93
edits
Changes
→Conclusion
==== Conclusion ====
We decided to go with image profiling. It is a pretty simple parallelization since the transformation functions are matrix transformations which don't care about which element is processed first.
There are some possible issues with working with the simple-lzw-compression-algorithm and CUDA. You cannot use the C++ string type in a kernel because CUDA does not include a device version of the C++ String library that would be able run on the GPU. Even if it was possible to use string in a kernel, it's not something you would want to do because string handles memory dynamically, which would be likely to be slow.
Char explanation (replace tmr) - https://stackoverflow.com/questions/26993351/is-there-a-penalty-to-using-char-variables-in-cuda-kernels?utm_medium=organic&utm_source=google_rich_qa&utm_campaign=google_rich_qa
Essentially, using chars on the gpu would require we use 8-bit arithmetic and need to convert from 32-bit arithmetic for operations. On top of that, the compress function refers to a map of strings/int pairs to shrink the size of the file. Even if we did manage the character operations, it would need to somehow use the string map to get the corresponding integer which could lead to being unable to use device memory for optimization.
=== Assignment 2 ===
=== Assignment 3 ===
