66
edits
Changes
→Conclusion
Solution:
----
One way to address low compute utilization is to attempt to increase occupancy of each SM. According to Cuda's occupancy calculator the machine we were using for testing had a compute capability of 6.1. This means that each SM had 32 resident blocks and 2048 resident threads. To achieve maximum occupancy you would have 2048/32 = 64 threads/ block. To determine an appropriate grid size we would divide the total number of pixels by the 64 threads/block. This allows us to use dynamic grid sizing depending on the size of the image passed in.
<br><br>
Problem:
----
We considered shared memory when optimizing our kernel. When attempting to implement shared memory we realized that it would be a difficult task to complete because every pixel in a block needs access to a different range of pixels for averaging. One major problem was that neighborhood pixels may fall out of range of the block. We also attempted to store the entire image in shared memory but this solution is not salable scalable to larger image sizes as shared memory is a limited resource.
<br><br>
Below you'll see that our optimizations although show slight improvements sometimes, it was not effective. We are currently still looking for a way to implement shared memory which will surely improve efficiency and execution time.
With further optimization, we managed to slightly improve the execution time of the blur effect.
<br><br>
Below are the final results of all the test runs as well as the corresponding graph.
==== Results ====
[[File:boxFilterFinalTable.png | 500px]]
