CMPS 297S/396AA: GPU COMPUTING
ASSIGNMENT 4
In this assignment, you will implement a convolution kernel that uses shared memory tiling. Your kernel
is expected to work for any set of input dimensionsso make sure to handle boundary conditions correctly.
You must use shared memory tiling to receive credit for this assignment.
Instructions
1. Place the files provided with this assignment in a single directory. The files are:
main.cu: contains setup code, sequential code, and naïve kernel code without tiling
kernel.cu: where you will implement your code (you should only modify this file)
common.h: for shared declarations across main.cu and kernel.cu
timer.h: to assist with timing
Makefile: used for compilation
2. Edit kernel.cu where TODO is indicated to implement the following:
Copy the convolution mask to constant memory
Configure and call the kernel
o Be careful: you need enough threads per block to load an entire input tile, but
enough blocks in the grid to process every output tile
Perform the computation in the kernel:
o Find the output element that each thread is computing
Be careful: blockDim corresponds to the input tile size
o Find the input element that each thread in the block is loading
o Use all the threads in the block to load the input tile to shared memory
o Use a subset of the threads in the block to compute the output tile
3. Compile your code by running: make
4. Test your code by running: ./convolution
If you are using the HPC cluster, do not forget to use the submission system. Do not run
on the head node!
For testing on different input sizes, you can provide your own values for the input
dimensions as follows: ./convolution <height> <width>
Submission
Submit your modified kernel.cu file via Moodle by the due date. Do not submit any other files or
compressed folders.
