+- Cuberite Forum (https://forum.cuberite.org)
+-- Forum: Cuberite (https://forum.cuberite.org/forum-4.html)
+--- Forum: Development (https://forum.cuberite.org/forum-13.html)
+--- Thread: GPU code (/thread-1480.html)
RE: GPU code - worktycho - 06-18-2014
Two things, one did you install the AMD SDK? Two is the value of OPENCL_INCLUDE_DIR NOT-FOUND? if so you need to set it.
RE: GPU code - NiLSPACE - 06-18-2014
Do I need the AMD SDK if I have an NVidia?
RE: GPU code - worktycho - 06-18-2014
If you've got an NVIDA card you need the cuda toolkit but I haven't tried building with that.
RE: GPU code - worktycho - 06-18-2014
Now doing Biomal Without the linear interpolation at 1ms a chunk.
RE: GPU code - bearbin - 06-18-2014
Cool Is that better than before?
RE: GPU code - worktycho - 06-18-2014
Haven't checked. One thing is that it is only using 65% of the GPUs capability. If we implemented batching it would double throughput.
What I'm trying to do is implement OpenCl for all the height Generators where it makes sense so I can change the interface when using openCl.
RE: GPU code - worktycho - 06-19-2014
Did test with the mountains height Gen. Using OpenCL is 75% slower then using the CPU generator. I suspect that this is due to the huge latency at start-up, about 450 microseconds ~ 500,000 CPU cycles. If I can move more things on to the GPU and make the generator asynchronous I should be able to remove that from the throughput though.
RE: GPU code - NiLSPACE - 06-19-2014
But the HeightGen isn't really the bottleneck is it?
RE: GPU code - LO1ZB - 06-19-2014
Cant build your branch with thr NV cuda sdk, because it don't got a "cl.hpp" included.
Even the CUDA SDK 3.1 (with open cl ? 1.1 ? support)
RE: GPU code - worktycho - 06-19-2014
Did you set the OPENCL_INCLUDE_DIR and OPENCL_LIBRARY_PATH vars? It may be cmake can't find cuda.
@STR_Warrior, no but if part of the pipeline is on the CPU and part GPU it will cause latencies so I need to transfer most parts of the pipeline.