User:Arash: Difference between revisions
Jump to navigation
Jump to search
Created page with ''''Arash Sadrieh''' is working on developing GPU-based solvers for ASCEND. He is a PhD student at Murdoch University in Western Australia. == Goals == * reinstate bintoken func…' |
No edit summary |
||
| Line 3: | Line 3: | ||
== Goals == | == Goals == | ||
* reinstate bintoken functionality | * Make ASCEND to export models(residuals and jacobian) evaluators to bintokens. | ||
* add gradient calculation support to bintoken stuff | ** reinstate bintoken functionality | ||
* ... | ** add gradient calculation support to bintoken stuff | ||
* Prepare a large model (preferably 100,000+) and a unit test for verifying and benchmarking the NLA solver using this model. | |||
* Develop a CUDA code generator that creates GPU-based bintokens. | |||
* Create a new library in ascend (GPU_manager) which is responsible to manage all the GPU related tasks. Including data transfer between host and GPU, launching bintoken CUDA kernels and parallel calculation of the residuals normal (required in line-search algorithm). | |||
* Fork a new NLA solver from current solver: In the new solver when the solver needs to evaluate a block residual or Jacobian, the call is redirected to GPU_manager. | |||
* Wrapping appropriate functionality in ascend solver interface that decouples GPU manager from the solver. (The interface should provide batch residual (and Jacobian) evaluation for group of relations). | |||
* Benchmark the results and probably switch to other many (or multi) core architectures and languages. | |||
== Progress == | == Progress == | ||
''fill in here'' | ''fill in here'' | ||
Revision as of 06:49, 23 February 2011
Arash Sadrieh is working on developing GPU-based solvers for ASCEND. He is a PhD student at Murdoch University in Western Australia.
Goals
- Make ASCEND to export models(residuals and jacobian) evaluators to bintokens.
- reinstate bintoken functionality
- add gradient calculation support to bintoken stuff
- Prepare a large model (preferably 100,000+) and a unit test for verifying and benchmarking the NLA solver using this model.
- Develop a CUDA code generator that creates GPU-based bintokens.
- Create a new library in ascend (GPU_manager) which is responsible to manage all the GPU related tasks. Including data transfer between host and GPU, launching bintoken CUDA kernels and parallel calculation of the residuals normal (required in line-search algorithm).
- Fork a new NLA solver from current solver: In the new solver when the solver needs to evaluate a block residual or Jacobian, the call is redirected to GPU_manager.
- Wrapping appropriate functionality in ascend solver interface that decouples GPU manager from the solver. (The interface should provide batch residual (and Jacobian) evaluation for group of relations).
- Benchmark the results and probably switch to other many (or multi) core architectures and languages.
Progress
fill in here