Matrix Multiplication with OpenMP Offloading Quiz¶

1. Which OpenMP directive should be used to offload matrix multiplication to the GPU while collapsing both outer loops (row and col)?

• A. #pragma omp target teams collapse(2)
• B. #pragma omp target parallel for collapse(2)
• C. #pragma omp target loop collapse(2)
• D. #pragma omp target parallel collapse(2)

2. In matrix multiplication with #pragma omp target teams distribute parallel for, what is the main advantage of using teams?

• A. It specifies that the code should run on the CPU.
• B. It creates multiple teams on the GPU, each responsible for a portion of the work, improving workload distribution.
• C. It forces all computations to be sequential.
• D. It prevents data from being transferred to the GPU.

3. True or False: The collapse(2) clause is used to merge two loops (e.g., row and col loops) so they can be executed in parallel as a single loop.

4. Which clause would you use to ensure each thread has a private copy of row, col, and i within a parallelized matrix multiplication loop?

• A. collapse
• B. num_teams
• C. private(row, col, i)
• D. map(tofrom: row, col, i)

5. In the option #pragma omp target teams distribute parallel for num_teams(5) collapse(2), what does num_teams(5) do?

• A. Specifies that each team should have 5 threads.
• B. Creates exactly 5 teams to distribute the work on the GPU.
• C. Allocates 5 memory spaces on the GPU.
• D. Limits the number of iterations each thread can execute to 5.

6. In the C/C++ example, what does the map(to: a[0:n*n], b[0:n*n]) map(from: c[0:n*n]) clause do?

• A. Copies a and b from the device to the host and c from the host to the device.
• B. Allocates memory for a, b, and c on the device but does not initialize them.
• C. Copies a and b to the device and brings c back from the device to the host after computation.
• D. Allocates memory for c on the host but initializes a and b on the device.

7. Which of the following is a benefit of using teams distribute parallel for over parallel for alone in GPU offloading?

• A. It prevents memory access conflicts by limiting memory access.
• B. It supports multi-level parallelism by creating teams and allowing parallel execution within each team.
• C. It enables execution on the CPU instead of the GPU.
• D. It disables the need for memory mapping.

8. What is the purpose of collapse(2) in #pragma omp target teams distribute parallel for collapse(2) when performing matrix multiplication?

• A. It divides the workload into two halves.
• B. It forces two copies of each variable to be created.
• C. It merges the row and col loops to treat them as a single loop, improving parallelism.
• D. It prevents any data races by creating private copies of variables.

9. True or False: #pragma omp target teams distribute parallel for can create both teams and threads to handle a distributed workload on the GPU.

10. Which construct would you use if you want to control the number of teams created on the GPU during matrix multiplication?

• A. collapse(2)
• B. private
• C. num_teams(5)
• D. map(to: a[0:n*n], b[0:n*n])