Fork of Chroma developed for the use with the MicroBooNE neutrino experiment.
Actually, this is a fork of Simon C. Blyth's fork of Chroma.

• Modified gpu/tools.py to be able to load context with cuda context flags.
  By default, loads the flag for device-mapped host memory. Needed for running on an older card (C1080).
• Fixed what seemed like a bug in detector.py where the pdf was defined with one too many bins.
• Fixed a bug in gpu/tools.py: make_gpu_struct. There were problems with the numpy buffer interface.
• Modifications have been made to test functions that were used to help explore program behavior. Some functions have been added.
• Added abilit to run with OpenCL, using pyOpenCL
• Providing new module, importgeo, that uses tools written by S.C. Blyth to import geometries described in COLLADA file output by G4DAE C++ library for Geant4
• Having ROOT with PyROOT extension will allow some tests to make plots. Not required (or at least tried not to make it required).

• API, CUDA or OpenCL, is selectable using functions in api.py
• api.py also has functions to test which API
• opencl has some big differences that affect chroma: no device-mapped memory, no built-in random number generator, and different struct creation/definition.
• because of these differences, port was not done so cleanly. many modules have been reorganized and classes changed. merging this fork may require coordination.
• however, I did try to keep changes to orginal chroma modules minial. Typically, what was added were one or more keyword arguments.

• I see some precision effects which causes differences in the BVH tree between opencl and cuda. Tests used the companion cube mesh.
• Get differences in quantize centroid values which then propagates to rest of tree structure.
• From my studies, differences are only off by one quantum along any axis
• Example Mismatch: Triangle ID=286, morton codes are cuda= 114847925804877 cl= 114847925804876
• Bounding box for mismatch is: CUDA=[37663, 55944, 4589], CL=[37662, 55944, 4589], diff= [1 0 0]
• CUDA value always seems to be bigger
• For companion cube mesh about 1.9% of all morton codes are different

• Unlike CUDA, OpenCL does not have a built-in random number generator
• Implemented RNG using Random123 package: http://www.deshawresearch.com/downloads/download_random123.cgi/. Included in repo.
• Random123 is in principle stateless. Created state vector in gpu/clrandstate.py. Implemented uniform sampler in cl/random.cl. Require state, because being able to reproduce results is a requirement for particle physics MC.
• tests, cltest/test_randomgen.py and cltest/test_sample_cdf.py, generate flat and gaussian distribution, respectively. Seems to work.

See the original chroma website: http://chroma.bitbucket.org

I really love the idea of ShrinkWrap, I really do. But I am neither lucky enough nor smart enough to get it to work out-of-the-box. For me, what worked was I reduced the number of components to install in the chroma_deps setup.py script and manually installed the other dependencies. For a diary summarizing what I installed (and short notes on how) during my from-scratch effort on linux machines where I did not have root priveledges, see INSTALLATION_DIARY.md.