec_c_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(i, k, 0)],
+ ref_div_vec_c[div_vec_c_at(i, k, 0)], tol)
<< "Results differ at i=" << i << ", k=" << k;
}
}
@@ -229,36 +261,55 @@ TYPED_TEST(HorizontalDivTest, TestDiv3D2FSpecific) {
const auto &f4d_at = at<nproma, nlev, nblks_e, dim4d>;
const auto &f4dout_at = at<nproma, nlev, nblks_c, dim4d>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto f4din_h = Kokkos::create_mirror_view(this->f4din);
+ auto f4dout_h = Kokkos::create_mirror_view(this->f4dout);
+
// Initialization with specific values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
- this->f4din[f4d_at(i, k, 0, 0)] =
- (i + 1) * (k + 2); // Different pattern for second field
+ vec_e_h[vec_e_at(i, k, 0)] =
+ static_cast<TypeParam>((i + 1) * (k + 1)); // Simple pattern
+ f4din_h[f4d_at(i, k, 0, 0)] = static_cast<TypeParam>(
+ (i + 1) * (k + 2)); // Different pattern for second field
}
// Set edge indices to point to specific cells (including self)
- this->cell_edge_idx[cell_edge_at(i, 0, 0)] = i;
- this->cell_edge_idx[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_edge_idx[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 0)] = i;
+ cell_edge_idx_h[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
// All edges are in the same block for this test
for (int j = 0; j < 3; ++j) {
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
}
// Geometric factors
- this->geofac_div[geofac_div_at(i, 0, 0)] = 0.5;
- this->geofac_div[geofac_div_at(i, 1, 0)] = 0.3;
- this->geofac_div[geofac_div_at(i, 2, 0)] = 0.2;
+ geofac_div_h[geofac_div_at(i, 0, 0)] = static_cast<TypeParam>(0.5);
+ geofac_div_h[geofac_div_at(i, 1, 0)] = static_cast<TypeParam>(0.3);
+ geofac_div_h[geofac_div_at(i, 2, 0)] = static_cast<TypeParam>(0.2);
// Initialize div_vec_c and f4dout to zero
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = 0.0;
- this->f4dout[f4dout_at(i, k, 0, 0)] = 0.0;
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
+ f4dout_h[f4dout_at(i, k, 0, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->f4din, f4din_h);
+ Kokkos::deep_copy(this->f4dout, f4dout_h);
+
// Call the div3d_2field function
div3d_2field<TypeParam>(this->vec_e.data(), this->cell_edge_idx.data(),
this->cell_edge_blk.data(), this->geofac_div.data(),
@@ -268,21 +319,37 @@ TYPED_TEST(HorizontalDivTest, TestDiv3D2FSpecific) {
this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(f4dout_h, this->f4dout);
+
// Check first field (same as in div3d test)
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 1, 0)], 3.4, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 0, 0)], 2.1, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 1, 0)], 4.2, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 0, 0)], 2.2, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 1, 0)], 4.4, 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 0, 0)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 1, 0)], static_cast<TypeParam>(3.4),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 0, 0)], static_cast<TypeParam>(2.1),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 1, 0)], static_cast<TypeParam>(4.2),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 0, 0)], static_cast<TypeParam>(2.2),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 1, 0)], static_cast<TypeParam>(4.4),
+ 1e-6);
// Check second field (expected values calculated manually)
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 0, 0, 0)], 3.4, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 1, 0, 0)], 5.1, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 0, 0, 0)], 4.2, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 1, 0, 0)], 6.3, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 0, 0, 0)], 4.4, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 1, 0, 0)], 6.6, 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 0, 0, 0)], static_cast<TypeParam>(3.4),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 1, 0, 0)], static_cast<TypeParam>(5.1),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 0, 0, 0)], static_cast<TypeParam>(4.2),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 1, 0, 0)], static_cast<TypeParam>(6.3),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 0, 0, 0)], static_cast<TypeParam>(4.4),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 1, 0, 0)], static_cast<TypeParam>(6.6),
+ 1e-6);
}
TYPED_TEST(HorizontalDivTest, TestDiv3D2FRandom) {
@@ -299,38 +366,56 @@ TYPED_TEST(HorizontalDivTest, TestDiv3D2FRandom) {
const auto &f4d_at = at<nproma, nlev, nblks_e, dim4d>;
const auto &f4dout_at = at<nproma, nlev, nblks_c, dim4d>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto f4din_h = Kokkos::create_mirror_view(this->f4din);
+ auto f4dout_h = Kokkos::create_mirror_view(this->f4dout);
+
// Set up random number generators
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
// Initialization with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
- this->f4din[f4d_at(i, k, 0, 0)] = real_distrib(gen);
+ vec_e_h[vec_e_at(i, k, 0)] = real_distrib(gen);
+ f4din_h[f4d_at(i, k, 0, 0)] = real_distrib(gen);
}
// Set random edge indices
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = int_distrib(gen);
- this->cell_edge_blk[cell_edge_at(i, 0, j)] =
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = int_distrib(gen);
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] =
0; // Keep in same block for simplicity
}
// Random geometric factors
for (int j = 0; j < 3; ++j) {
- this->geofac_div[geofac_div_at(i, j, 0)] = real_distrib(gen);
+ geofac_div_h[geofac_div_at(i, j, 0)] = real_distrib(gen);
}
// Initialize div_vec_c and f4dout to random values
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = real_distrib(gen);
- this->f4dout[f4dout_at(i, k, 0, 0)] = real_distrib(gen);
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = real_distrib(gen);
+ f4dout_h[f4dout_at(i, k, 0, 0)] = real_distrib(gen);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->f4din, f4din_h);
+ Kokkos::deep_copy(this->f4dout, f4dout_h);
+
// Call the div3d_2field function
div3d_2field<TypeParam>(this->vec_e.data(), this->cell_edge_idx.data(),
this->cell_edge_blk.data(), this->geofac_div.data(),
@@ -340,55 +425,56 @@ TYPED_TEST(HorizontalDivTest, TestDiv3D2FRandom) {
this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(f4dout_h, this->f4dout);
+
// Calculate reference values separately and verify results
std::vector<TypeParam> ref_div_vec_c(nproma * nlev * nblks_c, 0.0);
std::vector<TypeParam> ref_f4dout(nproma * nlev * nblks_c * dim4d, 0.0);
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
// Calculate reference value for first field
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)])] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)])] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)])] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)])] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)])] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)])] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
// Calculate reference value for second field
ref_f4dout[f4dout_at(jc, jk, jb, 0)] =
- this->f4din[f4d_at(this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)],
- 0)] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->f4din[f4d_at(this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)],
- 0)] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->f4din[f4d_at(this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)],
- 0)] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ f4din_h[f4d_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)], 0)] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ f4din_h[f4d_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)], 0)] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ f4din_h[f4d_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)], 0)] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results for first field
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(i, k, 0)],
- ref_div_vec_c[div_vec_c_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(i, k, 0)],
+ ref_div_vec_c[div_vec_c_at(i, k, 0)], tol)
<< "First field results differ at i=" << i << ", k=" << k;
}
}
@@ -396,8 +482,8 @@ TYPED_TEST(HorizontalDivTest, TestDiv3D2FRandom) {
// Verify results for second field
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->f4dout[f4dout_at(i, k, 0, 0)],
- ref_f4dout[f4dout_at(i, k, 0, 0)], 1e-5)
+ EXPECT_NEAR(f4dout_h[f4dout_at(i, k, 0, 0)],
+ ref_f4dout[f4dout_at(i, k, 0, 0)], tol)
<< "Second field results differ at i=" << i << ", k=" << k;
}
}
@@ -415,22 +501,37 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DSpecific) {
const auto &f4din_at = at<nproma, nlev, nblks_e, dim4d>;
const auto &f4dout_at = at<nproma, nlev, nblks_c, dim4d>;
+ // Create mirror views to store data on the host
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto f4din_h = Kokkos::create_mirror_view(this->f4din);
+ auto f4dout_h = Kokkos::create_mirror_view(this->f4dout);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = (i + j) % nproma;
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
- this->geofac_div[geofac_div_at(i, j, 0)] = 0.1 * (j + 1);
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = (i + j) % nproma;
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
+ geofac_div_h[geofac_div_at(i, j, 0)] =
+ static_cast<TypeParam>(0.1 * (j + 1));
}
for (int k = 0; k < nlev; ++k) {
for (int d = 0; d < dim4d; ++d) {
- this->f4din[f4din_at(i, k, 0, d)] = 1.0 + i + k + d;
- this->f4dout[f4dout_at(i, k, 0, d)] = 0.0;
+ f4din_h[f4din_at(i, k, 0, d)] = static_cast<TypeParam>(1.0 + i + k + d);
+ f4dout_h[f4dout_at(i, k, 0, d)] = static_cast<TypeParam>(0.0);
}
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->f4din, f4din_h);
+ Kokkos::deep_copy(this->f4dout, f4dout_h);
+
// Test function
div4d<TypeParam>(this->cell_edge_idx.data(), this->cell_edge_blk.data(),
this->geofac_div.data(), this->f4din.data(),
@@ -439,18 +540,33 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DSpecific) {
this->slev.data(), this->elev.data(), this->nproma,
this->lacc, this->nlev, this->nblks_c, this->nblks_e);
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 0, 0, 0)], 1.4, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 0, 0, 0)], 1.1, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 0, 0, 0)], 1.1, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 1, 0, 0)], 2.0, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 1, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 1, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 0, 0, 1)], 2.0, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 0, 0, 1)], 1.7, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 0, 0, 1)], 1.7, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(0, 1, 0, 1)], 2.6, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(1, 1, 0, 1)], 2.3, 1e-6);
- EXPECT_NEAR(this->f4dout[f4dout_at(2, 1, 0, 1)], 2.3, 1e-6);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(f4dout_h, this->f4dout);
+
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 0, 0, 0)], static_cast<TypeParam>(1.4),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 0, 0, 0)], static_cast<TypeParam>(1.1),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 0, 0, 0)], static_cast<TypeParam>(1.1),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 1, 0, 0)], static_cast<TypeParam>(2.0),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 1, 0, 0)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 1, 0, 0)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 0, 0, 1)], static_cast<TypeParam>(2.0),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 0, 0, 1)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 0, 0, 1)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(0, 1, 0, 1)], static_cast<TypeParam>(2.6),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(1, 1, 0, 1)], static_cast<TypeParam>(2.3),
+ 1e-6);
+ EXPECT_NEAR(f4dout_h[f4dout_at(2, 1, 0, 1)], static_cast<TypeParam>(2.3),
+ 1e-6);
}
TYPED_TEST(HorizontalDivTest, TestDiv4DRandom) {
@@ -465,6 +581,13 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DRandom) {
const auto &f4din_at = at<nproma, nlev, nblks_e, dim4d>;
const auto &f4dout_at = at<nproma, nlev, nblks_c, dim4d>;
+ // Create mirror views to store data on the host
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto f4din_h = Kokkos::create_mirror_view(this->f4din);
+ auto f4dout_h = Kokkos::create_mirror_view(this->f4dout);
+
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
@@ -473,19 +596,26 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DRandom) {
// Initialize with random values
for (int i = 0; i < nproma; ++i) {
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = int_distrib(gen);
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
- this->geofac_div[geofac_div_at(i, j, 0)] = real_distrib(gen);
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = int_distrib(gen);
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
+ geofac_div_h[geofac_div_at(i, j, 0)] = real_distrib(gen);
}
for (int k = 0; k < nlev; ++k) {
for (int d = 0; d < dim4d; ++d) {
- this->f4din[f4din_at(i, k, 0, d)] = real_distrib(gen);
- this->f4dout[f4dout_at(i, k, 0, d)] = 0.0;
+ f4din_h[f4din_at(i, k, 0, d)] = real_distrib(gen);
+ f4dout_h[f4dout_at(i, k, 0, d)] = static_cast<TypeParam>(0.0);
}
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->f4din, f4din_h);
+ Kokkos::deep_copy(this->f4dout, f4dout_h);
+
// Test function
div4d<TypeParam>(this->cell_edge_idx.data(), this->cell_edge_blk.data(),
this->geofac_div.data(), this->f4din.data(),
@@ -494,25 +624,30 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DRandom) {
this->slev.data(), this->elev.data(), this->nproma,
this->lacc, this->nlev, this->nblks_c, this->nblks_e);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(f4dout_h, this->f4dout);
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Compute reference result and check
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
for (int ji = 0; ji < dim4d; ++ji) {
- for (int jk = this->slev[ji]; jk < this->elev[ji]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[ji]; jk <= this->elev[ji]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
TypeParam expected = 0.0;
for (int je = 0; je < 3; ++je) {
- expected +=
- this->f4din[f4din_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, je)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, je)], ji)] *
- this->geofac_div[geofac_div_at(jc, je, jb)];
+ expected += f4din_h[f4din_at(
+ cell_edge_idx_h[cell_edge_at(jc, jb, je)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, je)], ji)] *
+ geofac_div_h[geofac_div_at(jc, je, jb)];
}
- EXPECT_NEAR(this->f4dout[f4dout_at(jc, jk, jb, ji)], expected, 1e-5)
+ EXPECT_NEAR(f4dout_h[f4dout_at(jc, jk, jb, ji)], expected, tol)
<< "Random test fails at jc=" << jc << ", jk=" << jk
<< ", jb=" << jb << ", ji=" << ji;
}
@@ -521,77 +656,103 @@ TYPED_TEST(HorizontalDivTest, TestDiv4DRandom) {
}
}
-TYPED_TEST_SUITE(HorizontalDivTest, ValueTypes);
-
TYPED_TEST(HorizontalDivTest, TestDivAvgSpecific) {
constexpr int nproma = this->nproma;
constexpr int nlev = this->nlev;
constexpr int nblks_c = this->nblks_c;
constexpr int nblks_e = this->nblks_e;
- constexpr int dim4d = this->dim4d;
const auto &vec_e_at = at<nproma, nlev, nblks_e>;
const auto &cell_edge_at = at<nproma, nblks_c, 3>;
const auto &geofac_div_at = at<nproma, 3, nblks_c>;
const auto &div_vec_c_at = at<nproma, nlev, nblks_c>;
+ const auto &cell_neighbor_at = at<nproma, nblks_c, 3>;
+ const auto &avg_coeff_at = at<nproma, 4, nblks_c>;
// Vectors for additional parameters
// Vectors for block and index ranges
std::vector<int> i_startblk_in(3, 0);
- std::vector<int> i_endblk_in(3, nblks_c);
+ std::vector<int> i_endblk_in(3, nblks_c - 1);
std::vector<int> i_startidx_in(3, 0);
- std::vector<int> i_endidx_in(3, nproma);
+ std::vector<int> i_endidx_in(3, nproma - 1);
// Parameters for the test
int patch_id = 1;
bool l_limited_area = true;
bool l2fields = true;
- const auto &cell_neighbor_at = at<nproma, nblks_c, 3>;
- const auto &avg_coeff_at = at<nproma, 4, nblks_c>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto opt_in2_h = Kokkos::create_mirror_view(this->opt_in2);
+ auto opt_out2_h = Kokkos::create_mirror_view(this->opt_out2);
// Initialize the vectors with specific values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
- this->opt_in2[vec_e_at(i, k, 0)] =
- (i + 1) * (k + 1) * 0.5; // Half of vec_e
+ vec_e_h[vec_e_at(i, k, 0)] =
+ static_cast<TypeParam>((i + 1) * (k + 1)); // Simple pattern
+ opt_in2_h[vec_e_at(i, k, 0)] =
+ static_cast<TypeParam>((i + 1) * (k + 1) * 0.5); // Half of vec_e
}
// Set edge indices to point to specific cells
- this->cell_edge_idx[cell_edge_at(i, 0, 0)] = i;
- this->cell_edge_idx[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_edge_idx[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 0)] = i;
+ cell_edge_idx_h[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
// Set neighbor indices similarly
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = i;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 2)] = (i + 2) % nproma;
// All edges and neighbors are in the same block for this test
for (int j = 0; j < 3; ++j) {
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
// Geometric factors
- this->geofac_div[geofac_div_at(i, 0, 0)] = 0.5;
- this->geofac_div[geofac_div_at(i, 1, 0)] = 0.3;
- this->geofac_div[geofac_div_at(i, 2, 0)] = 0.2;
+ geofac_div_h[geofac_div_at(i, 0, 0)] = static_cast<TypeParam>(0.5);
+ geofac_div_h[geofac_div_at(i, 1, 0)] = static_cast<TypeParam>(0.3);
+ geofac_div_h[geofac_div_at(i, 2, 0)] = static_cast<TypeParam>(0.2);
// Average coefficients
- this->avg_coeff[avg_coeff_at(i, 0, 0)] = 0.4; // Self
- this->avg_coeff[avg_coeff_at(i, 1, 0)] = 0.2; // First neighbor
- this->avg_coeff[avg_coeff_at(i, 2, 0)] = 0.2; // Second neighbor
- this->avg_coeff[avg_coeff_at(i, 3, 0)] = 0.2; // Third neighbor
+ avg_coeff_h[avg_coeff_at(i, 0, 0)] = static_cast<TypeParam>(0.4); // Self
+ avg_coeff_h[avg_coeff_at(i, 1, 0)] =
+ static_cast<TypeParam>(0.2); // First neighbor
+ avg_coeff_h[avg_coeff_at(i, 2, 0)] =
+ static_cast<TypeParam>(0.2); // Second neighbor
+ avg_coeff_h[avg_coeff_at(i, 3, 0)] =
+ static_cast<TypeParam>(0.2); // Third neighbor
// Initialize div_vec_c and opt_out2 to zero
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = 0.0;
- this->opt_out2[div_vec_c_at(i, k, 0)] = 0.0;
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
+ opt_out2_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->opt_in2, opt_in2_h);
+ Kokkos::deep_copy(this->opt_out2, opt_out2_h);
+
// Call the div_avg function
div_avg<TypeParam>(
this->vec_e.data(), this->cell_neighbor_idx.data(),
@@ -603,19 +764,37 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgSpecific) {
this->nproma, patch_id, l_limited_area, l2fields, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 0, 0)], 1.88, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 1, 0)], 3.76, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 0, 0)], 2.04, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 1, 0)], 4.08, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 0, 0)], 2.08, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 1, 0)], 4.16, 1e-6);
-
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(0, 0, 0)], 0.94, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(0, 1, 0)], 1.88, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(1, 0, 0)], 1.02, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(1, 1, 0)], 2.04, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(2, 0, 0)], 1.04, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(2, 1, 0)], 2.08, 1e-6);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(opt_out2_h, this->opt_out2);
+
+ // Verify first field results
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 0, 0)], static_cast<TypeParam>(1.88),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 1, 0)], static_cast<TypeParam>(3.76),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 0, 0)], static_cast<TypeParam>(2.04),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 1, 0)], static_cast<TypeParam>(4.08),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 0, 0)], static_cast<TypeParam>(2.08),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 1, 0)], static_cast<TypeParam>(4.16),
+ 1e-6);
+
+ // Verify second field results
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(0, 0, 0)], static_cast<TypeParam>(0.94),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(0, 1, 0)], static_cast<TypeParam>(1.88),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(1, 0, 0)], static_cast<TypeParam>(1.02),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(1, 1, 0)], static_cast<TypeParam>(2.04),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(2, 0, 0)], static_cast<TypeParam>(1.04),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(2, 1, 0)], static_cast<TypeParam>(2.08),
+ 1e-6);
}
TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
@@ -631,9 +810,9 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
// Vectors for block and index ranges
std::vector<int> i_startblk_in(3, 0);
- std::vector<int> i_endblk_in(3, nblks_c);
+ std::vector<int> i_endblk_in(3, nblks_c - 1);
std::vector<int> i_startidx_in(3, 0);
- std::vector<int> i_endidx_in(3, nproma);
+ std::vector<int> i_endidx_in(3, nproma - 1);
// Parameters for the test
int patch_id = 1;
@@ -643,47 +822,73 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
const auto &cell_neighbor_at = at<nproma, nblks_c, 3>;
const auto &avg_coeff_at = at<nproma, 4, nblks_c>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto opt_in2_h = Kokkos::create_mirror_view(this->opt_in2);
+ auto opt_out2_h = Kokkos::create_mirror_view(this->opt_out2);
+
// Set up random number generators
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
// Initialize with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
- this->opt_in2[vec_e_at(i, k, 0)] = real_distrib(gen);
+ vec_e_h[vec_e_at(i, k, 0)] = real_distrib(gen);
+ opt_in2_h[vec_e_at(i, k, 0)] = real_distrib(gen);
}
// Set random edge indices
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = int_distrib(gen);
- this->cell_edge_blk[cell_edge_at(i, 0, j)] =
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = int_distrib(gen);
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] =
0; // Keep in same block for simplicity
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] =
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] =
0; // Keep in same block for simplicity
}
// Random geometric factors
for (int j = 0; j < 3; ++j) {
- this->geofac_div[geofac_div_at(i, j, 0)] = real_distrib(gen);
+ geofac_div_h[geofac_div_at(i, j, 0)] = real_distrib(gen);
}
// Random average coefficients
for (int j = 0; j < 4; ++j) {
- this->avg_coeff[avg_coeff_at(i, j, 0)] = real_distrib(gen);
+ avg_coeff_h[avg_coeff_at(i, j, 0)] = real_distrib(gen);
}
// Random initial values for div_vec_c and opt_out2
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = real_distrib(gen);
- this->opt_out2[div_vec_c_at(i, k, 0)] = real_distrib(gen);
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
+ opt_out2_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->opt_in2, opt_in2_h);
+ Kokkos::deep_copy(this->opt_out2, opt_out2_h);
+
// Call the div_avg function
div_avg<TypeParam>(
this->vec_e.data(), this->cell_neighbor_idx.data(),
@@ -695,6 +900,10 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
this->nproma, patch_id, l_limited_area, l2fields, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(opt_out2_h, this->opt_out2);
+
// Calculate reference values manually
std::vector<TypeParam> aux_c(dim_combine(nproma, nlev, nblks_c));
std::vector<TypeParam> aux_c2(dim_combine(nproma, nlev, nblks_c));
@@ -702,52 +911,46 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
std::vector<TypeParam> ref_opt_out2(dim_combine(nproma, nlev, nblks_c));
// Step 1: Calculate aux_c and aux_c2
- for (int jb = i_startblk_in[0]; jb < i_endblk_in[0]; ++jb) {
+ for (int jb = i_startblk_in[0]; jb <= i_endblk_in[0]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[0], i_endidx_in[0], nproma, jb,
i_startblk_in[0], i_endblk_in[0], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
aux_c[div_vec_c_at(jc, jk, jb)] =
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)])] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)])] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)])] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)])] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)])] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)])] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
aux_c2[div_vec_c_at(jc, jk, jb)] =
- this->opt_in2[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)])] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->opt_in2[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)])] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->opt_in2[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)])] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ opt_in2_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)])] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ opt_in2_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)])] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ opt_in2_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)])] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
}
}
}
// Step 2: Assign aux_c to div_vec_c and aux_c2 to opt_out2 for patch_id > 0
- for (int jb = i_startblk_in[1]; jb < i_endblk_in[1]; ++jb) {
+ for (int jb = i_startblk_in[1]; jb <= i_endblk_in[1]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[1], i_endidx_in[1], nproma, jb,
i_startblk_in[1], i_endblk_in[1], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
aux_c[div_vec_c_at(jc, jk, jb)];
ref_opt_out2[div_vec_c_at(jc, jk, jb)] =
@@ -757,57 +960,60 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandom) {
}
// Step 3: Perform averaging for the rest of the blocks
- for (int jb = i_startblk_in[2]; jb < i_endblk_in[2]; ++jb) {
+ for (int jb = i_startblk_in[2]; jb <= i_endblk_in[2]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[2], i_endidx_in[2], nproma, jb,
i_startblk_in[2], i_endblk_in[2], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
aux_c[div_vec_c_at(jc, jk, jb)] *
- this->avg_coeff[avg_coeff_at(jc, 0, jb)] +
+ avg_coeff_h[avg_coeff_at(jc, 0, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 0)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 0)])] *
- this->avg_coeff[avg_coeff_at(jc, 1, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 0)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 0)])] *
+ avg_coeff_h[avg_coeff_at(jc, 1, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 1)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 1)])] *
- this->avg_coeff[avg_coeff_at(jc, 2, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 1)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 1)])] *
+ avg_coeff_h[avg_coeff_at(jc, 2, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 2)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 2)])] *
- this->avg_coeff[avg_coeff_at(jc, 3, jb)];
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 2)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 2)])] *
+ avg_coeff_h[avg_coeff_at(jc, 3, jb)];
ref_opt_out2[div_vec_c_at(jc, jk, jb)] =
aux_c2[div_vec_c_at(jc, jk, jb)] *
- this->avg_coeff[avg_coeff_at(jc, 0, jb)] +
+ avg_coeff_h[avg_coeff_at(jc, 0, jb)] +
aux_c2[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 0)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 0)])] *
- this->avg_coeff[avg_coeff_at(jc, 1, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 0)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 0)])] *
+ avg_coeff_h[avg_coeff_at(jc, 1, jb)] +
aux_c2[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 1)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 1)])] *
- this->avg_coeff[avg_coeff_at(jc, 2, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 1)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 1)])] *
+ avg_coeff_h[avg_coeff_at(jc, 2, jb)] +
aux_c2[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 2)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 2)])] *
- this->avg_coeff[avg_coeff_at(jc, 3, jb)];
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 2)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 2)])] *
+ avg_coeff_h[avg_coeff_at(jc, 3, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(i, k, 0)],
- ref_div_vec_c[div_vec_c_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(i, k, 0)],
+ ref_div_vec_c[div_vec_c_at(i, k, 0)], tol)
<< "div_vec_c results differ at i=" << i << ", k=" << k;
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(i, k, 0)],
- ref_opt_out2[div_vec_c_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(i, k, 0)],
+ ref_opt_out2[div_vec_c_at(i, k, 0)], tol)
<< "opt_out2 results differ at i=" << i << ", k=" << k;
}
}
@@ -818,7 +1024,6 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgSpecificNoL2fields) {
constexpr int nlev = this->nlev;
constexpr int nblks_c = this->nblks_c;
constexpr int nblks_e = this->nblks_e;
- constexpr int dim4d = this->dim4d;
const auto &vec_e_at = at<nproma, nlev, nblks_e>;
const auto &cell_edge_at = at<nproma, nblks_c, 3>;
@@ -827,9 +1032,9 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgSpecificNoL2fields) {
// Vectors for block and index ranges
std::vector<int> i_startblk_in(3, 0);
- std::vector<int> i_endblk_in(3, nblks_c);
+ std::vector<int> i_endblk_in(3, nblks_c - 1);
std::vector<int> i_startidx_in(3, 0);
- std::vector<int> i_endidx_in(3, nproma);
+ std::vector<int> i_endidx_in(3, nproma - 1);
// Parameters for the test
int patch_id = 1;
@@ -839,48 +1044,78 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgSpecificNoL2fields) {
const auto &cell_neighbor_at = at<nproma, nblks_c, 3>;
const auto &avg_coeff_at = at<nproma, 4, nblks_c>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto opt_in2_h = Kokkos::create_mirror_view(this->opt_in2);
+ auto opt_out2_h = Kokkos::create_mirror_view(this->opt_out2);
+
// Initialize the vectors with specific values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
- this->opt_in2[vec_e_at(i, k, 0)] =
- (i + 1) * (k + 1) * 0.5; // Half of vec_e
+ vec_e_h[vec_e_at(i, k, 0)] =
+ static_cast<TypeParam>((i + 1) * (k + 1)); // Simple pattern
+ opt_in2_h[vec_e_at(i, k, 0)] =
+ static_cast<TypeParam>((i + 1) * (k + 1) * 0.5); // Half of vec_e
}
// Set edge indices to point to specific cells
- this->cell_edge_idx[cell_edge_at(i, 0, 0)] = i;
- this->cell_edge_idx[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_edge_idx[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 0)] = i;
+ cell_edge_idx_h[cell_edge_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_edge_idx_h[cell_edge_at(i, 0, 2)] = (i + 2) % nproma;
// Set neighbor indices similarly
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 1)] = (i + 1) % nproma;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 2)] = (i + 2) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = i;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 1)] = (i + 1) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 2)] = (i + 2) % nproma;
// All edges and neighbors are in the same block for this test
for (int j = 0; j < 3; ++j) {
- this->cell_edge_blk[cell_edge_at(i, 0, j)] = 0;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] = 0;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
// Geometric factors
- this->geofac_div[geofac_div_at(i, 0, 0)] = 0.5;
- this->geofac_div[geofac_div_at(i, 1, 0)] = 0.3;
- this->geofac_div[geofac_div_at(i, 2, 0)] = 0.2;
+ geofac_div_h[geofac_div_at(i, 0, 0)] = static_cast<TypeParam>(0.5);
+ geofac_div_h[geofac_div_at(i, 1, 0)] = static_cast<TypeParam>(0.3);
+ geofac_div_h[geofac_div_at(i, 2, 0)] = static_cast<TypeParam>(0.2);
// Average coefficients
- this->avg_coeff[avg_coeff_at(i, 0, 0)] = 0.4; // Self
- this->avg_coeff[avg_coeff_at(i, 1, 0)] = 0.2; // First neighbor
- this->avg_coeff[avg_coeff_at(i, 2, 0)] = 0.2; // Second neighbor
- this->avg_coeff[avg_coeff_at(i, 3, 0)] = 0.2; // Third neighbor
+ avg_coeff_h[avg_coeff_at(i, 0, 0)] = static_cast<TypeParam>(0.4); // Self
+ avg_coeff_h[avg_coeff_at(i, 1, 0)] =
+ static_cast<TypeParam>(0.2); // First neighbor
+ avg_coeff_h[avg_coeff_at(i, 2, 0)] =
+ static_cast<TypeParam>(0.2); // Second neighbor
+ avg_coeff_h[avg_coeff_at(i, 3, 0)] =
+ static_cast<TypeParam>(0.2); // Third neighbor
// Initialize div_vec_c and opt_out2 to zero
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = 0.0;
- this->opt_out2[div_vec_c_at(i, k, 0)] = 0.0;
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
+ opt_out2_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->opt_in2, opt_in2_h);
+ Kokkos::deep_copy(this->opt_out2, opt_out2_h);
+
// Call the div_avg function
div_avg<TypeParam>(
this->vec_e.data(), this->cell_neighbor_idx.data(),
@@ -892,19 +1127,37 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgSpecificNoL2fields) {
this->nproma, patch_id, l_limited_area, l2fields, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 0, 0)], 1.88, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(0, 1, 0)], 3.76, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 0, 0)], 2.04, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(1, 1, 0)], 4.08, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 0, 0)], 2.08, 1e-6);
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(2, 1, 0)], 4.16, 1e-6);
-
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(0, 0, 0)], 0.0, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(0, 1, 0)], 0.0, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(1, 0, 0)], 0.0, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(1, 1, 0)], 0.0, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(2, 0, 0)], 0.0, 1e-6);
- EXPECT_NEAR(this->opt_out2[div_vec_c_at(2, 1, 0)], 0.0, 1e-6);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(opt_out2_h, this->opt_out2);
+
+ // Verify first field results
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 0, 0)], static_cast<TypeParam>(1.88),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(0, 1, 0)], static_cast<TypeParam>(3.76),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 0, 0)], static_cast<TypeParam>(2.04),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(1, 1, 0)], static_cast<TypeParam>(4.08),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 0, 0)], static_cast<TypeParam>(2.08),
+ 1e-6);
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(2, 1, 0)], static_cast<TypeParam>(4.16),
+ 1e-6);
+
+ // Since l2fields=false, opt_out2 should not be modified
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(0, 0, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(0, 1, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(1, 0, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(1, 1, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(2, 0, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
+ EXPECT_NEAR(opt_out2_h[div_vec_c_at(2, 1, 0)], static_cast<TypeParam>(0.0),
+ 1e-6);
}
TYPED_TEST(HorizontalDivTest, TestDivAvgRandomNoL2fields) {
@@ -920,9 +1173,9 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandomNoL2fields) {
// Vectors for block and index ranges
std::vector<int> i_startblk_in(3, 0);
- std::vector<int> i_endblk_in(3, nblks_c);
+ std::vector<int> i_endblk_in(3, nblks_c - 1);
std::vector<int> i_startidx_in(3, 0);
- std::vector<int> i_endidx_in(3, nproma);
+ std::vector<int> i_endidx_in(3, nproma - 1);
// Parameters for the test
int patch_id = 1;
@@ -932,49 +1185,75 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandomNoL2fields) {
const auto &cell_neighbor_at = at<nproma, nblks_c, 3>;
const auto &avg_coeff_at = at<nproma, 4, nblks_c>;
+ // Create mirror views to store data on the host
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto geofac_div_h = Kokkos::create_mirror_view(this->geofac_div);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto div_vec_c_h = Kokkos::create_mirror_view(this->div_vec_c);
+ auto opt_in2_h = Kokkos::create_mirror_view(this->opt_in2);
+ auto opt_out2_h = Kokkos::create_mirror_view(this->opt_out2);
+
// Set up random number generators
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
// Initialize with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
- this->opt_in2[vec_e_at(i, k, 0)] =
+ vec_e_h[vec_e_at(i, k, 0)] = real_distrib(gen);
+ opt_in2_h[vec_e_at(i, k, 0)] =
real_distrib(gen); // Not used but initialize anyway
}
// Set random edge indices
for (int j = 0; j < 3; ++j) {
- this->cell_edge_idx[cell_edge_at(i, 0, j)] = int_distrib(gen);
- this->cell_edge_blk[cell_edge_at(i, 0, j)] =
+ cell_edge_idx_h[cell_edge_at(i, 0, j)] = int_distrib(gen);
+ cell_edge_blk_h[cell_edge_at(i, 0, j)] =
0; // Keep in same block for simplicity
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] =
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] =
0; // Keep in same block for simplicity
}
// Random geometric factors
for (int j = 0; j < 3; ++j) {
- this->geofac_div[geofac_div_at(i, j, 0)] = real_distrib(gen);
+ geofac_div_h[geofac_div_at(i, j, 0)] = real_distrib(gen);
}
// Random average coefficients
for (int j = 0; j < 4; ++j) {
- this->avg_coeff[avg_coeff_at(i, j, 0)] = real_distrib(gen);
+ avg_coeff_h[avg_coeff_at(i, j, 0)] = real_distrib(gen);
}
// Random initial values for div_vec_c and opt_out2
for (int k = 0; k < nlev; ++k) {
- this->div_vec_c[div_vec_c_at(i, k, 0)] = real_distrib(gen);
- this->opt_out2[div_vec_c_at(i, k, 0)] =
- real_distrib(gen); // Not used but initialize anyway
+ div_vec_c_h[div_vec_c_at(i, k, 0)] = static_cast<TypeParam>(0.0);
+ opt_out2_h[div_vec_c_at(i, k, 0)] =
+ static_cast<TypeParam>(0.0); // Not used but initialize anyway
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_div, geofac_div_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->div_vec_c, div_vec_c_h);
+ Kokkos::deep_copy(this->opt_in2, opt_in2_h);
+ Kokkos::deep_copy(this->opt_out2, opt_out2_h);
+
// Call the div_avg function with l2fields=false
div_avg<TypeParam>(
this->vec_e.data(), this->cell_neighbor_idx.data(),
@@ -986,44 +1265,45 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandomNoL2fields) {
this->nproma, patch_id, l_limited_area, l2fields, this->lacc, this->nlev,
this->nblks_c, this->nblks_e);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(div_vec_c_h, this->div_vec_c);
+ Kokkos::deep_copy(opt_out2_h, this->opt_out2);
+
// Calculate reference values manually
std::vector<TypeParam> aux_c(dim_combine(nproma, nlev, nblks_c));
std::vector<TypeParam> ref_div_vec_c(dim_combine(nproma, nlev, nblks_c));
// Step 1: Calculate aux_c (but not aux_c2 since l2fields=false)
- for (int jb = i_startblk_in[0]; jb < i_endblk_in[0]; ++jb) {
+ for (int jb = i_startblk_in[0]; jb <= i_endblk_in[0]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[0], i_endidx_in[0], nproma, jb,
i_startblk_in[0], i_endblk_in[0], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
aux_c[div_vec_c_at(jc, jk, jb)] =
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 0)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 0)])] *
- this->geofac_div[geofac_div_at(jc, 0, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 1)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 1)])] *
- this->geofac_div[geofac_div_at(jc, 1, jb)] +
- this->vec_e[vec_e_at(
- this->cell_edge_idx[cell_edge_at(jc, jb, 2)], jk,
- this->cell_edge_blk[cell_edge_at(jc, jb, 2)])] *
- this->geofac_div[geofac_div_at(jc, 2, jb)];
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 0)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 0)])] *
+ geofac_div_h[geofac_div_at(jc, 0, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 1)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 1)])] *
+ geofac_div_h[geofac_div_at(jc, 1, jb)] +
+ vec_e_h[vec_e_at(cell_edge_idx_h[cell_edge_at(jc, jb, 2)], jk,
+ cell_edge_blk_h[cell_edge_at(jc, jb, 2)])] *
+ geofac_div_h[geofac_div_at(jc, 2, jb)];
}
}
}
// Step 2: Assign aux_c to div_vec_c for patch_id > 0 (opt_out2 not updated
// since l2fields=false)
- for (int jb = i_startblk_in[1]; jb < i_endblk_in[1]; ++jb) {
+ for (int jb = i_startblk_in[1]; jb <= i_endblk_in[1]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[1], i_endidx_in[1], nproma, jb,
i_startblk_in[1], i_endblk_in[1], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
aux_c[div_vec_c_at(jc, jk, jb)];
}
@@ -1032,38 +1312,41 @@ TYPED_TEST(HorizontalDivTest, TestDivAvgRandomNoL2fields) {
// Step 3: Perform averaging for the rest of the blocks (only for div_vec_c,
// not opt_out2)
- for (int jb = i_startblk_in[2]; jb < i_endblk_in[2]; ++jb) {
+ for (int jb = i_startblk_in[2]; jb <= i_endblk_in[2]; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(i_startidx_in[2], i_endidx_in[2], nproma, jb,
i_startblk_in[2], i_endblk_in[2], i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
ref_div_vec_c[div_vec_c_at(jc, jk, jb)] =
aux_c[div_vec_c_at(jc, jk, jb)] *
- this->avg_coeff[avg_coeff_at(jc, 0, jb)] +
+ avg_coeff_h[avg_coeff_at(jc, 0, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 0)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 0)])] *
- this->avg_coeff[avg_coeff_at(jc, 1, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 0)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 0)])] *
+ avg_coeff_h[avg_coeff_at(jc, 1, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 1)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 1)])] *
- this->avg_coeff[avg_coeff_at(jc, 2, jb)] +
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 1)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 1)])] *
+ avg_coeff_h[avg_coeff_at(jc, 2, jb)] +
aux_c[div_vec_c_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, 2)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, 2)])] *
- this->avg_coeff[avg_coeff_at(jc, 3, jb)];
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, 2)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, 2)])] *
+ avg_coeff_h[avg_coeff_at(jc, 3, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results - only check div_vec_c since l2fields=false means opt_out2
// isn't updated
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->div_vec_c[div_vec_c_at(i, k, 0)],
- ref_div_vec_c[div_vec_c_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(div_vec_c_h[div_vec_c_at(i, k, 0)],
+ ref_div_vec_c[div_vec_c_at(i, k, 0)], tol)
<< "div_vec_c results differ at i=" << i << ", k=" << k;
}
}
diff --git a/test/c/test_horizontal_recon.cpp b/test/c/test_horizontal_recon.cpp
index 8938a101bd2a350da12a0def450472d1e0c9ec9f..b83886c3d5cea1726fb4618ef66407ef06bb6d3a 100644
--- a/test/c/test_horizontal_recon.cpp
+++ b/test/c/test_horizontal_recon.cpp
@@ -14,8 +14,8 @@
#include <vector>
#include <Kokkos_Core.hpp>
-#include <gtest/gtest.h>
#include <dim_helper.hpp>
+#include <gtest/gtest.h>
#include <horizontal/mo_lib_divrot.hpp>
#include <support/mo_lib_loopindices.hpp>
@@ -26,8 +26,8 @@ enum class ReconstructionMethod {
cubic,
};
-/// Base test class for the horizontal reconstruct tests. Templated for the ValueType
-/// and ReconMethod for the reconstruction method.
