This is a reference implementation of the Mind Mappings Framework. Mind Mappings performs a gradient-based mapping space search for hardware accelerators.
To Get started, follow the below steps:
Install the Mind Mappings package: python3 -m pip install -e .
To test if timeloop and mind mappings setup are fine, run python3 costModel/timeloop/model_timeloop.py. This should randomly choose a valid mapping and print its cost.
Now, everything is setup. Take some time to explore the mind mappings package. costModel directory has models.py, which describes the mandatory functions that any cost model should implement. example/ directory contains a simple example cost model for finding minimum of a quadratic equation. timeloop/ directory shows the mind mappings and timeloop interface.
Everything related to performing mapping space search is handled with optimize.py. As understood from the paper, here are two key phases:
Train a Surrogate
Use Surrogate for Search
For convenience, two trained surrogate models are already provided for you (in gradSearch/saved_models/): model_CNN-layer.save and model_MTTKRP.save. Each of them are specific to the architecture described in the paper and the related algorithm. In case, you want to target a different architecture/algorithm, they need to be re-trained (steps are provided later).
--algorithm can be set to CNN-layer or MTTKRP, --problem should be set to the problem shape (N C K R S P Q/I J K L, see paper for description), --maxsteps can be set to the maximum number of steps you would like the search to run.
This prints out the best mapping and its predicted cost.
Check the parameters.py file, which controls most of the parameters for the run. Play around with them.
In case, you would like to train a different surrogate model, follow the steps shown below:
NOTE: To get the best out of mind mappings, you will need to tune search related parameters listed in parameters.py file.
If you would like to reproduce the results from the paper, you can run:
python3 optimize.py --command reproduce
If this was useful in your research, please cite:
Hegde, Kartik, Po-An Tsai, Sitao Huang, Vikas Chandra, Angshuman Parashar, and Christopher W. Fletcher. "Mind mappings: enabling efficient algorithm-accelerator mapping space search." In Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, pp. 943-958. 2021.
To Get started, follow the below steps:
python3 -m pip install -e .parameters.py:self.SCRATCHself.TIMELOOP_PATHto point to timeloop path.python3 costModel/timeloop/model_timeloop.py. This should randomly choose a valid mapping and print its cost.Now, everything is setup. Take some time to explore the mind mappings package.
costModeldirectory hasmodels.py, which describes the mandatory functions that any cost model should implement.example/directory contains a simple example cost model for finding minimum of a quadratic equation.timeloop/directory shows the mind mappings and timeloop interface.Everything related to performing mapping space search is handled with
optimize.py. As understood from the paper, here are two key phases:For convenience, two trained surrogate models are already provided for you (in
gradSearch/saved_models/):model_CNN-layer.saveandmodel_MTTKRP.save. Each of them are specific to the architecture described in the paper and the related algorithm. In case, you want to target a different architecture/algorithm, they need to be re-trained (steps are provided later).To perform mapping space search, run:
--algorithmcan be set to CNN-layer or MTTKRP,--problemshould be set to the problem shape (N C K R S P Q/I J K L, see paper for description),--maxstepscan be set to the maximum number of steps you would like the search to run.This prints out the best mapping and its predicted cost.
parameters.pyfile, which controls most of the parameters for the run. Play around with them.In case, you would like to train a different surrogate model, follow the steps shown below:
python3 optimize.py --command datagen --path <PATH> --algorithm <ALG> --costmodel <your new cost model>python3 optimize.py --command dataprocess --path <PATH> --algorithm <ALG> --costmodel <your new cost model>python3 optimize.py --command train --path <PATH> --algorithm <ALG>python3 optimize.py --command search --algorithm <ALG> --problem <DIMS> --maxsteps <STEPS>NOTE: To get the best out of mind mappings, you will need to tune search related parameters listed in
parameters.pyfile.If you would like to reproduce the results from the paper, you can run:
If this was useful in your research, please cite:
Hegde, Kartik, Po-An Tsai, Sitao Huang, Vikas Chandra, Angshuman Parashar, and Christopher W. Fletcher. "Mind mappings: enabling efficient algorithm-accelerator mapping space search." In Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, pp. 943-958. 2021.