Semi-Procedural Textures using Point Process Texture Basis Functions

Mixing Graphics and Compute

We use OpenGL compute shaders for all algorithms and traditional graphics shaders for rendering.

PPTBF Synthesis Timing for a Mega-Kernel Approach

We use one unique megakernel based on OpenGL compute shader. The PPTBF is stored in a 1D texture of 1-channel floating point (R32F). This approach requires no additional GPU memory allocation. But requires a lot of registers, giving rise to register pressure, latency... Complex code => branching...

For OpenGL compute kernels, we use blocks of 8x8 threads. Memory is limited by graphics card. For each output image size, we test without and with repulsive forces (5 iterations in Lloyd algorithm). These are the results for Voronoi window and 1 to 5 Gabor kernels in Feature function.

We show speed according to image size. Target device: NVIDIA GeForce 1060 GTX 6Go.

PPTBF Synthesis Timing for a Wavefront Approach

Using a wavefront approach split algorithm into multi-passes. [Laine et al. 2013] It follows node-based tools way of thinking. User can then choose and customize components. The deformation component D could be fBm (fractional brownian motion); turbulence, etc... The point process component PP could be normal or add forces such as repulsions. The wavefront approach requires to store results at each step. We rely on imageStore to write into textures, and bindless texture extensions to be able to access a lot of textures/images in read or write mode. Bindless textures requires an GPU device at least Kepler (but some old Kepler are not working).

For OpenGL compute kernels, we use blocks of 8x8 threads. Memory is limited by graphics card. For each output image size, we test without and with repulsive forces (5 iterations in Lloyd algorithm). These are the results for Voronoi window and 1 to 5 Gabor kernels in Feature function.

We show speed according to image size. Target device: NVIDIA GeForce 1060 GTX 6Go.

PPTBF Rendering

Rendering relies on a classical fullscreen triangle with a fragment shader reading the PPTBF texture.

For OpenGL compute kernels, we use blocks of 8x8 threads. Memory is limited by graphics card. For each output image size, we test without and with repulsive forces (5 iterations in Lloyd algorithm). These are the results for Voronoi window and 1 to 5 Gabor kernels in Feature function.

We show speed according to image size. Target device: NVIDIA GeForce 1060 GTX 6Go.