OSPRay MPI v2.4.0

• Significant improvements have been made to loading performance in
  the MPI Offload device. Applications which make large numbers of API
  calls or create many smaller geometries or volumes should see
  substantial load time improvements.
• Strided data arrays are compacted on the app rank before sending

OSPRay MPI v2.2.0

• Improve parallelism of framebuffer compression & decompression when
  collecting the final framebuffer to the head rank. This provides a
  substantial performance improvement when using just a few ranks or
  large framebuffers (both in pixel count or channel count).
• The MPI module will now default to setting thread affinity off, if
  no option is selected. This improves thread usage and core
  assignment of threads in most cases, where no specific options are
  provided to the MPI runtime.
• Fix bug where OSPObject handles where not translated to worker-local
  pointers when committing an OSPData in the MPIOffloadDevice.
• Fix handling of OSP_STRING parameters
• Move from ospcommon to rkcommon v1.4.2

OSPRay MPI v2.1.0

• Add support for ospGetTaskDuration to query the render time of a
  asynchronous (or synchronous) renderFrame call
• Use flush bcasts to allow us to use non-owning views for data
  transfer. Note that shared ospData with strides is currently
  transmitted as whole
• Fix member variable type for bcast
• Fix incorrect data size computation in offload device
• Fix large data chunking support for MPI Bcast

OSPRay MPI v2.0.1

This release of the MPI module requires OSPRay 2.0.1 or higher

• The MPI module is now built outside of OSPRay source, and can be built as part of the super build

This is the first release of the MPI module for OSPRay v2.x, and updates the API to match:

• Asynchronous rendering is now supported in both distributed and offload devices
• The regions parameter on the OSPWorld has been simplified to be a list of boxes. Sharing data is now determined by two ranks specifying the same bounding box
• When using the distributed device, if all ranks specify the exact same data the renderer will switch to perform image-parallel rendering, enabling use of the path tracer see ospMPIDistributedTutorialReplicatedData

See the README and tutorials for more information on updates to the API, command line options and object parameters.