LinearSolverImpl

CUDA and Python Image Reconstruction Algorithms Implementation

Introduction

This repository majorly features a CUDA Implementation of several Image Reconstruction Algorithms, including ISTA, FISTA, SAFISTA and ADMM. All of them support 2-D image reconstruction with the Gaussian blur as point spread function. 1-D (vector) reconstruction with matrix-vector multiplication is supported for ISTA and FISTA. Compared to their original implementations in python (which is also provided in repository), the CUDA parallelized version could be 1600x faster in speed.

Content Overview

CUDAImpls

In this folder there are two files: FISTA.cu and ADMM.cu.

FISTA.cu contains all algorithms except ADMM. In FISTA.cu, ista() and fista() are 1-D vector reconstruction, and functions with name like run_xxx_image() is the 2-D image reconstruction. The parameters used in the algorithm could be set at the top, which are defined as constants.

ADMM.cu contains the ADMM algorithm. The run_admm_image() is the reconstruction function for 2-D image reconstrucion.

You can customize the image and psf used by digging into the functions and find the relevant line. This will be moved out as an argument in later versions.

PythonImpls

In this folder there are python implementations of above algorithms. The structure is similar to above.

PythonCUDA

Folder expresses some of the trials with Python-CUDA compliant software. ADMM-based Reconstructioni ADMM update steps is based on the work done by the Waller Lab for 3D construction.

Requirements

Compiler

CUDA 10.1

Libraries

CUSolver

CUFFT

CURand

CUBLAS

libpng

chrono

Compilation Command

nvcc xxx.cu -o xxx.o -lcufft -lcurand -lcusolver -lcublas -lpng

./xxx.o