Widmer, Sven (2018)
Foundations and Methods for GPU based Image Synthesis.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
Effects such as global illumination, caustics, defocus and motion blur are an integral part of generating images that are perceived as realistic pictures and cannot be distinguished from photographs. In general, two different approaches exist to render images: ray tracing and rasterization. Ray tracing is a widely used technique for production quality rendering of images. The image quality and physical correctness are more important than the time needed for rendering. Generating these effects is a very compute and memory intensive process and can take minutes to hours for a single camera shot. Rasterization on the other hand is used to render images if real-time constraints have to be met (e.g. computer games). Often specialized algorithms are used to approximate these complex effects to achieve plausible results while sacrificing image quality for performance.
This thesis is split into two parts. In the first part we look at algorithms and load-balancing schemes for general purpose computing on graphics processing units (GPUs). Most of the ray tracing related algorithms (e.g. KD-tree construction or bidirectional path tracing) have unpredictable memory requirements. Dynamic memory allocation on GPUs suffers from global synchronization required to keep the state of current allocations. We present a method to reduce this overhead on massively parallel hardware architectures. In particular, we merge small parallel allocation requests from different threads that can occur while exploiting SIMD style parallelism. We speed-up the dynamic allocation using a set of constraints that can be applied to a large class of parallel algorithms. To achieve the image quality needed for feature films GPU-cluster are often used to cope with the amount of computation needed. We present a framework that employs a dynamic load balancing approach and applies fair scheduling to minimize the average execution time of spawned computational tasks. The load balancing capabilities are shown by handling irregular workloads: a bidirectional path tracer allowing renderings of complex effects at near interactive frame rates.
In the second part of the thesis we try to reduce the image quality gap between production and real-time rendering. Therefore, an adaptive acceleration structure for screen-space ray tracing is presented that represents the scene geometry by planar approximations. The benefit is a fast method to skip empty space and compute exact intersection points based on the planar approximation. This technique allows simulating complex phenomena including depth-of-field rendering and ray traced reflections at real-time frame rates. To handle motion blur in combination with transparent objects we present a unified rendering approach that decouples space and time sampling. Thereby, we can achieve interactive frame rates by reusing fragments during the sampling step. The scene geometry that is potentially visible at any point in time for the duration of a frame is rendered in a rasterization step and stored in temporally varying fragments. We perform spatial sampling to determine all temporally varying fragments that intersect with a specific viewing ray at any point in time. Viewing rays can be sampled according to the lens uv-sampling to incorporate depth-of-field. In a final temporal sampling step, we evaluate the pre-determined viewing ray/fragment intersections for one or multiple points in time. This allows incorporating standard shading effects including and resulting in a physically plausible motion and defocus blur for transparent and opaque objects.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2018 | ||||
Autor(en): | Widmer, Sven | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Foundations and Methods for GPU based Image Synthesis | ||||
Sprache: | Englisch | ||||
Referenten: | Goesele, Prof. Dr. Michael ; Dachsbacher, Prof. Dr. Carsten | ||||
Publikationsjahr: | 2018 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 14 Juli 2017 | ||||
URL / URN: | https://tuprints.ulb.tu-darmstadt.de/7516 | ||||
Kurzbeschreibung (Abstract): | Effects such as global illumination, caustics, defocus and motion blur are an integral part of generating images that are perceived as realistic pictures and cannot be distinguished from photographs. In general, two different approaches exist to render images: ray tracing and rasterization. Ray tracing is a widely used technique for production quality rendering of images. The image quality and physical correctness are more important than the time needed for rendering. Generating these effects is a very compute and memory intensive process and can take minutes to hours for a single camera shot. Rasterization on the other hand is used to render images if real-time constraints have to be met (e.g. computer games). Often specialized algorithms are used to approximate these complex effects to achieve plausible results while sacrificing image quality for performance. This thesis is split into two parts. In the first part we look at algorithms and load-balancing schemes for general purpose computing on graphics processing units (GPUs). Most of the ray tracing related algorithms (e.g. KD-tree construction or bidirectional path tracing) have unpredictable memory requirements. Dynamic memory allocation on GPUs suffers from global synchronization required to keep the state of current allocations. We present a method to reduce this overhead on massively parallel hardware architectures. In particular, we merge small parallel allocation requests from different threads that can occur while exploiting SIMD style parallelism. We speed-up the dynamic allocation using a set of constraints that can be applied to a large class of parallel algorithms. To achieve the image quality needed for feature films GPU-cluster are often used to cope with the amount of computation needed. We present a framework that employs a dynamic load balancing approach and applies fair scheduling to minimize the average execution time of spawned computational tasks. The load balancing capabilities are shown by handling irregular workloads: a bidirectional path tracer allowing renderings of complex effects at near interactive frame rates. In the second part of the thesis we try to reduce the image quality gap between production and real-time rendering. Therefore, an adaptive acceleration structure for screen-space ray tracing is presented that represents the scene geometry by planar approximations. The benefit is a fast method to skip empty space and compute exact intersection points based on the planar approximation. This technique allows simulating complex phenomena including depth-of-field rendering and ray traced reflections at real-time frame rates. To handle motion blur in combination with transparent objects we present a unified rendering approach that decouples space and time sampling. Thereby, we can achieve interactive frame rates by reusing fragments during the sampling step. The scene geometry that is potentially visible at any point in time for the duration of a frame is rendered in a rasterization step and stored in temporally varying fragments. We perform spatial sampling to determine all temporally varying fragments that intersect with a specific viewing ray at any point in time. Viewing rays can be sampled according to the lens uv-sampling to incorporate depth-of-field. In a final temporal sampling step, we evaluate the pre-determined viewing ray/fragment intersections for one or multiple points in time. This allows incorporating standard shading effects including and resulting in a physically plausible motion and defocus blur for transparent and opaque objects. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-75164 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik | ||||
Fachbereich(e)/-gebiet(e): | 20 Fachbereich Informatik 20 Fachbereich Informatik > Graphics, Capture and Massively Parallel Computing |
||||
Hinterlegungsdatum: | 16 Sep 2018 19:55 | ||||
Letzte Änderung: | 16 Sep 2018 19:55 | ||||
PPN: | |||||
Referenten: | Goesele, Prof. Dr. Michael ; Dachsbacher, Prof. Dr. Carsten | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 14 Juli 2017 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |