This project ports parts of the libacc library to CUDA. While construction of the acceleration structures is out of scope, the library provides constructs to upload and use the libacc data structures within CUDA. Data structures to handle kernel input and output are provided additionally.
Add relevant headers and .cu files to your project and compile with
--relocatable-device-code=true
The location of data structures is encoded as template argument. Uploading data to the GPU is done via the assignment operator:
Tree<HOST> htree = ...
Tree<DEVICE> dtree = htree;
Since it is preferable to access some data structures as textures it is necessary to bind those textures before executing the kernel.
tracing::bind_textures(dtree.cdata());
Per default, parts of the traversal stack are cached in shared memory. Be sure
to reduce the block size or short stack size
(SSTACK_SIZE in
tracing.h)
if you want to use shared memory within your kernel.
For very large trees one might have to increase the overall stack size
(GSTACK_SIZE + SSTACK_SIZE in
tracing.h)
NVCC 7.x - C++11 support
The software is licensed under the BSD 3-Clause license, for more details see the LICENSE.txt file.