+/// Base test class for the horizontal reconstruct tests. Templated for the
+/// ValueType and ReconMethod for the reconstruction method.
template <typename ValueType, int ReconMethod>
class HorizontalReconTest : public ::testing::Test {
protected:
@@ -41,13 +41,15 @@ protected:
return std::make_tuple(9, 5);
case ReconstructionMethod::cubic:
return std::make_tuple(9, 9);
+ default:
+ return std::make_tuple(0, 0); // or throw/assert if appropriate
}
}
// Constant dimensions.
static constexpr int nproma = 3; // inner loop length
- static constexpr int nlev = 1; // number of vertical levels
- static constexpr int nblks_c = 1; // number of cell blocks (for p_e_in)
+ static constexpr int nlev = 2; // number of vertical levels
+ static constexpr int nblks_c = 2; // number of cell blocks (for p_e_in)
static constexpr std::tuple<int, int> lsq_dim =
init_lsq_dim(static_cast<ReconstructionMethod>(ReconMethod));
static constexpr int lsq_dim_c = std::get<0>(lsq_dim);
@@ -55,39 +57,51 @@ protected:
// Parameter values.
int i_startblk = 0;
- int i_endblk = nblks_c; // Test blocks [0 .. nblks_c-1]
+ int i_endblk = nblks_c - 1; // Test blocks [0 .. nblks_c-1]
int i_startidx_in = 0;
- int i_endidx_in = nproma; // Full range: 0 .. nproma-1
+ int i_endidx_in = nproma - 1; // Full range: 0 .. nproma-1
int slev = 0;
- int elev = nlev; // Full vertical range (0 .. nlev-1)
+ int elev = nlev - 1; // Full vertical range (0 .. nlev-1)
int patch_id = 0;
bool lacc = false; // Not using ACC-specific behavior.
bool acc_async = false; // No asynchronous execution.
bool l_consv = true; // With conservative correction.
bool l_limited_area = true; // Limited area setup
- std::vector<ValueType> p_cc;
- std::vector<int> cell_neighbor_idx;
- std::vector<int> cell_neighbor_blk;
- std::vector<ValueType> lsq_qtmat_c;
- std::vector<ValueType> lsq_rmat_rdiag_c;
- std::vector<ValueType> lsq_rmat_utri_c;
- std::vector<ValueType> lsq_moments;
- std::vector<ValueType> lsq_pseudoinv;
- std::vector<ValueType> p_coeff;
-
- HorizontalReconTest() {
- p_cc.resize(dim_combine(nproma, nlev, nblks_c));
- cell_neighbor_idx.resize(dim_combine(nproma, nblks_c, lsq_dim_c));
- cell_neighbor_blk.resize(dim_combine(nproma, nblks_c, lsq_dim_c));
- lsq_qtmat_c.resize(dim_combine(nproma, lsq_dim_unk, lsq_dim_c, nblks_c));
- lsq_rmat_rdiag_c.resize(dim_combine(nproma, lsq_dim_unk, nblks_c));
- lsq_rmat_utri_c.resize(dim_combine(
- nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c));
- lsq_moments.resize(dim_combine(nproma, nblks_c, lsq_dim_unk));
- lsq_pseudoinv.resize(dim_combine(nproma, lsq_dim_unk, lsq_dim_c, nblks_c));
- p_coeff.resize(dim_combine(lsq_dim_unk + 1, nproma, nlev, nblks_c));
- }
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Kokkos Views for test data
+ Kokkos::View<ValueType *, memory_space> p_cc;
+ Kokkos::View<int *, memory_space> cell_neighbor_idx;
+ Kokkos::View<int *, memory_space> cell_neighbor_blk;
+ Kokkos::View<ValueType *, memory_space> lsq_qtmat_c;
+ Kokkos::View<ValueType *, memory_space> lsq_rmat_rdiag_c;
+ Kokkos::View<ValueType *, memory_space> lsq_rmat_utri_c;
+ Kokkos::View<ValueType *, memory_space> lsq_moments;
+ Kokkos::View<ValueType *, memory_space> lsq_pseudoinv;
+ Kokkos::View<ValueType *, memory_space> p_coeff;
+
+ HorizontalReconTest()
+ : p_cc("p_cc", dim_combine(nproma, nlev, nblks_c)),
+ cell_neighbor_idx("cell_neighbor_idx",
+ dim_combine(nproma, nblks_c, lsq_dim_c)),
+ cell_neighbor_blk("cell_neighbor_blk",
+ dim_combine(nproma, nblks_c, lsq_dim_c)),
+ lsq_qtmat_c("lsq_qtmat_c",
+ dim_combine(nproma, lsq_dim_unk, lsq_dim_c, nblks_c)),
+ lsq_rmat_rdiag_c("lsq_rmat_rdiag_c",
+ dim_combine(nproma, lsq_dim_unk, nblks_c)),
+ lsq_rmat_utri_c(
+ "lsq_rmat_utri_c",
+ dim_combine(nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2,
+ nblks_c)),
+ lsq_moments("lsq_moments", dim_combine(nproma, nblks_c, lsq_dim_unk)),
+ lsq_pseudoinv("lsq_pseudoinv",
+ dim_combine(nproma, lsq_dim_unk, lsq_dim_c, nblks_c)),
+ p_coeff("p_coeff",
+ dim_combine(lsq_dim_unk + 1, nproma, nlev, nblks_c)) {}
};
/// Test class for the horizontal tests. The reconstruction method is specified
@@ -134,30 +148,54 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCell) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
+
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 1)] = i;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 2)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 1)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 2)] = i;
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
- this->lsq_qtmat_c[qtmat_at(i, 0, j, 0)] = 1.0;
- this->lsq_qtmat_c[qtmat_at(i, 1, j, 0)] = 0.5;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
+ lsq_qtmat_c_h[qtmat_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_qtmat_c_h[qtmat_at(i, 1, j, 0)] = static_cast<TypeParam>(0.5);
}
+
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, 0, 0)] = 2.0;
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, 1, 0)] = 2.0;
- this->lsq_rmat_utri_c[rmat_utri_at(i, 0, 0)] = 0.1;
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, 0, 0)] = static_cast<TypeParam>(2.0);
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, 1, 0)] = static_cast<TypeParam>(2.0);
+ lsq_rmat_utri_c_h[rmat_utri_at(i, 0, 0)] = static_cast<TypeParam>(0.1);
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_l<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -168,16 +206,19 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCell) {
this->elev, this->nproma, this->l_consv, this->lacc, this->acc_async,
this->nblks_c, this->nlev, this->lsq_dim_unk, this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- 0.34, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(0.34), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 1.8, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(1.8), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- 1.0, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(1.0), 1e-6);
}
TYPED_TEST(HorizontalReconLinearTest, TestLsqCellRandom) {
@@ -196,33 +237,56 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellRandom) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(0.0, 3.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.0, 1.0);
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
- this->lsq_qtmat_c[qtmat_at(i, 0, j, 0)] = real_distrib(gen);
- this->lsq_qtmat_c[qtmat_at(i, 1, j, 0)] = real_distrib(gen);
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
+ lsq_qtmat_c_h[qtmat_at(i, 0, j, 0)] = real_distrib(gen);
+ lsq_qtmat_c_h[qtmat_at(i, 1, j, 0)] = real_distrib(gen);
}
+
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, 0, 0)] = real_distrib(gen);
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, 1, 0)] = real_distrib(gen);
- this->lsq_rmat_utri_c[rmat_utri_at(i, 0, 0)] = real_distrib(gen);
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, 0, 0)] = real_distrib(gen);
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, 1, 0)] = real_distrib(gen);
+ lsq_rmat_utri_c_h[rmat_utri_at(i, 0, 0)] = real_distrib(gen);
- this->lsq_moments[moments_at(i, 0, 0)] = real_distrib(gen);
- this->lsq_moments[moments_at(i, 0, 1)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, 0)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, 1)] = real_distrib(gen);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_l<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -233,50 +297,59 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellRandom) {
this->elev, this->nproma, this->l_consv, this->lacc, this->acc_async,
this->nblks_c, this->nlev, this->lsq_dim_unk, this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
std::vector<TypeParam> z_qt_times_d(lsq_dim_unk);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+
+ // doing the calculation only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+ // calculation only for jk = 0
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
z_qt_times_d[0] = 0.0;
z_qt_times_d[1] = 0.0;
for (int i = 0; i < lsq_dim_c; ++i) {
- z_qt_times_d[0] += this->lsq_qtmat_c[qtmat_at(jc, 0, i, jb)] * z_d[i];
- z_qt_times_d[1] += this->lsq_qtmat_c[qtmat_at(jc, 1, i, jb)] * z_d[i];
+ z_qt_times_d[0] += lsq_qtmat_c_h[qtmat_at(jc, 0, i, jb)] * z_d[i];
+ z_qt_times_d[1] += lsq_qtmat_c_h[qtmat_at(jc, 1, i, jb)] * z_d[i];
}
p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] =
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(jc, 1, jb)] * z_qt_times_d[1];
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 1, jb)] * z_qt_times_d[1];
p_result[at<lsq_dim_unk + 1, nproma>(1, jc)] =
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(jc, 0, jb)] *
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 0, jb)] *
(z_qt_times_d[0] -
- this->lsq_rmat_utri_c[rmat_utri_at(jc, 0, jb)] *
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 0, jb)] *
p_result[at<lsq_dim_unk + 1, nproma>(2, jc)]);
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)] -
+ p_cc_h[p_cc_at(jc, jk, jb)] -
p_result[at<lsq_dim_unk + 1, nproma>(1, jc)] *
- this->lsq_moments[moments_at(jc, jb, 0)] -
+ lsq_moments_h[moments_at(jc, jb, 0)] -
p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] *
- this->lsq_moments[moments_at(jc, jb, 1)];
+ lsq_moments_h[moments_at(jc, jb, 1)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(i, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], 1e-5)
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
<< "For loop result fails for i = " << i << ", jc = " << jc;
}
}
@@ -295,26 +368,46 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellSVD) {
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
+
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 1)] = i;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 2)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 1)] = i;
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 2)] = i;
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
- this->lsq_pseudoinv[pseudoinv_at(i, 0, j, 0)] = 1.0;
- this->lsq_pseudoinv[pseudoinv_at(i, 1, j, 0)] = 0.5;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
+ lsq_pseudoinv_h[pseudoinv_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_pseudoinv_h[pseudoinv_at(i, 1, j, 0)] = static_cast<TypeParam>(0.5);
}
+
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_l_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -324,16 +417,19 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellSVD) {
this->elev, this->nproma, this->l_consv, this->lacc, this->acc_async,
this->nblks_c, this->nlev, this->lsq_dim_unk, this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- 0.65, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(0.65), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 1.0, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(1.0), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- 0.5, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(0.5), 1e-6);
}
TYPED_TEST(HorizontalReconLinearTest, TestLsqCellSVDRandom) {
@@ -349,29 +445,48 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellSVDRandom) {
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(0.0, 3.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.0, 1.0);
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
- this->lsq_pseudoinv[pseudoinv_at(i, 0, j, 0)] = real_distrib(gen);
- this->lsq_pseudoinv[pseudoinv_at(i, 1, j, 0)] = real_distrib(gen);
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
+ lsq_pseudoinv_h[pseudoinv_at(i, 0, j, 0)] = real_distrib(gen);
+ lsq_pseudoinv_h[pseudoinv_at(i, 1, j, 0)] = real_distrib(gen);
}
+
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = real_distrib(gen);
- this->lsq_moments[moments_at(i, 0, 1)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, 0)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, 1)] = real_distrib(gen);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_l_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -381,45 +496,53 @@ TYPED_TEST(HorizontalReconLinearTest, TestLsqCellSVDRandom) {
this->elev, this->nproma, this->l_consv, this->lacc, this->acc_async,
this->nblks_c, this->nlev, this->lsq_dim_unk, this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+ // doing the calculation only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+ // calculation only for jk = 0
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] =
- this->lsq_pseudoinv[pseudoinv_at(jc, 1, 0, jb)] * z_d[0] +
- this->lsq_pseudoinv[pseudoinv_at(jc, 1, 1, jb)] * z_d[1] +
- this->lsq_pseudoinv[pseudoinv_at(jc, 1, 2, jb)] * z_d[2];
+ lsq_pseudoinv_h[pseudoinv_at(jc, 1, 0, jb)] * z_d[0] +
+ lsq_pseudoinv_h[pseudoinv_at(jc, 1, 1, jb)] * z_d[1] +
+ lsq_pseudoinv_h[pseudoinv_at(jc, 1, 2, jb)] * z_d[2];
p_result[at<lsq_dim_unk + 1, nproma>(1, jc)] =
- this->lsq_pseudoinv[pseudoinv_at(jc, 0, 0, jb)] * z_d[0] +
- this->lsq_pseudoinv[pseudoinv_at(jc, 0, 1, jb)] * z_d[1] +
- this->lsq_pseudoinv[pseudoinv_at(jc, 0, 2, jb)] * z_d[2];
+ lsq_pseudoinv_h[pseudoinv_at(jc, 0, 0, jb)] * z_d[0] +
+ lsq_pseudoinv_h[pseudoinv_at(jc, 0, 1, jb)] * z_d[1] +
+ lsq_pseudoinv_h[pseudoinv_at(jc, 0, 2, jb)] * z_d[2];
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)] -
+ p_cc_h[p_cc_at(jc, jk, jb)] -
p_result[at<lsq_dim_unk + 1, nproma>(1, jc)] *
- this->lsq_moments[moments_at(jc, jb, 0)] -
+ lsq_moments_h[moments_at(jc, jb, 0)] -
p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] *
- this->lsq_moments[moments_at(jc, jb, 1)];
+ lsq_moments_h[moments_at(jc, jb, 1)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(i, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], 1e-5)
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
<< "For loop result fails for i = " << i << ", jc = " << jc;
}
}
@@ -443,43 +566,65 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCell) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, 0)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, 0)] = 0;
for (int j = 1; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = i;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = i;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_c; ++j) {
- this->lsq_qtmat_c[qtmat_at(i, 0, j, 0)] = 1.0;
- this->lsq_qtmat_c[qtmat_at(i, 1, j, 0)] = 0.5;
- this->lsq_qtmat_c[qtmat_at(i, 2, j, 0)] = 0.2;
- this->lsq_qtmat_c[qtmat_at(i, 3, j, 0)] = 0.7;
- this->lsq_qtmat_c[qtmat_at(i, 4, j, 0)] = 1.3;
+ lsq_qtmat_c_h[qtmat_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_qtmat_c_h[qtmat_at(i, 1, j, 0)] = static_cast<TypeParam>(0.5);
+ lsq_qtmat_c_h[qtmat_at(i, 2, j, 0)] = static_cast<TypeParam>(0.2);
+ lsq_qtmat_c_h[qtmat_at(i, 3, j, 0)] = static_cast<TypeParam>(0.7);
+ lsq_qtmat_c_h[qtmat_at(i, 4, j, 0)] = static_cast<TypeParam>(1.3);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
for (int j = 0; j < lsq_dim_unk; ++j) {
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, j, 0)] = 2.0;
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, j, 0)] = static_cast<TypeParam>(2.0);
}
for (int j = 0; j < (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2; ++j) {
- this->lsq_rmat_utri_c[rmat_utri_at(i, j, 0)] = 1.0;
+ lsq_rmat_utri_c_h[rmat_utri_at(i, j, 0)] = static_cast<TypeParam>(1.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
- this->lsq_moments[moments_at(i, 0, 2)] = 0.4;
- this->lsq_moments[moments_at(i, 0, 3)] = 0.5;
- this->lsq_moments[moments_at(i, 0, 4)] = 0.6;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
+ lsq_moments_h[moments_at(i, 0, 2)] = static_cast<TypeParam>(0.4);
+ lsq_moments_h[moments_at(i, 0, 3)] = static_cast<TypeParam>(0.5);
+ lsq_moments_h[moments_at(i, 0, 4)] = static_cast<TypeParam>(0.6);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_q<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -491,25 +636,28 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCell) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- 0.24, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(0.24), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 3.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(3.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- -2.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(-2.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
- 2.8, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
+ static_cast<TypeParam>(2.8), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
- -3.8, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
+ static_cast<TypeParam>(-3.8), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
- 2.6, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
+ static_cast<TypeParam>(2.6), 1e-6);
}
TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellRandom) {
@@ -528,6 +676,19 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellRandom) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
+ // Use fixed seed for reproducibility
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
@@ -535,29 +696,39 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellRandom) {
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_unk; ++j) {
for (int k = 0; k < lsq_dim_c; ++k) {
- this->lsq_qtmat_c[qtmat_at(i, j, k, 0)] = real_distrib(gen);
+ lsq_qtmat_c_h[qtmat_at(i, j, k, 0)] = real_distrib(gen);
}
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, j, 0)] = real_distrib(gen);
- this->lsq_moments[moments_at(i, 0, j)] = real_distrib(gen);
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, j, 0)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, j)] = real_distrib(gen);
}
for (int j = 0; j < (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2; ++j) {
- this->lsq_rmat_utri_c[rmat_utri_at(i, j, 0)] = real_distrib(gen);
+ lsq_rmat_utri_c_h[rmat_utri_at(i, j, 0)] = real_distrib(gen);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_q<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -569,56 +740,91 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellRandom) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
std::vector<TypeParam> z_qt_times_d(lsq_dim_unk);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+
+ for (int i = 0; i < nproma; ++i) {
+ for (int j = 0; j < lsq_dim_unk + 1; ++j) {
+ p_result[(at<lsq_dim_unk + 1, nproma>(j, i))] = static_cast<TypeParam>(0.0);
+ }
+ }
+
+ // doing the calculation only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ // Step 1: Calculate z_d values
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
+
+ // Matrix multiplication (Q^T * d)
for (int j = 0; j < lsq_dim_unk; ++j) {
z_qt_times_d[j] = 0.0;
for (int i = 0; i < lsq_dim_c; ++i) {
z_qt_times_d[j] +=
- this->lsq_qtmat_c[qtmat_at(jc, j, i, jb)] * z_d[i];
- }
- }
- int utri_id = 0;
- for (int j = lsq_dim_unk; j > 0; --j) {
- p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] = z_qt_times_d[j - 1];
- for (int k = j + 1; k <= lsq_dim_unk; ++k) {
- p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] -=
- this->lsq_rmat_utri_c[rmat_utri_at(jc, utri_id++, jb)] *
- p_result[at<lsq_dim_unk + 1, nproma>(k, jc)];
+ lsq_qtmat_c_h[qtmat_at(jc, j, i, jb)] * z_d[i];
}
- p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] *=
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(jc, j - 1, jb)];
}
+
+ // Back-substitution (mirrors the order in the GPU implementation)
+ p_result[at<lsq_dim_unk + 1, nproma>(5, jc)] =
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 4, jb)] * z_qt_times_d[4];
+
+ p_result[at<lsq_dim_unk + 1, nproma>(4, jc)] =
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 3, jb)] *
+ (z_qt_times_d[3] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 0, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(5, jc)]);
+ p_result[at<lsq_dim_unk + 1, nproma>(3, jc)] =
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 2, jb)] *
+ (z_qt_times_d[2] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 1, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(4, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 2, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(5, jc)]);
+ p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] =
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 1, jb)] *
+ (z_qt_times_d[1] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 3, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(3, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 4, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(4, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 5, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(5, jc)]);
+ p_result[at<lsq_dim_unk + 1, nproma>(1, jc)] =
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, 0, jb)] *
+ (z_qt_times_d[0] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 6, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(2, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 7, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(3, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 8, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(4, jc)] -
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, 9, jb)] * p_result[at<lsq_dim_unk + 1, nproma>(5, jc)]);
+ // Conservation correction
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)];
+ p_cc_h[p_cc_at(jc, jk, jb)];
for (int j = 0; j < lsq_dim_unk; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] -=
p_result[at<lsq_dim_unk + 1, nproma>(j + 1, jc)] *
- this->lsq_moments[moments_at(jc, jb, j)];
+ lsq_moments_h[moments_at(jc, jb, j)];
}
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(i, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], 1e-5)
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
<< "For loop result fails for i = " << i << ", jc = " << jc;
}
}
@@ -637,35 +843,53 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellSVD) {
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, 0)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, 0)] = 0;
for (int j = 1; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = i;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = i;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_c; ++j) {
- this->lsq_pseudoinv[pseudoinv_at(i, 0, j, 0)] = 1.0;
- this->lsq_pseudoinv[pseudoinv_at(i, 1, j, 0)] = 0.5;
- this->lsq_pseudoinv[pseudoinv_at(i, 2, j, 0)] = 0.2;
- this->lsq_pseudoinv[pseudoinv_at(i, 3, j, 0)] = 0.7;
- this->lsq_pseudoinv[pseudoinv_at(i, 4, j, 0)] = 1.3;
+ lsq_pseudoinv_h[pseudoinv_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_pseudoinv_h[pseudoinv_at(i, 1, j, 0)] = static_cast<TypeParam>(0.5);
+ lsq_pseudoinv_h[pseudoinv_at(i, 2, j, 0)] = static_cast<TypeParam>(0.2);
+ lsq_pseudoinv_h[pseudoinv_at(i, 3, j, 0)] = static_cast<TypeParam>(0.7);
+ lsq_pseudoinv_h[pseudoinv_at(i, 4, j, 0)] = static_cast<TypeParam>(1.3);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
- this->lsq_moments[moments_at(i, 0, 2)] = 0.4;
- this->lsq_moments[moments_at(i, 0, 3)] = 0.5;
- this->lsq_moments[moments_at(i, 0, 4)] = 0.6;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
+ lsq_moments_h[moments_at(i, 0, 2)] = static_cast<TypeParam>(0.4);
+ lsq_moments_h[moments_at(i, 0, 3)] = static_cast<TypeParam>(0.5);
+ lsq_moments_h[moments_at(i, 0, 4)] = static_cast<TypeParam>(0.6);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_q_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -676,25 +900,28 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellSVD) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- -0.56, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(-0.56), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 1.0, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(1.0), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- 0.5, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(0.5), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
- 0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
+ static_cast<TypeParam>(0.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
- 0.7, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
+ static_cast<TypeParam>(0.7), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
- 1.3, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
+ static_cast<TypeParam>(1.3), 1e-6);
}
TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellSVDRandom) {
@@ -708,37 +935,53 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellSVDRandom) {
const auto &cell_neighbor_at = at<nproma, nblks_c, lsq_dim_c>;
const auto &pseudoinv_at = at<nproma, lsq_dim_unk, lsq_dim_c, nblks_c>;
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
- const auto &rmat_rdiag_at = at<nproma, lsq_dim_unk, nblks_c>;
- const auto &rmat_utri_at =
- at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
+ // Use fixed seed for reproducibility
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
std::uniform_real_distribution<TypeParam> real_distrib(0.0, 1.0);
- // Initialization is done only for iblk = 0 and ilev = 0
+ // Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_unk; ++j) {
for (int k = 0; k < lsq_dim_c; ++k) {
- this->lsq_pseudoinv[pseudoinv_at(i, j, k, 0)] = real_distrib(gen);
+ lsq_pseudoinv_h[pseudoinv_at(i, j, k, 0)] = real_distrib(gen);
}
- this->lsq_moments[moments_at(i, 0, j)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, j)] = real_distrib(gen);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_q_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -749,46 +992,56 @@ TYPED_TEST(HorizontalReconQuadraticTest, TestLsqCellSVDRandom) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
- std::vector<TypeParam> z_qt_times_d(lsq_dim_unk);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+ // calculating only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+ // calculating only for jk = 0
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
for (int j = 1; j < lsq_dim_unk + 1; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] = 0.0;
for (int i = 0; i < lsq_dim_c; ++i) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] +=
- this->lsq_pseudoinv[pseudoinv_at(jc, j - 1, i, jb)] * z_d[i];
+ lsq_pseudoinv_h[pseudoinv_at(jc, j - 1, i, jb)] * z_d[i];
}
+ // p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] *=
+ // lsq_moments_h[moments_at(jc, jb, j - 1)];
}
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)];
- for (int j = 0; j < lsq_dim_unk; ++j) {
+ p_cc_h[p_cc_at(jc, jk, jb)];
+ for (int j = 0; j < lsq_dim_unk + 1; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] -=
+ // p_result[at<lsq_dim_unk + 1, nproma>(j, jc)];
p_result[at<lsq_dim_unk + 1, nproma>(j + 1, jc)] *
- this->lsq_moments[moments_at(jc, jb, j)];
+ lsq_moments_h[moments_at(jc, jb, j)];
}
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
- for (int j = 0; j < lsq_dim_unk + 1; ++j) {
+ for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(j, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(j, jc))], 1e-5)
- << "For loop result fails for j = " << j << ", jc = " << jc;
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
+ << "For loop result fails for i = " << i << ", jc = " << jc;
}
}
}
@@ -811,51 +1064,73 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCell) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, 0)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, 0)] = 0;
for (int j = 1; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = i;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = i;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_c; ++j) {
- this->lsq_qtmat_c[qtmat_at(i, 0, j, 0)] = 1.0;
- this->lsq_qtmat_c[qtmat_at(i, 1, j, 0)] = 0.9;
- this->lsq_qtmat_c[qtmat_at(i, 2, j, 0)] = 0.8;
- this->lsq_qtmat_c[qtmat_at(i, 3, j, 0)] = 0.7;
- this->lsq_qtmat_c[qtmat_at(i, 4, j, 0)] = 0.6;
- this->lsq_qtmat_c[qtmat_at(i, 5, j, 0)] = 0.5;
- this->lsq_qtmat_c[qtmat_at(i, 6, j, 0)] = 0.4;
- this->lsq_qtmat_c[qtmat_at(i, 7, j, 0)] = 0.3;
- this->lsq_qtmat_c[qtmat_at(i, 8, j, 0)] = 0.2;
+ lsq_qtmat_c_h[qtmat_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_qtmat_c_h[qtmat_at(i, 1, j, 0)] = static_cast<TypeParam>(0.9);
+ lsq_qtmat_c_h[qtmat_at(i, 2, j, 0)] = static_cast<TypeParam>(0.8);
+ lsq_qtmat_c_h[qtmat_at(i, 3, j, 0)] = static_cast<TypeParam>(0.7);
+ lsq_qtmat_c_h[qtmat_at(i, 4, j, 0)] = static_cast<TypeParam>(0.6);
+ lsq_qtmat_c_h[qtmat_at(i, 5, j, 0)] = static_cast<TypeParam>(0.5);
+ lsq_qtmat_c_h[qtmat_at(i, 6, j, 0)] = static_cast<TypeParam>(0.4);
+ lsq_qtmat_c_h[qtmat_at(i, 7, j, 0)] = static_cast<TypeParam>(0.3);
+ lsq_qtmat_c_h[qtmat_at(i, 8, j, 0)] = static_cast<TypeParam>(0.2);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
for (int j = 0; j < lsq_dim_unk; ++j) {
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, j, 0)] = 2.0;
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, j, 0)] = static_cast<TypeParam>(2.0);
}
for (int j = 0; j < (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2; ++j) {
- this->lsq_rmat_utri_c[rmat_utri_at(i, j, 0)] = 1.0;
+ lsq_rmat_utri_c_h[rmat_utri_at(i, j, 0)] = static_cast<TypeParam>(1.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
- this->lsq_moments[moments_at(i, 0, 2)] = 0.4;
- this->lsq_moments[moments_at(i, 0, 3)] = 0.5;
- this->lsq_moments[moments_at(i, 0, 4)] = 0.6;
- this->lsq_moments[moments_at(i, 0, 5)] = 0.7;
- this->lsq_moments[moments_at(i, 0, 6)] = 0.8;
- this->lsq_moments[moments_at(i, 0, 7)] = 0.9;
- this->lsq_moments[moments_at(i, 0, 8)] = 1.0;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
+ lsq_moments_h[moments_at(i, 0, 2)] = static_cast<TypeParam>(0.4);
+ lsq_moments_h[moments_at(i, 0, 3)] = static_cast<TypeParam>(0.5);
+ lsq_moments_h[moments_at(i, 0, 4)] = static_cast<TypeParam>(0.6);
+ lsq_moments_h[moments_at(i, 0, 5)] = static_cast<TypeParam>(0.7);
+ lsq_moments_h[moments_at(i, 0, 6)] = static_cast<TypeParam>(0.8);
+ lsq_moments_h[moments_at(i, 0, 7)] = static_cast<TypeParam>(0.9);
+ lsq_moments_h[moments_at(i, 0, 8)] = static_cast<TypeParam>(1.0);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_c<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -867,37 +1142,40 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCell) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- 0.28, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(0.28), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- -0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(-0.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
- -0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
+ static_cast<TypeParam>(-0.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(6, 0, 0, 0))],
- -0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(6, 0, 0, 0))],
+ static_cast<TypeParam>(-0.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(7, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(7, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(8, 0, 0, 0))],
- -0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(8, 0, 0, 0))],
+ static_cast<TypeParam>(-0.2), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(9, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(9, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
}
TYPED_TEST(HorizontalReconCubicTest, TestLsqCellRandom) {
@@ -916,6 +1194,17 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellRandom) {
at<nproma, (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_qtmat_c_h = Kokkos::create_mirror_view(this->lsq_qtmat_c);
+ auto lsq_rmat_rdiag_c_h = Kokkos::create_mirror_view(this->lsq_rmat_rdiag_c);
+ auto lsq_rmat_utri_c_h = Kokkos::create_mirror_view(this->lsq_rmat_utri_c);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
@@ -923,29 +1212,39 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellRandom) {
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_unk; ++j) {
for (int k = 0; k < lsq_dim_c; ++k) {
- this->lsq_qtmat_c[qtmat_at(i, j, k, 0)] = real_distrib(gen);
+ lsq_qtmat_c_h[qtmat_at(i, j, k, 0)] = real_distrib(gen);
}
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(i, j, 0)] = real_distrib(gen);
- this->lsq_moments[moments_at(i, 0, j)] = real_distrib(gen);
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(i, j, 0)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, j)] = real_distrib(gen);
}
for (int j = 0; j < (lsq_dim_unk * lsq_dim_unk - lsq_dim_unk) / 2; ++j) {
- this->lsq_rmat_utri_c[rmat_utri_at(i, j, 0)] = real_distrib(gen);
