iFDK: A Scalable Framework for Instant High-resolution Image Reconstruction (SC' 19)

We use ABCI (https://abci.ai/) HPC to solve the high-resolution image reconstruction problem, e.g. 2048^3, 4096^3, 8192^3. This repository contains the artifact, e.g. job scripts, benchmarks.

DEPENDENCIES

We tested on ABCI, using Nvidia Volta V100 GPUs, with GCC 4.8 and NVCC 9.0. The following libraries and tools are requirements:

cmake = 3.1
CUDA = 9.0
python >= 2.7
Intel MPI 2018.2.199
Intel IPP 2018.2.199
Malab R2018a
Insight Segmentation and Registration Toolkit (ITK)
Reconstruction Toolkit (RTK)

BUILD/RUN-TIME ENVIRONMENT

SOURCE CODE

Our Source Code is Closed.
RTK-1.4.0 BP kernel with single precision the improved kernel can be found in

src/RTK/rtkCudaFDKBackProjectionImageFilter.cu

DATA

Generate 3D shepp-logan phantom by Matlab script as
```
  tools/phantom3d/phantom3d.m
```
a sample (size 512^3, single precison, raw data) can be downloaded from dropbox as

https://www.dropbox.com/s/o0xgt4igipdve2l/Shepp-Logan-512x512x512.vol?dl=0
Generate 2D projection
```
  ./generate-projections.sh    
```

How to run

all modules are in folder bin

Generating job script by run scrip in folder jobs/generate-jobs as

  python gen-jobs.py strong2k
  python gen-jobs.py strong4k
  python gen-jobs.py strong8k
  python gen-jobs.py weak2k
  python gen-jobs.py weak4k
  python gen-jobs.py weak8k

Run jobs in the Root-folder of iFDK-archifact

  ./run.sh all
  ./run.sh strong2k
  ./run.sh strong4k
  ./run.sh strong8k
  ./run.sh weak2k
  ./run.sh weak4k
  ./run.sh weak8k

Run benchmarks

BENCHMARKS

The related benchmarks can be found in follows:

https://www.openrtk.org/

https://github.com/LLNL/ior

https://docs.nvidia.com/cuda/cuda-samples/index.html

tools

https://jp.mathworks.com/matlabcentral/fileexchange/9416-3d-shepp-logan-phantom?s_tid=mwa_osa_a