high performance computing on graphics processing units: hgpu.org

Modern graphics Application Programming Interfaces (APIs) provide first-class support for ray tracing. Hardware vendors implement drivers for the graphics API including a black-box compiler. The black-box compiler creates architecture-specific binaries that leverage ray-tracing hardware acceleration. Ray-tracing support in modern APIs allows all geometry and shaders to be specified for a single execution. Thus, ray tracing is more complex and difficult to reason about than rasterization, a traditional rendering method. Ray-tracing developers must contend with the unknowns of an inscrutable GPU binary and a monolithic execution model. The increase in complexity from rasterization to ray tracing has not been accompanied by commensurate tooling. This thesis first presents Vulkan Vision (V-Vision). V-Vision is a framework for developing instrumentation passes for shaders in the Vulkan graphics API. V-Vision handles the commonalities of generating, analyzing, and presenting instrumentation data. Specifically, V-Vision provides instrumentation primitives to capture a complete inter-shader and intra-shader ray-tracing execution trace. Instrumentation utilities implemented using V-Vision are operating-system, vendor, and architecture agnostic. V-Vision does not require source-code modification or recompilation. V-Vision’s out-of-the-box instrumentation utilities demonstrate the ability to gather fine-grained execution data. Moreover, VVision’s utilities are capable of measuring microarchitectural effects, such as independent thread scheduling. The execution data enables limit studies at hardware, compiler, and application levels. V-Vision’s annotated shader and heatmap representations enable productive debugging and profiling. V-Vision has been accepted into the MindInsight tool family. Next, this thesis presents RayScope. RayScope automatically captures application-agnostic ray-tracing execution data and geometry data from Vulkan applications. RayScope provides an interactive visualizer, populated using the ray tracing and geometry data. Therefore, RayScope can be understood as a set of tools that enables understanding, debugging, profiling, and designing through visualization of application execution data. RayScope implements an instrumentation pass and analysis using V-Vision but also implements Vulkan-API call instrumentation. RayScope’s outputs are human-readable to encourage integration with other visualization and debugging tools. RayScope assisted in identifying longstanding bugs in Vulkan ray-tracing applications. RayScope further assisted in uncovering poorly defined minimum collision distances causing wasted computations in multiple ray-tracing applications. RayScope also helped identify geometry construction problems causing visual artifacts and wasteful computation in the well-known model Sponza. Finally, RayScope automatically identified a misconfiguration of Vulkan geometry flags and recommended a solution for one ray-tracing application. Applying the recommendation results in a reduction of 96.8% of any-hit shader executions. The level of information provided to the developer has a large impact on the quality of the application that they develop. Changes motivated by the information provided by V-Vision and RayScope are often minimal but have tangible implications for performance and correctness. The effectiveness of V-Vision and RayScope indicates that tooling, and the knowledge it provides, was lacking for real-time hardware-accelerated ray tracing in Vulkan. The work presented in this thesis improves the tooling landscape by releasing V-Vision and RayScope as open-source projects, and improves the body of knowledge by sharing common pitfalls in real-time hardware-acceleration ray tracing.