Skip to content

Matrix Transpose Tutorial Cleanup #1917

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
MrBurmark wants to merge 16 commits into
base: develop
Choose a base branch
from
+95 −82
Open

Matrix Transpose Tutorial Cleanup #1917

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
3ac295a
Fix indenting
MrBurmark Sep 21, 2025
8060a32
Use unchecked policies
MrBurmark Sep 21, 2025
d0c36ce
Fix team synchronization
MrBurmark Sep 21, 2025
1e1d347
Fix device synchronization
MrBurmark Sep 21, 2025
410f2bb
Remap launch thread mapping
MrBurmark Sep 21, 2025
e5d6004
clean up pass for launch
artv3 Sep 21, 2025
93b5a2a
use row/col_t
artv3 Sep 22, 2025
ec57de9
missed updating variable names
artv3 Sep 22, 2025
53346ca
Merge branch 'develop' into bugfix/burmark1/transpose
artv3 Sep 30, 2025
72cf1c9
Merge branch 'develop' into bugfix/burmark1/transpose
artv3 Oct 1, 2025
258fea0
Merge branch 'develop' into bugfix/burmark1/transpose
artv3 Oct 2, 2025
4ed8196
fix tranpose c example
artv3 Oct 2, 2025
a9e8dc4
Merge branch 'bugfix/burmark1/transpose' of github.com:LLNL/RAJA into…
artv3 Oct 2, 2025
586b594
Merge branch 'develop' into bugfix/burmark1/transpose
rhornung67 Nov 3, 2025
6f489d4
Merge branch 'develop' into bugfix/burmark1/transpose
rhornung67 Dec 1, 2025
627f418
Fix compilation errors
rhornung67 Dec 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 7 additions & 12 deletions exercises/kernel-matrix-transpose-local-array_solution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -399,8 +399,8 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
// These loops iterate over the number of
// tiles needed to carry out the transpose
//
RAJA::statement::Tile<1, RAJA::tile_fixed<TILE_DIM>, RAJA::cuda_block_y_loop,
RAJA::statement::Tile<0, RAJA::tile_fixed<TILE_DIM>, RAJA::cuda_block_x_loop,
RAJA::statement::Tile<1, RAJA::tile_fixed<TILE_DIM>, RAJA::cuda_block_y_direct_unchecked,
RAJA::statement::Tile<0, RAJA::tile_fixed<TILE_DIM>, RAJA::cuda_block_x_direct_unchecked,
// This statement will initalize local array memory inside a
// kernel. The cpu_tile_mem policy specifies that memory should be
// allocated on the stack. The entries in the RAJA::ParamList
Expand Down Expand Up @@ -431,10 +431,7 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
RAJA::statement::ForICount<1, RAJA::statement::Param<0>, RAJA::cuda_thread_x_direct,
RAJA::statement::Lambda<1>
>
>,
// Synchronize threads to ensure all reads
// from the local array are complete
RAJA::statement::CudaSyncThreads
>
>
>
>
Expand Down Expand Up @@ -494,8 +491,8 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
// These loops iterate over the number of
// tiles needed to carry out the transpose
//
RAJA::statement::Tile<1, RAJA::tile_fixed<TILE_DIM>, RAJA::hip_block_y_loop,
RAJA::statement::Tile<0, RAJA::tile_fixed<TILE_DIM>, RAJA::hip_block_x_loop,
RAJA::statement::Tile<1, RAJA::tile_fixed<TILE_DIM>, RAJA::hip_block_y_direct_unchecked,
RAJA::statement::Tile<0, RAJA::tile_fixed<TILE_DIM>, RAJA::hip_block_x_direct_unchecked,
// This statement will initalize local array memory inside a
// kernel. The cpu_tile_mem policy specifies that memory should be
// allocated on the stack. The entries in the RAJA::ParamList
Expand Down Expand Up @@ -526,10 +523,7 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
RAJA::statement::ForICount<1, RAJA::statement::Param<0>, RAJA::hip_thread_x_direct,
RAJA::statement::Lambda<1>
>
>,
// Synchronize threads to ensure all reads
// from the local array are complete
RAJA::statement::HipSyncThreads
>
>
>
>
Expand All @@ -556,6 +550,7 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
);

