samuel-emrys
commented
2 years ago It looks like this is substantially more complicated than I first anticipated. Most of these are very low level libraries that introduce substantial boilerplate and would require a refactor of existing maths as it appears that only the Eigen::Tensor object is supported in SYCL.
Some resources of interest:
- https://vulkan-tutorial.com/Introduction (focuses mostly on graphics pipelines rather than compute pipelines)
- https://www.vulkan.org/learn
- Accelerating machine learning using vulkan
- Compute shader 101
- https://www.reddit.com/r/vulkan/
- https://sycl.tech/learn/
- https://github.com/ville-k/sycl_starter
- https://www.codeplay.com/portal/blogs/2020/05/19/ray-tracing-in-a-weekend-with-sycl-basic-sphere-tracing.html
- https://www.codeplay.com/portal/blogs/2020/06/19/ray-tracing-in-a-weekend-with-sycl-part-2-pixel-sampling-and-material-tracing.html
- https://github.com/charles-lunarg/vk-bootstrap
- https://www.khronos.org/registry/vulkan/specs/1.2-khr-extensions/html/index.html
- https://shader-playground.timjones.io/
It looks like using SYCL will require the usage of intels DPC++ compiler. It looks like using opencl or vulkan directly will require authoring shaders. Both look like they may require a substantial refactoring of current application architecture.
DPC++ seems to have some conan support, but is a proprietary compiler. Availability on arch linux seems patchy.
Add the ability to offload the raytracing tasks to the GPU with a framework such as opencl, vulkan, vuda or sycl. Selection of this api will require further research.