/ReFloat

Low-Cost Floating-Point Processing in ReRAM for Accelerating Iterative Linear Solvers

Primary LanguageC++MIT LicenseMIT

DOI

ReFloat

ReFloat is a ReRAM based accelerator for accelerating large-scale sparse linear solvers. The refloat format is a new floating-point format which fits the ReRAM crossbars with hardware awareness. We provided the implemnetations of the conjugate gradient (CG) and the biconjugate gradient stabilized (BiCGSTAB) solvers for different platforms including Nvidia GPU, CPU, and ReFloat. The matrices used in the evaluation are from SuiteSparse collection in the martrix market format.

You can find the techinical details in our paper ReFloat: Low-Cost Floating-Point Processing in ReRAM for Accelerating Iterative Linear Solvers.

If you find this code useful, please cite:

@inproceedings{song2023refloat,
    title={ReFloat: Low-Cost Floating-Point Processing in ReRAM for Accelerating Iterative Linear Solvers},
    author={Song, Linghao and Chen, Fan and Li, Hai and Chen, Yiran},
    booktitle={The International Conference for High Performance Computing, Networking, Storage and Analysis (SC '23)},
    year = {2023}
}

The input matrix

We provided one sample matrix crystm03 in the matrices directory. To download and decompress all the matrices, run

cd matrices
sh download.sh

The GPU implemnetation

We have tested the GPU implemneation on a Nvidia V100 GPU with CCuSparse (CUDA version 11.7). To compile the GPU code,

cd gpu
make

We provided a script to run all the evaluations,

sh run_gpu.sh

The CPU implemnetation

We provided the CPU implemneation and the simulation code under the src directory. We suggest that your CPU paltform has OpenMP installed. To compile,

cd src
make

To run the CPU implementation,

cd run/cpu
sh run_cpu.sh

The simulation

To run the simulation,

cd run/refloat
sh run_refloat.sh

All the scripts can be configured to run CG by setting SOLVER=0 or BiCG by setting SOLVER=1.

For SC'23 Artifact Evaluation:

Step 1. To obtain a copy of the source code

git clone https://github.com/linghaosong/ReFloat.git

Step 2. To download the matrices

cd ReFloat/matrices 
sh download.sh

Step 3. We provided a GPU implementation which requires a Nvidia P100 or similar HBM GPU. It is optional to run the GPU implementation. To compile the GPU code

cd ReFloat/gpu 
make

Step 4. To run the GPU baseline and obtain the results

4.1 To obtain the GPU CG execution time

First, edit the script run_gpu.sh to set

PRINT_RES= 
RUN=gpucg

Then, then run the script

sh run_gpu.sh

4.2 To obtain the GPU BiCG execution time

First, edit the script run_gpu.sh to set

PRINT_RES= 
RUN=gpubicg

Then, then run the script

sh run_gpu.sh

4.3 To obtain the GPU CG execution residual

First, edit the script run_gpu.sh to set

PRINT_RES=1 
RUN=gpucg

Then, then run the script

sh run_gpu.sh

4.4 To obtain the GPU BiCG execution residual

First, edit the script run_gpu.sh to set

PRINT_RES=1 
RUN=gpubicg

Then, then run the script

sh run_gpu.sh

Step 5. To compile the simulation code

cd ReFloat/src
make

5.1 To run the Feinberg baseline with the assumption that the funtionality is the same as FP64, go to

cd ReFloat/src/run/baselinefc

To run CG, edit the script run_baselinefc.sh to set

SOLVER=0

Then run

sh run_baselinefc.sh

To run BiCG, edit the script run_baselinefc.sh to set

SOLVER=1

Then run

sh run_baselinefc.sh

5.2 To run the ReFloat, go to

cd ReFloat/src/run/refloat

To run CG, edit the script run_refloat.sh to set

SOLVER=0

Then run

sh run_refloat.sh

To run BiCG, edit the script run_refloat.sh to set

SOLVER=1

Then run

sh run_refloat.sh