Skip to content

Draft: Matrix Transpose Lesson 11 for Kernel #40

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

Closed
rchen20 wants to merge 2 commits into from
Closed

Draft: Matrix Transpose Lesson 11 for Kernel #40

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f76c892
Preliminary kernel transpose for lesson 11.
rchen20 Jun 30, 2025
7198498
Use global policies.
rchen20 Jul 3, 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
50 changes: 50 additions & 0 deletions Intro_Tutorial/lessons/11_raja_device_kernel/11_raja_transpose_kernel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,50 @@
#include <iostream>

#include "RAJA/RAJA.hpp"
#include "umpire/Umpire.hpp"
#include "umpire/strategy/QuickPool.hpp"

int main()
{
constexpr int N{10000};
constexpr int M{7000};
double* a{nullptr};
double* a_t{nullptr};

auto& rm = umpire::ResourceManager::getInstance();

auto allocator = rm.getAllocator("UM");
auto pool = rm.makeAllocator<umpire::strategy::QuickPool>("POOL", allocator);

a = static_cast<double *>(pool.allocate(N*M*sizeof(double)));
a_t = static_cast<double *>(pool.allocate(N*M*sizeof(double)));

constexpr int DIM = 2;

RAJA::View<double, RAJA::Layout<DIM>> A(a, N, M);
RAJA::View<double, RAJA::Layout<DIM>> A_t(a_t, M, N);

RAJA::TypedRangeSegment<int> row_range(0, N);
RAJA::TypedRangeSegment<int> col_range(0, M);
Copy link
Member

@artv3 artv3 Jul 1, 2025

Choose a reason for hiding this comment

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

Where is the TODO part of the exercise?


using EXEC_POL =
RAJA::KernelPolicy<
RAJA::statement::CudaKernel<
RAJA::statement::For<1, RAJA::cuda_thread_y_loop,
RAJA::statement::For<0, RAJA::cuda_thread_x_loop,
RAJA::statement::Lambda<0>
>
>
>
>;

RAJA::kernel<EXEC_POL>(RAJA::make_tuple(col_range, row_range),
[=] RAJA_DEVICE (int col, int row) {
A_t(col, row) = A(row, col);
});

pool.deallocate(a);
pool.deallocate(a_t);

return 0;
}
4 changes: 4 additions & 0 deletions Intro_Tutorial/lessons/11_raja_device_kernel/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,4 +3,8 @@ if (ENABLE_CUDA)
NAME 11_raja_device_kernel
SOURCES 11_raja_device_kernel.cpp
DEPENDS_ON RAJA umpire cuda)
blt_add_executable(
NAME 11_raja_transpose_kernel
SOURCES 11_raja_transpose_kernel.cpp
DEPENDS_ON RAJA umpire cuda)
endif()
50 changes: 50 additions & 0 deletions Intro_Tutorial/lessons/11_raja_device_kernel/solution/11_raja_transpose_kernel_solution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,50 @@
#include <iostream>

#include "RAJA/RAJA.hpp"
#include "umpire/Umpire.hpp"
#include "umpire/strategy/QuickPool.hpp"

int main()
{
constexpr int N{10000};
constexpr int M{7000};
double* a{nullptr};
double* a_t{nullptr};

auto& rm = umpire::ResourceManager::getInstance();

auto allocator = rm.getAllocator("UM");
auto pool = rm.makeAllocator<umpire::strategy::QuickPool>("POOL", allocator);

a = static_cast<double *>(pool.allocate(N*M*sizeof(double)));
a_t = static_cast<double *>(pool.allocate(N*M*sizeof(double)));

constexpr int DIM = 2;

RAJA::View<double, RAJA::Layout<DIM>> A(a, N, M);
RAJA::View<double, RAJA::Layout<DIM>> A_t(a_t, M, N);

RAJA::TypedRangeSegment<int> row_range(0, N);
RAJA::TypedRangeSegment<int> col_range(0, M);

using EXEC_POL =
RAJA::KernelPolicy<
RAJA::statement::CudaKernel<
RAJA::statement::For<1, RAJA::cuda_thread_y_loop,
RAJA::statement::For<0, RAJA::cuda_thread_x_loop,
Copy link
Member

@artv3 artv3 Jul 1, 2025

Choose a reason for hiding this comment

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

we should coordinate with @johnbowen42 to make sure we use the same policies

Copy link
Member

@rhornung67 rhornung67 Jul 1, 2025

Choose a reason for hiding this comment

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

I also think that it would be useful to explain the basic differences between 'loop' and 'direct' policies. We've had issues in the past when users apply loop policies following our basic examples and see performance that is less than expected. I would recommend to not use loop policies for an introduction.

Copy link
Member

@artv3 artv3 Jul 1, 2025

Choose a reason for hiding this comment

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

I think we should use the global policies here, and part of slides could be an overview of different options. I was planning include that in my slide deck.

Copy link
Member

@rhornung67 rhornung67 Jul 1, 2025

Choose a reason for hiding this comment

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

I like that idea.

RAJA::statement::Lambda<0>
>
>
>
>;

RAJA::kernel<EXEC_POL>(RAJA::make_tuple(col_range, row_range),
[=] RAJA_DEVICE (int col, int row) {
A_t(col, row) = A(row, col);
});

pool.deallocate(a);
pool.deallocate(a_t);

return 0;
}