TITRE DU STAGE  :
Studying and profiling the GPU performances of 3D ray-marchers

FONCTIONS ET TACHES : [ subjet ]
GPU shaders looks like threaded C, but hardware constraints ( SIMD, warps, groups, small caches, slow memory access ) hugely impact performances on unintuitives ways. In the case of ray-marchers, neighbor rays having different fate locally breaks the parallelism, at places where the algorithm is already the slowest (e.g. near silhouettes).
We want to study various simple test case and implementation variants, getting cost heatmaps and timings profilings, to understand how bad some cases can be, interpret what happen, and get some good practice recommandations.

OBJECTIFS ET RESULTATS ATTENDUS : [ delivrables ]
Several OpenGL/GLSL programs on Linux or online shaders in Shadertoy, getting profiling info (timings, maps) from simple configurations (ray-marching of a sphere) using various algorithmic variants (distance field, voxels, octree, octree+grid, or equivalent stubs). Plus possibly some acid-test programs or shaders to bench the GPU.