CAMP_HIP_API_INVOKE_AND_CHECK(hipMemcpy, At, d_At, N_r * N_c * sizeof(int), hipMemcpyDeviceToHost);
CAMP_HIP_API_INVOKE_AND_CHECK(hipDeviceSynchronize);
checkResult<int>(Atview, N_c, N_r);
// printResult<int>(Atview, N_c, N_r);
#endif
Expand Down
45 changes: 24 additions & 21 deletions exercises/launch-matrix-transpose-local-array.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,18 +145,17 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
//
// (2) Inner loops to write array data into output array tile
//
// Note: loop order is swapped from above so that output matrix
// data access is stride-1.
//
for (int tx = 0; tx < TILE_DIM; ++tx) {
for (int ty = 0; ty < TILE_DIM; ++ty) {

int col = bx * TILE_DIM + tx; // Matrix column index
int row = by * TILE_DIM + ty; // Matrix row index
// Tranpose tile offset
int col = by * TILE_DIM + tx; // Matrix column index
int row = bx * TILE_DIM + ty; // Matrix row index

// Bounds check
if (row < N_r && col < N_c) {
Atview(col, row) = Tile[ty][tx];
if (row < N_c && col < N_r) {
Atview(row, col) = Tile[tx][ty];
}
}
}
Expand Down Expand Up @@ -195,10 +194,10 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
/// input matrix into the RAJA_TEAM_SHARED memory array
///

RAJA::loop_icount<loop_pol_1>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<loop_pol_1>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_1>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_1>(ctx, row_tile, [&] (int col, int tx) {

Atview(col, row) = Tile_Array[ty][tx];
Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand Down Expand Up @@ -244,18 +243,18 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))

RAJA_TEAM_SHARED double Tile_Array[TILE_DIM][TILE_DIM];

RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int col, int tx) {

Tile_Array[ty][tx] = Aview(row, col);
Tile_Array[ty][tx] = Aview(row, col);

});
});

RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int col, int tx) {

Atview(col, row) = Tile_Array[ty][tx];
Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand Down Expand Up @@ -307,10 +306,12 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<cuda_threads_x>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<cuda_threads_y>(ctx, row_tile, [&] (int row, int ty) {
ctx.teamSync();

Atview(col, row) = Tile_Array[ty][tx];
RAJA::loop_icount<cuda_threads_y>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<cuda_threads_x>(ctx, row_tile, [&] (int col, int tx) {

Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand Down Expand Up @@ -379,10 +380,12 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<hip_threads_x>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<hip_threads_y>(ctx, row_tile, [&] (int row, int ty) {
ctx.teamSync();

RAJA::loop_icount<hip_threads_y>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<hip_threads_x>(ctx, row_tile, [&] (int col, int tx) {

d_Atview(col, row) = Tile_Array[ty][tx];
d_Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand Down
97 changes: 55 additions & 42 deletions exercises/launch-matrix-transpose-local-array_solution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -145,18 +145,17 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
//
// (2) Inner loops to write array data into output array tile
//
// Note: loop order is swapped from above so that output matrix
// data access is stride-1.
//
for (int tx = 0; tx < TILE_DIM; ++tx) {
for (int ty = 0; ty < TILE_DIM; ++ty) {
for (int ty = 0; ty < TILE_DIM; ++ty) {
for (int tx = 0; tx < TILE_DIM; ++tx) {

int col = bx * TILE_DIM + tx; // Matrix column index
int row = by * TILE_DIM + ty; // Matrix row index
// Tranpose tile offset
int col = by * TILE_DIM + tx; // Matrix column index
int row = bx * TILE_DIM + ty; // Matrix row index