+ lsq_rmat_utri_c_h[rmat_utri_at(i, j, 0)] = real_distrib(gen);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_qtmat_c, lsq_qtmat_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_rdiag_c, lsq_rmat_rdiag_c_h);
+ Kokkos::deep_copy(this->lsq_rmat_utri_c, lsq_rmat_utri_c_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_c<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -957,27 +1256,31 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellRandom) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
std::vector<TypeParam> z_qt_times_d(lsq_dim_unk);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+ // calculating only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+ // calculating only for jk = 0
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
for (int j = 0; j < lsq_dim_unk; ++j) {
z_qt_times_d[j] = 0.0;
for (int i = 0; i < lsq_dim_c; ++i) {
- z_qt_times_d[j] +=
- this->lsq_qtmat_c[qtmat_at(jc, j, i, jb)] * z_d[i];
+ z_qt_times_d[j] += lsq_qtmat_c_h[qtmat_at(jc, j, i, jb)] * z_d[i];
}
}
int utri_id = 0;
@@ -985,28 +1288,31 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellRandom) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] = z_qt_times_d[j - 1];
for (int k = j + 1; k <= lsq_dim_unk; ++k) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] -=
- this->lsq_rmat_utri_c[rmat_utri_at(jc, utri_id++, jb)] *
+ lsq_rmat_utri_c_h[rmat_utri_at(jc, utri_id++, jb)] *
p_result[at<lsq_dim_unk + 1, nproma>(k, jc)];
}
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] *=
- this->lsq_rmat_rdiag_c[rmat_rdiag_at(jc, j - 1, jb)];
+ lsq_rmat_rdiag_c_h[rmat_rdiag_at(jc, j - 1, jb)];
}
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)];
+ p_cc_h[p_cc_at(jc, jk, jb)];
for (int j = 0; j < lsq_dim_unk; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] -=
p_result[at<lsq_dim_unk + 1, nproma>(j + 1, jc)] *
- this->lsq_moments[moments_at(jc, jb, j)];
+ lsq_moments_h[moments_at(jc, jb, j)];
}
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(i, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], 1e-5)
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
<< "For loop result fails for i = " << i << ", jc = " << jc;
}
}
@@ -1025,43 +1331,61 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVD) {
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = (i + 1);
+ p_cc_h[p_cc_at(i, 0, 0)] = static_cast<TypeParam>(i + 1);
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, 0)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, 0)] = (i + 1) % nproma;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, 0)] = 0;
for (int j = 1; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = i;
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = i;
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_c; ++j) {
- this->lsq_pseudoinv[pseudoinv_at(i, 0, j, 0)] = 1.0;
- this->lsq_pseudoinv[pseudoinv_at(i, 1, j, 0)] = 0.9;
- this->lsq_pseudoinv[pseudoinv_at(i, 2, j, 0)] = 0.8;
- this->lsq_pseudoinv[pseudoinv_at(i, 3, j, 0)] = 0.7;
- this->lsq_pseudoinv[pseudoinv_at(i, 4, j, 0)] = 0.6;
- this->lsq_pseudoinv[pseudoinv_at(i, 5, j, 0)] = 0.5;
- this->lsq_pseudoinv[pseudoinv_at(i, 6, j, 0)] = 0.4;
- this->lsq_pseudoinv[pseudoinv_at(i, 7, j, 0)] = 0.3;
- this->lsq_pseudoinv[pseudoinv_at(i, 8, j, 0)] = 0.2;
+ lsq_pseudoinv_h[pseudoinv_at(i, 0, j, 0)] = static_cast<TypeParam>(1.0);
+ lsq_pseudoinv_h[pseudoinv_at(i, 1, j, 0)] = static_cast<TypeParam>(0.9);
+ lsq_pseudoinv_h[pseudoinv_at(i, 2, j, 0)] = static_cast<TypeParam>(0.8);
+ lsq_pseudoinv_h[pseudoinv_at(i, 3, j, 0)] = static_cast<TypeParam>(0.7);
+ lsq_pseudoinv_h[pseudoinv_at(i, 4, j, 0)] = static_cast<TypeParam>(0.6);
+ lsq_pseudoinv_h[pseudoinv_at(i, 5, j, 0)] = static_cast<TypeParam>(0.5);
+ lsq_pseudoinv_h[pseudoinv_at(i, 6, j, 0)] = static_cast<TypeParam>(0.4);
+ lsq_pseudoinv_h[pseudoinv_at(i, 7, j, 0)] = static_cast<TypeParam>(0.3);
+ lsq_pseudoinv_h[pseudoinv_at(i, 8, j, 0)] = static_cast<TypeParam>(0.2);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = 0.0;
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
- this->lsq_moments[moments_at(i, 0, 0)] = 0.2;
- this->lsq_moments[moments_at(i, 0, 1)] = 0.3;
- this->lsq_moments[moments_at(i, 0, 2)] = 0.4;
- this->lsq_moments[moments_at(i, 0, 3)] = 0.5;
- this->lsq_moments[moments_at(i, 0, 4)] = 0.6;
- this->lsq_moments[moments_at(i, 0, 5)] = 0.7;
- this->lsq_moments[moments_at(i, 0, 6)] = 0.8;
- this->lsq_moments[moments_at(i, 0, 7)] = 0.9;
- this->lsq_moments[moments_at(i, 0, 8)] = 1.0;
+ lsq_moments_h[moments_at(i, 0, 0)] = static_cast<TypeParam>(0.2);
+ lsq_moments_h[moments_at(i, 0, 1)] = static_cast<TypeParam>(0.3);
+ lsq_moments_h[moments_at(i, 0, 2)] = static_cast<TypeParam>(0.4);
+ lsq_moments_h[moments_at(i, 0, 3)] = static_cast<TypeParam>(0.5);
+ lsq_moments_h[moments_at(i, 0, 4)] = static_cast<TypeParam>(0.6);
+ lsq_moments_h[moments_at(i, 0, 5)] = static_cast<TypeParam>(0.7);
+ lsq_moments_h[moments_at(i, 0, 6)] = static_cast<TypeParam>(0.8);
+ lsq_moments_h[moments_at(i, 0, 7)] = static_cast<TypeParam>(0.9);
+ lsq_moments_h[moments_at(i, 0, 8)] = static_cast<TypeParam>(1.0);
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_c_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -1072,37 +1396,40 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVD) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Check result
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
- -1.64, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(0, 0, 0, 0))],
+ static_cast<TypeParam>(-1.64), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
- 1.0, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(1, 0, 0, 0))],
+ static_cast<TypeParam>(1.0), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
- 0.9, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(2, 0, 0, 0))],
+ static_cast<TypeParam>(0.9), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
- 0.8, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(3, 0, 0, 0))],
+ static_cast<TypeParam>(0.8), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
- 0.7, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(4, 0, 0, 0))],
+ static_cast<TypeParam>(0.7), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
- 0.6, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(5, 0, 0, 0))],
+ static_cast<TypeParam>(0.6), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(6, 0, 0, 0))],
- 0.5, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(6, 0, 0, 0))],
+ static_cast<TypeParam>(0.5), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(7, 0, 0, 0))],
- 0.4, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(7, 0, 0, 0))],
+ static_cast<TypeParam>(0.4), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(8, 0, 0, 0))],
- 0.3, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(8, 0, 0, 0))],
+ static_cast<TypeParam>(0.3), 1e-6);
EXPECT_NEAR(
- this->p_coeff[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(9, 0, 0, 0))],
- 0.2, 1e-6);
+ p_coeff_h[(at<lsq_dim_unk + 1, nproma, nlev, nblks_c>(9, 0, 0, 0))],
+ static_cast<TypeParam>(0.2), 1e-6);
}
TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVDRandom) {
@@ -1118,6 +1445,16 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVDRandom) {
const auto &p_coeff_at = at<lsq_dim_unk + 1, nproma, nlev, nblks_c>;
const auto &moments_at = at<nproma, nblks_c, lsq_dim_unk>;
+ // Create host mirror views
+ auto p_cc_h = Kokkos::create_mirror_view(this->p_cc);
+ auto cell_neighbor_idx_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h =
+ Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto lsq_pseudoinv_h = Kokkos::create_mirror_view(this->lsq_pseudoinv);
+ auto lsq_moments_h = Kokkos::create_mirror_view(this->lsq_moments);
+ auto p_coeff_h = Kokkos::create_mirror_view(this->p_coeff);
+
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
@@ -1125,25 +1462,33 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVDRandom) {
// Initialization
for (int i = 0; i < nproma; ++i) {
- this->p_cc[p_cc_at(i, 0, 0)] = real_distrib(gen);
+ p_cc_h[p_cc_at(i, 0, 0)] = real_distrib(gen);
for (int j = 0; j < lsq_dim_c; ++j) {
- this->cell_neighbor_idx[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
- this->cell_neighbor_blk[cell_neighbor_at(i, 0, j)] = 0;
+ cell_neighbor_idx_h[cell_neighbor_at(i, 0, j)] = int_distrib(gen);
+ cell_neighbor_blk_h[cell_neighbor_at(i, 0, j)] = 0;
}
for (int j = 0; j < lsq_dim_unk; ++j) {
for (int k = 0; k < lsq_dim_c; ++k) {
- this->lsq_pseudoinv[pseudoinv_at(i, j, k, 0)] = real_distrib(gen);
+ lsq_pseudoinv_h[pseudoinv_at(i, j, k, 0)] = real_distrib(gen);
}
- this->lsq_moments[moments_at(i, 0, j)] = real_distrib(gen);
+ lsq_moments_h[moments_at(i, 0, j)] = real_distrib(gen);
}
for (int j = 0; j < lsq_dim_unk + 1; ++j) {
- this->p_coeff[p_coeff_at(j, i, 0, 0)] = real_distrib(gen);
+ p_coeff_h[p_coeff_at(j, i, 0, 0)] = static_cast<TypeParam>(0.0);
}
}
+ // Copy to device
+ Kokkos::deep_copy(this->p_cc, p_cc_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->lsq_pseudoinv, lsq_pseudoinv_h);
+ Kokkos::deep_copy(this->lsq_moments, lsq_moments_h);
+ Kokkos::deep_copy(this->p_coeff, p_coeff_h);
+
// Test function
recon_lsq_cell_c_svd<TypeParam>(
this->p_cc.data(), this->cell_neighbor_idx.data(),
@@ -1154,45 +1499,54 @@ TYPED_TEST(HorizontalReconCubicTest, TestLsqCellSVDRandom) {
this->lacc, this->nblks_c, this->nlev, this->lsq_dim_unk,
this->lsq_dim_c);
+ // Copy results back to host
+ Kokkos::deep_copy(p_coeff_h, this->p_coeff);
+
// Compute reference result
std::vector<TypeParam> z_d(lsq_dim_c);
std::vector<TypeParam> z_qt_times_d(lsq_dim_unk);
std::vector<TypeParam> p_result((lsq_dim_unk + 1) * nproma);
+ // calculating only for jb=0
for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+ // calculating only for jk = 0
for (int jk = this->slev; jk < this->elev; ++jk) {
- for (int jc = i_startidx; jc < i_endidx; ++jc) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
for (int i = 0; i < lsq_dim_c; ++i) {
- z_d[i] = this->p_cc[p_cc_at(
- this->cell_neighbor_idx[cell_neighbor_at(jc, jb, i)], jk,
- this->cell_neighbor_blk[cell_neighbor_at(jc, jb, i)])] -
- this->p_cc[p_cc_at(jc, jk, jb)];
+ z_d[i] = p_cc_h[p_cc_at(
+ cell_neighbor_idx_h[cell_neighbor_at(jc, jb, i)], jk,
+ cell_neighbor_blk_h[cell_neighbor_at(jc, jb, i)])] -
+ p_cc_h[p_cc_at(jc, jk, jb)];
}
for (int j = 1; j < lsq_dim_unk + 1; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] = 0.0;
for (int i = 0; i < lsq_dim_c; ++i) {
p_result[at<lsq_dim_unk + 1, nproma>(j, jc)] +=
- this->lsq_pseudoinv[pseudoinv_at(jc, j - 1, i, jb)] * z_d[i];
+ lsq_pseudoinv_h[pseudoinv_at(jc, j - 1, i, jb)] * z_d[i];
}
}
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] =
- this->p_cc[p_cc_at(jc, jk, jb)];
+ p_cc_h[p_cc_at(jc, jk, jb)];
for (int j = 0; j < lsq_dim_unk; ++j) {
p_result[at<lsq_dim_unk + 1, nproma>(0, jc)] -=
p_result[at<lsq_dim_unk + 1, nproma>(j + 1, jc)] *
- this->lsq_moments[moments_at(jc, jb, j)];
+ lsq_moments_h[moments_at(jc, jb, j)];
}
}
}
}
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Check result
for (int i = 0; i < lsq_dim_unk + 1; ++i) {
for (int jc = 0; jc < nproma; ++jc) {
- EXPECT_NEAR(this->p_coeff[(p_coeff_at(i, jc, 0, 0))],
- p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], 1e-5)
+ EXPECT_NEAR(p_coeff_h[(p_coeff_at(i, jc, 0, 0))],
+ p_result[(at<lsq_dim_unk + 1, nproma>(i, jc))], tol)
<< "For loop result fails for i = " << i << ", jc = " << jc;
}
}
diff --git a/test/c/test_horizontal_rot.cpp b/test/c/test_horizontal_rot.cpp
index 68e80245f2fa4ea19f173f1f8ac095fda5505775..ca675a494302ccac8df0ba5f01417cc36808153c 100644
--- a/test/c/test_horizontal_rot.cpp
+++ b/test/c/test_horizontal_rot.cpp
@@ -14,8 +14,8 @@
#include <vector>
#include <Kokkos_Core.hpp>
-#include <gtest/gtest.h>
#include <dim_helper.hpp>
+#include <gtest/gtest.h>
#include <horizontal/mo_lib_divrot.hpp>
#include <support/mo_lib_loopindices.hpp>
@@ -30,34 +30,36 @@ protected:
static constexpr int dim4d = 2; // 4th dimension size
int i_startblk = 0;
- int i_endblk = nblks_v; // Test blocks [0 .. nblks_v-1]
+ int i_endblk = nblks_v - 1; // Test blocks [0 .. nblks_v-1]
int i_startidx_in = 0;
- int i_endidx_in = nproma; // Full range: 0 .. nproma-1
+ int i_endidx_in = nproma - 1; // Full range: 0 .. nproma-1
std::vector<int> slev;
std::vector<int> elev;
bool lacc = false; // Not using ACC-specific behavior.
bool acc_async = false; // Not using ACC-specific behavior.
- std::vector<ValueType> vec_e;
- std::vector<int> vert_edge_idx;
- std::vector<int> vert_edge_blk;
- std::vector<ValueType> geofac_rot;
- std::vector<ValueType> rot_vec;
- std::vector<ValueType> f4din;
- std::vector<ValueType> f4dout;
-
- HorizontalRotVertexTest() {
- slev.resize(dim4d, 0);
- elev.resize(dim4d, nlev); // Full vertical range (0 .. nlev-1)
-
- vec_e.resize(dim_combine(nproma, nlev, nblks_e));
- vert_edge_idx.resize(dim_combine(nproma, nblks_v, 6));
- vert_edge_blk.resize(dim_combine(nproma, nblks_v, 6));
- geofac_rot.resize(dim_combine(nproma, 6, nblks_v));
- rot_vec.resize(dim_combine(nproma, nlev, nblks_v));
- f4din.resize(dim_combine(nproma, nlev, nblks_e, dim4d));
- f4dout.resize(dim_combine(nproma, nlev, nblks_v, dim4d));
- }
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Views for the test data. All the data is assigned as one-dimensional arrays
+ Kokkos::View<ValueType *, memory_space> vec_e;
+ Kokkos::View<int *, memory_space> vert_edge_idx;
+ Kokkos::View<int *, memory_space> vert_edge_blk;
+ Kokkos::View<ValueType *, memory_space> geofac_rot;
+ Kokkos::View<ValueType *, memory_space> rot_vec;
+ Kokkos::View<ValueType *, memory_space> f4din;
+ Kokkos::View<ValueType *, memory_space> f4dout;
+
+ HorizontalRotVertexTest()
+ : slev(dim4d, 0),
+ elev(dim4d, nlev - 1), // Full vertical range (0 .. nlev-1)
+ vec_e("vec_e", dim_combine(nproma, nlev, nblks_e)),
+ vert_edge_idx("vert_edge_idx", dim_combine(nproma, nblks_v, 6)),
+ vert_edge_blk("vert_edge_blk", dim_combine(nproma, nblks_v, 6)),
+ geofac_rot("geofac_rot", dim_combine(nproma, 6, nblks_v)),
+ rot_vec("rot_vec", dim_combine(nproma, nlev, nblks_v)),
+ f4din("f4din", dim_combine(nproma, nlev, nblks_e, dim4d)),
+ f4dout("f4dout", dim_combine(nproma, nlev, nblks_v, dim4d)) {}
};
/// ValueTypes which the divrot tests should run with
@@ -76,33 +78,46 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexAtmosSpecific) {
const auto &geofac_rot_at = at<nproma, 6, nblks_v>;
const auto &rot_vec_at = at<nproma, nlev, nblks_v>;
+ // Create host mirror views
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto vert_edge_idx_h = Kokkos::create_mirror_view(this->vert_edge_idx);
+ auto vert_edge_blk_h = Kokkos::create_mirror_view(this->vert_edge_blk);
+ auto geofac_rot_h = Kokkos::create_mirror_view(this->geofac_rot);
+ auto rot_vec_h = Kokkos::create_mirror_view(this->rot_vec);
+
// Initialization with specific values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
+ vec_e_h(vec_e_at(i, k, 0)) = (i + 1) * (k + 1); // Simple pattern
}
// Set edge indices to point to specific edges
for (int j = 0; j < 6; ++j) {
- this->vert_edge_idx[vert_edge_at(i, 0, j)] = (i + j) % nproma;
+ vert_edge_idx_h(vert_edge_at(i, 0, j)) = (i + j) % nproma;
// All edges are in the same block for this test
- this->vert_edge_blk[vert_edge_at(i, 0, j)] = 0;
+ vert_edge_blk_h(vert_edge_at(i, 0, j)) = 0;
}
- // Geometric factors for rotation
- this->geofac_rot[geofac_rot_at(i, 0, 0)] = 0.3;
- this->geofac_rot[geofac_rot_at(i, 1, 0)] = 0.2;
- this->geofac_rot[geofac_rot_at(i, 2, 0)] = 0.1;
- this->geofac_rot[geofac_rot_at(i, 3, 0)] = 0.2;
- this->geofac_rot[geofac_rot_at(i, 4, 0)] = 0.1;
- this->geofac_rot[geofac_rot_at(i, 5, 0)] = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 0, 0)) = 0.3;
+ geofac_rot_h(geofac_rot_at(i, 1, 0)) = 0.2;
+ geofac_rot_h(geofac_rot_at(i, 2, 0)) = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 3, 0)) = 0.2;
+ geofac_rot_h(geofac_rot_at(i, 4, 0)) = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 5, 0)) = 0.1;
// Initialize rot_vec to zero
for (int k = 0; k < nlev; ++k) {
- this->rot_vec[rot_vec_at(i, k, 0)] = 0.0;
+ rot_vec_h(rot_vec_at(i, k, 0)) = 0.0;
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->vert_edge_idx, vert_edge_idx_h);
+ Kokkos::deep_copy(this->vert_edge_blk, vert_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_rot, geofac_rot_h);
+ Kokkos::deep_copy(this->rot_vec, rot_vec_h);
+
// Call the rot_vertex_atmos function
rot_vertex_atmos<TypeParam>(
this->vec_e.data(), this->vert_edge_idx.data(),
@@ -111,13 +126,22 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexAtmosSpecific) {
this->slev[0], this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_e, this->nblks_v);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(rot_vec_h, this->rot_vec);
+
// Expected values based on the initialization pattern
- EXPECT_NEAR(this->rot_vec[rot_vec_at(0, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(0, 1, 0)], 3.4, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(1, 0, 0)], 2.1, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(1, 1, 0)], 4.2, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(2, 0, 0)], 2.2, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(2, 1, 0)], 4.4, 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(0, 0, 0)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(0, 1, 0)], static_cast<TypeParam>(3.4),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(1, 0, 0)], static_cast<TypeParam>(2.1),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(1, 1, 0)], static_cast<TypeParam>(4.2),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(2, 0, 0)], static_cast<TypeParam>(2.2),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(2, 1, 0)], static_cast<TypeParam>(4.4),
+ 1e-6);
}
TYPED_TEST(HorizontalRotVertexTest, TestRotVertexAtmosRandom) {
@@ -131,36 +155,50 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexAtmosRandom) {
const auto &geofac_rot_at = at<nproma, 6, nblks_v>;
const auto &rot_vec_at = at<nproma, nlev, nblks_v>;
+ // Create host mirror views
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto vert_edge_idx_h = Kokkos::create_mirror_view(this->vert_edge_idx);
+ auto vert_edge_blk_h = Kokkos::create_mirror_view(this->vert_edge_blk);
+ auto geofac_rot_h = Kokkos::create_mirror_view(this->geofac_rot);
+ auto rot_vec_h = Kokkos::create_mirror_view(this->rot_vec);
+
// Set up random number generators
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
// Initialization with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
+ vec_e_h(vec_e_at(i, k, 0)) = real_distrib(gen);
}
// Set random edge indices
for (int j = 0; j < 6; ++j) {
- this->vert_edge_idx[vert_edge_at(i, 0, j)] = int_distrib(gen);
- this->vert_edge_blk[vert_edge_at(i, 0, j)] =
+ vert_edge_idx_h(vert_edge_at(i, 0, j)) = int_distrib(gen);
+ vert_edge_blk_h(vert_edge_at(i, 0, j)) =
0; // Keep in same block for simplicity
}
// Random geometric factors
for (int j = 0; j < 6; ++j) {
- this->geofac_rot[geofac_rot_at(i, j, 0)] = real_distrib(gen);
+ geofac_rot_h(geofac_rot_at(i, j, 0)) = real_distrib(gen);
}
- // Initialize rot_vec to random values
+ // Initialize rot_vec to zero
for (int k = 0; k < nlev; ++k) {
- this->rot_vec[rot_vec_at(i, k, 0)] = real_distrib(gen);
+ rot_vec_h(rot_vec_at(i, k, 0)) = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->vert_edge_idx, vert_edge_idx_h);
+ Kokkos::deep_copy(this->vert_edge_blk, vert_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_rot, geofac_rot_h);
+ Kokkos::deep_copy(this->rot_vec, rot_vec_h);
+
// Call the rot_vertex_atmos function
rot_vertex_atmos<TypeParam>(
this->vec_e.data(), this->vert_edge_idx.data(),
@@ -169,50 +207,50 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexAtmosRandom) {
this->slev[0], this->elev[0], this->nproma, this->lacc, this->nlev,
this->nblks_e, this->nblks_v);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(rot_vec_h, this->rot_vec);
+
// Calculate reference values separately and verify results
std::vector<TypeParam> ref_rot_vec(nproma * nlev * nblks_v, 0.0);
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_v_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jv = i_startidx; jv < i_endidx; ++jv) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
ref_rot_vec[rot_vec_at(jv, jk, jb)] =
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 0)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 0)])] *
- this->geofac_rot[geofac_rot_at(jv, 0, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 1)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 1)])] *
- this->geofac_rot[geofac_rot_at(jv, 1, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 2)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 2)])] *
- this->geofac_rot[geofac_rot_at(jv, 2, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 3)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 3)])] *
- this->geofac_rot[geofac_rot_at(jv, 3, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 4)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 4)])] *
- this->geofac_rot[geofac_rot_at(jv, 4, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 5)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 5)])] *
- this->geofac_rot[geofac_rot_at(jv, 5, jb)];
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 0)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 0)])] *
+ geofac_rot_h[geofac_rot_at(jv, 0, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 1)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 1)])] *
+ geofac_rot_h[geofac_rot_at(jv, 1, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 2)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 2)])] *
+ geofac_rot_h[geofac_rot_at(jv, 2, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 3)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 3)])] *
+ geofac_rot_h[geofac_rot_at(jv, 3, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 4)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 4)])] *
+ geofac_rot_h[geofac_rot_at(jv, 4, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 5)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 5)])] *
+ geofac_rot_h[geofac_rot_at(jv, 5, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->rot_vec[rot_vec_at(i, k, 0)],
- ref_rot_vec[rot_vec_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(i, k, 0)],
+ ref_rot_vec[rot_vec_at(i, k, 0)], tol)
<< "Results differ at i=" << i << ", k=" << k;
}
}
@@ -231,33 +269,47 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexRISpecific) {
const auto &geofac_rot_at = at<nproma, 6, nblks_v>;
const auto &rot_vec_at = at<nproma, nlev, nblks_v>;
+ // Create host mirror views
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto vert_edge_idx_h = Kokkos::create_mirror_view(this->vert_edge_idx);
+ auto vert_edge_blk_h = Kokkos::create_mirror_view(this->vert_edge_blk);
+ auto geofac_rot_h = Kokkos::create_mirror_view(this->geofac_rot);
+ auto rot_vec_h = Kokkos::create_mirror_view(this->rot_vec);
+
// Initialization with specific values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = (i + 1) * (k + 1); // Simple pattern
+ vec_e_h(vec_e_at(i, k, 0)) = (i + 1) * (k + 1); // Simple pattern
}
// Set edge indices to point to specific edges
for (int j = 0; j < 6; ++j) {
- this->vert_edge_idx[vert_edge_at(i, 0, j)] = (i + j) % nproma;
+ vert_edge_idx_h(vert_edge_at(i, 0, j)) = (i + j) % nproma;
// All edges are in the same block for this test
- this->vert_edge_blk[vert_edge_at(i, 0, j)] = 0;
+ vert_edge_blk_h(vert_edge_at(i, 0, j)) = 0;
}
// Geometric factors for rotation
- this->geofac_rot[geofac_rot_at(i, 0, 0)] = 0.3;
- this->geofac_rot[geofac_rot_at(i, 1, 0)] = 0.2;
- this->geofac_rot[geofac_rot_at(i, 2, 0)] = 0.1;
- this->geofac_rot[geofac_rot_at(i, 3, 0)] = 0.2;
- this->geofac_rot[geofac_rot_at(i, 4, 0)] = 0.1;
- this->geofac_rot[geofac_rot_at(i, 5, 0)] = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 0, 0)) = 0.3;
+ geofac_rot_h(geofac_rot_at(i, 1, 0)) = 0.2;
+ geofac_rot_h(geofac_rot_at(i, 2, 0)) = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 3, 0)) = 0.2;
+ geofac_rot_h(geofac_rot_at(i, 4, 0)) = 0.1;
+ geofac_rot_h(geofac_rot_at(i, 5, 0)) = 0.1;
// Initialize rot_vec to zero
for (int k = 0; k < nlev; ++k) {
- this->rot_vec[rot_vec_at(i, k, 0)] = 0.0;
+ rot_vec_h(rot_vec_at(i, k, 0)) = 0.0;
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->vert_edge_idx, vert_edge_idx_h);
+ Kokkos::deep_copy(this->vert_edge_blk, vert_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_rot, geofac_rot_h);
+ Kokkos::deep_copy(this->rot_vec, rot_vec_h);
+
// Call the rot_vertex_ri function
rot_vertex_ri<TypeParam>(
this->vec_e.data(), this->vert_edge_idx.data(),
@@ -266,13 +318,22 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexRISpecific) {
this->slev[0], this->elev[0], this->nproma, this->lacc, this->acc_async,
this->nlev, this->nblks_e, this->nblks_v);
+ // Copy results back to host for verification
+ Kokkos::deep_copy(rot_vec_h, this->rot_vec);
+
// Expected values based on the initialization pattern
- EXPECT_NEAR(this->rot_vec[rot_vec_at(0, 0, 0)], 1.7, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(0, 1, 0)], 3.4, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(1, 0, 0)], 2.1, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(1, 1, 0)], 4.2, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(2, 0, 0)], 2.2, 1e-6);
- EXPECT_NEAR(this->rot_vec[rot_vec_at(2, 1, 0)], 4.4, 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(0, 0, 0)], static_cast<TypeParam>(1.7),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(0, 1, 0)], static_cast<TypeParam>(3.4),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(1, 0, 0)], static_cast<TypeParam>(2.1),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(1, 1, 0)], static_cast<TypeParam>(4.2),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(2, 0, 0)], static_cast<TypeParam>(2.2),
+ 1e-6);
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(2, 1, 0)], static_cast<TypeParam>(4.4),
+ 1e-6);
}
TYPED_TEST(HorizontalRotVertexTest, TestRotVertexRIRandom) {
@@ -286,36 +347,50 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexRIRandom) {
const auto &geofac_rot_at = at<nproma, 6, nblks_v>;
const auto &rot_vec_at = at<nproma, nlev, nblks_v>;
+ // Create host mirror views
+ auto vec_e_h = Kokkos::create_mirror_view(this->vec_e);
+ auto vert_edge_idx_h = Kokkos::create_mirror_view(this->vert_edge_idx);
+ auto vert_edge_blk_h = Kokkos::create_mirror_view(this->vert_edge_blk);
+ auto geofac_rot_h = Kokkos::create_mirror_view(this->geofac_rot);
+ auto rot_vec_h = Kokkos::create_mirror_view(this->rot_vec);
+
// Set up random number generators
std::random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<int> int_distrib(0, nproma - 1);
- std::uniform_real_distribution<TypeParam> real_distrib(-10.0, 10.0);
+ std::uniform_real_distribution<TypeParam> real_distrib(-1.0, 1.0);
// Initialization with random values
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- this->vec_e[vec_e_at(i, k, 0)] = real_distrib(gen);
+ vec_e_h(vec_e_at(i, k, 0)) = real_distrib(gen);
}
// Set random edge indices
for (int j = 0; j < 6; ++j) {
- this->vert_edge_idx[vert_edge_at(i, 0, j)] = int_distrib(gen);
- this->vert_edge_blk[vert_edge_at(i, 0, j)] =
+ vert_edge_idx_h(vert_edge_at(i, 0, j)) = int_distrib(gen);
+ vert_edge_blk_h(vert_edge_at(i, 0, j)) =
0; // Keep in same block for simplicity
}
// Random geometric factors
for (int j = 0; j < 6; ++j) {
- this->geofac_rot[geofac_rot_at(i, j, 0)] = real_distrib(gen);
+ geofac_rot_h(geofac_rot_at(i, j, 0)) = real_distrib(gen);
}
- // Initialize rot_vec to random values
+ // Initialize rot_vec to zero
for (int k = 0; k < nlev; ++k) {
- this->rot_vec[rot_vec_at(i, k, 0)] = real_distrib(gen);
+ rot_vec_h(rot_vec_at(i, k, 0)) = static_cast<TypeParam>(0.0);
}
}
+ // Copy initialized data to device
+ Kokkos::deep_copy(this->vec_e, vec_e_h);
+ Kokkos::deep_copy(this->vert_edge_idx, vert_edge_idx_h);
+ Kokkos::deep_copy(this->vert_edge_blk, vert_edge_blk_h);
+ Kokkos::deep_copy(this->geofac_rot, geofac_rot_h);
+ Kokkos::deep_copy(this->rot_vec, rot_vec_h);
+
// Call the rot_vertex_ri function
rot_vertex_ri<TypeParam>(
this->vec_e.data(), this->vert_edge_idx.data(),
@@ -324,55 +399,51 @@ TYPED_TEST(HorizontalRotVertexTest, TestRotVertexRIRandom) {
this->slev[0], this->elev[0], this->nproma, this->lacc, this->acc_async,
this->nlev, this->nblks_e, this->nblks_v);
- // Ensure computation is complete for both modes
- Kokkos::fence();
+ // Copy results back to host for verification
+ Kokkos::deep_copy(rot_vec_h, this->rot_vec);
// Calculate reference values separately and verify results
std::vector<TypeParam> ref_rot_vec(nproma * nlev * nblks_v, 0.0);
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
int i_startidx, i_endidx;
get_indices_v_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
this->i_startblk, this->i_endblk, i_startidx, i_endidx);
- for (int jk = this->slev[0]; jk < this->elev[0]; ++jk) {
- for (int jv = i_startidx; jv < i_endidx; ++jv) {
+ for (int jk = this->slev[0]; jk <= this->elev[0]; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
ref_rot_vec[rot_vec_at(jv, jk, jb)] =
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 0)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 0)])] *
- this->geofac_rot[geofac_rot_at(jv, 0, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 1)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 1)])] *
- this->geofac_rot[geofac_rot_at(jv, 1, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 2)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 2)])] *
- this->geofac_rot[geofac_rot_at(jv, 2, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 3)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 3)])] *
- this->geofac_rot[geofac_rot_at(jv, 3, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 4)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 4)])] *
- this->geofac_rot[geofac_rot_at(jv, 4, jb)] +
- this->vec_e[vec_e_at(
- this->vert_edge_idx[vert_edge_at(jv, jb, 5)], jk,
- this->vert_edge_blk[vert_edge_at(jv, jb, 5)])] *
- this->geofac_rot[geofac_rot_at(jv, 5, jb)];
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 0)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 0)])] *
+ geofac_rot_h[geofac_rot_at(jv, 0, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 1)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 1)])] *
+ geofac_rot_h[geofac_rot_at(jv, 1, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 2)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 2)])] *
+ geofac_rot_h[geofac_rot_at(jv, 2, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 3)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 3)])] *
+ geofac_rot_h[geofac_rot_at(jv, 3, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 4)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 4)])] *
+ geofac_rot_h[geofac_rot_at(jv, 4, jb)] +
+ vec_e_h[vec_e_at(vert_edge_idx_h[vert_edge_at(jv, jb, 5)], jk,
+ vert_edge_blk_h[vert_edge_at(jv, jb, 5)])] *
+ geofac_rot_h[geofac_rot_at(jv, 5, jb)];
}
}
}
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
// Verify results
for (int i = 0; i < nproma; ++i) {
for (int k = 0; k < nlev; ++k) {
- EXPECT_NEAR(this->rot_vec[rot_vec_at(i, k, 0)],
- ref_rot_vec[rot_vec_at(i, k, 0)], 1e-5)
+ EXPECT_NEAR(rot_vec_h[rot_vec_at(i, k, 0)],
+ ref_rot_vec[rot_vec_at(i, k, 0)], tol)
<< "Results differ at i=" << i << ", k=" << k << ")";
}
}
}
-
diff --git a/test/c/test_interpolation_scalar.cpp b/test/c/test_interpolation_scalar.cpp
index 507ec3f01ae5e79b6053cd52fc0d1af41cc67f41..20ccf4f205cfc8f9945aa2ae69a8daf10cf5ea76 100644
--- a/test/c/test_interpolation_scalar.cpp
+++ b/test/c/test_interpolation_scalar.cpp
@@ -10,18 +10,12 @@
// ---------------------------------------------------------------
#include "mo_lib_interpolation_scalar.hpp"
+#include "mo_lib_loopindices.hpp"
#include <Kokkos_Core.hpp>
#include <gtest/gtest.h>
#include <vector>
-
-// Free-function helpers for 3D and 4D array sizes (assumed column-major)
-template <typename T> size_t num_elements_3d(int d1, int d2, int d3) {
- return static_cast<size_t>(d1) * d2 * d3;
-}
-
-template <typename T> size_t num_elements_4d(int d1, int d2, int d3, int d4) {
- return static_cast<size_t>(d1) * d2 * d3 * d4;
-}
+#include <random>
+#include "dim_helper.hpp"
// Define a helper struct that holds the two types.
