Visual computing tasks such as computational imaging, image/video understanding, and real-time 3D graphics are key responsibilities of modern computer systems ranging from sensor-rich smart phones, autonomous robots, and large datacenters. These workloads demand exceptional system efficiency and this course examines the key ideas, techniques, and challenges associated with the design of parallel, heterogeneous systems that accelerate visual computing applications. This course is intended for systems students interested in architecting efficient graphics, image processing, and computer vision platforms (both new hardware architectures and domain-optimized programming frameworks for these platforms) and for graphics, vision, and machine learning students that wish to understand throughput computing principles to design new algorithms that map efficiently to these machines.
Review superscalar, multi-core, SIMD, and multi-threaded CPU/GPU designs, + understanding latency and bandwidth
Algorithms for taking raw sensor pixels to an RGB image: demosaicing, sharpening, correcting lens aberrations, multi-shot alignment/merging, image filtering
Multi-scale processing with Gaussian and Laplacian pyramids, HDR (local tone mapping), portrait mode
Digital Camera Processing Pipeline (Part III)
Finishing and reviewing concepts from prior lectures (see prior slides)
Balancing locality, parallelism, and work, fusion and tiling, design of the Halide domain-specific language, automatically scheduling image processing pipelines
Popular DNN trunks and topologies, efficient topologies like MobileNet, where the compute lies in modern networks, DNN pruning, neural architecture search
GPUs, TPUs, special instructions for DNN evaluation (and their efficiency vs custom ASIC), choice of precision in arithmetic, recent ISCA/MICRO papers on DNN acceleration, flexibility vs efficiency trade-offs
Footprint challenges of training, model vs. data parallelism, asynchronous vs. synchronous training debate, parameter server designs, key systems optimizations for parallel training
If the most important step of ML is acquiring labeled data for training and validation, why don't we have better systems for it?
Systems for specifying models at a higher level of abstraction than DNN architecture graphs (Overton, Ludwig). Goal: removing the need for a low-level ML engineer.
Exploiting temporal coherence in video, specialization to camera viewpoint, scene appearance, and task.
H.264 video representation/encoding, parallel encoding, motivations for ASIC acceleration, emerging opportunities for compression when machines, not humans, will observe most images
System design issues for building a video conferencing system: reducing latency, bandwidth, etc.
Motivations for ray tracing vs. rasterization. Ray coherence during traversal, parallelizing BVH build, multi-level BVHs.
Modern hardware acceleration (RTX GPUs), conversing noisy images to clean images using neural techniques.
Rendering realistic scenes with many lights
How might systems for rendering and simulating virtual worlds be architected differently to support the needs of training machines instead of video games? (a.k.a. rendering for machine eyes, not human eyes)
Key rendering issues for VR and AR devices
Guest Lecture (Chris Wyman, NVIDIA)
algorithmic innovation (and future challenges) for real-time ray tracing
student project presentations
In addition to expectation that all students attend and participate in discussions in live lecture, there will be two short programming assignments and a self-selected term project.
|Apr 13||Burst Mode HDR Camera RAW Processing|
|April 27||Optimizing a Conv Layer in Halide (Making Students Appreciate cuBLAS)|
|Jun 4||Term Project|