This project was the final project for The University of Washington's Professional Master's Program GPU Programming course.
GPU Parallel Processing allows for the computational speed up of many problems, one problem of interest to me is parallel processing of image encoding and decoding. I think this is an interesting application space as data transfer of images can take a lot of time, and if the images need to be compressed on the way out and decompressed on the way in, having this process be an efficient one is important.
Because of this need for fast data transfer of images, we need a way to parallelize the compression. For this project, I chose JPEG compression and decompression (JPEG codec) for this application. JPEG codec is well suited for GPU computing for a few reasons. In Figure 1 below we see the multiple compression stages of a JPEG Encoder, and in Figure 2 we can note the steps of the image reconstruction via the JPEG Decoder. Because the mathematical process of the DCT, quantization and zig zag algorithms are repeated for all 8x8 blocks in the image, this brings about the opportunity for parallel processing.
The serial CPU JPEG codec was my first task and from there I utilized the mathematical and algorithmic functions to convert the code to my parallel GPU version. Creating the GPU kernels from the CPU functions was harder than expected for certain kernels as access patterns were tricky. My parallel implementation has two versions (variations to dct and idct kernels), one where I compute the cosine math in the kernel, and one where I precompute and then read from global memory instead, an analysis of why and how it performed can be found in the report.
The dat/ folder contains the input images used, and the CPU/GPU output versions. As shown below, the CPU/GPU versions are identical. Check out page 10 of the report to see various timing results and the speed up of the DCT and iDCT kernels when using a precomputed method to lower the register count and increase thread count!
Figure 3: CPU Version
Figure 4: GPU Version
The codec.cu file contains the functions to run the serial and parallel codecs, and different units tests that were used to ensure proper functionality at each step, as well as my main function. Inside utils/ you can find the supporting functions for the JPEG serial and parallel codecs.