template <typename InT, typename OutT> struct MixedPrecision {
@@ -44,116 +38,99 @@ typedef ::testing::Types<MixedPrecision<double, double>,
class interp_dimensions {
public:
// Constant dimensions.
- static constexpr int nproma = 16; // inner loop length
- static constexpr int nlev = 7; // number of vertical levels
+ static constexpr int nproma = 2; // inner loop length
+ static constexpr int nlev = 3; // number of vertical levels
static constexpr int nblks_c = 2; // number of cell blocks
static constexpr int nblks_e = 2; // number of edge blocks (for p_e_in)
static constexpr int nblks_v = 2; // number of vertex blocks
+ static constexpr int cell_type = 6;
+ static constexpr int npromz_c = 2;
+
// Parameter values.
const int i_startblk = 0;
- const int i_endblk = 1; // Test blocks [0, 1]
- const int i_startidx = 2;
- const int i_endidx = nproma - 3; // Partial range: 2 .. nproma-3
- const int slev = 1;
+ const int i_endblk = nblks_c - 1; // Test blocks [0, 1]
+ const int i_startidx = 0;
+ const int i_endidx = nproma - 1; // Partial range: 2 .. nproma-3
+ const int slev = 0;
const int elev = nlev - 1; // Partial vertical range (1 .. nlev-1)
const bool lacc = false; // Not using ACC-specific behavior.
const bool acc_async = false; // No asynchronous execution.
};
-template <typename T>
-class InterpolationScalarTypedTestFixture : public ::testing::Test,
+template <typename ValueType>
+class InterpolationScalarSingleParamTest : public ::testing::Test,
public interp_dimensions {
-public:
- // Arrays used for verts2edges
- std::vector<T> p_vertex_in; // Dimensions: (nproma, nlev, nblks_v)
- std::vector<int> edge_vertex_idx; // Dimensions: (nproma, nblks_e, 4)
- std::vector<int> edge_vertex_blk; // Dimensions: (nproma, nblks_e, 4)
- std::vector<T> coeff_int_edges; // Dimensions: (nproma, 2, nblks_e)
- std::vector<T> p_edge_out; // Dimensions: (nproma, nlev, nblks_e)
-
- // Arrays used for edges2verts
- std::vector<T> p_edge_in; // Dimensions: (nproma, nlev, nblks_e)
- std::vector<int> edge_vert_idx; // Dimensions: (nproma, nblks_e, 6)
- std::vector<int> edge_vert_blk; // Dimensions: (nproma, nblks_e, 6)
- std::vector<T> v_int; // Dimensions: (nproma, 6, nblks_v)
- std::vector<T> p_vert_out; // Dimensions: (nproma, nlev, nblks_v)
-
- // Arrays used for edges2cells
- // std::vector<T> p_edge_in; // Dimensions: (nproma, nlev, nblks_e)
- std::vector<int> edge_idx; // Dimensions: (nproma, nblks_c, 3)
- std::vector<int> edge_blk; // Dimensions: (nproma, nblks_c, 3)
- std::vector<T> coeff_int_cells; // Dimensions: (nproma, 3, nblks_c)
- std::vector<T> p_cell_out; // Dimensions: (nproma, nlev, nblks_c)
+ protected:
+
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ Kokkos::View<ValueType*, memory_space> p_vertex_in;
+ Kokkos::View<int*, memory_space> edge_vertex_idx;
+ Kokkos::View<int*, memory_space> edge_vertex_blk;
+ Kokkos::View<ValueType*, memory_space> coeff_int_edges;
+ Kokkos::View<ValueType*, memory_space> p_edge_out;
+
+ // // Arrays used for edges2verts
+ Kokkos::View<ValueType*, memory_space> p_edge_in;
+ Kokkos::View<int*, memory_space> edge_vert_idx;
+ Kokkos::View<int*, memory_space> edge_vert_blk;
+ Kokkos::View<ValueType*, memory_space> v_int;
+ Kokkos::View<ValueType*, memory_space> p_vert_out;
+
+ // // Arrays used for edges2cells
+ Kokkos::View<int*, memory_space> edge_idx; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<int*, memory_space> edge_blk; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<ValueType*, memory_space> coeff_int_cells; // Dimensions: (nproma, 3, nblks_c)
+ Kokkos::View<ValueType*, memory_space> p_cell_out; // Dimensions: (nproma, nlev, nblks_c)
// Arrays used for verts2cells
- std::vector<T> p_vert_in; // Dimensions: (nproma, nlev, nblks_v)
- std::vector<int> cell_index_idx; // Dimensions: (nproma, nblks_c, 3)
- std::vector<int> cell_index_blk; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<ValueType*, memory_space> p_vert_in; // Dimensions: (nproma, nlev, nblks_v)
+ Kokkos::View<int*, memory_space> cell_index_idx; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<int*, memory_space> cell_index_blk; // Dimensions: (nproma, nblks_c, 3)
// Arrays used for avg_lib
- std::vector<T> psi_c; // Dimensions: (nproma, nlev, nblks_c)
- std::vector<int> cell_neighbor_idx; // Dimensions: (nproma, nblks_c, 3)
- std::vector<int> cell_neighbor_blk; // Dimensions: (nproma, nblks_c, 3)
- std::vector<T> avg_coeff; // Dimensions: (nproma, nlev, nblks_c)
- std::vector<T> avg_psi_c; // Dimensions: (nproma, nlev, nblks_c)
-
- const int cell_type = 6;
- const int npromz_c = 32;
-
- InterpolationScalarTypedTestFixture() {
- // Allocate and initialize arrays needed for verts2edges
- p_vertex_in.resize(num_elements_3d<T>(nproma, nlev, nblks_v),
- static_cast<T>(1));
- edge_vertex_idx.resize(num_elements_3d<int>(nproma, nblks_e, 4), 1);
- edge_vertex_blk.resize(num_elements_3d<int>(nproma, nblks_e, 4), 0);
- coeff_int_edges.resize(num_elements_3d<T>(nproma, 2, nblks_e),
- static_cast<T>(1));
-
- p_edge_out.resize(num_elements_3d<T>(nproma, nlev, nblks_e),
- static_cast<T>(0));
-
- // Allocate & Initialize arrays needed for edges2verts
- p_edge_in.resize(num_elements_3d<T>(nproma, nlev, nblks_e),
- static_cast<T>(1));
- edge_vert_idx.resize(num_elements_3d<int>(nproma, nblks_e, 6), 1);
- edge_vert_blk.resize(num_elements_3d<int>(nproma, nblks_e, 6), 0);
- v_int.resize(num_elements_3d<T>(nproma, 6, nblks_v), static_cast<T>(1));
-
- p_vert_out.resize(num_elements_3d<T>(nproma, nlev, nblks_v),
- static_cast<T>(0));
-
- // Allocate & Initialize arrays needed for edges2cells
- edge_idx.resize(num_elements_3d<int>(nproma, nblks_c, 3), 1);
- edge_blk.resize(num_elements_3d<int>(nproma, nblks_c, 3), 0);
- coeff_int_cells.resize(num_elements_3d<T>(nproma, 3, nblks_c),
- static_cast<T>(1));
-
- p_cell_out.resize(num_elements_3d<T>(nproma, nlev, nblks_c),
- static_cast<T>(0));
-
- // Allocate and initialize arrays needed for verts2cells
- p_vert_in.resize(num_elements_3d<T>(nproma, nlev, nblks_v),
- static_cast<T>(1));
- cell_index_idx.resize(num_elements_3d<int>(nproma, nblks_c, 3), 1);
- cell_index_blk.resize(num_elements_3d<int>(nproma, nblks_c, 3), 0);
-
- // Allocate and initialize arrays needed for avg_lib
- psi_c.resize(num_elements_3d<T>(nproma, nlev, nblks_c), static_cast<T>(1));
- cell_neighbor_idx.resize(num_elements_3d<int>(nproma, nblks_c, 3), 1);
- cell_neighbor_blk.resize(num_elements_3d<int>(nproma, nblks_c, 3), 0);
- avg_coeff.resize(num_elements_3d<T>(nproma, nlev, nblks_c),
- static_cast<T>(1));
-
- // Allocate output arrays and initialize to zero.
- avg_psi_c.resize(num_elements_3d<T>(nproma, nlev, nblks_c),
- static_cast<T>(0));
- }
+ Kokkos::View<ValueType*, memory_space> psi_c; // Dimensions: (nproma, nlev, nblks_c)
+ Kokkos::View<int*, memory_space> cell_neighbor_idx; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<int*, memory_space> cell_neighbor_blk; // Dimensions: (nproma, nblks_c, 3)
+ Kokkos::View<ValueType*, memory_space> avg_coeff; // Dimensions: (nproma, 4, nblks_c)
+ Kokkos::View<ValueType*, memory_space> avg_psi_c; // Dimensions: (nproma, nlev, nblks_c)
+
+ InterpolationScalarSingleParamTest()
+ : p_vertex_in("p_vertex_in", nproma * nlev * nblks_v),
+ edge_vertex_idx("edge_vertex_idx", nproma * nblks_e * 4),
+ edge_vertex_blk("edge_vertex_blk", nproma * nblks_e * 4),
+ coeff_int_edges("coeff_int_edges", nproma * 2 * nblks_e),
+ p_edge_out("p_edge_out", nproma * nlev * nblks_e),
+
+ p_edge_in("p_edge_in", nproma * nlev * nblks_e),
+ edge_vert_idx("edge_vert_idx", nproma * nblks_e * 6),
+ edge_vert_blk("edge_vert_blk", nproma * nblks_e * 6),
+ v_int("v_int", nproma * 6 * nblks_v),
+ p_vert_out("p_vert_out", nproma * nlev * nblks_v),
+
+ edge_idx("edge_idx", nproma * nblks_c * 3),
+ edge_blk("edge_blk", nproma * nblks_c * 3),
+ coeff_int_cells("coeff_int_cells", nproma * 3 * nblks_c),
+ p_cell_out("p_cell_out", nproma * nlev * nblks_c),
+
+ p_vert_in("p_vert_in", nproma * nlev * nblks_v),
+ cell_index_idx("cell_index_idx", nproma * nblks_c * 3),
+ cell_index_blk("cell_index_blk", nproma * nblks_c * 3),
+
+ psi_c("psi_c", nproma * nlev * nblks_c),
+ cell_neighbor_idx("cell_neighbor_idx", nproma * nblks_c * 3),
+ cell_neighbor_blk("cell_neighbor_blk", nproma * nblks_c * 3),
+ avg_coeff("avg_coeff", nproma * 4 * nblks_c), // 4 coefficients (self + 3 neighbors)
+ avg_psi_c("avg_psi_c", nproma * nlev * nblks_c)
+ {}
};
typedef ::testing::Types<float, double> SingleType;
-TYPED_TEST_SUITE(InterpolationScalarTypedTestFixture, SingleType);
+TYPED_TEST_SUITE(InterpolationScalarSingleParamTest, SingleType);
////////////////////////////////////////////////////////////////////////////////
//
@@ -161,29 +138,203 @@ TYPED_TEST_SUITE(InterpolationScalarTypedTestFixture, SingleType);
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarTypedTestFixture, Verts2Edges) {
+TYPED_TEST(InterpolationScalarSingleParamTest, Verts2EdgesSpecific) {
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int nblks_e = this->nblks_e;
+
+ // Define indexing helpers
+ const auto &vertex_at = at<nproma, nlev, nblks_v>;
+ const auto &idx_at = at<nproma, nblks_e, 4>;
+ const auto &blk_at = at<nproma, nblks_e, 4>;
+ const auto &coeff_at = at<nproma, 2, nblks_e>;
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_vertex_in_h = Kokkos::create_mirror_view(this->p_vertex_in);
+ auto edge_vertex_idx_h = Kokkos::create_mirror_view(this->edge_vertex_idx);
+ auto edge_vertex_blk_h = Kokkos::create_mirror_view(this->edge_vertex_blk);
+ auto coeff_int_edges_h = Kokkos::create_mirror_view(this->coeff_int_edges);
+ auto p_edge_out_h = Kokkos::create_mirror_view(this->p_edge_out);
+
+ // Initialize with specific test values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vertex_in_h[vertex_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);;
+ }
+ }
+ }
+
+ // Initialize edge connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each edge connects to two specific vertices
+ edge_vertex_idx_h[idx_at(ic, ib, 0)] = ic % (nproma - 1); // First vertex index
+ edge_vertex_idx_h[idx_at(ic, ib, 1)] = (ic + 1) % nproma; // Second vertex index
+ edge_vertex_idx_h[idx_at(ic, ib, 2)] = 0; // Not used
+ edge_vertex_idx_h[idx_at(ic, ib, 3)] = 0; // Not used
+
+ edge_vertex_blk_h[blk_at(ic, ib, 0)] = ib % nblks_v; // First vertex block
+ edge_vertex_blk_h[blk_at(ic, ib, 1)] = (ib + 1) % nblks_v; // Second vertex block
+ edge_vertex_blk_h[blk_at(ic, ib, 2)] = 0; // Not used
+ edge_vertex_blk_h[blk_at(ic, ib, 3)] = 0; // Not used
+
+ coeff_int_edges_h[coeff_at(ic, 0, ib)] = static_cast<TypeParam>(0.5 + ic * 0.01);
+ coeff_int_edges_h[coeff_at(ic, 1, ib)] = static_cast<TypeParam>(0.5 + ib * 0.01);
+
+ // Initialize output to zero and calculate expected results
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_edge_out_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // copy data to device
+ Kokkos::deep_copy(this->p_vertex_in, p_vertex_in_h);
+ Kokkos::deep_copy(this->edge_vertex_idx, edge_vertex_idx_h);
+ Kokkos::deep_copy(this->edge_vertex_blk, edge_vertex_blk_h);
+ Kokkos::deep_copy(this->coeff_int_edges, coeff_int_edges_h);
+ Kokkos::deep_copy(this->p_edge_out, p_edge_out_h);
verts2edges_scalar_lib<TypeParam>(
this->p_vertex_in.data(), this->edge_vertex_idx.data(),
this->edge_vertex_blk.data(), this->coeff_int_edges.data(),
this->p_edge_out.data(), this->i_startblk, this->i_endblk,
- this->i_startidx, this->i_endidx, this->slev, this->elev, this->nproma,
- this->nlev, this->nblks_v, this->nblks_e, this->lacc);
-
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx, i_endidx] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 2 stencil points,
- // expect 2.
- EXPECT_NEAR(this->p_edge_out[idx], static_cast<TypeParam>(2),
- static_cast<TypeParam>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ this->i_startidx, this->i_endidx, this->slev, this->elev, nproma,
+ nlev, nblks_v, nblks_e, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_edge_out_h, this->p_edge_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_edges = {
+ 1.505, 1.015, 1.605, 1.116, 1.705, 1.217,
+ 1.525, 1.0251, 1.626, 1.1271, 1.727, 1.2291
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = this->i_startidx; jv <= this->i_endidx; ++jv) {
+ EXPECT_NEAR(p_edge_out_h[edge_at(jv, jk, jb)],
+ expected_edges[edge_at(jv, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+ }
+ }
+ }
+}
+
+// Repeat the same test with randomized data
+TYPED_TEST(InterpolationScalarSingleParamTest, Verts2EdgesRandom) {
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int nblks_e = this->nblks_e;
+
+ // Define indexing helpers
+ const auto &vertex_at = at<nproma, nlev, nblks_v>;
+ const auto &idx_at = at<nproma, nblks_e, 4>;
+ const auto &blk_at = at<nproma, nblks_e, 4>;
+ const auto &coeff_at = at<nproma, 2, nblks_e>;
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_vertex_in_h = Kokkos::create_mirror_view(this->p_vertex_in);
+ auto edge_vertex_idx_h = Kokkos::create_mirror_view(this->edge_vertex_idx);
+ auto edge_vertex_blk_h = Kokkos::create_mirror_view(this->edge_vertex_blk);
+ auto coeff_int_edges_h = Kokkos::create_mirror_view(this->coeff_int_edges);
+ auto p_edge_out_h = Kokkos::create_mirror_view(this->p_edge_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_v - 1);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vertex_in_h[vertex_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // Initialize edge connectivity indices with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ edge_vertex_idx_h[idx_at(ic, ib, 0)] = edge_distrib(gen);
+ edge_vertex_idx_h[idx_at(ic, ib, 1)] = edge_distrib(gen);
+ edge_vertex_idx_h[idx_at(ic, ib, 2)] = edge_distrib(gen);
+ edge_vertex_idx_h[idx_at(ic, ib, 3)] = edge_distrib(gen);
+
+ edge_vertex_blk_h[blk_at(ic, ib, 0)] = block_distrib(gen);
+ edge_vertex_blk_h[blk_at(ic, ib, 1)] = block_distrib(gen);
+ edge_vertex_blk_h[blk_at(ic, ib, 2)] = block_distrib(gen);
+ edge_vertex_blk_h[blk_at(ic, ib, 3)] = block_distrib(gen);
+
+ coeff_int_edges_h[coeff_at(ic, 0, ib)] = real_distrib(gen);
+ coeff_int_edges_h[coeff_at(ic, 1, ib)] = real_distrib(gen);
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_edge_out_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // copy data to device
+ Kokkos::deep_copy(this->p_vertex_in, p_vertex_in_h);
+ Kokkos::deep_copy(this->edge_vertex_idx, edge_vertex_idx_h);
+ Kokkos::deep_copy(this->edge_vertex_blk, edge_vertex_blk_h);
+ Kokkos::deep_copy(this->coeff_int_edges, coeff_int_edges_h);
+ Kokkos::deep_copy(this->p_edge_out, p_edge_out_h);
+
+ // Call the function
+ verts2edges_scalar_lib<TypeParam>(
+ this->p_vertex_in.data(), this->edge_vertex_idx.data(),
+ this->edge_vertex_blk.data(), this->coeff_int_edges.data(),
+ this->p_edge_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev, nproma,
+ nlev, nblks_v, nblks_e, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_edge_out_h, this->p_edge_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_edges("expected_edges", nproma, nlev, nblks_e);
+
+ for (int ib = this->i_startblk; ib <= this->i_endblk; ++ib) {
+ for (int ik = this->slev; ik <= this->elev; ++ik) {
+ for (int ic = this->i_startidx; ic <= this->i_endidx; ++ic) {
+ // Compute expected values
+ expected_edges(ic, ik, ib) =
+ coeff_int_edges_h[coeff_at(ic, 0, ib)] *
+ p_vertex_in_h[vertex_at(edge_vertex_idx_h[idx_at(ic, ib, 0)], ik,
+ edge_vertex_blk_h[blk_at(ic, ib, 0)])] +
+ coeff_int_edges_h[coeff_at(ic, 1, ib)] *
+ p_vertex_in_h[vertex_at(edge_vertex_idx_h[idx_at(ic, ib, 1)], ik,
+ edge_vertex_blk_h[blk_at(ic, ib, 1)])];
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = this->i_startidx; jv <= this->i_endidx; ++jv) {
+ EXPECT_NEAR(p_edge_out_h[edge_at(jv, jk, jb)],
+ expected_edges(jv, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
}
}
}
@@ -195,29 +346,196 @@ TYPED_TEST(InterpolationScalarTypedTestFixture, Verts2Edges) {
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarTypedTestFixture, Edges2Verts) {
+TYPED_TEST(InterpolationScalarSingleParamTest, Edges2VertsSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_edge_in_h = Kokkos::create_mirror_view(this->p_edge_in);
+ auto edge_vert_idx_h = Kokkos::create_mirror_view(this->edge_vert_idx);
+ auto edge_vert_blk_h = Kokkos::create_mirror_view(this->edge_vert_blk);
+ auto v_int_h = Kokkos::create_mirror_view(this->v_int);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_edge_in_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex connects to 6 edges
+ for (int j = 0; j < 6; ++j) {
+ // Edge indices with a pattern
+ edge_vert_idx_h[idx_at(ic, ib, j)] = (ic + j) % nproma;
+ edge_vert_blk_h[blk_at(ic, ib, j)] = (ib + j % 2) % nblks_e;
+
+ // Interpolation coefficients that depend on indices
+ v_int_h[coeff_at(ic, j, ib)] = static_cast<TypeParam>(1.0 / 6.0 + j * 0.01);
+}
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_edge_in, p_edge_in_h);
+ Kokkos::deep_copy(this->edge_vert_idx, edge_vert_idx_h);
+ Kokkos::deep_copy(this->edge_vert_blk, edge_vert_blk_h);
+ Kokkos::deep_copy(this->v_int, v_int_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+ // Call the function under test
edges2verts_scalar_lib<TypeParam>(
this->p_edge_in.data(), this->edge_vert_idx.data(),
- this->edge_vert_blk.data(), this->v_int.data(), this->p_vert_out.data(),
- this->i_startblk, this->i_endblk, this->i_startidx, this->i_endidx,
- this->slev, this->elev, this->nproma, this->nlev, this->nblks_e,
- this->nblks_v, this->lacc);
-
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 6 stencil points,
- // expect 6.
- EXPECT_NEAR(this->p_vert_out[idx], static_cast<TypeParam>(6),
- static_cast<TypeParam>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ this->edge_vert_blk.data(), this->v_int.data(),
+ this->p_vert_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_e, nblks_v, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_verts = {
+ 1.7459, 1.7159, 1.8609, 1.8309, 1.9759, 1.9459,
+ 1.7456, 1.7156, 1.8606, 1.8306, 1.9756, 1.9456
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = this->i_startidx; jv <= this->i_endidx; ++jv) {
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts[vert_at(jv, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+ }
+ }
+ }
+}
+
+TYPED_TEST(InterpolationScalarSingleParamTest, Edges2VertsRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_edge_in_h = Kokkos::create_mirror_view(this->p_edge_in);
+ auto edge_vert_idx_h = Kokkos::create_mirror_view(this->edge_vert_idx);
+ auto edge_vert_blk_h = Kokkos::create_mirror_view(this->edge_vert_blk);
+ auto v_int_h = Kokkos::create_mirror_view(this->v_int);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_e - 1);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_edge_in_h[edge_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex connects to 6 edges
+ for (int j = 0; j < 6; ++j) {
+ edge_vert_idx_h[idx_at(ic, ib, j)] = edge_distrib(gen);
+ edge_vert_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+
+ // Random interpolation coefficients
+ v_int_h[coeff_at(ic, j, ib)] = real_distrib(gen) / 6.0; // Scaled to ensure reasonable sums
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_edge_in, p_edge_in_h);
+ Kokkos::deep_copy(this->edge_vert_idx, edge_vert_idx_h);
+ Kokkos::deep_copy(this->edge_vert_blk, edge_vert_blk_h);
+ Kokkos::deep_copy(this->v_int, v_int_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+
+ // Call the function under test
+ edges2verts_scalar_lib<TypeParam>(
+ this->p_edge_in.data(), this->edge_vert_idx.data(),
+ this->edge_vert_blk.data(), this->v_int.data(),
+ this->p_vert_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_e, nblks_v, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_verts("expected_verts", nproma, nlev, nblks_v);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = this->i_startidx; jv <= this->i_endidx; ++jv) {
+ expected_verts(jv, jk, jb) = 0.0;
+
+ for (int j = 0; j < 6; ++j) {
+ int edge_idx = edge_vert_idx_h[idx_at(jv, jb, j)];
+ int edge_blk = edge_vert_blk_h[blk_at(jv, jb, j)];
+ TypeParam coeff = v_int_h[coeff_at(jv, j, jb)];
+
+ expected_verts(jv, jk, jb) += coeff * p_edge_in_h[edge_at(edge_idx, jk, edge_blk)];
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = this->i_startidx; jv <= this->i_endidx; ++jv) {
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts(jv, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
}
}
}
@@ -229,55 +547,403 @@ TYPED_TEST(InterpolationScalarTypedTestFixture, Edges2Verts) {
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarTypedTestFixture, Edges2Cells) {
+TYPED_TEST(InterpolationScalarSingleParamTest, Edges2CellsSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_c = this->nblks_c;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 3, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_edge_in_h = Kokkos::create_mirror_view(this->p_edge_in);
+ auto edge_idx_h = Kokkos::create_mirror_view(this->edge_idx);
+ auto edge_blk_h = Kokkos::create_mirror_view(this->edge_blk);
+ auto coeff_int_cells_h = Kokkos::create_mirror_view(this->coeff_int_cells);
+ auto p_cell_out_h = Kokkos::create_mirror_view(this->p_cell_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_edge_in_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+ // Initialize cell connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell connects to 3 edges
+ for (int j = 0; j < 3; ++j) {
+ // Edge indices with a pattern
+ edge_idx_h[idx_at(ic, ib, j)] = (ic + j) % nproma;
+ edge_blk_h[blk_at(ic, ib, j)] = (ib + j % 2) % nblks_e;
+
+ // Interpolation coefficients that depend on indices
+ coeff_int_cells_h[coeff_at(ic, j, ib)] = static_cast<TypeParam>(1.0 / 3.0 + j * 0.01);
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_cell_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_edge_in, p_edge_in_h);
+ Kokkos::deep_copy(this->edge_idx, edge_idx_h);
+ Kokkos::deep_copy(this->edge_blk, edge_blk_h);
+ Kokkos::deep_copy(this->coeff_int_cells, coeff_int_cells_h);
+ Kokkos::deep_copy(this->p_cell_out, p_cell_out_h);
+
+ // Call the function under test
edges2cells_scalar_lib<TypeParam>(
- this->p_edge_in.data(), this->edge_idx.data(), this->edge_blk.data(),
- this->coeff_int_cells.data(), this->p_cell_out.data(), this->i_startblk,
- this->i_endblk, this->i_startidx, this->i_endidx, this->slev, this->elev,
- this->nproma, this->nlev, this->nblks_e, this->nblks_c, this->lacc);
-
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 3 stencil points,
- // expect 3.
