DataHandlerGPU: Documentation (#21)

docs/data_handler_GPU.md

+elsa offers a fully GPU based DataContainer. This option requires compilation with
+```
+-DELSA_CUDA_VECTOR=ON
+```
+Additionally, a matching version of CUDA as well as clang has to be available. This is necessary because of C++17 in device code which is not supported by nvcc. Two tested combinations on Ubuntu 18.04 are:
+
+| clang | CUDA | compute capability |
+| ------ | ------ | ----- |
+| 8 | 10.0 | up to 7.5 |
+| 8 | 9.2 | up to 7 |
+
+Similarly to the `DataHandlerGPU` which uses Eigen, the `DataHandlerGPU` uses the [Quickvec](https://gitlab.lrz.de/IP/quickvec) project. The compile flags will be set automatically.
+
+Enabling the compile-time option `ELSA_CUDA_VECTOR` will set the default `DataContainer` type to use the GPU.
+
+Keep in mind that running full reconstruction examples with large volumes requires large amounts of memory (which the GPU might not have depending on your hardware). As a rough estimate, running a reconstruction task with $700^3$ volume elements requires around 10 GB of memory, as multiple copies have to be created.
+
+On newer GPUs, oversubscription of the GPU memory through the unified memory architecture is possible. However, this will slow down the task at hand considerably to the point that a CPU-only version in main memory might be faster.

docs/modules/core.rst

@@ -63,6 +63,12 @@ DataHandlerCPU
 
 .. doxygenclass:: elsa::DataHandlerCPU
 
+DataHandlerGPU
+--------------
+
+.. mdinclude:: data_handler_GPU.md
+.. doxygenclass:: elsa::DataHandlerGPU
+
 Expression
 ----------
 .. mdinclude:: expression_templates.md