high performance computing on graphics processing units: hgpu.org

Graphics processors of today are highly efficient, parallel processors, capable of rendering complex scenes consisting of millions of polygons on the screen each and every second. They are highly specialized towards game graphics and similar, polygon based graphics. In the past, however, they have not been very efficient at rendering volumetric data, and especially not with ray tracing. As the demand for a more open and flexible rendering pipeline has grown stronger each year, the graphics units of today has evolved into very flexible and programmable computing machines, and the possibilities of using the graphics card outside the realm of polygon graphics has increased dramatically. The main problem with ray tracing is the complexity of the algorithm, as it requires a plethora of rays to be traced in order to render the final image. This is however also the origin of one of its main strengths, as each ray can be easily processed independently and thus be efficiently parallelized on a concurrent architecture such as a modern graphics processing unit. This report examine the possibilities of implementing an efficient, real time ray tracer for volumetric data sets using such hardware. The focus is on the computing language OpenCL, and the goal is to create a volume renderer that can be used interactively on commodity graphics hardware readily available today. The techniques presented herein will be very generic and highly applicable on different computing architectures as well, such as NVIDIAs CUDA.