- EXPECT_NEAR(this->p_cell_out[idx], static_cast<TypeParam>(3),
+ this->p_edge_in.data(), this->edge_idx.data(),
+ this->edge_blk.data(), this->coeff_int_cells.data(),
+ this->p_cell_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_e, nblks_c, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_cell_out_h, this->p_cell_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_cells = {
+ 1.37677, 1.7201, 1.47977, 1.8231, 1.58277, 1.9261,
+ 1.3802, 1.72353, 1.4832, 1.82653, 1.5862, 1.92953
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ EXPECT_NEAR(p_cell_out_h[cell_at(jc, jk, jb)],
+ expected_cells[cell_at(jc, jk, jb)],
static_cast<TypeParam>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
}
}
}
}
-TYPED_TEST(InterpolationScalarTypedTestFixture, Verts2Cells) {
+TYPED_TEST(InterpolationScalarSingleParamTest, Edges2CellsRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_c = this->nblks_c;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 3, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_edge_in_h = Kokkos::create_mirror_view(this->p_edge_in);
+ auto edge_idx_h = Kokkos::create_mirror_view(this->edge_idx);
+ auto edge_blk_h = Kokkos::create_mirror_view(this->edge_blk);
+ auto coeff_int_cells_h = Kokkos::create_mirror_view(this->coeff_int_cells);
+ auto p_cell_out_h = Kokkos::create_mirror_view(this->p_cell_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_e - 1);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_edge_in_h[edge_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // Initialize cell connectivity indices with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell connects to 3 edges
+ for (int j = 0; j < 3; ++j) {
+ edge_idx_h[idx_at(ic, ib, j)] = edge_distrib(gen);
+ edge_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+
+ // Random interpolation coefficients
+ coeff_int_cells_h[coeff_at(ic, j, ib)] = real_distrib(gen) / 3.0; // Scaled to ensure reasonable sums
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_cell_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_edge_in, p_edge_in_h);
+ Kokkos::deep_copy(this->edge_idx, edge_idx_h);
+ Kokkos::deep_copy(this->edge_blk, edge_blk_h);
+ Kokkos::deep_copy(this->coeff_int_cells, coeff_int_cells_h);
+ Kokkos::deep_copy(this->p_cell_out, p_cell_out_h);
+
+ // Call the function under test
+ edges2cells_scalar_lib<TypeParam>(
+ this->p_edge_in.data(), this->edge_idx.data(),
+ this->edge_blk.data(), this->coeff_int_cells.data(),
+ this->p_cell_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_e, nblks_c, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_cell_out_h, this->p_cell_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_cells("expected_cells", nproma, nlev, nblks_c);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ expected_cells(jc, jk, jb) = 0.0;
+
+ for (int j = 0; j < 3; ++j) {
+ int edge_index = edge_idx_h[idx_at(jc, jb, j)];
+ int edge_block = edge_blk_h[blk_at(jc, jb, j)];
+ TypeParam coeff = coeff_int_cells_h[coeff_at(jc, j, jb)];
+
+ expected_cells(jc, jk, jb) += coeff * p_edge_in_h[edge_at(edge_index, jk, edge_block)];
+ }
+ }
+ }
+ }
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ EXPECT_NEAR(p_cell_out_h[cell_at(jc, jk, jb)],
+ expected_cells(jc, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
+ }
+ }
+ }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// ! verts2cells
+//
+////////////////////////////////////////////////////////////////////////////////
+
+TYPED_TEST(InterpolationScalarSingleParamTest, Verts2CellsSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int npromz_c = this->npromz_c;
+
+ // Define indexing helpers
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 3, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vert_in_h = Kokkos::create_mirror_view(this->p_vert_in);
+ auto cell_index_idx_h = Kokkos::create_mirror_view(this->cell_index_idx);
+ auto cell_index_blk_h = Kokkos::create_mirror_view(this->cell_index_blk);
+ auto coeff_int_cells_h = Kokkos::create_mirror_view(this->coeff_int_cells);
+ auto p_cell_out_h = Kokkos::create_mirror_view(this->p_cell_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_vert_in_h[vert_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize cell connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell connects to 3 vertices
+ for (int j = 0; j < 3; ++j) {
+ // Vertex indices with a pattern
+ cell_index_idx_h[idx_at(ic, ib, j)] = (ic + j) % nproma;
+ cell_index_blk_h[blk_at(ic, ib, j)] = (ib + j % 2) % nblks_v;
+
+ // Interpolation coefficients that depend on indices
+ coeff_int_cells_h[coeff_at(ic, j, ib)] = static_cast<TypeParam>(1.0 / 3.0 + j * 0.01);
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_cell_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vert_in, p_vert_in_h);
+ Kokkos::deep_copy(this->cell_index_idx, cell_index_idx_h);
+ Kokkos::deep_copy(this->cell_index_blk, cell_index_blk_h);
+ Kokkos::deep_copy(this->coeff_int_cells, coeff_int_cells_h);
+ Kokkos::deep_copy(this->p_cell_out, p_cell_out_h);
+
+ // Call the function under test
verts2cells_scalar_lib<TypeParam>(
this->p_vert_in.data(), this->cell_index_idx.data(),
this->cell_index_blk.data(), this->coeff_int_cells.data(),
- this->p_cell_out.data(), this->nblks_c, this->npromz_c, this->slev,
- this->elev, this->nproma, this->nlev, this->nblks_v, this->lacc);
-
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 3 stencil points,
- // expect 3.
- EXPECT_NEAR(this->p_cell_out[idx], static_cast<TypeParam>(3),
+ this->p_cell_out.data(), nblks_c, npromz_c, this->slev, this->elev,
+ nproma, nlev, nblks_v, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_cell_out_h, this->p_cell_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_cells = {
+ 1.37677, 1.7201, 1.47977, 1.8231, 1.58277, 1.9261,
+ 1.3802, 1.72353, 1.4832, 1.82653, 1.5862, 1.92953
+ };
+
+ // Verify results - check the same ranges as in the expected calculation
+ for (int jb = 0; jb < nblks_c; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ int nlen = (jb != nblks_c - 1) ? nproma : npromz_c;
+ int start_idx = (jb >= this->i_startblk && jb <= this->i_endblk) ? this->i_startidx : 0;
+ int end_idx = (jb >= this->i_startblk && jb <= this->i_endblk) ? this->i_endidx : nlen - 1;
+
+ for (int jc = start_idx; jc <= end_idx; ++jc) {
+ EXPECT_NEAR(p_cell_out_h[cell_at(jc, jk, jb)],
+ expected_cells[cell_at(jc, jk, jb)],
static_cast<TypeParam>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
+ }
+ }
+ }
+}
+
+TYPED_TEST(InterpolationScalarSingleParamTest, Verts2CellsRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int npromz_c = this->npromz_c;
+
+ // Define indexing helpers
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 3, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vert_in_h = Kokkos::create_mirror_view(this->p_vert_in);
+ auto cell_index_idx_h = Kokkos::create_mirror_view(this->cell_index_idx);
+ auto cell_index_blk_h = Kokkos::create_mirror_view(this->cell_index_blk);
+ auto coeff_int_cells_h = Kokkos::create_mirror_view(this->coeff_int_cells);
+ auto p_cell_out_h = Kokkos::create_mirror_view(this->p_cell_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> vert_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_v - 1);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vert_in_h[vert_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // Initialize cell connectivity indices with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell connects to 3 vertices
+ for (int j = 0; j < 3; ++j) {
+ cell_index_idx_h[idx_at(ic, ib, j)] = vert_distrib(gen);
+ cell_index_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+
+ // Random interpolation coefficients
+ coeff_int_cells_h[coeff_at(ic, j, ib)] = real_distrib(gen) / 3.0; // Scaled to ensure reasonable sums
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_cell_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vert_in, p_vert_in_h);
+ Kokkos::deep_copy(this->cell_index_idx, cell_index_idx_h);
+ Kokkos::deep_copy(this->cell_index_blk, cell_index_blk_h);
+ Kokkos::deep_copy(this->coeff_int_cells, coeff_int_cells_h);
+ Kokkos::deep_copy(this->p_cell_out, p_cell_out_h);
+
+ // Call the function under test
+ verts2cells_scalar_lib<TypeParam>(
+ this->p_vert_in.data(), this->cell_index_idx.data(),
+ this->cell_index_blk.data(), this->coeff_int_cells.data(),
+ this->p_cell_out.data(), nblks_c, npromz_c, this->slev, this->elev,
+ nproma, nlev, nblks_v, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_cell_out_h, this->p_cell_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_cells("expected_cells", nproma, nlev, nblks_c);
+
+ // Compute expected values
+ for (int jb = 0; jb < nblks_c; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ int nlen = (jb != nblks_c - 1) ? nproma : npromz_c;
+ for (int jc = 0; jc < nlen; ++jc) {
+ expected_cells(jc, jk, jb) = 0.0;
+
+ for (int j = 0; j < 3; ++j) {
+ int vert_index = cell_index_idx_h[idx_at(jc, jb, j)];
+ int vert_block = cell_index_blk_h[blk_at(jc, jb, j)];
+ TypeParam coeff = coeff_int_cells_h[coeff_at(jc, j, jb)];
+
+ expected_cells(jc, jk, jb) += coeff * p_vert_in_h[vert_at(vert_index, jk, vert_block)];
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = 0; jb < nblks_c; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ int nlen = (jb != nblks_c - 1) ? nproma : npromz_c;
+ for (int jc = 0; jc < nlen; ++jc) {
+ EXPECT_NEAR(p_cell_out_h[cell_at(jc, jk, jb)],
+ expected_cells(jc, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
}
}
}
@@ -289,48 +955,226 @@ TYPED_TEST(InterpolationScalarTypedTestFixture, Verts2Cells) {
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarTypedTestFixture, AvgLib) {
+TYPED_TEST(InterpolationScalarSingleParamTest, CellAvgLibSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+
+ // Define indexing helpers
+ const auto &psi_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 4, nblks_c>; // 4 coefficients (self + 3 neighbors)
+ const auto &avg_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto psi_c_h = Kokkos::create_mirror_view(this->psi_c);
+ auto cell_neighbor_idx_h = Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h = Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto avg_psi_c_h = Kokkos::create_mirror_view(this->avg_psi_c);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ psi_c_h[psi_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
- // Call the function
- cell_avg_lib<TypeParam>(this->psi_c.data(), this->cell_neighbor_idx.data(),
- this->cell_neighbor_blk.data(),
- this->avg_coeff.data(), this->avg_psi_c.data(),
- this->i_startblk, this->i_endblk, this->i_startidx,
- this->i_endidx, this->slev, this->elev, this->nproma,
- this->nlev, this->nblks_c, this->lacc);
-
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 4 stencil points,
- // expect 4.
- EXPECT_NEAR(this->avg_psi_c[idx], static_cast<TypeParam>(4),
+ // Initialize cell neighbor indices with specific pattern
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell has 3 neighboring cells
+ for (int j = 0; j < 3; ++j) {
+ // Neighbor indices with a pattern
+ cell_neighbor_idx_h[idx_at(ic, ib, j)] = (ic + j + 1) % nproma;
+ cell_neighbor_blk_h[blk_at(ic, ib, j)] = (ib + j % 2) % nblks_c;
+ }
+
+ // Averaging coefficients - one for the cell itself and one for each neighbor
+ avg_coeff_h[coeff_at(ic, 0, ib)] = static_cast<TypeParam>(0.4); // Self weight
+ avg_coeff_h[coeff_at(ic, 1, ib)] = static_cast<TypeParam>(0.2); // First neighbor
+ avg_coeff_h[coeff_at(ic, 2, ib)] = static_cast<TypeParam>(0.2); // Second neighbor
+ avg_coeff_h[coeff_at(ic, 3, ib)] = static_cast<TypeParam>(0.2); // Third neighbor
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ avg_psi_c_h[avg_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->psi_c, psi_c_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->avg_psi_c, avg_psi_c_h);
+
+ // Call the function under test
+ cell_avg_lib<TypeParam>(
+ this->psi_c.data(), this->cell_neighbor_idx.data(),
+ this->cell_neighbor_blk.data(), this->avg_coeff.data(),
+ this->avg_psi_c.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_c, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(avg_psi_c_h, this->avg_psi_c);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_avg = {
+ 1.402, 1.602, 1.502, 1.702, 1.602, 1.802,
+ 1.408, 1.608, 1.508, 1.708, 1.608, 1.808
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ EXPECT_NEAR(avg_psi_c_h[avg_at(jc, jk, jb)],
+ expected_avg[avg_at(jc, jk, jb)],
static_cast<TypeParam>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
+ }
+ }
+ }
+}
+
+TYPED_TEST(InterpolationScalarSingleParamTest, CellAvgLibRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+
+ // Define indexing helpers
+ const auto &psi_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_c, 3>;
+ const auto &blk_at = at<nproma, nblks_c, 3>;
+ const auto &coeff_at = at<nproma, 4, nblks_c>; // 4 coefficients (self + 3 neighbors)
+ const auto &avg_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto psi_c_h = Kokkos::create_mirror_view(this->psi_c);
+ auto cell_neighbor_idx_h = Kokkos::create_mirror_view(this->cell_neighbor_idx);
+ auto cell_neighbor_blk_h = Kokkos::create_mirror_view(this->cell_neighbor_blk);
+ auto avg_coeff_h = Kokkos::create_mirror_view(this->avg_coeff);
+ auto avg_psi_c_h = Kokkos::create_mirror_view(this->avg_psi_c);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> cell_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_c - 1);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<TypeParam> coeff_distrib(0.01, 0.5); // Keep coefficients reasonable
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ psi_c_h[psi_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // Initialize cell neighbor indices with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell has 3 neighboring cells
+ for (int j = 0; j < 3; ++j) {
+ cell_neighbor_idx_h[idx_at(ic, ib, j)] = cell_distrib(gen);
+ cell_neighbor_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+ }
+
+ avg_coeff_h[coeff_at(ic, 0, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ avg_coeff_h[coeff_at(ic, 1, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ avg_coeff_h[coeff_at(ic, 2, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ avg_coeff_h[coeff_at(ic, 3, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ avg_psi_c_h[avg_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->psi_c, psi_c_h);
+ Kokkos::deep_copy(this->cell_neighbor_idx, cell_neighbor_idx_h);
+ Kokkos::deep_copy(this->cell_neighbor_blk, cell_neighbor_blk_h);
+ Kokkos::deep_copy(this->avg_coeff, avg_coeff_h);
+ Kokkos::deep_copy(this->avg_psi_c, avg_psi_c_h);
+
+ // Call the function under test
+ cell_avg_lib<TypeParam>(
+ this->psi_c.data(), this->cell_neighbor_idx.data(),
+ this->cell_neighbor_blk.data(), this->avg_coeff.data(),
+ this->avg_psi_c.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_c, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(avg_psi_c_h, this->avg_psi_c);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_avg("expected_avg", nproma, nlev, nblks_c);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ // Self contribution
+ expected_avg(jc, jk, jb) =
+ psi_c_h[psi_at(jc, jk, jb)] * avg_coeff_h[coeff_at(jc, 0, jb)];
+
+ // Neighbor contributions
+ for (int j = 0; j < 3; ++j) {
+ int neighbor_idx = cell_neighbor_idx_h[idx_at(jc, jb, j)];
+ int neighbor_blk = cell_neighbor_blk_h[blk_at(jc, jb, j)];
+ TypeParam coeff = avg_coeff_h[coeff_at(jc, j+1, jb)];
+
+ expected_avg(jc, jk, jb) +=
+ psi_c_h[psi_at(neighbor_idx, jk, neighbor_blk)] * coeff;
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx; jc <= this->i_endidx; ++jc) {
+ EXPECT_NEAR(avg_psi_c_h[avg_at(jc, jk, jb)],
+ expected_avg(jc, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jc;
}
}
}
}
template <typename TypePair>
-class InterpolationScalarMixedTestFixture : public ::testing::Test,
+class InterpolationScalarDoubleParamTest : public ::testing::Test,
public interp_dimensions {
-public:
+ protected:
using InType = typename TypePair::in_type;
using OutType = typename TypePair::out_type;
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
// Arrays used for cells2edges
- std::vector<InType> p_cell_in; // Dimensions: (nproma, nlev, nblks_c)
- std::vector<int> edge_cell_idx; // Dimensions: (nproma, nblks_e, 2)
- std::vector<int> edge_cell_blk; // Dimensions: (nproma, nblks_e, 2)
- std::vector<OutType> coeff_int_edges; // Dimensions: (nproma, 2, nblks_e)
- std::vector<OutType> p_edge_out; // Dimensions: (nproma, nlev, nblks_e)
+ Kokkos::View<InType*, memory_space> p_cell_in;
+ Kokkos::View<int*, memory_space> edge_cell_idx;
+ Kokkos::View<int*, memory_space> edge_cell_blk;
+ Kokkos::View<OutType*, memory_space> coeff_int_edges;
+ Kokkos::View<OutType*, memory_space> p_edge_out;
// Further parameters for cells2edges
const int patch_id = 0;
@@ -342,41 +1186,31 @@ public:
std::vector<int> i_endidx_in; // Dimensions: (2)
// Arrays used for cells2verts
- std::vector<int> vert_cell_idx; // Dimensions: (nproma, nblks_v, 6)
- std::vector<int> vert_cell_blk; // Dimensions: (nproma, nblks_v, 6)
- std::vector<OutType> coeff_int_verts; // Dimensions: (nproma, 6, nblks_v)
- std::vector<OutType> p_vert_out; // Dimensions: (nproma, nlev, nblks_v)
-
- InterpolationScalarMixedTestFixture() {
- // Allocate and initialize arrays needed for cells2edges
- p_cell_in.resize(num_elements_3d<InType>(nproma, nlev, nblks_c),
- static_cast<InType>(1));
- edge_cell_idx.resize(num_elements_3d<int>(nproma, nblks_e, 2), 1);
- edge_cell_blk.resize(num_elements_3d<int>(nproma, nblks_e, 2), 0);
- coeff_int_edges.resize(num_elements_3d<InType>(nproma, 2, nblks_e),
- static_cast<OutType>(1));
-
- p_edge_out.resize(num_elements_3d<OutType>(nproma, nlev, nblks_e),
- static_cast<OutType>(0));
-
+ Kokkos::View<int*, memory_space> vert_cell_idx;
+ Kokkos::View<int*, memory_space> vert_cell_blk;
+ Kokkos::View<OutType*, memory_space> coeff_int_verts;
+ Kokkos::View<OutType*, memory_space> p_vert_out;
+
+ InterpolationScalarDoubleParamTest()
+ : p_cell_in("p_cell_in", nproma * nlev * nblks_c),
+ edge_cell_idx("edge_cell_idx", nproma * nblks_e * 2),
+ edge_cell_blk("edge_cell_blk", nproma * nblks_e * 2),
+ coeff_int_edges("coeff_int_edges", nproma * 2 * nblks_e),
+ p_edge_out("p_edge_out", nproma * nlev * nblks_e),
+ vert_cell_idx("vert_cell_idx", nproma * nblks_v * 6),
+ vert_cell_blk("vert_cell_blk", nproma * nblks_v * 6),
+ coeff_int_verts("coeff_int_verts", nproma * 6 * nblks_v),
+ p_vert_out("p_vert_out", nproma * nlev * nblks_v)
+ {
// Allocate neighbour indexes for cells2edges
i_startblk_in.resize(2, i_startblk);
i_endblk_in.resize(2, i_endblk);
i_startidx_in.resize(2, i_startidx);
i_endidx_in.resize(2, i_endidx);
-
- // Allocate & Initialize arrays needed for cells2verts
- vert_cell_idx.resize(num_elements_3d<int>(nproma, nblks_v, 6), 1);
- vert_cell_blk.resize(num_elements_3d<int>(nproma, nblks_v, 6), 0);
- coeff_int_verts.resize(num_elements_3d<InType>(nproma, 6, nblks_v),
- static_cast<OutType>(1));
-
- p_vert_out.resize(num_elements_3d<OutType>(nproma, nlev, nblks_v),
- static_cast<OutType>(0));
}
};
-TYPED_TEST_SUITE(InterpolationScalarMixedTestFixture, MixedTypesSP2DP);
+TYPED_TEST_SUITE(InterpolationScalarDoubleParamTest, MixedTypesSP2DP);
////////////////////////////////////////////////////////////////////////////////
//
@@ -384,34 +1218,234 @@ TYPED_TEST_SUITE(InterpolationScalarMixedTestFixture, MixedTypesSP2DP);
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarMixedTestFixture, Cells2Edges) {
+TYPED_TEST(InterpolationScalarDoubleParamTest, Cells2EdgesSpecific) {
using InType = typename TestFixture::InType;
using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_e = this->nblks_e;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &edge_idx_at = at<nproma, nblks_e, 2>;
+ const auto &edge_blk_at = at<nproma, nblks_e, 2>;
+ const auto &coeff_at = at<nproma, 2, nblks_e>;
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto edge_cell_idx_h = Kokkos::create_mirror_view(this->edge_cell_idx);
+ auto edge_cell_blk_h = Kokkos::create_mirror_view(this->edge_cell_blk);
+ auto coeff_int_edges_h = Kokkos::create_mirror_view(this->coeff_int_edges);
+ auto p_edge_out_h = Kokkos::create_mirror_view(this->p_edge_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
- // Call the function
+ // Initialize edge connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each edge connects to 2 cells
+ edge_cell_idx_h[edge_idx_at(ic, ib, 0)] = ic % nproma; // First cell index
+ edge_cell_idx_h[edge_idx_at(ic, ib, 1)] = (ic + 1) % nproma; // Second cell index
+
+ edge_cell_blk_h[edge_blk_at(ic, ib, 0)] = ib % nblks_c; // First cell block
+ edge_cell_blk_h[edge_blk_at(ic, ib, 1)] = (ib + 1) % nblks_c; // Second cell block
+
+ // Interpolation coefficients that depend on indices
+ coeff_int_edges_h[coeff_at(ic, 0, ib)] = static_cast<OutType>(0.5 + ic * 0.01);
+ coeff_int_edges_h[coeff_at(ic, 1, ib)] = static_cast<OutType>(0.5 - ic * 0.01);
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_edge_out_h[edge_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->edge_cell_idx, edge_cell_idx_h);
+ Kokkos::deep_copy(this->edge_cell_blk, edge_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int_edges, coeff_int_edges_h);
+ Kokkos::deep_copy(this->p_edge_out, p_edge_out_h);
+
+ // Call the function under test
cells2edges_scalar_lib<InType, OutType>(
this->p_cell_in.data(), this->edge_cell_idx.data(),
this->edge_cell_blk.data(), this->coeff_int_edges.data(),
this->p_edge_out.data(), this->i_startblk_in.data(),
this->i_endblk_in.data(), this->i_startidx_in.data(),
- this->i_endidx_in.data(), this->slev, this->elev, this->nproma,
- this->nlev, this->nblks_c, this->nblks_e, this->patch_id,
+ this->i_endidx_in.data(), this->slev, this->elev, nproma,
+ nlev, nblks_c, nblks_e, this->patch_id,
this->l_limited_area, this->lfill_latbc, this->lacc);
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 2 stencil points,
- // expect 2.
- EXPECT_NEAR(this->p_edge_out[idx], static_cast<OutType>(2),
+ // Copy results back to host
+ Kokkos::deep_copy(p_edge_out_h, this->p_edge_out);
+
+ int i_startblk = this->i_startblk_in[1];
+ int i_endblk = this->i_endblk_in[1];
+ int i_startidx_range = this->i_startidx_in[1];
+ int i_endidx_range = this->i_endidx_in[1];
+
+ // Expected results based on the specific test values
+ std::vector<OutType> expected_edges = {
+ 1.505, 1.5149, 1.605, 1.6149, 1.705, 1.7149,
+ 1.505, 1.5151, 1.605, 1.6151, 1.705, 1.7151
+ };
+
+ // Verify results
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_e_lib(i_startidx_range, i_endidx_range, nproma, jb,
+ i_startblk, i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ EXPECT_NEAR(p_edge_out_h[edge_at(je, jk, jb)],
+ expected_edges[edge_at(je, jk, jb)],
static_cast<OutType>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ << "Failure at block " << jb << ", level " << jk << ", index " << je;
+ }
+ }
+ }
+}
+
+TYPED_TEST(InterpolationScalarDoubleParamTest, Cells2EdgesRandom) {
+ using InType = typename TestFixture::InType;
+ using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_e = this->nblks_e;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &edge_idx_at = at<nproma, nblks_e, 2>;
+ const auto &edge_blk_at = at<nproma, nblks_e, 2>;
+ const auto &coeff_at = at<nproma, 2, nblks_e>;
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto edge_cell_idx_h = Kokkos::create_mirror_view(this->edge_cell_idx);
+ auto edge_cell_blk_h = Kokkos::create_mirror_view(this->edge_cell_blk);
+ auto coeff_int_edges_h = Kokkos::create_mirror_view(this->coeff_int_edges);
+ auto p_edge_out_h = Kokkos::create_mirror_view(this->p_edge_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> cell_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_c - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize edge connectivity indices with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each edge connects to 2 cells
+ edge_cell_idx_h[edge_idx_at(ic, ib, 0)] = cell_distrib(gen);
+ edge_cell_idx_h[edge_idx_at(ic, ib, 1)] = cell_distrib(gen);
+
+ edge_cell_blk_h[edge_blk_at(ic, ib, 0)] = block_distrib(gen);
+ edge_cell_blk_h[edge_blk_at(ic, ib, 1)] = block_distrib(gen);
+
+ coeff_int_edges_h[coeff_at(ic, 0, ib)] = static_cast<OutType>(real_distrib(gen));
+ coeff_int_edges_h[coeff_at(ic, 1, ib)] = static_cast<OutType>(real_distrib(gen));
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_edge_out_h[edge_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->edge_cell_idx, edge_cell_idx_h);
+ Kokkos::deep_copy(this->edge_cell_blk, edge_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int_edges, coeff_int_edges_h);
+ Kokkos::deep_copy(this->p_edge_out, p_edge_out_h);
+
+ // Call the function under test
+ cells2edges_scalar_lib<InType, OutType>(
+ this->p_cell_in.data(), this->edge_cell_idx.data(),
+ this->edge_cell_blk.data(), this->coeff_int_edges.data(),
+ this->p_edge_out.data(), this->i_startblk_in.data(),
+ this->i_endblk_in.data(), this->i_startidx_in.data(),
+ this->i_endidx_in.data(), this->slev, this->elev, nproma,
+ nlev, nblks_c, nblks_e, this->patch_id,
+ this->l_limited_area, this->lfill_latbc, this->lacc);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_edge_out_h, this->p_edge_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<OutType***, host_space> expected_edges("expected_edges", nproma, nlev, nblks_e);
+
+ // Since we're not testing the lateral boundary condition filling
+ // (this->l_limited_area == false && this->lfill_latbc == false),
+ // we only need to check the blocks in i_startblk_in[1] to i_endblk_in[1]
+ int i_startblk = this->i_startblk_in[1];
+ int i_endblk = this->i_endblk_in[1];
+ int i_startidx_range = this->i_startidx_in[1];
+ int i_endidx_range = this->i_endidx_in[1];
+
+ // Compute expected values
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
+ // Get the actual indices to process for this block
+ int i_startidx, i_endidx;
+ get_indices_e_lib(i_startidx_range, i_endidx_range, nproma, jb,
+ i_startblk, i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ expected_edges(je, jk, jb) =
+ static_cast<OutType>(coeff_int_edges_h[coeff_at(je, 0, jb)] *
+ p_cell_in_h[cell_at(edge_cell_idx_h[edge_idx_at(je, jb, 0)],
+ jk,
+ edge_cell_blk_h[edge_blk_at(je, jb, 0)])]) +
+ static_cast<OutType>(coeff_int_edges_h[coeff_at(je, 1, jb)] *
+ p_cell_in_h[cell_at(edge_cell_idx_h[edge_idx_at(je, jb, 1)],
+ jk,
+ edge_cell_blk_h[edge_blk_at(je, jb, 1)])]);
+ }
+ }
+ }
+
+ OutType tol = std::is_same<OutType, float>::value ?
+ static_cast<OutType>(1e-5) : static_cast<OutType>(1e-13);
+
+ // Verify results
+ for (int jb = i_startblk; jb <= i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_e_lib(i_startidx_range, i_endidx_range, nproma, jb,
+ i_startblk, i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ EXPECT_NEAR(p_edge_out_h[edge_at(je, jk, jb)],
+ expected_edges(je, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << je;
}
}
}
@@ -423,31 +1457,217 @@ TYPED_TEST(InterpolationScalarMixedTestFixture, Cells2Edges) {
//
////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(InterpolationScalarMixedTestFixture, Cells2Verts) {
+TYPED_TEST(InterpolationScalarDoubleParamTest, Cells2VertsSpecific) {
using InType = typename TestFixture::InType;
using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto vert_cell_idx_h = Kokkos::create_mirror_view(this->vert_cell_idx);
+ auto vert_cell_blk_h = Kokkos::create_mirror_view(this->vert_cell_blk);
+ auto coeff_int_verts_h = Kokkos::create_mirror_view(this->coeff_int_verts);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+ // Initialize vertex connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex is connected to 6 cells
+ for (int j = 0; j < 6; ++j) {
+ // Cell indices with a pattern
+ vert_cell_idx_h[idx_at(ic, ib, j)] = (ic + j) % nproma;
+ vert_cell_blk_h[blk_at(ic, ib, j)] = (ib + j % nblks_c) % nblks_c;
+
+ // Interpolation coefficients that depend on indices
+ coeff_int_verts_h[coeff_at(ic, j, ib)] = static_cast<OutType>(1.0 / 6.0 + j * 0.01);
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->vert_cell_idx, vert_cell_idx_h);
+ Kokkos::deep_copy(this->vert_cell_blk, vert_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int_verts, coeff_int_verts_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+
+ // Call the function under test
cells2verts_scalar_lib<InType, OutType>(
this->p_cell_in.data(), this->vert_cell_idx.data(),
this->vert_cell_blk.data(), this->coeff_int_verts.data(),
this->p_vert_out.data(), this->i_startblk, this->i_endblk,
- this->i_startidx, this->i_endidx, this->slev, this->elev, this->nproma,
- this->nlev, this->nblks_c, this->nblks_v, this->lacc, this->acc_async);
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_c, nblks_v, this->lacc, this->acc_async);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Expected results based on the specific test values
+ std::vector<OutType> expected_verts = {
+ 1.7459, 1.7159, 1.8609, 1.8309, 1.9759, 1.9459,
+ 1.7456, 1.7156, 1.8606, 1.8306, 1.9756, 1.9456
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts[vert_at(jv, jk, jb)],
+ static_cast<OutType>(1e-5))
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+ }
+ }
+ }
+}
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 6 stencil points,
- // expect 6.
