bh107/bohrium

📋 Changes

• Introducing a new Python package, ``bh107``, which is a pure Python front-end that imitate NumPy.

📋 Changes

• User kernels https://bohrium.readthedocs.io/users/python/index.html#userkernel
• Extended the C++ interface significantly https://bohrium.readthedocs.io/users/cpp.html

📋 Changes

• Separation of the versions of Bohrium and Bohrium API.
• Simplified how to choose where the OpenCL backend should run.
• Set `BH_OPENCL_DEVICE_NUMBER` to the device number you want.
• Run `python -m bohrium_api --info` to see all available devices (and their numbers).

📋 Changes

• Major code clean up
• Divided the Python fronted (bridge) into two parts: `bohrium_api` and `bohrium`.
• Added wheel packages with Python3.7

📋 Changes

• Implement some GPU optimizations
• Dynamic views is changed (@skjoenberg):
• Now the sliding view information is transferred to the cxx bridge right after the view is flushed, instead of right after indexing. The sliding view information is now stored in a Python object called dynamic_view_info.
• Sliding views now allow changes to shape between iterations
• Sliding views now support broadcasting
• Sliding views now support nesting sliding views (such as a=[i:i+5], b=a[i])
• Iterators can now be in grids corresponding to nested loops. The iterators can not be mixed when indexing into the same dimension.
• Temporary arrays now inherit the dynamic changes of the views involved in the operation

📋 Changes

• Implemented a malloc cache, which should speedup most applications
• Implemented lossy channels in the Proxy VEM
• Implemented `hstack()` and friends

📋 Changes

• Now do_while() loop, and the new for_loop() supports sliding views (@skjoenberg): http://bohrium.readthedocs.io/users/python/index.html#sliding-views-between-iterations
• Building PyPI `manylinux1` packages for Python 2.7, 3.4, 3.5, and 3.6
• Building PyPI packages for Python 2.7 and 3.6 on Mac OSX 10.12 and 10.13
• The homebrew package is now deprecated
• Re-implemented signal handling -- we now use [libsigsegv](https://www.gnu.org/software/libsigsegv/).

📋 Changes

• Now Bohrium will clean up the cache folder so that total number of cache files is less than 50000 files by default.
• Bohrium now has a `do_while()` function to replace Python for and while loop for better performance: http://bohrium.readthedocs.io/users/python/index.html#accelerate-loops
• Added Bohrium pip packages
• Fixed `np.mean()`
• Re-written the C/Python API

📋 Changes

• Now backends cache the JIT code generation, which should speedup the execution with small data sets (see #436).

📋 Changes

• Implemented a proxy component, which makes it possible to send array operations to a backend server over TCP/IP.
• `convolve()` and `correlate()` now works on pure numpy arrays as well as bohrium array.
• Now using a simpler fuser in special cases, which significantly reduce the fusion overhead (issue #425).

📋 Changes

• `clang` can now use OpenMP with latest version of CMake
• `clang` will run with `-O3` when executing kernels
• Removed the `BH_SYNC` byte code

📋 Changes

• Implemented interoperability (see http://bohrium.readthedocs.io/users/python/index.html#interoperability)
• Cython interoperability
• PyOpenCL interoperability
• PyCUDA interoperability
• Implemented `numpy.bincount()` using Cython, PyOpenCL, and PyCUDA
• Implemented Nd `convolve()` and `correlate()` (by @Thunderl3ear)

📋 Changes

• Re-implementation of the C++ frontend
• Implemented profile context in Python e.g.:
• Implemented `correlate()` and `convolve()` (by Thorbjørn Koch)
• Implemented `argmin()` and `argmax()` for 1d arrays

📋 Changes

• Fixed a `lib64` issue (#317)
• Fixed a `numpy.matlib` issue (#319)
• Henceforth, the VERSION file has to be manually updated when creating new releases

📋 Changes

• Fixed a fatal bug when the version consist of three components.
• Fixed a segfault on shutdown when using OpenMP (@mfherbst)

📋 Changes

• Added a CUDA backend (experimental)
• Added a VERSION file
• Clean up and Bug fixes

📋 Changes

• Added support of fancy indexing of multi dimension arrays
• Fixed BLAS & LAPACK dependencies in the Conda build (@dionhaefner)
• Added Ubuntu 17.04 package build
• Now `bohrium.where()` handles `NaN` values (@dionhaefner)
• Fixed the CG solver see #311 (@dionhaefner)

📋 Changes

• Fixed signed integer division so it matches Python/NumPy
• Added a TDMA solver (by @dionhaefner)
• Fixed many bugs and added python tests to verify the fixes
• Added support of `np.take()` and `np.put()`
• Added support of fancy indexing of vectors such as `a[[3,2,1]] = a[[1,2,5]]`
• Added OpenMP and OpenCL implementation of *some* LAPACK functions (as extension methods)
• Added Conjugate gradient
• Removed dependency of Cheetah
• + 1 more

📋 Changes

• Fully MacOS support (both OpenMP and OpenCL)
• Conda/Anaconda support (linux-64)
• Reduced the overhead of the fusers significantly
• Improved the compiler cache hit rate significantly
• Implemented scalar assignment to masked array e.g.: `a[a>42] = 7`
• Implemented np.where(), which can be used instead of masked array in most cases
• Implemented OpenMP and OpenCL implementation of BLAS (as extension methods)
• Fixed many bugs and added python tests to verify the fixes

📋 Changes

• Fixed MAJOR memory leak
• Cleanup of both the OpenCL and OpenMP backend
• Fixed many bugs and added python tests to verify the fixes

📋 Changes

• Introducing two new backends: OpenMP and OpenCL.
• The OpenMP backend is now the default backend
• Python v3 support