Skip to content

Commit

Permalink
Merge branch 'main' into FOR_REDUCE
Browse files Browse the repository at this point in the history 
ENH: Merge with main.
  • Loading branch information
@jacob-moore22
jacob-moore22 committed Oct 7, 2024
2 parents f32f0ad + 0bb2d0c commit 03da254
Show file tree
Hide file tree
Showing 6 changed files with 1,195 additions and 364 deletions.
115 changes: 110 additions & 5 deletions examples/kokkos_for.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,10 +38,10 @@
using namespace mtr; // matar namespace

// main
int main()
int main(int argc, char* argv[])
{
Kokkos::initialize();
{
MATAR_KOKKOS_INIT
{ // kokkos scope
printf("starting test of loop macros \n");

// Kokkos::View<int *> arr("ARR", 10);
Expand Down Expand Up @@ -348,9 +348,114 @@ int main()
printf(" %d %d %d \n", i, j, k);
});

printf("done\n");
// Hierarchical

printf("\n\n\nHierarchical\n");
size_t hiersize = 4;
auto hierTest1D = CArrayKokkos<double>(hiersize);
auto hierTest2D = CArrayKokkos<double>(hiersize, hiersize);
auto hierTest3D = CArrayKokkos<double>(hiersize, hiersize, hiersize);
FOR_ALL(i_i, 0, hiersize, j_j, 0, hiersize, k_k, 0, hiersize, {
hierTest3D(i_i, j_j, k_k) = 0.0;
});
FOR_FIRST(i_i, 0, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
FOR_SECOND(j_j, i_i, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
FOR_THIRD(k_k, i_i, j_j, {
printf("%d,%d,%d\n", i_i, j_j, k_k);
// hierTest3D(i_i,j_j,k_k) = i_i*j_j*k_k;
});
});
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
// printf("%f\n", hierTest3D(0,0,ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
Kokkos::finalize();
//printf("\n\n");

// Hierarchical reductions

FOR_ALL(i_i, 0, hiersize, j_j, 0, hiersize, k_k, 0, hiersize, {
hierTest3D(i_i, j_j, k_k) = i_i*hiersize*hiersize+j_j*hiersize+k_k;
});

printf("\n\n\nHierarchical Reduce\n");
//2D nesting
FOR_FIRST(i_i,0,hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
double result = 0;
double lsum;
FOR_REDUCE_SUM_SECOND(j_j, i_i, hiersize, lsum, {
lsum += hierTest3D(i_i,j_j,0);
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
}, result);
hierTest1D(i_i)= result;
//printf("value at %d is %f\n", i_i, hierTest1D(i_i));
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
//printf("%f\n", hierTest1D(ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
printf("\n\n");

printf("\n\n\nHierarchical Vectorized Reduce\n");
//3D vector nesting
FOR_FIRST(i_i,0,hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
double result = 0;
double lsum;
FOR_SECOND(j_j, i_i, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
FOR_REDUCE_SUM_THIRD(k_k, i_i, j_j, lsum, {
lsum += hierTest3D(i_i,j_j,k_k);
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
}, result);
hierTest2D(i_i,j_j)= result;
//printf("value at %d , %d is %f\n", i_i, j_j, hierTest2D(i_i,j_j));
});
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
//printf("%f\n", hierTest1D(ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
printf("\n\n");

printf("done\n");

} // end kokkos scope
MATAR_KOKKOS_FINALIZE

return 0;
}
104 changes: 0 additions & 104 deletions examples/main_kokkos.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -865,110 +865,6 @@ int main(int argc, char* argv[])
printf("\nalias name value [0] on device = %d \n", a_carray_device(0));
});

// Hierarchical

printf("\n\n\nHierarchical\n");
size_t hiersize = 4;
auto hierTest1D = CArrayKokkos<double>(hiersize);
auto hierTest2D = CArrayKokkos<double>(hiersize, hiersize);
auto hierTest3D = CArrayKokkos<double>(hiersize, hiersize, hiersize);
FOR_ALL(i_i, 0, hiersize, j_j, 0, hiersize, k_k, 0, hiersize, {
hierTest3D(i_i, j_j, k_k) = 0.0;
});
FOR_FIRST(i_i, 0, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
FOR_SECOND(j_j, i_i, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
FOR_THIRD(k_k, i_i, j_j, {
printf("%d,%d,%d\n", i_i, j_j, k_k);
// hierTest3D(i_i,j_j,k_k) = i_i*j_j*k_k;
});
});
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
// printf("%f\n", hierTest3D(0,0,ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
//printf("\n\n");