- EXPECT_NEAR(this->p_vert_out[idx], static_cast<OutType>(6),
- static_cast<OutType>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+TYPED_TEST(InterpolationScalarDoubleParamTest, Cells2VertsRandom) {
+ using InType = typename TestFixture::InType;
+ using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto vert_cell_idx_h = Kokkos::create_mirror_view(this->vert_cell_idx);
+ auto vert_cell_blk_h = Kokkos::create_mirror_view(this->vert_cell_blk);
+ auto coeff_int_verts_h = Kokkos::create_mirror_view(this->coeff_int_verts);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> cell_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_c - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<double> coeff_distrib(0.01, 0.3); // Keep coefficients reasonable
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex is connected to 6 cells
+ for (int j = 0; j < 6; ++j) {
+ vert_cell_idx_h[idx_at(ic, ib, j)] = cell_distrib(gen);
+ vert_cell_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+
+ // Normalized coefficients
+ coeff_int_verts_h[coeff_at(ic, j, ib)] = static_cast<OutType>(coeff_distrib(gen));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->vert_cell_idx, vert_cell_idx_h);
+ Kokkos::deep_copy(this->vert_cell_blk, vert_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int_verts, coeff_int_verts_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+
+ // Call the function under test
+ cells2verts_scalar_lib<InType, OutType>(
+ this->p_cell_in.data(), this->vert_cell_idx.data(),
+ this->vert_cell_blk.data(), this->coeff_int_verts.data(),
+ this->p_vert_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_c, nblks_v, this->lacc, this->acc_async);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<OutType***, host_space> expected_verts("expected_verts", nproma, nlev, nblks_v);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ // Get the actual indices to process for this block
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ expected_verts(jv, jk, jb) = static_cast<OutType>(0.0);
+
+ for (int j = 0; j < 6; ++j) {
+ int cell_idx = vert_cell_idx_h[idx_at(jv, jb, j)];
+ int cell_blk = vert_cell_blk_h[blk_at(jv, jb, j)];
+ OutType coeff = coeff_int_verts_h[coeff_at(jv, j, jb)];
+
+ expected_verts(jv, jk, jb) +=
+ static_cast<OutType>(coeff * p_cell_in_h[cell_at(cell_idx, jk, cell_blk)]);
+ }
+ }
+ }
+ }
+
+ OutType tol = std::is_same<OutType, float>::value ?
+ static_cast<OutType>(1e-5) : static_cast<OutType>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts(jv, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
}
}
}
@@ -460,7 +1680,7 @@ TYPED_TEST(InterpolationScalarMixedTestFixture, Cells2Verts) {
////////////////////////////////////////////////////////////////////////////////
// The test for cells2verts_ri is similar to cells2verts, but is done here
-// separtely to avoid as a differebt template instantiation is needed for the
+// separtely to avoid as a different template instantiation is needed for the
// function call
template <typename Types>
class Cells2vertsriScalarLibTestFixture : public testing::Test,
@@ -469,36 +1689,102 @@ public:
using InType = typename Types::in_type;
using OutType = typename Types::out_type;
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
// Arrays stored in std::vector.
- std::vector<InType> p_cell_in; // Dimensions: (nproma, nlev, nblks_c)
- std::vector<int> vert_cell_idx; // Dimensions: (nproma, nblks_v, 6)
- std::vector<int> vert_cell_blk; // Dimensions: (nproma, nblks_v, 6)
- std::vector<InType> coeff_int; // Dimensions: (nproma, 6, nblks_v)
- std::vector<OutType> p_vert_out; // Dimensions: (nproma, nlev, nblks_v)
-
- Cells2vertsriScalarLibTestFixture() {
- // Allocate and initialize inputs.
- p_cell_in.resize(num_elements_3d<InType>(nproma, nlev, nblks_c),
- static_cast<InType>(1));
- vert_cell_idx.resize(num_elements_3d<int>(nproma, nblks_v, 6), 1);
- vert_cell_blk.resize(num_elements_3d<int>(nproma, nblks_v, 6), 0);
- coeff_int.resize(num_elements_3d<InType>(nproma, 6, nblks_v),
- static_cast<InType>(1));
-
- // Allocate output arrays and initialize to zero.
- p_vert_out.resize(num_elements_3d<OutType>(nproma, nlev, nblks_v),
- static_cast<OutType>(0));
- }
+ Kokkos::View<InType*, memory_space> p_cell_in;
+ Kokkos::View<int*, memory_space> vert_cell_idx;
+ Kokkos::View<int*, memory_space> vert_cell_blk;
+ Kokkos::View<InType*, memory_space> coeff_int;
+ Kokkos::View<OutType*, memory_space> p_vert_out;
+
+ Cells2vertsriScalarLibTestFixture()
+ : p_cell_in("p_cell_in", nproma * nlev * nblks_c),
+ vert_cell_idx("vert_cell_idx", nproma * nblks_v * 6),
+ vert_cell_blk("vert_cell_blk", nproma * nblks_v * 6),
+ coeff_int("coeff_int", nproma * 6 * nblks_v),
+ p_vert_out("p_vert_out", nproma * nlev * nblks_v)
+ {}
};
// Add test suite
TYPED_TEST_SUITE(Cells2vertsriScalarLibTestFixture, MixedTypes);
// Add test
-TYPED_TEST(Cells2vertsriScalarLibTestFixture, Cells2VertsRI) {
+TYPED_TEST(Cells2vertsriScalarLibTestFixture, Cells2VertsRISpecific) {
using InType = typename TestFixture::InType;
using OutType = typename TestFixture::OutType;
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+
+ // For output, we need to handle different layouts depending on __LOOP_EXCHANGE
+ // This is a special case for this function
+#ifdef __LOOP_EXCHANGE
+ const auto &vert_at = at<nproma, nlev, nblks_c>; // jv, jk, jb order
+#else
+ const auto &vert_at = at<nlev, nproma, nblks_c>; // jk, jv, jb order
+#endif
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto vert_cell_idx_h = Kokkos::create_mirror_view(this->vert_cell_idx);
+ auto vert_cell_blk_h = Kokkos::create_mirror_view(this->vert_cell_blk);
+ auto coeff_int_h = Kokkos::create_mirror_view(this->coeff_int);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Initialize with index-based test values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex is connected to 6 cells
+ for (int j = 0; j < 6; ++j) {
+ // Cell indices with a pattern
+ vert_cell_idx_h[idx_at(ic, ib, j)] = (ic + j) % nproma;
+ vert_cell_blk_h[blk_at(ic, ib, j)] = (ib + j % nblks_c) % nblks_c;
+
+ // Interpolation coefficients that depend on indices
+ coeff_int_h[coeff_at(ic, j, ib)] = static_cast<OutType>(1.0 / 6.0 + j * 0.01);
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ // Handle different indexing depending on __LOOP_EXCHANGE
+#ifdef __LOOP_EXCHANGE
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+#else
+ p_vert_out_h[vert_at(ik, ic, ib)] = static_cast<OutType>(0.0);
+#endif
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->vert_cell_idx, vert_cell_idx_h);
+ Kokkos::deep_copy(this->vert_cell_blk, vert_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int, coeff_int_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+
// Call the function
cells2verts_scalar_ri_lib<InType, OutType>(
this->p_cell_in.data(), this->vert_cell_idx.data(),
@@ -507,25 +1793,180 @@ TYPED_TEST(Cells2vertsriScalarLibTestFixture, Cells2VertsRI) {
this->i_startidx, this->i_endidx, this->slev, this->elev, this->nproma,
this->nlev, this->nblks_c, this->nblks_v, this->lacc, this->acc_async);
- // Check the outputs only for blocks in the range
- // { [i_startblk, i_endblk], [slev,elev], [i_startidx_in, i_endidx_in] }
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = this->slev; level < this->elev; ++level) {
- for (int i = this->i_startidx; i < this->i_endidx; ++i) {
- // Compute the linear index for a 3D array in column-major order:
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Expected results based on the specific test values
+ std::vector<OutType> expected_verts = {
#ifdef __LOOP_EXCHANGE
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
+ 1.7459, 1.7159, 1.8609, 1.8309, 1.9759, 1.9459,
+ 1.7456, 1.7156, 1.8606, 1.8306, 1.9756, 1.9456
#else
- size_t idx = level + i * this->nlev + block * this->nproma * this->nlev;
+ 1.7459, 1.8609, 1.9759, 1.7159, 1.8309, 1.9459,
+ 1.7456, 1.8606, 1.9756, 1.7156, 1.8306, 1.9456
+#endif
+ };
+
+ // Verify results - using the appropriate indexing depending on __LOOP_EXCHANGE
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+#ifdef __LOOP_EXCHANGE
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts[vert_at(jv, jk, jb)],
+ static_cast<OutType>(1e-5))
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+#else
+ EXPECT_NEAR(p_vert_out_h[vert_at(jk, jv, jb)],
+ expected_verts[vert_at(jk, jv, jb)],
+ static_cast<OutType>(1e-5))
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+#endif
+ }
+ }
+ }
+}
+
+TYPED_TEST(Cells2vertsriScalarLibTestFixture, Cells2VertsRIRandom) {
+ using InType = typename TestFixture::InType;
+ using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_v = this->nblks_v;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<nproma, nblks_v, 6>;
+ const auto &blk_at = at<nproma, nblks_v, 6>;
+ const auto &coeff_at = at<nproma, 6, nblks_v>;
+
+ // For output, we need to handle different layouts depending on __LOOP_EXCHANGE
+#ifdef __LOOP_EXCHANGE
+ const auto &vert_at = at<nproma, nlev, nblks_v>; // jv, jk, jb order
+#else
+ const auto &vert_at = at<nlev, nproma, nblks_v>; // jk, jv, jb order
+#endif
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto vert_cell_idx_h = Kokkos::create_mirror_view(this->vert_cell_idx);
+ auto vert_cell_blk_h = Kokkos::create_mirror_view(this->vert_cell_blk);
+ auto coeff_int_h = Kokkos::create_mirror_view(this->coeff_int);
+ auto p_vert_out_h = Kokkos::create_mirror_view(this->p_vert_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> cell_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_c - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<double> coeff_distrib(0.01, 0.3); // Keep coefficients reasonable
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<InType>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex is connected to 6 cells
+ for (int j = 0; j < 6; ++j) {
+ vert_cell_idx_h[idx_at(ic, ib, j)] = cell_distrib(gen);
+ vert_cell_blk_h[blk_at(ic, ib, j)] = block_distrib(gen);
+
+ // Normalized coefficients
+ coeff_int_h[coeff_at(ic, j, ib)] = static_cast<InType>(coeff_distrib(gen));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ // Handle different indexing depending on __LOOP_EXCHANGE
+#ifdef __LOOP_EXCHANGE
+ p_vert_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+#else
+ p_vert_out_h[vert_at(ik, ic, ib)] = static_cast<OutType>(0.0);
+#endif
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->vert_cell_idx, vert_cell_idx_h);
+ Kokkos::deep_copy(this->vert_cell_blk, vert_cell_blk_h);
+ Kokkos::deep_copy(this->coeff_int, coeff_int_h);
+ Kokkos::deep_copy(this->p_vert_out, p_vert_out_h);
+
+ // Call the function
+ cells2verts_scalar_ri_lib<InType, OutType>(
+ this->p_cell_in.data(), this->vert_cell_idx.data(),
+ this->vert_cell_blk.data(), this->coeff_int.data(),
+ this->p_vert_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx, this->i_endidx, this->slev, this->elev,
+ nproma, nlev, nblks_c, nblks_v, this->lacc, this->acc_async);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vert_out_h, this->p_vert_out);
+
+ // Prepare expected results storage
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<OutType***, host_space> expected_verts("expected_verts", nproma, nlev, nblks_v);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ // Get the actual indices to process for this block
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ expected_verts(jv, jk, jb) = static_cast<OutType>(0.0);
+
+ for (int j = 0; j < 6; ++j) {
+ int cell_idx = vert_cell_idx_h[idx_at(jv, jb, j)];
+ int cell_blk = vert_cell_blk_h[blk_at(jv, jb, j)];
+ InType coeff = coeff_int_h[coeff_at(jv, j, jb)];
+
+ expected_verts(jv, jk, jb) +=
+ static_cast<OutType>(coeff * p_cell_in_h[cell_at(cell_idx, jk, cell_blk)]);
+ }
+ }
+ }
+ }
+
+ OutType tol = std::is_same<OutType, float>::value ?
+ static_cast<OutType>(1e-5) : static_cast<OutType>(1e-13);
+
+ // Verify results - using the appropriate indexing depending on __LOOP_EXCHANGE
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx, this->i_endidx, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+#ifdef __LOOP_EXCHANGE
+ EXPECT_NEAR(p_vert_out_h[vert_at(jv, jk, jb)],
+ expected_verts(jv, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
+#else
+ EXPECT_NEAR(p_vert_out_h[vert_at(jk, jv, jb)],
+ expected_verts(jv, jk, jb), tol)
+ << "Failure at block " << jb << ", level " << jk << ", index " << jv;
#endif
- // Since every contribution is 1 and there are 6 stencil points,
- // expect 6.
- EXPECT_NEAR(this->p_vert_out[idx], static_cast<OutType>(6),
- static_cast<OutType>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
}
}
}
}
+
diff --git a/test/c/test_interpolation_vector.cpp b/test/c/test_interpolation_vector.cpp
index 680fb6e5ac669549b7e96f3fd5c94ba7a69edd3e..0806e3575f1d4eb12c4b0f9cc10a4b51497aa30e 100644
--- a/test/c/test_interpolation_vector.cpp
+++ b/test/c/test_interpolation_vector.cpp
@@ -12,104 +12,375 @@
#include <Kokkos_Core.hpp>
#include <gtest/gtest.h>
#include <vector>
+#include <random>
#include "mo_lib_interpolation_vector.hpp"
+#include "dim_helper.hpp"
-// Dimensions for the test (small, trivial test).
-// We assume Fortran ordering: column-major, but our C wrappers will wrap raw
-// pointers into Kokkos::Views with LayoutLeft.
-constexpr int nproma = 2;
-constexpr int nlev = 3;
-constexpr int nblks_e = 2; // For the edge arrays (p_vn_in, p_vt_in)
-constexpr int nblks_c = 2; // For the cell arrays and interpolation coefficients
-
-// For the get_indices_c_lib inputs.
-constexpr int i_startblk = 0;
-constexpr int i_endblk = 1; // two blocks: indices 0 and 1
-constexpr int i_startidx_in = 0;
-constexpr int i_endidx_in = nproma - 1; // 0 and 1
-constexpr int slev = 0;
-constexpr int elev = nlev - 1; // 0 .. 2
-
-// Helper to compute total number of elements for a 3D array stored in
-// column-major order.
-template <typename T> size_t num_elements(int dim1, int dim2, int dim3) {
- return static_cast<size_t>(dim1) * dim2 * dim3;
-}
+/// Base test class for the edges2cells tests. Templated for the ValueType.
+template <typename ValueType>
+class InterpolationVectorTest : public ::testing::Test {
+protected:
+ // Constant dimensions
+ static constexpr int nproma = 2; // inner loop length
+ static constexpr int nlev = 3; // number of vertical levels
+ static constexpr int nblks_e = 2; // number of edge blocks
+ static constexpr int nblks_c = 2; // number of cell blocks
+ static constexpr int num_edges = 3; // number of edges per cell
+
+ // Parameter values
+ int i_startblk = 0;
+ int i_endblk = nblks_c - 1; // Test blocks [0 .. nblks_c-1]
+ int i_startidx_in = 0;
+ int i_endidx_in = nproma - 1; // Full range: 0 .. nproma-1
+ int slev = 0;
+ int elev = nlev - 1; // Full vertical range (0 .. nlev-1)
+
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Kokkos Views for test data
+ Kokkos::View<ValueType*, memory_space> p_vn_in;
+ Kokkos::View<ValueType*, memory_space> p_vt_in;
+ Kokkos::View<int*, memory_space> cell_edge_idx;
+ Kokkos::View<int*, memory_space> cell_edge_blk;
+ Kokkos::View<ValueType*, memory_space> e_bln_c_u;
+ Kokkos::View<ValueType*, memory_space> e_bln_c_v;
+ Kokkos::View<ValueType*, memory_space> p_u_out;
+ Kokkos::View<ValueType*, memory_space> p_v_out;
+
+ InterpolationVectorTest()
+ : p_vn_in("p_vn_in", dim_combine(nproma, nlev, nblks_e)),
+ p_vt_in("p_vt_in", dim_combine(nproma, nlev, nblks_e)),
+ cell_edge_idx("cell_edge_idx", dim_combine(nproma, nblks_c, num_edges)),
+ cell_edge_blk("cell_edge_blk", dim_combine(nproma, nblks_c, num_edges)),
+ e_bln_c_u("e_bln_c_u", dim_combine(nproma, 6, nblks_c)),
+ e_bln_c_v("e_bln_c_v", dim_combine(nproma, 6, nblks_c)),
+ p_u_out("p_u_out", dim_combine(nproma, nlev, nblks_c)),
+ p_v_out("p_v_out", dim_combine(nproma, nlev, nblks_c))
+ {}
+};
+
+/// ValueTypes to test with
+typedef ::testing::Types<float, double> ValueTypes;
+
+TYPED_TEST_SUITE(InterpolationVectorTest, ValueTypes);
-// Test for the double precision (dp) version.
-TEST(Edges2CellsTest, DPTest) {
- // Allocate and fill input arrays.
- std::vector<double> p_vn_in(num_elements<double>(nproma, nlev, nblks_e), 1.0);
- std::vector<double> p_vt_in(num_elements<double>(nproma, nlev, nblks_e), 1.0);
- // cell_edge_idx and cell_edge_blk: dimensions [nproma, nblks_c, 3]
- std::vector<int> cell_edge_idx(num_elements<int>(nproma, nblks_c, 3), 1);
- std::vector<int> cell_edge_blk(num_elements<int>(nproma, nblks_c, 3), 1);
-
- // Here we set cell_edge_idx to 1, 2, 1 for every triple.
- for (int i = 0; i < num_elements<int>(nproma, nblks_c, 3); i += 3) {
- cell_edge_idx[i] = 1;
- cell_edge_idx[i + 1] = 2;
- cell_edge_idx[i + 2] = 1;
+TYPED_TEST(InterpolationVectorTest, Edges2CellsSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int num_edges = this->num_edges;
+
+ // Define indexing helpers
+ const auto &vn_at = at<nproma, nlev, nblks_e>;
+ const auto &vt_at = at<nproma, nlev, nblks_e>;
+ const auto &edge_idx_at = at<nproma, nblks_c, num_edges>;
+ const auto &edge_blk_at = at<nproma, nblks_c, num_edges>;
+ const auto &bln_at = at<nproma, 6, nblks_c>;
+ const auto &out_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto p_vt_in_h = Kokkos::create_mirror_view(this->p_vt_in);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto e_bln_c_u_h = Kokkos::create_mirror_view(this->e_bln_c_u);
+ auto e_bln_c_v_h = Kokkos::create_mirror_view(this->e_bln_c_v);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Initialize with simple values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vn_in_h[vn_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik + ib);
+ p_vt_in_h[vt_at(ic, ik, ib)] = static_cast<TypeParam>(2.0 + ic + ik + ib);
+ }
+ }
}
- // Similarly, set cell_edge_blk to all ones (valid since nblks_e=2, so index 1
- // means block 0 after subtracting 1). e_bln_c_u and e_bln_c_v: dimensions
- // [nproma, 6, nblks_c]
- std::vector<double> e_bln_c_u(num_elements<double>(nproma, 6, nblks_c), 1.0);
- std::vector<double> e_bln_c_v(num_elements<double>(nproma, 6, nblks_c), 1.0);
- // Output arrays: dimensions [nproma, nlev, nblks_c]
- std::vector<double> p_u_out(num_elements<double>(nproma, nlev, nblks_c), 0.0);
- std::vector<double> p_v_out(num_elements<double>(nproma, nlev, nblks_c), 0.0);
-
- std::vector<double> p_u_ref(num_elements<double>(nproma, nlev, nblks_c), 6.0);
- std::vector<double> p_v_ref(num_elements<double>(nproma, nlev, nblks_c), 6.0);
-
- // Call the dp (double precision) version.
- edges2cells_vector_lib<double>(
- p_vn_in.data(), p_vt_in.data(), cell_edge_idx.data(),
- cell_edge_blk.data(), e_bln_c_u.data(), e_bln_c_v.data(), p_u_out.data(),
- p_v_out.data(), i_startblk, i_endblk, i_startidx_in, i_endidx_in, slev,
- elev, nproma, nlev, nblks_e, nblks_c);
-
- // Check that for each computed cell in p_u_out and p_v_out, the value is 6.
- // This is because for each cell, the kernel adds 6 terms of 1*1.
- for (size_t idx = 0; idx < p_u_out.size(); ++idx) {
- EXPECT_NEAR(p_u_out[idx], p_u_ref[idx], 1e-12);
- EXPECT_NEAR(p_v_out[idx], p_v_ref[idx], 1e-12);
+
+ // Set each cell to connect to 3 edges
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Edge indices are 1-indexed in the function
+ cell_edge_idx_h[edge_idx_at(ic, ib, 0)] = 1;
+ cell_edge_idx_h[edge_idx_at(ic, ib, 1)] = 2;
+ cell_edge_idx_h[edge_idx_at(ic, ib, 2)] = 3;
+
+ // Edge blocks are 1-indexed in the function
+ cell_edge_blk_h[edge_blk_at(ic, ib, 0)] = 1;
+ cell_edge_blk_h[edge_blk_at(ic, ib, 1)] = 1;
+ cell_edge_blk_h[edge_blk_at(ic, ib, 2)] = 1;
+
+ // Initialize bilinear coefficients
+ for (int j = 0; j < 6; ++j) {
+ e_bln_c_u_h[bln_at(ic, j, ib)] = static_cast<TypeParam>(0.1 * (j + 1));
+ e_bln_c_v_h[bln_at(ic, j, ib)] = static_cast<TypeParam>(0.05 * (j + 1));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[out_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ p_v_out_h[out_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->p_vt_in, p_vt_in_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->e_bln_c_u, e_bln_c_u_h);
+ Kokkos::deep_copy(this->e_bln_c_v, e_bln_c_v_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ // Call the function
+ edges2cells_vector_lib<TypeParam>(
+ this->p_vn_in.data(), this->p_vt_in.data(),
+ this->cell_edge_idx.data(), this->cell_edge_blk.data(),
+ this->e_bln_c_u.data(), this->e_bln_c_v.data(),
+ this->p_u_out.data(), this->p_v_out.data(),
+ this->i_startblk, this->i_endblk, this->i_startidx_in, this->i_endidx_in,
+ this->slev, this->elev, nproma, nlev, nblks_e, nblks_c);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Compute expected results on host
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam**, host_space> expected_u("expected_u", nproma, nlev);
+ Kokkos::View<TypeParam**, host_space> expected_v("expected_v", nproma, nlev);
+
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ // Compute expected values
+ expected_u(jc, jk) =
+ e_bln_c_u_h[bln_at(jc, 0, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 1, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 2, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 3, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 4, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 5, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)];
+
+ expected_v(jc, jk) =
+ e_bln_c_v_h[bln_at(jc, 0, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 1, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 2, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 3, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 4, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 5, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)];
+ }
+ }
+ }
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx_in; jc <= this->i_endidx_in; ++jc) {
+ EXPECT_NEAR(p_u_out_h[out_at(jc, jk, jb)], expected_u(jc, jk), 1e-5)
+ << "u value mismatch at jc=" << jc << ", jk=" << jk;
+ EXPECT_NEAR(p_v_out_h[out_at(jc, jk, jb)], expected_v(jc, jk), 1e-5)
+ << "v value mismatch at jc=" << jc << ", jk=" << jk;
+ }
+ }
}
}
-// Test for the single precision (sp) version.
-TEST(Edges2CellsTest, SPTest) {
- // Allocate and fill input arrays.
- std::vector<float> p_vn_in(num_elements<float>(nproma, nlev, nblks_e), 1.0f);
- std::vector<float> p_vt_in(num_elements<float>(nproma, nlev, nblks_e), 1.0f);
- std::vector<int> cell_edge_idx(num_elements<int>(nproma, nblks_c, 3), 1);
- std::vector<int> cell_edge_blk(num_elements<int>(nproma, nblks_c, 3), 1);
- // Set cell_edge_idx values to 1, 2, 1.
- for (int i = 0; i < num_elements<int>(nproma, nblks_c, 3); i += 3) {
- cell_edge_idx[i] = 1;
- cell_edge_idx[i + 1] = 2;
- cell_edge_idx[i + 2] = 1;
+TYPED_TEST(InterpolationVectorTest, Edges2CellsRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int num_edges = this->num_edges;
+
+ // Define indexing helpers
+ const auto &vn_at = at<nproma, nlev, nblks_e>;
+ const auto &vt_at = at<nproma, nlev, nblks_e>;
+ const auto &edge_idx_at = at<nproma, nblks_c, num_edges>;
+ const auto &edge_blk_at = at<nproma, nblks_c, num_edges>;
+ const auto &bln_at = at<nproma, 6, nblks_c>;
+ const auto &out_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto p_vt_in_h = Kokkos::create_mirror_view(this->p_vt_in);
+ auto cell_edge_idx_h = Kokkos::create_mirror_view(this->cell_edge_idx);
+ auto cell_edge_blk_h = Kokkos::create_mirror_view(this->cell_edge_blk);
+ auto e_bln_c_u_h = Kokkos::create_mirror_view(this->e_bln_c_u);
+ auto e_bln_c_v_h = Kokkos::create_mirror_view(this->e_bln_c_v);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> edge_distrib(1, num_edges);
+ std::uniform_int_distribution<int> block_distrib(1, nblks_e);
+ std::uniform_real_distribution<TypeParam> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vn_in_h[vn_at(ic, ik, ib)] = real_distrib(gen);
+ p_vt_in_h[vt_at(ic, ik, ib)] = real_distrib(gen);
+ }
+ }
+ }
+
+ // // Set each cell to connect to random edges
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Edge indices are 1-indexed in the function
+ cell_edge_idx_h[edge_idx_at(ic, ib, 0)] = edge_distrib(gen);
+ cell_edge_idx_h[edge_idx_at(ic, ib, 1)] = edge_distrib(gen);
+ cell_edge_idx_h[edge_idx_at(ic, ib, 2)] = edge_distrib(gen);
+
+ // Edge blocks are 1-indexed in the function
+ cell_edge_blk_h[edge_blk_at(ic, ib, 0)] = block_distrib(gen);
+ cell_edge_blk_h[edge_blk_at(ic, ib, 1)] = block_distrib(gen);
+ cell_edge_blk_h[edge_blk_at(ic, ib, 2)] = block_distrib(gen);
+
+ // Initialize random bilinear coefficients
+ for (int j = 0; j < 6; ++j) {
+ e_bln_c_u_h[bln_at(ic, j, ib)] = real_distrib(gen);
+ e_bln_c_v_h[bln_at(ic, j, ib)] = real_distrib(gen);
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[out_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ p_v_out_h[out_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->p_vt_in, p_vt_in_h);
+ Kokkos::deep_copy(this->cell_edge_idx, cell_edge_idx_h);
+ Kokkos::deep_copy(this->cell_edge_blk, cell_edge_blk_h);
+ Kokkos::deep_copy(this->e_bln_c_u, e_bln_c_u_h);
+ Kokkos::deep_copy(this->e_bln_c_v, e_bln_c_v_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ // Call the function
+ edges2cells_vector_lib<TypeParam>(
+ this->p_vn_in.data(), this->p_vt_in.data(),
+ this->cell_edge_idx.data(), this->cell_edge_blk.data(),
+ this->e_bln_c_u.data(), this->e_bln_c_v.data(),
+ this->p_u_out.data(), this->p_v_out.data(),
+ this->i_startblk, this->i_endblk, this->i_startidx_in, this->i_endidx_in,
+ this->slev, this->elev, nproma, nlev, nblks_e, nblks_c);
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Compute expected results on host
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_u("expected_u", nproma, nlev, nblks_c);
+ Kokkos::View<TypeParam***, host_space> expected_v("expected_v", nproma, nlev, nblks_c);
+
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ // Compute expected values
+ expected_u(jc, jk, jb) =
+ e_bln_c_u_h[bln_at(jc, 0, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 1, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 2, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 3, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 4, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)] +
+ e_bln_c_u_h[bln_at(jc, 5, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)];
+
+ expected_v(jc, jk, jb) =
+ e_bln_c_v_h[bln_at(jc, 0, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 1, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 0)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 0)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 2, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 3, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 1)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 1)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 4, jb)] *
+ p_vn_in_h[vn_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)] +
+ e_bln_c_v_h[bln_at(jc, 5, jb)] *
+ p_vt_in_h[vt_at(cell_edge_idx_h[edge_idx_at(jc, jb, 2)] - 1, jk,
+ cell_edge_blk_h[edge_blk_at(jc, jb, 2)] - 1)];
+ }
+ }
}
- std::vector<float> e_bln_c_u(num_elements<float>(nproma, 6, nblks_c), 1.0f);
- std::vector<float> e_bln_c_v(num_elements<float>(nproma, 6, nblks_c), 1.0f);
- std::vector<float> p_u_out(num_elements<float>(nproma, nlev, nblks_c), 0.0f);
- std::vector<float> p_v_out(num_elements<float>(nproma, nlev, nblks_c), 0.0f);
-
- std::vector<float> p_u_ref(num_elements<float>(nproma, nlev, nblks_c), 6.0f);
- std::vector<float> p_v_ref(num_elements<float>(nproma, nlev, nblks_c), 6.0f);
-
- // Call the sp (float precision) version.
- edges2cells_vector_lib<float>(
- p_vn_in.data(), p_vt_in.data(), cell_edge_idx.data(),
- cell_edge_blk.data(), e_bln_c_u.data(), e_bln_c_v.data(), p_u_out.data(),
- p_v_out.data(), i_startblk, i_endblk, i_startidx_in, i_endidx_in, slev,
- elev, nproma, nlev, nblks_e, nblks_c);
-
- // Verify that every computed output equals 6.
- for (size_t idx = 0; idx < p_u_out.size(); ++idx) {
- EXPECT_NEAR(p_u_out[idx], p_u_ref[idx], 1e-5f);
- EXPECT_NEAR(p_v_out[idx], p_v_ref[idx], 1e-5f);
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = this->i_startidx_in; jc <= this->i_endidx_in; ++jc) {
+ EXPECT_NEAR(p_u_out_h[out_at(jc, jk, 0)], expected_u(jc, jk, 0), tol)
+ << "u value mismatch at jc=" << jc << ", jk=" << jk;
+ EXPECT_NEAR(p_v_out_h[out_at(jc, jk, 0)], expected_v(jc, jk, 0), tol)
+ << "v value mismatch at jc=" << jc << ", jk=" << jk;
+ }
+ }
}
}
diff --git a/test/c/test_intp_rbf.cpp b/test/c/test_intp_rbf.cpp
index 040d440223c683407a585de764511e5e2b384aea..a67480de0746d9564a5f3baf323e78fa1314ccfd 100644
--- a/test/c/test_intp_rbf.cpp
+++ b/test/c/test_intp_rbf.cpp
@@ -15,15 +15,9 @@
#include <gtest/gtest.h>
#include <numeric>
#include <vector>
-
-// Free-function helpers for 3D and 4D array sizes (assumed column-major)
-template <typename T> size_t num_elements_3d(int d1, int d2, int d3) {
- return static_cast<size_t>(d1) * d2 * d3;
-}
-
-template <typename T> size_t num_elements_4d(int d1, int d2, int d3, int d4) {
- return static_cast<size_t>(d1) * d2 * d3 * d4;
-}
+#include <random>
+#include <iostream>
+#include "dim_helper.hpp"
// Define a helper struct that holds the two types.
template <typename InT, typename OutT> struct MixedPrecision {
@@ -31,6 +25,9 @@ template <typename InT, typename OutT> struct MixedPrecision {
using out_type = OutT;
};
+// Define the list of types we want to test.
+typedef ::testing::Types<float, double> MyTypes;
+
// Define the list of type pairs we want to test.
typedef ::testing::Types<MixedPrecision<double, double>,
MixedPrecision<double, float>,
@@ -48,6 +45,7 @@ public:
static constexpr int rbf_c2grad_dim = 10; // fixed dimension
static constexpr int rbf_vec_dim_c = 9;
static constexpr int rbf_vec_dim_e = 4;
+ static constexpr int rbf_vec_dim_v = 6; // Fixed dimension for RBF
// Parameter values.
const int i_startblk = 0;
@@ -60,244 +58,1103 @@ public:
const bool acc_async = false; // No asynchronous execution.
};
-// Define a typed test fixture for the functions which have the same input and
-// output types
template <typename T>
-class RbfInterpolTypedTestFixture : public ::testing::Test,
+class RbfInterpolSingleParamTest : public ::testing::Test,
public interp_dimensions {
public:
- // Data arrays.