// Bounds check
if (row < N_r && col < N_c) {
Atview(col, row) = Tile[ty][tx];
if (row < N_c && col < N_r) {
Atview(row, col) = Tile[tx][ty];
}
}
}
Expand All @@ -182,9 +181,12 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
RAJA::LaunchParams(), //LaunchParams may be empty when only running on the cpu
[=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx) {

RAJA::tile<loop_pol_1>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r), [&] (RAJA::TypedRangeSegment<int> const &row_tile) {

RAJA::tile<loop_pol_1>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c), [&] (RAJA::TypedRangeSegment<int> const &col_tile) {
RAJA::tile<loop_pol_1>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r),
[&] (RAJA::TypedRangeSegment<int> const &row_tile) {

RAJA::tile<loop_pol_1>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c),
[&] (RAJA::TypedRangeSegment<int> const &col_tile) {

RAJA_TEAM_SHARED double Tile_Array[TILE_DIM][TILE_DIM];

Expand All @@ -196,10 +198,10 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<loop_pol_1>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<loop_pol_1>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_1>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_1>(ctx, row_tile, [&] (int col, int tx) {

Atview(col, row) = Tile_Array[ty][tx];
Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand Down Expand Up @@ -232,9 +234,11 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
RAJA::LaunchParams(), //LaunchParams may be empty when only running on the cpu
[=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx) {

RAJA::tile<omp_pol_2>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r), [&] (RAJA::TypedRangeSegment<int> const &row_tile) {
RAJA::tile<omp_pol_2>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r),
[&] (RAJA::TypedRangeSegment<int> const &row_tile) {

RAJA::tile<loop_pol_2>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c), [&] (RAJA::TypedRangeSegment<int> const &col_tile) {
RAJA::tile<loop_pol_2>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c),
[&] (RAJA::TypedRangeSegment<int> const &col_tile) {

RAJA_TEAM_SHARED double Tile_Array[TILE_DIM][TILE_DIM];

Expand All @@ -246,10 +250,10 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<loop_pol_2>(ctx, row_tile, [&] (int col, int tx) {

Atview(col, row) = Tile_Array[ty][tx];
Atview(row, col) = Tile_Array[tx][ty];

});
});
Expand All @@ -274,8 +278,8 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
const int n_blocks_c = RAJA_DIVIDE_CEILING_INT(N_c, c_block_sz);
const int n_blocks_r = RAJA_DIVIDE_CEILING_INT(N_r, r_block_sz);

using cuda_teams_y = RAJA::LoopPolicy<RAJA::cuda_block_y_direct>;
using cuda_teams_x = RAJA::LoopPolicy<RAJA::cuda_block_x_direct>;
using cuda_teams_y = RAJA::LoopPolicy<RAJA::cuda_block_y_direct_unchecked>;
using cuda_teams_x = RAJA::LoopPolicy<RAJA::cuda_block_x_direct_unchecked>;

using cuda_threads_y = RAJA::LoopPolicy<RAJA::cuda_thread_y_direct>;
using cuda_threads_x = RAJA::LoopPolicy<RAJA::cuda_thread_x_direct>;
Expand All @@ -285,12 +289,14 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))

RAJA::launch<cuda_launch_policy>(
RAJA::LaunchParams(RAJA::Teams(n_blocks_c, n_blocks_r),
RAJA::Threads(c_block_sz, r_block_sz)),
RAJA::Threads(c_block_sz, r_block_sz)),
[=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx) {

RAJA::tile<cuda_teams_y>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r), [&] (RAJA::TypedRangeSegment<int> const &row_tile) {
RAJA::tile<cuda_teams_y>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r),
[&] (RAJA::TypedRangeSegment<int> const &row_tile) {

RAJA::tile<cuda_teams_x>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c), [&] (RAJA::TypedRangeSegment<int> const &col_tile) {
RAJA::tile<cuda_teams_x>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c),
[&] (RAJA::TypedRangeSegment<int> const &col_tile) {

RAJA_TEAM_SHARED double Tile_Array[TILE_DIM][TILE_DIM];

