1 Objective
The boundary condition can be handled by filling “identity value (0 for sum)” into the shared memory of the last block when the length is not a multiple of the work group size.
2 Instructions
Edit the skeleton code to perform the following:
• Allocate device memory
• Copy host memory to device
• Initialize thread block and kernel grid dimensions
• Invoke CUDA kernel
• Copy results from device to host
• Deallocate device memory
• Implement the work efficient scan routine
• Use shared memory to reduce the number of global memory accesses, handle the boundary conditions when loading input elements into the shared memory
• Write the CUDA kernel
Compile the template with the provided Makefile. The executable generated as a result of compilation can be run using the following code:
./ListScan Template -e <expected.raw> -i <input.raw> -o <output.raw>
-t vector
where <expected.raw> is the expected output, <input.raw> is the input dataset, and <output.raw>
is an optional path to store the results.
README.md has details on how to build libgputk, template.cpp and the dataset generator.
3 What to Turn in
Submit a report that includes the following:
1. How many global memory reads are being performed by your kernel?
2. How many global memory writes are being performed by your kernel?
3. How many times does a single thread block synchronize to reduce its portion of the array to asingle value?
4. Suppose that you want to scan using a binary operator that is not commutative. Can you use aparallel scan for that?
5. Is it possible to get different results from running the serial version and parallel version of scan?Explain.
6. Your version of template.cpp.
7. The result as a table/graph of kernel execution times for different input data, with the systeminformation where you performed your evaluation. Run your implementation with the input generated by the provided dataset generator. For time measurement, use gpuTKTime start and gpuTKTime stop functions (You can find details in libgputk/README.md).