// Hierarchical reductions

FOR_ALL(i_i, 0, hiersize, j_j, 0, hiersize, k_k, 0, hiersize, {
hierTest3D(i_i, j_j, k_k) = i_i*hiersize*hiersize+j_j*hiersize+k_k;
});

printf("\n\n\nHierarchical Reduce\n");
//2D nesting
FOR_FIRST(i_i, 0, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
double result = 0;
double lsum;
FOR_REDUCE_SUM_SECOND(j_j, i_i, hiersize, lsum, {
lsum += hierTest3D(i_i,j_j,0);
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
}, result);
hierTest1D(i_i)= result;
//printf("value at %d is %f\n", i_i, hierTest1D(i_i));
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
//printf("%f\n", hierTest1D(ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
printf("\n\n");

printf("\n\n\nHierarchical Vectorized Reduce\n");
//3D vector nesting
FOR_FIRST(i_i, 1, hiersize+1, {
// Kokkos::parallel_for( \
//Kokkos::TeamPolicy<>( 32, Kokkos::AUTO, 32 ), \
//KOKKOS_LAMBDA ( const Kokkos::TeamPolicy<>::member_type &teamMember ) {
//const int i_i = TEAM_ID;
double result = 0;
double lsum;
FOR_SECOND(j_j, i_i-1, hiersize, {
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
FOR_REDUCE_SUM_THIRD(k_k, i_i-1, j_j, lsum, {
lsum += hierTest3D(i_i-1,j_j,k_k);
// Kokkos::parallel_for( \
//Kokkos::TeamThreadRange( teamMember, istart, iend ), [&] ( const int (j_j) ) {
// hierTest2D(i_i,j_j) = i_i * (j_j+1);
// int jstart = j_j*32;
// int jend = (j_j+1)*32;
}, result);
hierTest2D(i_i-1,j_j)= result;
printf("value at %d , %d is %f\n", i_i, j_j, hierTest2D(i_i-1,j_j));
});
});
Kokkos::fence();
for (int ppp = 0; ppp < hiersize; ppp++) {
//printf("%f\n", hierTest1D(ppp));
// printf("%f\n", hierTest2D(3,ppp));
// printf("%f\n", hierTest3D(3,3,ppp));
}
printf("\n\n");

} // end of kokkos scope

Kokkos::finalize();
Expand Down
84 changes: 36 additions & 48 deletions examples/test_set_values.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -184,98 +184,86 @@ int main()
Kokkos::initialize();
{
DFArrayKokkos <double> DFAtest (2, 3, 4);
DViewFArrayKokkos <double> DVFAtest (&DFAtest(0, 0, 0), 3, 4);
DFAtest.set_values(1.25);
DVFAtest.set_values(2.34);
printf("DViewFArrayKokkos set_values 2.34 writing over DFArrayKokkos set_values 1.25.\n");
FOR_ALL(i, 0, 4,
j, 0, 3,
k, 0, 2, {
printf("%.2f ", DFAtest(k,j,i));
});
printf("\n");
FOR_ALL(i, 0, 4,
j, 0, 3,{
printf("%.2f ", DVFAtest(j,i));
});
printf("\n");
DFMatrixKokkos <double> DFMtest (2, 3, 4);
DViewFMatrixKokkos <double> DVFMtest (&DFMtest(1, 1, 1), 3, 4);
DFMtest.set_values(1.33);
DVFMtest.set_values(3.24);
printf("DViewFMatrixKokkos set_values 3.24 writing over DFMatrixKokkos set_values 1.33.\n");
FOR_ALL(i, 1, 5,
j, 1, 4,
k, 1, 3, {
printf("%.2f ", DFMtest(k,j,i));
});
printf("\n");
FOR_ALL(i, 1, 5,
j, 1, 4,{
printf("%.2f ", DVFMtest(j,i));
});
printf("\n");
DCArrayKokkos <double> DCAtest (2, 3, 4);
DViewCArrayKokkos <double> DVCAtest (&DCAtest(0, 0, 0), 3, 4);
DCAtest.set_values(1.53);
DVCAtest.set_values(2.33);
printf("DViewCArrayKokkos set_values 2.33 writing over DCArrayKokkos set_values 1.53.\n");
FOR_ALL(i, 0, 4,
j, 0, 3,
k, 0, 2, {
printf("%.2f ", DCAtest(k,j,i));
});
printf("\n");
FOR_ALL(i, 0, 4,
j, 0, 3,{
printf("%.2f ", DVCAtest(j,i));
});
printf("\n");
DCMatrixKokkos <double> DCMtest (2, 3, 4);
DViewCMatrixKokkos <double> DVCMtest (&DCMtest(1, 1, 1), 3, 4);
DCMtest.set_values(1.77);
DVCMtest.set_values(2.17);
printf("DViewCMatrixKokkos set_values 2.17 writing over DCMatrixKokkos set_values 1.77.\n");
FOR_ALL(i, 1, 5,
j, 1, 4,
k, 1, 3, {
printf("%.2f ", DCMtest(k,j,i));
});
printf("\n");
FOR_ALL(i, 1, 5,
j, 1, 4,{
printf("%.2f ", DVCMtest(j,i));
});
printf("\n");

