|
Size: 2104
Comment:
|
Size: 2326
Comment:
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 3: | Line 3: |
| --- NOTE - Recent changes to CUDA broke the initial CUDA support we had in place. We are now (December 2010) revamping the CUDA support to take advantage of recently added features and make things work again. This process is likely to take a month or two. --- |
--- Note : We now have rewritten CUDA support to work with current CUDA versions. However, for most tasks, the current speedup is not worth the effort of using it. However, if you are doing single particle tomography, the speedup can be quite substantial. --- Note : We do not distribute CUDA binaries due to many many problems with cross-platform compilation, versioning, etc. This may change in future, but for now to experiment with CUDA, you will need to compile from source. |
GPGPU Computing
--- Note : We now have rewritten CUDA support to work with current CUDA versions. However, for most tasks, the current speedup is not worth the effort of using it. However, if you are doing single particle tomography, the speedup can be quite substantial.
--- Note : We do not distribute CUDA binaries due to many many problems with cross-platform compilation, versioning, etc. This may change in future, but for now to experiment with CUDA, you will need to compile from source.
This technique makes use of the GPU (graphics processing unit) common in most modern PC's, to dramatically accelerate many image processing algorithms. At present (summer 2010) we have implemented GPGPU support for a significant number of algorithms in the EMAN2 core (~20-30) using Nvidia's CUDA infrastructure. Implementing GPU support in a massive image processing library like EMAN2 is a major undertaking, and optimizing the entire image processing pipeline requires further efforts. At the present time, only a handful of these algorithms have been fully optimized. While you can build EMAN2 with CUDA support now, overall throughput on something like 3-D refinement is not significantly faster than just using multiple cores on the CPU. That is, if you have a typical quad-core CPU machine and a typical Nvidia graphics card, at the moment, you might as well just use your 4 cores with the --parallel option. We expect this situation to change over coming months...
If you wish to experiment with CUDA support in EMAN2, there are several steps:
- you will need to be able to build EMAN2 from source on your local machine
- run e2speedtest.py to establish a baseline for your CPU (1 core)
- you will need to install and test the CUDA drivers and SDK from Nvidia
- enable the CUDA flag using ccmake in EMAN2, and recompile
- run e2speedtest.py again to compare general performance with the GPU (1 GPU processor)
Eventually we may migrate to the more cross-GPU compatible OpenCL, but at the moment, reports are that performance suffers considerably as compared to CUDA.
Also note that as of July 2010, using CUDA on the Mac is quite tricky, since Snow Leopard (10.6) is 64 bit, and EMAN2 and dependencies are normally compiled as 64-bit, but CUDA at present only supports 32-bit.