- std::vector<T> p_cell_in; // size: nproma * nlev * nblks_c
- std::vector<int> rbf_c2grad_idx; // size: rbf_c2grad_dim * nproma * nblks_c
- std::vector<int> rbf_c2grad_blk; // size: rbf_c2grad_dim * nproma * nblks_c
- std::vector<int> rbf_vec_idx_c; // size: rbf_vec_dim_c * nproma * nblks_c
- std::vector<int> rbf_vec_blk_c; // size: rbf_vec_dim_c * nproma * nblks_c
- std::vector<T>
- rbf_c2grad_coeff; // size: rbf_c2grad_dim * 2 * nproma * nblks_c
- std::vector<T> grad_x; // size: nproma * nlev * nblks_c
- std::vector<T> grad_y; // size: nproma * nlev * nblks_c
- std::vector<T> p_vn_in;
- std::vector<T> rbf_vec_coeff_c;
- std::vector<T> p_u_out;
- std::vector<T> p_v_out;
-
- std::vector<int> rbf_vec_idx_e;
- std::vector<int> rbf_vec_blk_e;
- std::vector<T> rbf_vec_coeff_e;
- std::vector<T> p_vt_out;
-
- RbfInterpolTypedTestFixture() {
- size_t size3d = static_cast<size_t>(nproma) * nlev * nblks_c;
- size_t size3d_idx = static_cast<size_t>(rbf_c2grad_dim) * nproma * nblks_c;
- size_t size4d = static_cast<size_t>(rbf_c2grad_dim) * 2 * nproma * nblks_c;
-
- size_t size3d_vec_dim =
- static_cast<size_t>(rbf_vec_dim_c) * nproma * nblks_c;
- size_t size_4d_vec_dim =
- static_cast<size_t>(rbf_vec_dim_c) * 2 * nproma * nblks_c;
-
- size_t size3d_edge_lib =
- static_cast<size_t>(rbf_vec_dim_e) * nproma * nblks_c;
- size_t size_4d_edge_lib =
- static_cast<size_t>(rbf_vec_dim_e) * 2 * nproma * nblks_c;
-
- p_cell_in.resize(size3d, static_cast<T>(1));
- p_vn_in.resize(size3d, static_cast<T>(1));
-
- rbf_vec_idx_c.resize(size3d_vec_dim, 1);
- rbf_vec_blk_c.resize(size3d_vec_dim, 0);
- rbf_c2grad_idx.resize(size3d_idx, 1);
- rbf_c2grad_blk.resize(size3d_idx, 0); // Set block indices to 0 for testing.
- rbf_vec_idx_e.resize(size3d_vec_dim, 1);
- rbf_vec_blk_e.resize(size3d_vec_dim, 0);
-
- rbf_vec_coeff_c.resize(size_4d_vec_dim, static_cast<T>(1));
- rbf_c2grad_coeff.resize(size4d, static_cast<T>(1));
- rbf_vec_coeff_e.resize(size_4d_edge_lib, static_cast<T>(1));
-
- p_u_out.resize(size3d_vec_dim, static_cast<T>(0));
- p_v_out.resize(size3d_vec_dim, static_cast<T>(0));
- p_vt_out.resize(size3d_edge_lib, static_cast<T>(0));
-
- grad_x.resize(size3d, static_cast<T>(0));
- grad_y.resize(size3d, static_cast<T>(0));
- }
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Kokkos Views for test data
+ Kokkos::View<T*, memory_space> p_cell_in; // Dimensions: (nproma, nlev, nblks_c)
+ Kokkos::View<int*, memory_space> rbf_c2grad_idx; // Dimensions: (rbf_c2grad_dim, nproma, nblks_c)
+ Kokkos::View<int*, memory_space> rbf_c2grad_blk; // Dimensions: (rbf_c2grad_dim, nproma, nblks_c)
+ Kokkos::View<T*, memory_space> rbf_c2grad_coeff; // Dimensions: (rbf_c2grad_dim, 2, nproma, nblks_c)
+ Kokkos::View<T*, memory_space> grad_x; // Dimensions: (nproma, nlev, nblks_c)
+ Kokkos::View<T*, memory_space> grad_y; // Dimensions: (nproma, nlev, nblks_c)
+
+ // Additional arrays for other functions
+ Kokkos::View<T*, memory_space> p_vn_in;
+ Kokkos::View<int*, memory_space> rbf_vec_idx_c;
+ Kokkos::View<int*, memory_space> rbf_vec_blk_c;
+ Kokkos::View<T*, memory_space> rbf_vec_coeff_c;
+ Kokkos::View<T*, memory_space> p_u_out;
+ Kokkos::View<T*, memory_space> p_v_out;
+
+ Kokkos::View<int*, memory_space> rbf_vec_idx_e;
+ Kokkos::View<int*, memory_space> rbf_vec_blk_e;
+ Kokkos::View<T*, memory_space> rbf_vec_coeff_e;
+ Kokkos::View<T*, memory_space> p_vt_out;
+
+ RbfInterpolSingleParamTest()
+ : p_cell_in("p_cell_in", nproma * nlev * nblks_c),
+ rbf_c2grad_idx("rbf_c2grad_idx", rbf_c2grad_dim * nproma * nblks_c),
+ rbf_c2grad_blk("rbf_c2grad_blk", rbf_c2grad_dim * nproma * nblks_c),
+ rbf_c2grad_coeff("rbf_c2grad_coeff", rbf_c2grad_dim * 2 * nproma * nblks_c),
+ grad_x("grad_x", nproma * nlev * nblks_c),
+ grad_y("grad_y", nproma * nlev * nblks_c),
+
+ p_vn_in("p_vn_in", nproma * nlev * nblks_c),
+ rbf_vec_idx_c("rbf_vec_idx_c", rbf_vec_dim_c * nproma * nblks_c),
+ rbf_vec_blk_c("rbf_vec_blk_c", rbf_vec_dim_c * nproma * nblks_c),
+ rbf_vec_coeff_c("rbf_vec_coeff_c", rbf_vec_dim_c * 2 * nproma * nblks_c),
+ p_u_out("p_u_out", nproma * nlev * nblks_c),
+ p_v_out("p_v_out", nproma * nlev * nblks_c),
+
+ rbf_vec_idx_e("rbf_vec_idx_e", rbf_vec_dim_e * nproma * nblks_c),
+ rbf_vec_blk_e("rbf_vec_blk_e", rbf_vec_dim_e * nproma * nblks_c),
+ rbf_vec_coeff_e("rbf_vec_coeff_e", rbf_vec_dim_e * 2 * nproma * nblks_c),
+ p_vt_out("p_vt_out", nproma * nlev * nblks_c)
+ {}
};
-typedef ::testing::Types<float, double> MyTypes;
+TYPED_TEST_SUITE(RbfInterpolSingleParamTest, MyTypes);
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// ! rbf_interpol_c2grad
+//
+////////////////////////////////////////////////////////////////////////////////
+
+TYPED_TEST(RbfInterpolSingleParamTest, C2GradSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int rbf_c2grad_dim = this->rbf_c2grad_dim;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<rbf_c2grad_dim, nproma, nblks_c>;
+ const auto &blk_at = at<rbf_c2grad_dim, nproma, nblks_c>;
+ const auto &coeff_at = at<rbf_c2grad_dim, 2, nproma, nblks_c>;
+ const auto &grad_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto rbf_c2grad_idx_h = Kokkos::create_mirror_view(this->rbf_c2grad_idx);
+ auto rbf_c2grad_blk_h = Kokkos::create_mirror_view(this->rbf_c2grad_blk);
+ auto rbf_c2grad_coeff_h = Kokkos::create_mirror_view(this->rbf_c2grad_coeff);
+ auto grad_x_h = Kokkos::create_mirror_view(this->grad_x);
+ auto grad_y_h = Kokkos::create_mirror_view(this->grad_y);
-TYPED_TEST_SUITE(RbfInterpolTypedTestFixture, MyTypes);
+ // Initialize with index-based pattern for cell data
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // First index points to self
+ rbf_c2grad_idx_h[idx_at(0, ic, ib)] = ic;
+ rbf_c2grad_blk_h[blk_at(0, ic, ib)] = ib;
+
+ // Other indices follow a pattern
+ for (int j = 1; j < rbf_c2grad_dim; ++j) {
+ rbf_c2grad_idx_h[idx_at(j, ic, ib)] = (ic + j) % nproma;
+ rbf_c2grad_blk_h[blk_at(j, ic, ib)] = (ib + j % nblks_c) % nblks_c;
+ }
+
+ // Coefficients for x and y gradients - use a simple pattern that depends on ib, ic and j
+ for (int j = 0; j < rbf_c2grad_dim; ++j) {
+ rbf_c2grad_coeff_h[coeff_at(j, 0, ic, ib)] = static_cast<TypeParam>(1.0 + ib + 0.1*ic + 0.01 * j); // x coefficient
+ rbf_c2grad_coeff_h[coeff_at(j, 1, ic, ib)] = static_cast<TypeParam>(2.0 + ib + 0.1*ic + 0.01 * j); // y coefficient
+ }
+ }
+ }
-TYPED_TEST(RbfInterpolTypedTestFixture, C2Grad) {
- using T = TypeParam;
+ // Initialize gradients to zero
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ grad_x_h[grad_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ grad_y_h[grad_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->rbf_c2grad_idx, rbf_c2grad_idx_h);
+ Kokkos::deep_copy(this->rbf_c2grad_blk, rbf_c2grad_blk_h);
+ Kokkos::deep_copy(this->rbf_c2grad_coeff, rbf_c2grad_coeff_h);
+ Kokkos::deep_copy(this->grad_x, grad_x_h);
+ Kokkos::deep_copy(this->grad_y, grad_y_h);
+
+ Kokkos::fence();
+
+ // Call the function
rbf_interpol_c2grad_lib<TypeParam>(
this->p_cell_in.data(), this->rbf_c2grad_idx.data(),
this->rbf_c2grad_blk.data(), this->rbf_c2grad_coeff.data(),
this->grad_x.data(), this->grad_y.data(), this->i_startblk,
this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
- this->elev, this->nproma, this->rbf_c2grad_dim, this->nlev, this->nblks_c,
- this->lacc);
-
- // For each block from i_startblk to i_endblk-1, and for each (i, level)
- // the kernel sums rbf_c2grad_dim contributions, each equal to 1.
- // Therefore, we expect grad_x and grad_y to equal rbf_c2grad_dim (i.e., 10).
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
- for (int jk = 0; jk < this->nlev; ++jk) {
- for (int i = 0; i < this->nproma; ++i) {
- size_t idx = i + static_cast<size_t>(jk) * this->nproma +
- static_cast<size_t>(jb) * this->nproma * this->nlev;
- EXPECT_NEAR(this->grad_x[idx],
- static_cast<TypeParam>(this->rbf_c2grad_dim),
- static_cast<TypeParam>(1e-5))
- << "grad_x failure at block " << jb << ", level " << jk
- << ", index " << i;
- EXPECT_NEAR(this->grad_y[idx],
- static_cast<TypeParam>(this->rbf_c2grad_dim),
- static_cast<TypeParam>(1e-5))
- << "grad_y failure at block " << jb << ", level " << jk
- << ", index " << i;
+ this->elev, nproma, rbf_c2grad_dim, nlev, nblks_c, this->lacc);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(grad_x_h, this->grad_x);
+ Kokkos::deep_copy(grad_y_h, this->grad_y);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_grad_x = {
+ 19.9225, 22.9275, 26.2225, 20.9675, 24.0725, 27.4675,
+ 22.0125, 25.2175, 28.7125, 23.0575, 26.3625, 29.9575,
+ 38.972, 42.977, 47.272, 41.017, 45.122, 49.517,
+ 43.062, 47.267, 51.762, 45.107, 49.412, 54.007
+ };
+
+ std::vector<TypeParam> expected_grad_y = {
+ 38.9725, 42.9775, 47.2725, 41.0175, 45.1225, 49.5175,
+ 43.0625, 47.2675, 51.7625, 45.1075, 49.4125, 54.0075,
+ 58.022, 63.027, 68.322, 61.067, 66.172, 71.567,
+ 64.112, 69.317, 74.812, 67.157, 72.462, 78.057
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ EXPECT_NEAR(grad_x_h[grad_at(jc, jk, jb)],
+ expected_grad_x[grad_at(jc, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "grad_x failure at block " << jb << ", level " << jk << ", index " << jc;
+ EXPECT_NEAR(grad_y_h[grad_at(jc, jk, jb)],
+ expected_grad_y[grad_at(jc, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "grad_y failure at block " << jb << ", level " << jk << ", index " << jc;
}
}
}
}
-TYPED_TEST(RbfInterpolTypedTestFixture, Cell) {
- using T = TypeParam;
+TYPED_TEST(RbfInterpolSingleParamTest, C2GradRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int rbf_c2grad_dim = this->rbf_c2grad_dim;
+
+ // Define indexing helpers
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+ const auto &idx_at = at<rbf_c2grad_dim, nproma, nblks_c>;
+ const auto &blk_at = at<rbf_c2grad_dim, nproma, nblks_c>;
+ const auto &coeff_at = at<rbf_c2grad_dim, 2, nproma, nblks_c>;
+ const auto &grad_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_cell_in_h = Kokkos::create_mirror_view(this->p_cell_in);
+ auto rbf_c2grad_idx_h = Kokkos::create_mirror_view(this->rbf_c2grad_idx);
+ auto rbf_c2grad_blk_h = Kokkos::create_mirror_view(this->rbf_c2grad_blk);
+ auto rbf_c2grad_coeff_h = Kokkos::create_mirror_view(this->rbf_c2grad_coeff);
+ auto grad_x_h = Kokkos::create_mirror_view(this->grad_x);
+ auto grad_y_h = Kokkos::create_mirror_view(this->grad_y);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> cell_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_c - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<double> coeff_distrib(-0.2, 0.2); // Allow negative coefficients for gradients
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_cell_in_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize connectivity indices with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // First index points to self
+ rbf_c2grad_idx_h[idx_at(0, ic, ib)] = ic;
+ rbf_c2grad_blk_h[blk_at(0, ic, ib)] = ib;
+
+ // Other indices randomized
+ for (int j = 1; j < rbf_c2grad_dim; ++j) {
+ rbf_c2grad_idx_h[idx_at(j, ic, ib)] = cell_distrib(gen);
+ rbf_c2grad_blk_h[blk_at(j, ic, ib)] = block_distrib(gen);
+ }
+
+ // Random coefficients for gradient reconstruction
+ for (int j = 0; j < rbf_c2grad_dim; ++j) {
+ rbf_c2grad_coeff_h[coeff_at(j, 0, ic, ib)] = static_cast<TypeParam>(coeff_distrib(gen)); // x coefficient
+ rbf_c2grad_coeff_h[coeff_at(j, 1, ic, ib)] = static_cast<TypeParam>(coeff_distrib(gen)); // y coefficient
+ }
+ }
+ }
+
+ // Initialize gradients to zero
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ grad_x_h[grad_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ grad_y_h[grad_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_cell_in, p_cell_in_h);
+ Kokkos::deep_copy(this->rbf_c2grad_idx, rbf_c2grad_idx_h);
+ Kokkos::deep_copy(this->rbf_c2grad_blk, rbf_c2grad_blk_h);
+ Kokkos::deep_copy(this->rbf_c2grad_coeff, rbf_c2grad_coeff_h);
+ Kokkos::deep_copy(this->grad_x, grad_x_h);
+ Kokkos::deep_copy(this->grad_y, grad_y_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_interpol_c2grad_lib<TypeParam>(
+ this->p_cell_in.data(), this->rbf_c2grad_idx.data(),
+ this->rbf_c2grad_blk.data(), this->rbf_c2grad_coeff.data(),
+ this->grad_x.data(), this->grad_y.data(), this->i_startblk,
+ this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
+ this->elev, nproma, rbf_c2grad_dim, nlev, nblks_c, this->lacc);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(grad_x_h, this->grad_x);
+ Kokkos::deep_copy(grad_y_h, this->grad_y);
- rbf_vec_interpol_cell_lib<T>(
+ // Calculate expected values
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_grad_x("expected_grad_x", nproma, nlev, nblks_c);
+ Kokkos::View<TypeParam***, host_space> expected_grad_y("expected_grad_y", nproma, nlev, nblks_c);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ expected_grad_x(jc, jk, jb) = static_cast<TypeParam>(0.0);
+ expected_grad_y(jc, jk, jb) = static_cast<TypeParam>(0.0);
+
+ for (int j = 0; j < rbf_c2grad_dim; ++j) {
+ int cell_idx = rbf_c2grad_idx_h[idx_at(j, jc, jb)];
+ int cell_blk = rbf_c2grad_blk_h[blk_at(j, jc, jb)];
+ TypeParam coeff_x = rbf_c2grad_coeff_h[coeff_at(j, 0, jc, jb)];
+ TypeParam coeff_y = rbf_c2grad_coeff_h[coeff_at(j, 1, jc, jb)];
+
+ expected_grad_x(jc, jk, jb) +=
+ coeff_x * p_cell_in_h[cell_at(cell_idx, jk, cell_blk)];
+ expected_grad_y(jc, jk, jb) +=
+ coeff_y * p_cell_in_h[cell_at(cell_idx, jk, cell_blk)];
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ EXPECT_NEAR(grad_x_h[grad_at(jc, jk, jb)],
+ expected_grad_x(jc, jk, jb), tol)
+ << "grad_x failure at block " << jb << ", level " << jk << ", index " << jc;
+ EXPECT_NEAR(grad_y_h[grad_at(jc, jk, jb)],
+ expected_grad_y(jc, jk, jb), tol)
+ << "grad_y failure at block " << jb << ", level " << jk << ", index " << jc;
+ }
+ }
+ }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// ! rbf_vec_interpol_cell
+//
+////////////////////////////////////////////////////////////////////////////////
+
+TYPED_TEST(RbfInterpolSingleParamTest, CellSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int rbf_vec_dim_c = this->rbf_vec_dim_c;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_c, nproma, nblks_c>;
+ const auto &blk_at = at<rbf_vec_dim_c, nproma, nblks_c>;
+ const auto &coeff_at = at<rbf_vec_dim_c, 2, nproma, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto rbf_vec_idx_c_h = Kokkos::create_mirror_view(this->rbf_vec_idx_c);
+ auto rbf_vec_blk_c_h = Kokkos::create_mirror_view(this->rbf_vec_blk_c);
+ auto rbf_vec_coeff_c_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_c);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Initialize with index-based pattern for edge data
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_vn_in_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize cell connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each cell connects to rbf_vec_dim_c edges
+ for (int j = 0; j < rbf_vec_dim_c; ++j) {
+ // Edge indices with a pattern
+ rbf_vec_idx_c_h[idx_at(j, ic, ib)] = (ic + j) % nproma;
+ rbf_vec_blk_c_h[blk_at(j, ic, ib)] = (ib + j % nblks_e) % nblks_e;
+
+ // Interpolation coefficients that depend on indices
+ rbf_vec_coeff_c_h[coeff_at(j, 0, ic, ib)] = static_cast<TypeParam>(1.0 + ib + 0.1*ic + 0.01 * j); // x coefficient
+ rbf_vec_coeff_c_h[coeff_at(j, 1, ic, ib)] = static_cast<TypeParam>(2.0 + ib + 0.1*ic + 0.01 * j); // y coefficient
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ p_v_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_c, rbf_vec_idx_c_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_c, rbf_vec_blk_c_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_c, rbf_vec_coeff_c_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_vec_interpol_cell_lib<TypeParam>(
this->p_vn_in.data(), this->rbf_vec_idx_c.data(),
this->rbf_vec_blk_c.data(), this->rbf_vec_coeff_c.data(),
this->p_u_out.data(), this->p_v_out.data(), this->i_startblk,
this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
- this->elev, this->nproma, this->rbf_c2grad_dim, this->nlev, this->nblks_c,
- this->nblks_e, this->lacc, this->acc_async);
-
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
- for (int jk = 0; jk < this->nlev; ++jk) {
- for (int i = 0; i < this->nproma; ++i) {
- size_t idx = i + static_cast<size_t>(jk) * this->nproma +
- static_cast<size_t>(jb) * this->nproma * this->nlev;
- EXPECT_NEAR(this->p_u_out[idx], static_cast<T>(this->rbf_vec_dim_c),
- static_cast<T>(1e-5))
- << "p_u_out failure at block " << jb << ", level " << jk
- << ", index " << i;
+ this->elev, nproma, nlev, nblks_c, nblks_e, rbf_vec_dim_c,
+ this->lacc, this->acc_async);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_u = {
+ 18.8216, 20.5356, 22.3396, 19.7576, 21.5616, 23.4556,
+ 20.6936, 22.5876, 24.5716, 21.6296, 23.6136, 25.6876,
+ 36.882, 38.597, 40.402, 38.718, 40.523, 42.418,
+ 40.554, 42.449, 44.434, 42.39, 44.375, 46.45
+ };
+ std::vector<TypeParam> expected_v = {
+ 36.8616, 38.5756, 40.3796, 38.6976, 40.5016, 42.3956,
+ 40.5336, 42.4276, 44.4116, 42.3696, 44.3536, 46.4276,
+ 54.932, 56.647, 58.452, 57.668, 59.473, 61.368,
+ 60.404, 62.299, 64.284, 63.14, 65.125, 67.2
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ EXPECT_NEAR(p_u_out_h[cell_at(jc, jk, jb)],
+ expected_u[cell_at(jc, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "u failure at block " << jb << ", level " << jk << ", index " << jc;
+
+ EXPECT_NEAR(p_v_out_h[cell_at(jc, jk, jb)],
+ expected_v[cell_at(jc, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "v failure at block " << jb << ", level " << jk << ", index " << jc;
}
}
}
}
-TYPED_TEST(RbfInterpolTypedTestFixture, Edge) {
- using T = TypeParam;
+TYPED_TEST(RbfInterpolSingleParamTest, CellRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_c = this->nblks_c;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int rbf_vec_dim_c = this->rbf_vec_dim_c;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_c, nproma, nblks_c>;
+ const auto &blk_at = at<rbf_vec_dim_c, nproma, nblks_c>;
+ const auto &coeff_at = at<rbf_vec_dim_c, 2, nproma, nblks_c>;
+ const auto &cell_at = at<nproma, nlev, nblks_c>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto rbf_vec_idx_c_h = Kokkos::create_mirror_view(this->rbf_vec_idx_c);
+ auto rbf_vec_blk_c_h = Kokkos::create_mirror_view(this->rbf_vec_blk_c);
+ auto rbf_vec_coeff_c_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_c);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(43); // Different seed from other tests
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_e - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<double> coeff_distrib(0.01, 0.2);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vn_in_h[edge_at(ic, ik, ib)] = static_cast<TypeParam>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize cell connectivity indices with random values
+ for (int ib = 0; ib < nblks_c; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ for (int j = 0; j < rbf_vec_dim_c; ++j) {
+ // Random edge indices and blocks
+ rbf_vec_idx_c_h[idx_at(j, ic, ib)] = edge_distrib(gen);
+ rbf_vec_blk_c_h[blk_at(j, ic, ib)] = block_distrib(gen);
+ // Random coefficients for interpolation
+ rbf_vec_coeff_c_h[coeff_at(j, 0, ic, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ rbf_vec_coeff_c_h[coeff_at(j, 1, ic, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ p_v_out_h[cell_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_c, rbf_vec_idx_c_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_c, rbf_vec_blk_c_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_c, rbf_vec_coeff_c_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_vec_interpol_cell_lib<TypeParam>(
+ this->p_vn_in.data(), this->rbf_vec_idx_c.data(),
+ this->rbf_vec_blk_c.data(), this->rbf_vec_coeff_c.data(),
+ this->p_u_out.data(), this->p_v_out.data(), this->i_startblk,
+ this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
+ this->elev, nproma, nlev, nblks_c, nblks_e, rbf_vec_dim_c,
+ this->lacc, this->acc_async);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Calculate expected values
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_u("expected_u", nproma, nlev, nblks_c);
+ Kokkos::View<TypeParam***, host_space> expected_v("expected_v", nproma, nlev, nblks_c);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ expected_u(jc, jk, jb) = static_cast<TypeParam>(0.0);
+ expected_v(jc, jk, jb) = static_cast<TypeParam>(0.0);
+
+ for (int j = 0; j < rbf_vec_dim_c; ++j) {
+ int edge_idx = rbf_vec_idx_c_h[idx_at(j, jc, jb)];
+ int edge_blk = rbf_vec_blk_c_h[blk_at(j, jc, jb)];
+ TypeParam coeff_u = rbf_vec_coeff_c_h[coeff_at(j, 0, jc, jb)];
+ TypeParam coeff_v = rbf_vec_coeff_c_h[coeff_at(j, 1, jc, jb)];
+
+ expected_u(jc, jk, jb) +=
+ coeff_u * p_vn_in_h[edge_at(edge_idx, jk, edge_blk)];
+ expected_v(jc, jk, jb) +=
+ coeff_v * p_vn_in_h[edge_at(edge_idx, jk, edge_blk)];
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
- rbf_vec_interpol_edge_lib<T>(
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jc = i_startidx; jc <= i_endidx; ++jc) {
+ EXPECT_NEAR(p_u_out_h[cell_at(jc, jk, jb)],
+ expected_u(jc, jk, jb), tol)
+ << "u failure at block " << jb << ", level " << jk << ", index " << jc;
+
+ EXPECT_NEAR(p_v_out_h[cell_at(jc, jk, jb)],
+ expected_v(jc, jk, jb), tol)
+ << "v failure at block " << jb << ", level " << jk << ", index " << jc;
+ }
+ }
+ }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// ! rbf_vec_interpol_edge
+//
+////////////////////////////////////////////////////////////////////////////////
+
+TYPED_TEST(RbfInterpolSingleParamTest, EdgeSpecific) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int rbf_vec_dim_e = this->rbf_vec_dim_e;
+
+ // Define indexing helpers
+ const auto &vn_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &blk_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &coeff_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &vt_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto rbf_vec_idx_e_h = Kokkos::create_mirror_view(this->rbf_vec_idx_e);
+ auto rbf_vec_blk_e_h = Kokkos::create_mirror_view(this->rbf_vec_blk_e);
+ auto rbf_vec_coeff_e_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_e);
+ auto p_vt_out_h = Kokkos::create_mirror_view(this->p_vt_out);
+
+ // Initialize with index-based pattern for edge data
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_vn_in_h[vn_at(ic, ik, ib)] = static_cast<TypeParam>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize edge connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each edge uses rbf_vec_dim_e neighboring edges
+ for (int j = 0; j < rbf_vec_dim_e; ++j) {
+ // Edge indices with a pattern
+ rbf_vec_idx_e_h[idx_at(j, ic, ib)] = (ic + j) % nproma;
+ rbf_vec_blk_e_h[blk_at(j, ic, ib)] = (ib + j % nblks_e) % nblks_e;
+
+ // Interpolation coefficients that depend on indices
+ rbf_vec_coeff_e_h[coeff_at(j, ic, ib)] = static_cast<TypeParam>(1.0 + ib + 0.1*ic + 0.01 * j); // coefficient
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vt_out_h[vt_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_e, rbf_vec_idx_e_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_e, rbf_vec_blk_e_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_e, rbf_vec_coeff_e_h);
+ Kokkos::deep_copy(this->p_vt_out, p_vt_out_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_vec_interpol_edge_lib<TypeParam>(
this->p_vn_in.data(), this->rbf_vec_idx_e.data(),
this->rbf_vec_blk_e.data(), this->rbf_vec_coeff_e.data(),
this->p_vt_out.data(), this->i_startblk, this->i_endblk,
this->i_startidx_in, this->i_endidx_in, this->slev, this->elev,
- this->nlev, this->nproma, this->rbf_vec_dim_e, this->nblks_e, this->lacc,
- this->acc_async);
-
- for (int jb = this->i_startblk; jb < this->i_endblk; ++jb) {
- for (int jk = 0; jk < this->nlev; ++jk) {
- for (int i = 0; i < this->nproma; ++i) {
- size_t idx = i + static_cast<size_t>(jk) * this->nproma +
- static_cast<size_t>(jb) * this->nproma * this->nlev;
- EXPECT_NEAR(this->p_vt_out[idx], static_cast<T>(this->rbf_vec_dim_e),
- static_cast<T>(1e-5))
- << "p_vt_out failure at block " << jb << ", level " << jk
- << ", index " << i;
+ nlev, nproma, rbf_vec_dim_e, nblks_e, this->lacc, this->acc_async);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vt_out_h, this->p_vt_out);
+
+ // Expected results based on the specific test values
+ std::vector<TypeParam> expected_vt = {
+ 7.1304, 8.9324, 10.9644, 7.5364, 9.3784, 11.4504,
+ 7.9424, 9.8244, 11.9364, 8.3484, 10.2704, 12.4224,
+ 14.1502, 16.9522, 19.9842, 14.9562, 17.7982, 20.8702,
+ 15.7622, 18.6442, 21.7562, 16.5682, 19.4902, 22.6422,
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ EXPECT_NEAR(p_vt_out_h[vt_at(je, jk, jb)],
+ expected_vt[vt_at(je, jk, jb)],
+ static_cast<TypeParam>(1e-5))
+ << "Tangential velocity failure at block " << jb << ", level " << jk << ", index " << je;
+ }
+ }
+ }
+}
+
+TYPED_TEST(RbfInterpolSingleParamTest, EdgeRandom) {
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int rbf_vec_dim_e = this->rbf_vec_dim_e;
+
+ // Define indexing helpers
+ const auto &vn_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &blk_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &coeff_at = at<rbf_vec_dim_e, nproma, nblks_e>;
+ const auto &vt_at = at<nproma, nlev, nblks_e>;
+
+ // Create host mirror views
+ auto p_vn_in_h = Kokkos::create_mirror_view(this->p_vn_in);
+ auto rbf_vec_idx_e_h = Kokkos::create_mirror_view(this->rbf_vec_idx_e);
+ auto rbf_vec_blk_e_h = Kokkos::create_mirror_view(this->rbf_vec_blk_e);
+ auto rbf_vec_coeff_e_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_e);
+ auto p_vt_out_h = Kokkos::create_mirror_view(this->p_vt_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(44); // Different seed from other tests
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_e - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+ std::uniform_real_distribution<double> coeff_distrib(0.01, 0.5);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_vn_in_h[vn_at(ic, ik, ib)] = static_cast<TypeParam>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize edge connectivity indices with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+
+ for (int j = 0; j < rbf_vec_dim_e; ++j) {
+ // Random edge indices and blocks
+ rbf_vec_idx_e_h[idx_at(j, ic, ib)] = edge_distrib(gen);
+ rbf_vec_blk_e_h[blk_at(j, ic, ib)] = block_distrib(gen);
+ // Random coefficients for interpolation
+ rbf_vec_coeff_e_h[coeff_at(j, ic, ib)] = static_cast<TypeParam>(coeff_distrib(gen));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_vt_out_h[vt_at(ic, ik, ib)] = static_cast<TypeParam>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_vn_in, p_vn_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_e, rbf_vec_idx_e_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_e, rbf_vec_blk_e_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_e, rbf_vec_coeff_e_h);
+ Kokkos::deep_copy(this->p_vt_out, p_vt_out_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_vec_interpol_edge_lib<TypeParam>(
+ this->p_vn_in.data(), this->rbf_vec_idx_e.data(),
+ this->rbf_vec_blk_e.data(), this->rbf_vec_coeff_e.data(),
+ this->p_vt_out.data(), this->i_startblk, this->i_endblk,
+ this->i_startidx_in, this->i_endidx_in, this->slev, this->elev,
+ nlev, nproma, rbf_vec_dim_e, nblks_e, this->lacc, this->acc_async);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_vt_out_h, this->p_vt_out);
+
+ // Calculate expected values
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<TypeParam***, host_space> expected_vt("expected_vt", nproma, nlev, nblks_e);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ expected_vt(je, jk, jb) = static_cast<TypeParam>(0.0);
+
+ for (int j = 0; j < rbf_vec_dim_e; ++j) {
+ int edge_idx = rbf_vec_idx_e_h[idx_at(j, je, jb)];
+ int edge_blk = rbf_vec_blk_e_h[blk_at(j, je, jb)];
+ TypeParam coeff = rbf_vec_coeff_e_h[coeff_at(j, je, jb)];
+
+ expected_vt(je, jk, jb) +=
+ coeff * p_vn_in_h[vn_at(edge_idx, jk, edge_blk)];
+ }
+ }
+ }
+ }
+
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_c_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int je = i_startidx; je <= i_endidx; ++je) {
+ EXPECT_NEAR(p_vt_out_h[vt_at(je, jk, jb)],
+ expected_vt(je, jk, jb), tol)
+ << "Tangential velocity failure at block " << jb << ", level " << jk << ", index " << je;
}
}
}
}
-// Define a typed test fixture for the functions which have different input and
-// output types
template <typename TypePair>
-class RbfVecInterpolMixedTestFixture : public ::testing::Test,
+class RbfVecInterpolDoubleParamTest : public ::testing::Test,
public interp_dimensions {
public:
using InType = typename TypePair::in_type;
using OutType = typename TypePair::out_type;
- // Constant dimensions.