DynamicRaggedRightArrayKokkos <double> dynrightK (3,4);
dynrightK.stride(0) = 1;
dynrightK.stride(1) = 3;
dynrightK.stride(2) = 2;
//dynrightK.stride(0) = 1;
//dynrightK.stride(1) = 3;
//dynrightK.stride(2) = 2;
RUN({
dynrightK.stride(0) = 1;
dynrightK.stride(1) = 3;
dynrightK.stride(2) = 2;
});
dynrightK.set_values(2.14);
dynrightK.set_values_sparse(1.35);
printf("The values within the populated strides of the DynamicRaggedRight are set to 1.35 and the data in the rest of the array is set to 2.14.\n");
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 4; j++) {
printf("%.2f ", dynrightK(i,j));
}
printf("\n");
}
FOR_FIRST(i, 0, 3, {
FOR_SECOND(j, 0, dynrightK.stride(i), {
//printf("%.2f ", dynrightK(i,j));
});
});
printf("\n");
//for (int i = 0; i < 3; i++) {
// for (int j = 0; j < 4; j++) {
//printf("%.2f ", dynrightK(i,j));
// }
//printf("\n");
//}
DynamicRaggedDownArrayKokkos <double> dyndownK (3,4);
dyndownK.stride(0) = 1;
dyndownK.stride(1) = 3;
dyndownK.stride(2) = 2;
dyndownK.stride(3) = 1;
RUN({
dyndownK.stride(0) = 1;
dyndownK.stride(1) = 3;
dyndownK.stride(2) = 2;
dyndownK.stride(3) = 1;
});
dyndownK.set_values(2.14);
dyndownK.set_values_sparse(1.35);
printf("The values within the populated strides of the DynamicRaggedDown are set to 1.35 and the data in the rest of the array is set to 2.14.\n");
for (int i = 0; i < 3; i++) {
for (int j = 0; j < 4; j++) {
printf("%.2f ", dyndownK(i,j));
}
printf("\n");
}

FOR_FIRST(i, 0, 4, {
FOR_SECOND(j, 0, dyndownK.stride(i), {
//printf("%.2f ", dyndownK(i,j));
});
});
printf("\n");
//for (int i = 0; i < 4; i++) {
// for (int j = 0; j < 3; j++) {
// }
//}
}
Kokkos::finalize();
}
2 changes: 1 addition & 1 deletion scripts/matar-install.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@ fi

if [[ "$kokkos_build_type" = *"mpi"* ]]; then
cmake_options+=(
-D MPI=ON
-D Matar_ENABLE_MPI=ON
)
fi

Expand Down
2 changes: 1 addition & 1 deletion src/Kokkos/kokkos
Submodule kokkos updated 870 files
Loading

0 comments on commit 03da254

Please sign in to comment.