Efficient GPU-Accelerated Gaussian Filter Implementation for Image Processing

ECE 4822: ECE 4822: Final Project Final Project Luke Dewees

Part 0: Libraries Used Part 0: Libraries Used Numba Cuda kernels + grid striding For all kernels, I used the @cuda.jit decorator Considered @guvectorize for Gaussian Blur CuPy GPU-optimized NumPy features Used CuPy arrays to store image data on GPU Others tested: Dask -> 2D FFT function

Part 1: Gaussian Filter Part 1: Gaussian Filter- - Preprocessing Preprocessing Original plan: process multiple images with one kernel call Why won t this work? Images in database are very large, even when compressed (~0.38 GB each) Expanded into a NumPy array, I found that these can exceed 17 GB Modified plan: process as many windows as possible with one GPU kernel Before any preprocessing, I first sent the Gaussian Filter kernel to the GPU After this- for each image: 1. Extract 5000 windows and convert to grayscale with NumPy 2. Convert windows to CuPy arrays (load onto GPU) 3. Launch filter Kernel

Part 1: Gaussian Filter Part 1: Gaussian Filter- - GPU Kernel GPU Kernel Stored Gaussian filter coefficients and current windows batch in shared memory Grid striding X and Y track the current location in the window, Z tracks the current window being processed Used padding for the edges of each window Each thread computes one value of the blurred window

Part 1: Gaussian Filter Part 1: Gaussian Filter- - Timing Results Timing Results Timing results weren t great Very large amount of time spent on loading the image, preprocessing, and transferring data to GPU Time for a single image varied by minutes Significantly less time spent on calculations (0.00037x less) Runtimes varied from ~5 minutes to 9 minutes for a single image Example Linux Time Output for 1 Image

Part 1: What I Would Change Part 1: What I Would Change 1. Optimize preprocessing and GPU loading Streaming data? Bigger chunks of data 2. Remove redundant filter calculations 3. Flatten image chunks before applying filter ~7x speedup without GPU Would this multiplier increase or decrease with GPU usage? 4. Implement Numba s @guvectorize

Part 2: 2D FFT Part 2: 2D FFT Using CuPy to store image data on the GPU made this really simple CuPy has a built-in 2D FFT function This was very fast Dask s 2D FFT performed slightly better computation-wise in my tests However, this requires making a Dask array version of the blurred image data

Part 3: Histogram Computation Part 3: Histogram Computation Utilized grid striding FFT data in shared memory Each thread processes a chunk of the data and creates its own histogram Threads are synced and each chunk s histograms are combined

Timing Results Timing Results Code with histogram is still running For one image, timing results from test runs were not great Computation speeds are decent (<0.7 seconds for Gaussian Blur, FFT was very fast) Fastest time I have received so far is 542.784 seconds for Gaussian Blur and the 2D FFT

Thanks Thanks CREDITS: This presentation template was created by Slidesgo, and includes icons by Slidesgo Flaticon Flaticon, infographics & images by Freepik and contents by Swetha Tandri Freepik Please keep this slide for attribution