Expand All @@ -302,16 +308,18 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<cuda_threads_x>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<cuda_threads_y>(ctx, row_tile, [&] (int row, int ty) {
ctx.teamSync();

Atview(col, row) = Tile_Array[ty][tx];
RAJA::loop_icount<cuda_threads_y>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<cuda_threads_x>(ctx, row_tile, [&] (int col, int tx) {

});
});
Atview(row, col) = Tile_Array[tx][ty];

});
});

});
});
});
});

});

Expand Down Expand Up @@ -346,8 +354,8 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
const int n_blocks_c = RAJA_DIVIDE_CEILING_INT(N_c, c_block_sz);
const int n_blocks_r = RAJA_DIVIDE_CEILING_INT(N_r, r_block_sz);

using hip_teams_y = RAJA::LoopPolicy<RAJA::hip_block_y_direct>;
using hip_teams_x = RAJA::LoopPolicy<RAJA::hip_block_x_direct>;
using hip_teams_y = RAJA::LoopPolicy<RAJA::hip_block_y_direct_unchecked>;
using hip_teams_x = RAJA::LoopPolicy<RAJA::hip_block_x_direct_unchecked>;

using hip_threads_y = RAJA::LoopPolicy<RAJA::hip_thread_y_direct>;
using hip_threads_x = RAJA::LoopPolicy<RAJA::hip_thread_x_direct>;
Expand All @@ -357,12 +365,14 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))

RAJA::launch<hip_launch_policy>
(RAJA::LaunchParams(RAJA::Teams(n_blocks_c, n_blocks_r),
RAJA::Threads(c_block_sz, r_block_sz)),
RAJA::Threads(c_block_sz, r_block_sz)),
[=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx) {

RAJA::tile<hip_teams_y>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r), [&] (RAJA::TypedRangeSegment<int> const &row_tile) {
RAJA::tile<hip_teams_y>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_r),
[&] (RAJA::TypedRangeSegment<int> const &row_tile) {

RAJA::tile<hip_teams_x>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c), [&] (RAJA::TypedRangeSegment<int> const &col_tile) {
RAJA::tile<hip_teams_x>(ctx, TILE_DIM, RAJA::TypedRangeSegment<int>(0, N_c),
[&] (RAJA::TypedRangeSegment<int> const &col_tile) {

RAJA_TEAM_SHARED double Tile_Array[TILE_DIM][TILE_DIM];

Expand All @@ -374,20 +384,23 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
});
});

RAJA::loop_icount<hip_threads_x>(ctx, col_tile, [&] (int col, int tx) {
RAJA::loop_icount<hip_threads_y>(ctx, row_tile, [&] (int row, int ty) {
ctx.teamSync();

RAJA::loop_icount<hip_threads_y>(ctx, col_tile, [&] (int row, int ty) {
RAJA::loop_icount<hip_threads_x>(ctx, row_tile, [&] (int col, int tx) {

d_Atview(col, row) = Tile_Array[ty][tx];
d_Atview(row, col) = Tile_Array[tx][ty];
Copy link
Member

@artv3 artv3 Sep 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@MrBurmark , I switched it around so its clear that the x and y threads have been transposed in shared memory. I'm not too sure how to express that in Kernel.

Copy link
Member Author

@MrBurmark MrBurmark Sep 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Would it be more clear to call row and col here rowt and colt?

Copy link
Member

@artv3 artv3 Sep 22, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

good idea!

Copy link
Member

@rhornung67 rhornung67 Dec 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

using row_t and col_t was the source of the compilation error. I switched it back so code compiles. Please feel free to undo my change and fix the compile differently.


});
});
});
});

});
});
});
});

});

CAMP_HIP_API_INVOKE_AND_CHECK(hipMemcpy, At, d_At, N_r * N_c * sizeof(int), hipMemcpyDeviceToHost);
CAMP_HIP_API_INVOKE_AND_CHECK(hipDeviceSynchronize);
checkResult<int>(Atview, N_c, N_r);
// printResult<int>(Atview, N_c, N_r);
#endif
Expand Down
Loading