• anisotropic_diffusion(): Anisotropic diffusion filter.
• topk(): Returns top-K elements given an array.

• Fixed sift() and gloh(), which were improperly calling the library.

• Added len() method, which returns array.elements().

• Documented statistics API.
• Corrected sign() documentation.
• Modified helloworld example to match C++ lib.

Feature parity with ArrayFire 3.5.

• canny: Canny Edge detector
• Array.scalar: Return the first element of the array
• dot: Now support option to return scalar
• print_mem_info: Prints memory being used / locked by arrayfire memory manager.
• Array.allocated: Returs the amount of memory allocated for the given buffer.
• set_fft_plan_cache_size: Sets the size of the fft plan cache.

Bug Fixes:

• sort_by_key had key and value flipped in documentation.

Improvements and bugfixes from upstream include:

• CUDA backend uses nvrtc instead of nvvm
• Performance improvements to arrayfire.reorder
• Faster unified backend
• You can find more information at arrayfire's release notes

Feature parity with ArrayFire 3.4 libs

• Sparse matrix support

  • create_sparse
  • create_sparse_from_dense
  • create_sparse_from_host
  • convert_sparse_to_dense
  • convert_sparse
  • sparse_get_info
  • sparse_get_nnz
  • sparse_get_values
  • sparse_get_row_idx
  • sparse_get_col_idx
  • sparse_get_storage

• Random Engine support

  • Three new random engines, RANDOM_ENGINE.PHILOX, RANDOM_ENGINE.THREEFRY, and RANDOM_ENGINE.MERSENNE.
  • randu and randn now accept an additional engine parameter.
  • set_default_random_engine_type
  • get_default_random_engine

• New functions

  • scan
  • scan_by_key
  • clamp
  • medfilt1
  • medfilt2
  • moments
  • get_size_of
  • get_manual_eval_flag
  • set_manual_eval_flag

• Behavior changes

  • eval now supports fusing kernels.

• Graphics updates

  • plot updated to take new parameters.
  • plot2 added.
  • scatter updated to take new parameters.
  • scatter2 added.
  • vector_field added.
  • set_axes_limits added.

Bug fixes

• ArrayFire now has higher priority when numpy for mixed operations. ^{1 2}
• Numpy interoperability issues on Widnows. ¹
• Switch to a working backend by default. ¹
• Fixed incorrect behavior for Hermitian transpose and QR. ¹
• array[0:0] now returns empty arrays. ¹

Further Improvements from upstream can be read in the arrayfire release notes.

Bugfix: Increase arrayfire's priority over numpy for mixed operations

Added new library functions

• get_backend returns backend name

Bugfix to af.histogram

Added missing functions / methods

• gaussian_kernel

Added new array properties

• Array.T now returns transpose
• Array.H now returns hermitian transpose
• Array.shape now allows easier access individual dimensions

Feature parity with Arrayfire 3.3 libs

• Functions to interact with arryafire's internal data structures.
  • Array.offset
  • Array.strides
  • Array.is_owner
  • Array.is_linear
  • Array.raw_ptr
• Array constructor now takes offset and strides as optional parameters.
• New visualization functions: scatter and scatter3
• OpenCL backend specific functions:
  • get_device_type
  • get_platform
  • add_device_context
  • delete_device_context
  • set_device_context
• Functions to allocate and free memory on host and device
  • alloc_host and free_host
  • alloc_pinned and free_pinned
  • alloc_device and free_device
• Function to query which device and backend an array was created on
  • get_device_id
  • get_backend_id
• Miscellaneous functions
  • is_lapack_available
  • is_image_io_available

Interopability

• Transfer PyCUDA GPUArrays using af.pycuda_to_af_array
• Transfer PyOpenCL Arrays using af.pyopencl_to_af_array
• New helper function af.to_array added to convert a different array to arrayfire Array.
  • This function can be used in place of af.xyz_to_af_array functions mentioned above.

Deprecated functions list

• lock_device_ptr is deprecated. Use lock_array instead.
• unlock_device_ptr is deprecated. Use unlock_array instead.

Bug Fixes:

• Boolean indexing giving faulty results for multi dimensional arrays.
• Enum types comparision failures in Python 2.x
• Support loading SO versioned libraries in Linux and OSX.
• Fixed typo that prevented changing backend
• Fixed image processing functions that accepted floating point scalar paramters.
  • Affected functions include: translate, scale, skew, histogram, bilateral, mean_shift.

Bug fixes:

• A default AF_PATH is set if none is found as an environment variable.

Examples:

• Heston model example uses a smaller data set to help run on low end GPUs.

Bug fixes:

• get_version() now returns ints instead of c_int
• Fixed bug in tests/simple/device.py

The module now looks at additional paths when loading ArrayFire libraries.

• Link to the wiki is provided when ctypes.cdll.LoadLibrary fails.

New function:

• info_str() returns information similar to info() as a string.

Updated README.md with latest instructions

Feature parity with ArrayFire 3.2 libs

• New computer vision functions: sift, gloh, homography
• New graphics functions: plot3, surface
• Functions to load and save native images: load_image_native, save_image_native
• Use unified backend when possible

Added missing functions

• eval, init, convolve2_separable, as_type method
• cuda backend specific functions
• opencl backend specific functions
• timeit function to benchmark arrayfire functions

Added new examples

• getting_started: intro, convolve
• benchmarks: bench_blas, bench_fft
• financial: monte_carlo_options, black_scholes, heston_model
• graphics: fractal, histogram, plot3d, conway, surface

Bug fixes

• Fixed bug when array types were being reported incorrectly
• Fixed various bugs in graphics functions