- static constexpr int nproma = 3; // inner loop length
- static constexpr int nlev = 4; // number of vertical levels
- static constexpr int nblks_e = 2; // number of edge blocks (for p_e_in)
- static constexpr int nblks_v =
- 2; // number of vertex blocks (for rbf arrays and outputs)
- static constexpr int rbf_vec_dim =
- 6; // fixed dimension for rbf vector (stencil points)
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Kokkos Views for test data
+ Kokkos::View<InType*, memory_space> p_e_in; // Dimensions: (nproma, nlev, nblks_e)
+ Kokkos::View<int*, memory_space> rbf_vec_idx_v; // Dimensions: (rbf_vec_dim_v, nproma, nblks_v)
+ Kokkos::View<int*, memory_space> rbf_vec_blk_v; // Dimensions: (rbf_vec_dim_v, nproma, nblks_v)
+ Kokkos::View<InType*, memory_space> rbf_vec_coeff_v; // Dimensions: (rbf_vec_dim_v, 2, nproma, nblks_v)
+ Kokkos::View<OutType*, memory_space> p_u_out; // Dimensions: (nproma, nlev, nblks_v)
+ Kokkos::View<OutType*, memory_space> p_v_out; // Dimensions: (nproma, nlev, nblks_v)
- // Parameter values.
- int i_startblk = 0;
- int i_endblk = 1; // Test blocks [0, 1]
- int i_startidx_in = 0;
- int i_endidx_in = nproma - 1; // Full range: 0 .. nproma-1
- int slev = 0;
- int elev = nlev - 1; // Full vertical range (0 .. nlev-1)
- bool lacc = false; // Not using ACC-specific behavior.
- bool acc_async = false; // No asynchronous execution.
-
- // Arrays stored in std::vector.
- std::vector<InType> p_e_in; // Dimensions: (nproma, nlev, nblks_e)
- std::vector<int> rbf_vec_idx_v; // Dimensions: (rbf_vec_dim, nproma, nblks_v)
- std::vector<int> rbf_vec_blk_v; // Dimensions: (rbf_vec_dim, nproma, nblks_v)
- std::vector<InType>
- rbf_vec_coeff_v; // Dimensions: (rbf_vec_dim, 2, nproma, nblks_v)
- std::vector<OutType> p_u_out; // Dimensions: (nproma, nlev, nblks_v)
- std::vector<OutType> p_v_out; // Dimensions: (nproma, nlev, nblks_v)
-
- RbfVecInterpolMixedTestFixture() {
- // Allocate and initialize inputs.
- p_e_in.resize(num_elements_3d<InType>(nproma, nlev, nblks_e),
- static_cast<InType>(1));
- rbf_vec_idx_v.resize(num_elements_3d<int>(rbf_vec_dim, nproma, nblks_v), 1);
- rbf_vec_blk_v.resize(num_elements_3d<int>(rbf_vec_dim, nproma, nblks_v), 0);
- rbf_vec_coeff_v.resize(
- num_elements_4d<InType>(rbf_vec_dim, 2, nproma, nblks_v),
- static_cast<InType>(1));
-
- // Allocate output arrays and initialize to zero.
- p_u_out.resize(num_elements_3d<OutType>(nproma, nlev, nblks_v),
- static_cast<OutType>(0));
- p_v_out.resize(num_elements_3d<OutType>(nproma, nlev, nblks_v),
- static_cast<OutType>(0));
- }
+ RbfVecInterpolDoubleParamTest()
+ : p_e_in("p_e_in", nproma * nlev * nblks_e),
+ rbf_vec_idx_v("rbf_vec_idx_v", rbf_vec_dim_v * nproma * nblks_v),
+ rbf_vec_blk_v("rbf_vec_blk_v", rbf_vec_dim_v * nproma * nblks_v),
+ rbf_vec_coeff_v("rbf_vec_coeff_v", rbf_vec_dim_v * 2 * nproma * nblks_v),
+ p_u_out("p_u_out", nproma * nlev * nblks_v),
+ p_v_out("p_v_out", nproma * nlev * nblks_v)
+ {}
};
-TYPED_TEST_SUITE(RbfVecInterpolMixedTestFixture, MixedTypes);
+TYPED_TEST_SUITE(RbfVecInterpolDoubleParamTest, MixedTypes);
+
+////////////////////////////////////////////////////////////////////////////////
+//
+// ! rbf_vec_interpol_vertex
+//
+////////////////////////////////////////////////////////////////////////////////
-TYPED_TEST(RbfVecInterpolMixedTestFixture, Vertex) {
+TYPED_TEST(RbfVecInterpolDoubleParamTest, VertexSpecific) {
using InType = typename TestFixture::InType;
using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int rbf_vec_dim_v = this->rbf_vec_dim_v;
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_v, nproma, nblks_v>;
+ const auto &blk_at = at<rbf_vec_dim_v, nproma, nblks_v>;
+ const auto &coeff_at = at<rbf_vec_dim_v, 2, nproma, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_e_in_h = Kokkos::create_mirror_view(this->p_e_in);
+ auto rbf_vec_idx_v_h = Kokkos::create_mirror_view(this->rbf_vec_idx_v);
+ auto rbf_vec_blk_v_h = Kokkos::create_mirror_view(this->rbf_vec_blk_v);
+ auto rbf_vec_coeff_v_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_v);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Initialize with index-based pattern for edge data
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Value depends on all three indices
+ p_e_in_h[edge_at(ic, ik, ib)] = static_cast<InType>(1.0 + ic + ik * 0.1 + ib * 0.01);
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with specific pattern
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ // Each vertex connects to 6 edges
+ for (int j = 0; j < rbf_vec_dim_v; ++j) {
+ // Edge indices with a pattern
+ rbf_vec_idx_v_h[idx_at(j, ic, ib)] = (ic + j) % nproma;
+ rbf_vec_blk_v_h[blk_at(j, ic, ib)] = (ib + j % nblks_e) % nblks_e;
+
+ // Interpolation coefficients that depend on indices
+ rbf_vec_coeff_v_h[coeff_at(j, 0, ic, ib)] = static_cast<InType>(1.0 + ib + 0.1*ic + 0.01 * j); // x coefficient
+ rbf_vec_coeff_v_h[coeff_at(j, 1, ic, ib)] = static_cast<InType>(2.0 + ib + 0.1*ic + 0.01 * j); // y coefficient
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ p_v_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_e_in, p_e_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_v, rbf_vec_idx_v_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_v, rbf_vec_blk_v_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_v, rbf_vec_coeff_v_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ Kokkos::fence();
- // Call the function with mixed precision.
+ // Call the function
rbf_vec_interpol_vertex_lib<InType, OutType>(
this->p_e_in.data(), this->rbf_vec_idx_v.data(),
this->rbf_vec_blk_v.data(), this->rbf_vec_coeff_v.data(),
this->p_u_out.data(), this->p_v_out.data(), this->i_startblk,
this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
- this->elev, this->nproma, this->lacc, this->acc_async, this->nlev,
- this->nblks_e, this->nblks_v);
-
- // Check the outputs only for blocks in the range [i_startblk, i_endblk].
- for (int block = this->i_startblk; block <= this->i_endblk; ++block) {
- for (int level = 0; level < this->nlev; ++level) {
- for (int i = 0; i < this->nproma; ++i) {
- // Compute the linear index for a 3D array in column-major order:
- size_t idx =
- i + level * this->nproma + block * this->nproma * this->nlev;
- // Since every contribution is 1 and there are 6 stencil points,
- // expect 6.
- EXPECT_NEAR(this->p_u_out[idx], static_cast<OutType>(6),
- static_cast<OutType>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
- EXPECT_NEAR(this->p_v_out[idx], static_cast<OutType>(6),
- static_cast<OutType>(1e-5))
- << "Failure at block " << block << ", level " << level << ", index "
- << i;
+ this->elev, nproma, this->lacc, this->acc_async, nlev,
+ nblks_e, nblks_v);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Expected results based on the specific test values
+ std::vector<OutType> expected_u = {
+ 12.3709, 13.5139, 14.7169, 12.9859, 14.1889, 15.4519,
+ 13.6009, 14.8639, 16.1869, 14.2159, 15.5389, 16.9219,
+ 24.4006, 25.5436, 26.7466, 25.6156, 26.8186, 28.0816,
+ 26.8306, 28.0936, 29.4166, 28.0456, 29.3686, 30.7516
+ };
+ std::vector<OutType> expected_v = {
+ 24.4009, 25.5439, 26.7469, 25.6159, 26.8189, 28.0819,
+ 26.8309, 28.0939, 29.4169, 28.0459, 29.3689, 30.7519,
+ 36.4306, 37.5736, 38.7766, 38.2456, 39.4486, 40.7116,
+ 40.0606, 41.3236, 42.6466, 41.8756, 43.1986, 44.5816
+ };
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ EXPECT_NEAR(p_u_out_h[vert_at(jv, jk, jb)],
+ expected_u[vert_at(jv, jk, jb)],
+ static_cast<OutType>(1e-5))
+ << "u failure at block " << jb << ", level " << jk << ", index " << jv;
+ EXPECT_NEAR(p_v_out_h[vert_at(jv, jk, jb)],
+ expected_v[vert_at(jv, jk, jb)],
+ static_cast<OutType>(1e-5))
+ << "v failure at block " << jb << ", level " << jk << ", index " << jv;
+ }
+ }
+ }
+}
+
+TYPED_TEST(RbfVecInterpolDoubleParamTest, VertexRandom) {
+ using InType = typename TestFixture::InType;
+ using OutType = typename TestFixture::OutType;
+
+ constexpr int nproma = this->nproma;
+ constexpr int nlev = this->nlev;
+ constexpr int nblks_e = this->nblks_e;
+ constexpr int nblks_v = this->nblks_v;
+ constexpr int rbf_vec_dim_v = 6; // Fixed dimension for RBF
+
+ // Define indexing helpers
+ const auto &edge_at = at<nproma, nlev, nblks_e>;
+ const auto &idx_at = at<rbf_vec_dim_v, nproma, nblks_v>;
+ const auto &blk_at = at<rbf_vec_dim_v, nproma, nblks_v>;
+ const auto &coeff_at = at<rbf_vec_dim_v, 2, nproma, nblks_v>;
+ const auto &vert_at = at<nproma, nlev, nblks_v>;
+
+ // Create host mirror views
+ auto p_e_in_h = Kokkos::create_mirror_view(this->p_e_in);
+ auto rbf_vec_idx_v_h = Kokkos::create_mirror_view(this->rbf_vec_idx_v);
+ auto rbf_vec_blk_v_h = Kokkos::create_mirror_view(this->rbf_vec_blk_v);
+ auto rbf_vec_coeff_v_h = Kokkos::create_mirror_view(this->rbf_vec_coeff_v);
+ auto p_u_out_h = Kokkos::create_mirror_view(this->p_u_out);
+ auto p_v_out_h = Kokkos::create_mirror_view(this->p_v_out);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_int_distribution<int> edge_distrib(0, nproma - 1);
+ std::uniform_int_distribution<int> block_distrib(0, nblks_e - 1);
+ std::uniform_real_distribution<double> real_distrib(0.01, 1.0);
+
+ // Initialize with random values
+ for (int ib = 0; ib < nblks_e; ++ib) {
+ for (int ik = 0; ik < nlev; ++ik) {
+ for (int ic = 0; ic < nproma; ++ic) {
+ p_e_in_h[edge_at(ic, ik, ib)] = static_cast<InType>(real_distrib(gen));
+ }
+ }
+ }
+
+ // Initialize vertex connectivity indices with random values
+ for (int ib = 0; ib < nblks_v; ++ib) {
+ for (int ic = 0; ic < nproma; ++ic) {
+
+ for (int j = 0; j < rbf_vec_dim_v; ++j) {
+ // Random edge indices and blocks
+ rbf_vec_idx_v_h[idx_at(j, ic, ib)] = edge_distrib(gen);
+ rbf_vec_blk_v_h[blk_at(j, ic, ib)] = block_distrib(gen);
+ // Random coefficients for interpolation
+ rbf_vec_coeff_v_h[coeff_at(j, 0, ic, ib)] = static_cast<InType>(real_distrib(gen));
+ rbf_vec_coeff_v_h[coeff_at(j, 1, ic, ib)] = static_cast<InType>(real_distrib(gen));
+ }
+
+ // Initialize output to zero
+ for (int ik = 0; ik < nlev; ++ik) {
+ p_u_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ p_v_out_h[vert_at(ic, ik, ib)] = static_cast<OutType>(0.0);
+ }
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->p_e_in, p_e_in_h);
+ Kokkos::deep_copy(this->rbf_vec_idx_v, rbf_vec_idx_v_h);
+ Kokkos::deep_copy(this->rbf_vec_blk_v, rbf_vec_blk_v_h);
+ Kokkos::deep_copy(this->rbf_vec_coeff_v, rbf_vec_coeff_v_h);
+ Kokkos::deep_copy(this->p_u_out, p_u_out_h);
+ Kokkos::deep_copy(this->p_v_out, p_v_out_h);
+
+ Kokkos::fence();
+
+ // Call the function
+ rbf_vec_interpol_vertex_lib<InType, OutType>(
+ this->p_e_in.data(), this->rbf_vec_idx_v.data(),
+ this->rbf_vec_blk_v.data(), this->rbf_vec_coeff_v.data(),
+ this->p_u_out.data(), this->p_v_out.data(), this->i_startblk,
+ this->i_endblk, this->i_startidx_in, this->i_endidx_in, this->slev,
+ this->elev, nproma, this->lacc, this->acc_async, nlev,
+ nblks_e, nblks_v);
+
+ Kokkos::fence();
+
+ // Copy results back to host
+ Kokkos::deep_copy(p_u_out_h, this->p_u_out);
+ Kokkos::deep_copy(p_v_out_h, this->p_v_out);
+
+ // Calculate expected values
+ using host_space = Kokkos::HostSpace;
+ Kokkos::View<OutType***, host_space> expected_u("expected_u", nproma, nlev, nblks_v);
+ Kokkos::View<OutType***, host_space> expected_v("expected_v", nproma, nlev, nblks_v);
+
+ // Compute expected values
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ expected_u(jv, jk, jb) = static_cast<OutType>(0.0);
+ expected_v(jv, jk, jb) = static_cast<OutType>(0.0);
+
+ for (int j = 0; j < rbf_vec_dim_v; ++j) {
+ int edge_idx = rbf_vec_idx_v_h[idx_at(j, jv, jb)];
+ int edge_blk = rbf_vec_blk_v_h[blk_at(j, jv, jb)];
+ InType coeff_u = rbf_vec_coeff_v_h[coeff_at(j, 0, jv, jb)];
+ InType coeff_v = rbf_vec_coeff_v_h[coeff_at(j, 1, jv, jb)];
+
+ expected_u(jv, jk, jb) +=
+ static_cast<OutType>(coeff_u * p_e_in_h[edge_at(edge_idx, jk, edge_blk)]);
+ expected_v(jv, jk, jb) +=
+ static_cast<OutType>(coeff_v * p_e_in_h[edge_at(edge_idx, jk, edge_blk)]);
+ }
+ }
+ }
+ }
+
+ OutType tol = std::is_same<OutType, float>::value ?
+ static_cast<OutType>(1e-5) : static_cast<OutType>(1e-13);
+
+ // Verify results
+ for (int jb = this->i_startblk; jb <= this->i_endblk; ++jb) {
+ int i_startidx, i_endidx;
+ get_indices_v_lib(this->i_startidx_in, this->i_endidx_in, nproma, jb,
+ this->i_startblk, this->i_endblk, i_startidx, i_endidx);
+
+ for (int jk = this->slev; jk <= this->elev; ++jk) {
+ for (int jv = i_startidx; jv <= i_endidx; ++jv) {
+ EXPECT_NEAR(p_u_out_h[vert_at(jv, jk, jb)],
+ expected_u(jv, jk, jb), tol)
+ << "u failure at block " << jb << ", level " << jk << ", index " << jv;
+ EXPECT_NEAR(p_v_out_h[vert_at(jv, jk, jb)],
+ expected_v(jv, jk, jb), tol)
+ << "v failure at block " << jb << ", level " << jk << ", index " << jv;
}
}
}
diff --git a/test/c/test_tdma_solver.cpp b/test/c/test_tdma_solver.cpp
index 4e09ff335368b2cbc532fd2ade2aee8a8259415d..bb1dddff1d19f895fc0498f5abec69f260b9758a 100644
--- a/test/c/test_tdma_solver.cpp
+++ b/test/c/test_tdma_solver.cpp
@@ -13,76 +13,341 @@
#include <vector>
#include <algorithm>
#include "mo_math_utilities.hpp"
+#include "dim_helper.hpp"
+#include <Kokkos_Core.hpp>
+#include <random>
-// Helper function to compute the 1D index for column-major storage.
+// Helper function for column-major indexing
+template <typename T>
inline int idx(int i, int j, int nrows) {
return i + j * nrows;
}
-// Test fixture for the TDMA solver tests.
-class TDMASolverTestFixture : public ::testing::Test {
+template <typename T>
+class TDMASolverTypedTestFixture : public ::testing::Test {
protected:
- const int n = 10; // Matrix dimension.
- std::vector<double> a; // Input matrix a.
- std::vector<double> b; // Input matrix b.
- std::vector<double> c; // Input matrix c.
- std::vector<double> d; // Input matrix d.
- std::vector<double> x; // Output matrix.
-
- TDMASolverTestFixture()
- : a(n * n), b(n * n), c(n * n), d(n * n), x(n * n, 0.0) {}
-
- // SetUp is run before each test.
- void SetUp() override {
- // Fill arrays in column-major order.
+ const int n = 10; // Matrix dimension.
+
+ // Using Kokkos execution and memory spaces
+ using exec_space = Kokkos::DefaultExecutionSpace;
+ using memory_space = exec_space::memory_space;
+
+ // Kokkos Views for test data
+ Kokkos::View<T*, memory_space> a; // Input matrix a.
+ Kokkos::View<T*, memory_space> b; // Input matrix b.
+ Kokkos::View<T*, memory_space> c; // Input matrix c.
+ Kokkos::View<T*, memory_space> d; // Input matrix d.
+ Kokkos::View<T*, memory_space> x; // Output matrix.
+
+ TDMASolverTypedTestFixture()
+ : a("a", n * n),
+ b("b", n * n),
+ c("c", n * n),
+ d("d", n * n),
+ x("x", n * n)
+ {}
+
+ void SetUpSpecificTest() {
+ // Create host mirror views
+ auto a_h = Kokkos::create_mirror_view(a);
+ auto b_h = Kokkos::create_mirror_view(b);
+ auto c_h = Kokkos::create_mirror_view(c);
+ auto d_h = Kokkos::create_mirror_view(d);
+ auto x_h = Kokkos::create_mirror_view(x);
+
+ // Fill arrays in column-major order with the specific test values
for (int j = 0; j < n; j++) {
for (int i = 0; i < n; i++) {
- double value = (i + 1) + (j + 1);
- a[idx(i, j, n)] = 1.0 * value;
- b[idx(i, j, n)] = 2.0 * value;
- c[idx(i, j, n)] = 1.0 * value;
- d[idx(i, j, n)] = 1.0 * value;
+ T value = static_cast<T>((i + 1) + (j + 1));
+ a_h[idx<T>(i, j, n)] = static_cast<T>(1.0) * value;
+ b_h[idx<T>(i, j, n)] = static_cast<T>(2.0) * value;
+ c_h[idx<T>(i, j, n)] = static_cast<T>(1.0) * value;
+ d_h[idx<T>(i, j, n)] = static_cast<T>(1.0) * value;
+ x_h[idx<T>(i, j, n)] = static_cast<T>(0.0);
}
}
- // Clear the output vector.
- std::fill(x.begin(), x.end(), 0.0);
+
+ // Copy to device
+ Kokkos::deep_copy(a, a_h);
+ Kokkos::deep_copy(b, b_h);
+ Kokkos::deep_copy(c, c_h);
+ Kokkos::deep_copy(d, d_h);
+ Kokkos::deep_copy(x, x_h);
}
};
-TEST_F(TDMASolverTestFixture, FullTest) {
- // Call the solver over the full range:
- tdma_solver_vec<double>(a.data(), b.data(), c.data(), d.data(),
- 0, n, 0, n, n, n, x.data());
+// Define the types we want to test with
+typedef ::testing::Types<float, double> NumericTypes;
+TYPED_TEST_SUITE(TDMASolverTypedTestFixture, NumericTypes);
+
+// Specific test for the full matrix
+TYPED_TEST(TDMASolverTypedTestFixture, SpecificFull) {
+ const int n = this->n;
+
+ // Set up the test with specific values
+ this->SetUpSpecificTest();
+
+ // Call the solver over the full range
+ tdma_solver_vec<TypeParam>(
+ this->a.data(), this->b.data(), this->c.data(), this->d.data(),
+ 0, n, 0, n, n, n, this->x.data());
+
+ // Copy results back to host
+ auto x_h = Kokkos::create_mirror_view(this->x);
+ Kokkos::deep_copy(x_h, this->x);
- // Compute the sum of all elements in the output matrix.
- double sum = 0.0;
+ // Compute the sum of all elements in the output matrix
+ TypeParam sum = 0.0;
for (int j = 0; j < n; j++) {
for (int i = 0; i < n; i++) {
- sum += x[idx(i, j, n)];
+ sum += x_h[idx<TypeParam>(i, j, n)];
}
}
- // Expected reference sum
- double sum_ref = 27.2727272727272769;
- double tol = 1e-13;
+ // Expected reference sum (adjusted for precision)
+ TypeParam sum_ref = static_cast<TypeParam>(27.2727272727272769);
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
EXPECT_NEAR(sum, sum_ref, tol);
}
-TEST_F(TDMASolverTestFixture, PartialTest) {
- // Call the solver for a partial region:
- // For C++: slev = 1, elev = n-1, startidx = 1, endidx = n-1.
- tdma_solver_vec<double>(a.data(), b.data(), c.data(), d.data(),
- 1, n - 1, 1, n - 1, n, n, x.data());
+// Specific test for a partial region
+TYPED_TEST(TDMASolverTypedTestFixture, SpecificPartial) {
+ const int n = this->n;
+
+ // Set up the test with specific values
+ this->SetUpSpecificTest();
+
+ // Call the solver for a partial region
+ tdma_solver_vec<TypeParam>(
+ this->a.data(), this->b.data(), this->c.data(), this->d.data(),
+ 1, n - 1, 1, n - 1, n, n, this->x.data());
+
+ // Copy results back to host
+ auto x_h = Kokkos::create_mirror_view(this->x);
+ Kokkos::deep_copy(x_h, this->x);
// Compute the sum over a region
- double sum = 0.0;
+ TypeParam sum = 0.0;
for (int j = 1; j < n - 1; j++) {
for (int i = 1; i < n - 1; i++) {
- sum += x[idx(i, j, n)];
+ sum += x_h[idx<TypeParam>(i, j, n)];
}
}
- double sum_ref = 17.7777777777777679;
- double tol = 1e-13;
+ // Expected reference sum (adjusted for precision)
+ TypeParam sum_ref = static_cast<TypeParam>(17.7777777777777679);
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
EXPECT_NEAR(sum, sum_ref, tol);
}
+
+// Random test for the full matrix
+TYPED_TEST(TDMASolverTypedTestFixture, RandomFull) {
+ const int n = this->n;
+
+ // Create host mirror views
+ auto a_h = Kokkos::create_mirror_view(this->a);
+ auto b_h = Kokkos::create_mirror_view(this->b);
+ auto c_h = Kokkos::create_mirror_view(this->c);
+ auto d_h = Kokkos::create_mirror_view(this->d);
+ auto x_h = Kokkos::create_mirror_view(this->x);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(42);
+ std::uniform_real_distribution<double> diag_dist(5.0, 10.0); // For main diagonal
+ std::uniform_real_distribution<double> off_diag_dist(0.1, 2.0); // For off-diagonals
+ std::uniform_real_distribution<double> rhs_dist(-10.0, 10.0); // For right-hand side
+
+ // Fill arrays with random values
+ for (int j = 0; j < n; j++) {
+ for (int i = 0; i < n; i++) {
+ a_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-off_diag_dist(gen));
+ b_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(diag_dist(gen));
+ c_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-off_diag_dist(gen));
+ d_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(rhs_dist(gen));
+ x_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(0.0);
+ }
+ }
+
+ // Save copies for reference solution
+ std::vector<TypeParam> a_copy(n * n);
+ std::vector<TypeParam> b_copy(n * n);
+ std::vector<TypeParam> c_copy(n * n);
+ std::vector<TypeParam> d_copy(n * n);
+ std::vector<TypeParam> x_expected(n * n, 0.0);
+
+ for (int j = 0; j < n; j++) {
+ for (int i = 0; i < n; i++) {
+ a_copy[idx<TypeParam>(i, j, n)] = a_h[idx<TypeParam>(i, j, n)];
+ b_copy[idx<TypeParam>(i, j, n)] = b_h[idx<TypeParam>(i, j, n)];
+ c_copy[idx<TypeParam>(i, j, n)] = c_h[idx<TypeParam>(i, j, n)];
+ d_copy[idx<TypeParam>(i, j, n)] = d_h[idx<TypeParam>(i, j, n)];
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->a, a_h);
+ Kokkos::deep_copy(this->b, b_h);
+ Kokkos::deep_copy(this->c, c_h);
+ Kokkos::deep_copy(this->d, d_h);
+ Kokkos::deep_copy(this->x, x_h);
+
+ // Call the solver over the full range
+ tdma_solver_vec<TypeParam>(
+ this->a.data(), this->b.data(), this->c.data(), this->d.data(),
+ 0, n, 0, n, n, n, this->x.data());
+
+ // Copy results back to host
+ Kokkos::deep_copy(x_h, this->x);
+
+ // Calculate reference solution
+ for (int i = 0; i < n; i++) {
+ // Arrays for internal calculations
+ std::vector<TypeParam> c_prime(n, 0.0);
+ std::vector<TypeParam> d_prime(n, 0.0);
+
+ // Forward sweep
+ c_prime[0] = c_copy[idx<TypeParam>(i, 0, n)] / b_copy[idx<TypeParam>(i, 0, n)];
+ d_prime[0] = d_copy[idx<TypeParam>(i, 0, n)] / b_copy[idx<TypeParam>(i, 0, n)];
+
+ for (int j = 1; j < n; j++) {
+ TypeParam m = static_cast<TypeParam>(1.0) /
+ (b_copy[idx<TypeParam>(i, j, n)] - c_prime[j-1] * a_copy[idx<TypeParam>(i, j, n)]);
+ c_prime[j] = c_copy[idx<TypeParam>(i, j, n)] * m;
+ d_prime[j] = (d_copy[idx<TypeParam>(i, j, n)] - d_prime[j-1] * a_copy[idx<TypeParam>(i, j, n)]) * m;
+ }
+
+ // Back substitution
+ x_expected[idx<TypeParam>(i, n-1, n)] = d_prime[n-1];
+
+ for (int j = n-2; j >= 0; j--) {
+ x_expected[idx<TypeParam>(i, j, n)] = d_prime[j] - c_prime[j] * x_expected[idx<TypeParam>(i, j+1, n)];
+ }
+ }
+
+ // Set tolerance based on type
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify that individual values match
+ for (int j = 0; j < n; j++) {
+ for (int i = 0; i < n; i++) {
+ EXPECT_NEAR(x_h[idx<TypeParam>(i, j, n)], x_expected[idx<TypeParam>(i, j, n)], tol)
+ << "Mismatch at i=" << i << ", j=" << j;
+ }
+ }
+}
+
+// Random test for a partial region
+TYPED_TEST(TDMASolverTypedTestFixture, RandomPartial) {
+ const int n = this->n;
+ const int slev = 1;
+ const int elev = n - 1;
+ const int startidx = 1;
+ const int endidx = n - 1;
+
+ // Create host mirror views
+ auto a_h = Kokkos::create_mirror_view(this->a);
+ auto b_h = Kokkos::create_mirror_view(this->b);
+ auto c_h = Kokkos::create_mirror_view(this->c);
+ auto d_h = Kokkos::create_mirror_view(this->d);
+ auto x_h = Kokkos::create_mirror_view(this->x);
+
+ // Use fixed seed for reproducibility
+ std::mt19937 gen(43);
+ std::uniform_real_distribution<double> diag_dist(5.0, 10.0);
+ std::uniform_real_distribution<double> off_diag_dist(0.1, 2.0);
+ std::uniform_real_distribution<double> rhs_dist(-10.0, 10.0);
+
+ // Initialize all values to something that shouldn't be touched
+ for (int j = 0; j < n; j++) {
+ for (int i = 0; i < n; i++) {
+ a_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-999.0);
+ b_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-999.0);
+ c_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-999.0);
+ d_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-999.0);
+ x_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(0.0);
+ }
+ }
+
+ // Set random values only for the region to be processed
+ for (int j = slev; j < elev; j++) {
+ for (int i = startidx; i < endidx; i++) {
+ a_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-off_diag_dist(gen));
+ b_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(diag_dist(gen));
+ c_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(-off_diag_dist(gen));
+ d_h[idx<TypeParam>(i, j, n)] = static_cast<TypeParam>(rhs_dist(gen));
+ }
+ }
+
+ // Save copies for reference solution
+ std::vector<TypeParam> a_copy(n * n, static_cast<TypeParam>(-999.0));
+ std::vector<TypeParam> b_copy(n * n, static_cast<TypeParam>(-999.0));
+ std::vector<TypeParam> c_copy(n * n, static_cast<TypeParam>(-999.0));
+ std::vector<TypeParam> d_copy(n * n, static_cast<TypeParam>(-999.0));
+ std::vector<TypeParam> x_expected(n * n, static_cast<TypeParam>(0.0));
+
+ for (int j = slev; j < elev; j++) {
+ for (int i = startidx; i < endidx; i++) {
+ a_copy[idx<TypeParam>(i, j, n)] = a_h[idx<TypeParam>(i, j, n)];
+ b_copy[idx<TypeParam>(i, j, n)] = b_h[idx<TypeParam>(i, j, n)];
+ c_copy[idx<TypeParam>(i, j, n)] = c_h[idx<TypeParam>(i, j, n)];
+ d_copy[idx<TypeParam>(i, j, n)] = d_h[idx<TypeParam>(i, j, n)];
+ }
+ }
+
+ // Copy to device
+ Kokkos::deep_copy(this->a, a_h);
+ Kokkos::deep_copy(this->b, b_h);
+ Kokkos::deep_copy(this->c, c_h);
+ Kokkos::deep_copy(this->d, d_h);
+ Kokkos::deep_copy(this->x, x_h);
+
+ // Call the solver for the partial region
+ tdma_solver_vec<TypeParam>(
+ this->a.data(), this->b.data(), this->c.data(), this->d.data(),
+ slev, elev, startidx, endidx, n, n, this->x.data());
+
+ // Copy results back to host
+ Kokkos::deep_copy(x_h, this->x);
+
+ // Calculate reference solution for the partial region
+ for (int i = startidx; i < endidx; i++) {
+ // Arrays for internal calculations
+ std::vector<TypeParam> c_prime(n, 0.0);
+ std::vector<TypeParam> d_prime(n, 0.0);
+
+ // Forward sweep
+ c_prime[slev] = c_copy[idx<TypeParam>(i, slev, n)] / b_copy[idx<TypeParam>(i, slev, n)];
+ d_prime[slev] = d_copy[idx<TypeParam>(i, slev, n)] / b_copy[idx<TypeParam>(i, slev, n)];
+
+ for (int j = slev + 1; j < elev; j++) {
+ TypeParam m = static_cast<TypeParam>(1.0) /
+ (b_copy[idx<TypeParam>(i, j, n)] - c_prime[j-1] * a_copy[idx<TypeParam>(i, j, n)]);
+ c_prime[j] = c_copy[idx<TypeParam>(i, j, n)] * m;
+ d_prime[j] = (d_copy[idx<TypeParam>(i, j, n)] - d_prime[j-1] * a_copy[idx<TypeParam>(i, j, n)]) * m;
+ }
+
+ // Back substitution
+ x_expected[idx<TypeParam>(i, elev-1, n)] = d_prime[elev-1];
+
+ for (int j = elev-2; j >= slev; j--) {
+ x_expected[idx<TypeParam>(i, j, n)] = d_prime[j] - c_prime[j] * x_expected[idx<TypeParam>(i, j+1, n)];
+ }
+ }
+
+ // Set tolerance based on type
+ TypeParam tol = std::is_same<TypeParam, float>::value ?
+ static_cast<TypeParam>(1e-5) : static_cast<TypeParam>(1e-13);
+
+ // Verify that individual values match
+ for (int j = slev; j < elev; j++) {
+ for (int i = startidx; i < endidx; i++) {
+ EXPECT_NEAR(x_h[idx<TypeParam>(i, j, n)], x_expected[idx<TypeParam>(i, j, n)], tol)
+ << "Mismatch at i=" << i << ", j=" << j;
+ }
+ }
+}