imresh

Shrink-Wrap Imaginar Coefficient Reconstruction Algorithm

Compilation and Installation

Prerequisites

To compile this library you need

C++ Compiler (with c++11 support, e.g. GCC 4.7+)
CUDA (7.0+)
CMake (3.3.1+)
OpenMP

Optional Dependencies

FFTW3 (single precision build)
libSplash (for reading and writing HDF5)
PNGwriter (for storing reconstructed images as PNGs)

Build options

-DCMAKE_INSTALL_PREFIX

The prefix to install the library and additional files (such as headers) into.
-DRUN_TESTS (default off)

If true tests will be run.
-DBUILD_EXAMPLES (default off)

If true the examples from the examples directory will be built.
-DIMRESH_DEBUG (default off)

Adds at least debugging symbols to the code.
-DBUILD_DOC (default on)

Build the Doxygen documentation.
-DUSE_PNG (default off)

Enable PNG output.
-DUSE_SPLASH (default off)

Enable HDF5 in- and output.

Building

Create a build directory
```
 mkdir build
 cd build
```

Invoking CMake

 cmake ..

For a clean build with debugging information, try

 cmake .. -DIMRESH_DEBUG=on -DBUILD_DOC=off

To build and run everything, try

 cmake .. -DRUN_TESTS=on -DBUILD_EXAMPLES=on -DUSE_PNG=on -DUSE_SPLASH=on

Invoking make
```
 make
```
Installing
```
 make install
```

Usage

Basic usage

The usage of imresh is mainly divided into five parts:

Library initialization
Image loading
Image processing
Image writing
Library deinitialization

where image loading and writing can also be handled outside the library.

The library initialization is (from the user's perspective) just a single call to imresh::io::taskQueueInit( ). Internally this creates cudaStream_ts for each multiprocessor on each CUDA capable device found and stores them for later access.
Image loading can be done through imresh's own loading functions (found in imresh::io::readInFuncs) or with self-written functions.

Note:

Your self-written functions have to provide you both the image dimensions and the host memory containing the image. This memory has to be allocated via new if you're using the built-in write-out functions.
Image processing is just a call to imresh::io::addTask( ) (for explanation of the parameters please have a look at the Doxygen). This will start a thread (a C++ std::thread thread to be precise) handling your data transfers and image processing on the least recently used stream available. The given data write out function will be called inside of this thread, too.
Image writing can, just as the loading, be done via imresh's own write out functions (found in imresh::io::writeOutFuncs) or with self-written functions. These have to match the following signature:
```
 void writeOutFunc( float* memory, std::pair<unsigned int,unsigned int> size, std::string filename);
```
where memory is the raw image data, size the image dimension (size.first is horizontal, size.second is vertical) and filename the name of the file to store the image in.

Note:

If you're using imresh's own loading functions in combination with your own write-out functions be sure you're freeing the image memory with delete.

Note:

imresh's workflow is designed in a way that you'd free your memory inside of your write out function. It's never called before the algorithm finishs and therefore the ideal place for freeing the image data. imresh's built-in functions handle it that way.
Library deinitialization is again just a call to imresh::io::taskQueueDeinit( ). This will handle stream destroying, memory freeing and so on for you.

Note:

When you're using your own data reading and/or writing functions, you'll have to handle the memory inside of this functions yourself.

Advanced usage

There's a set of in-action examples in the examples directory. These can be compiled by appending the -DBUILD_EXAMPLES=on to your CMake call, e.g.

cmake .. -DBUILD_EXAMPLES=on

For a simple but complete example of how to use this library try miniExample and have a look at miniExample.cpp.
For a more complex example with batch processing please have a look at threadedExample and threadedExamples.cpp resp.
If you need more example data for your tests, please run outputExampleCreation

Folder structure

.
+-- benchmark: Contains older deprecated less optimized versions for comparison. Files in here should only be used by files in the tests folder
+-- cmake: CMake find package scripts
+-- examples: Executable examples showing how to use the library
|   +-- createTestData: Functions for generating example objects and diffraction intensity to test the algorithm on
|   +-- testData: Static examples with a fixed size to test e.g. `readInFuncs`
+-- src: The sources for the actual library. Only this folder is needed with the standard CMake options
|   +-- imresh
|       +-- algorithms
|       +-- io: File input/output and batch processing
|       +-- libs
+-- tests: Unit tests and benchmarks.

Authors

Maximilian Knespel (m.knespel at hzdr dot de)
Philipp Trommler (philipp.trommler at tu-dresden dot de)

Known Bugs

/usr/include/fftw3.h(373): error: identifier "__float128" is undefined

Update your fftw library to something higher than 3.3.4 (not yet released as of this writing) or manually apply the patch shown here, i.e. add || defined(__CUDACC__) to the faulty line in the header.
stddef.h(432): error: identifier "nullptr" is undefined

Your CMake version is too old.
/usr/include/host_config.h:105:2: error: #error -- unsupported GNU version! gcc 4.10 and up are not supported! #error -- unsupported GNU version! gcc 4.10 and up are not supported!

This is a common problem with debian and maybe its derivatives, because debian sid and stretch have by default GCC 5.x, but nvidia-cuda-toolkit 7.0.x. The latter wants GCC 4.10 or lower versions, though.

On Debian:
- apt-get install gcc-4.9
- cmake .. -DIMRESH_DEBUG=ON -DCMAKE_C_COMPILER=$(which gcc-4.9) -DCMAKE_CXX_COMPILER=$(which g++-4.9)

ComputationalRadiationPhysics / imresh