In quarterly Meetings of the Maintainers, we review the following roadmap setting a high-level direction for future work. The suggested version numbers and features are not binding and often change. The information is updated once per quarter.

• Improved oneAPI performance
• Continuous delivery
  • Support more distro packages
    • Based on Debian [2 hours]
    • Cent OS 7/8 [16 hours]
    • Perhaps Amazon Linux AMI (or more cloud providers) if it gives a better experience for users compared to using sh files
  • Python CI [16 hours]
• Debugging/Diagnostic functions [2 hours]
• Random Shuffle
  • https://numpy.org/doc/stable/reference/random/generated/numpy.random.permutation.html
• Stream function/Input streaming [80 hours]

• Data Frames
• JIT CPU
• Multi-GPU
• Multiple backend interoperability
• CPU Optimization/Parallelism
• OpenCL targeting multiple devices(CPU specific kernels)

• Support only float/double for pow
• ND Arrays? Is there interest that should be answered before implementation?
• Drop unnecessary type support for existing API; for example, drop unused types from image processing.
• Char -> bool types
• Merge index and index_gen
• Merge assign and assign_gen

Items from the list below are mapped to the

No API Changes

• Pan ArrayFire support to accept an existing array as output
• Strings? String operations on af::array
• Simpler graphics API

API Changes

• Add additional hardware to CI
• benchmarking regressions
• Add bf16 support
• Backend interop between backends

• [BUG] LAPACKE Error when using solve on OpenCL backend bug
• test_svd_dense_opencl fails on Ubuntu 20.04 LTS using NVIDIA OpenCL on AWS g3s.xlarge instance bug
• test_array_cpu hangs on Ubuntu 20.04 LTS using NVIDIA OpenCL bug
• Failed tests when compiling on AWS c6gd.2xlarge Instance bug
• [BUG] Unable to use forge/windows in arrayfire (MacOS Big Sur) bug
• [BUG] matmul nondeterministic result on CPU backend with gomp bug dependency
• [Perf] OpenCL performance issue with sparse matmul bug OpenCL perf
• arrayfire warning - exceptions over dll boundary bug
• 3D scatter-plot has too small extent bug dependency graphics
• No AF_ERR_NO_MEM error when afopencl array creation fails bug OpenCL
• Indexing an array using another array containing negative integers indexes wrong elements bug
• in v3.6.4 c2r FFT changes the value of the input bug CUDA dependency known issue
• Batched gemv very slow bug CUDA perf
• Swapped coordinates in FAST detector (OpenCL) bug
• Matmul throws CL_MEM_OBJECT_ALLOCATION_FAILURE on OpenCL bug OpenCL
• Sparse-dense matmul with AF_MAT_TRANS very slow in OpenCL on Nvidia card bug OpenCL perf
• Harris: OpenCL backend does not detect all features that CPU and CUDA does bug CPU
• ORB: CPU backend does not detect all features that CUDA and OpenCL does bug CPU
• OpenCL HelloWorld examples produce Segfault without display bug installer OpenCL
• canny_cuda test fails on Tegra TX2, AF 3.5.1 bug CUDA
• Internal error:998 after calling scan in multiple tests within OpenCL application bug OpenCL
• BUG: access to deleted memory in array_proxy index bug
• CUDA-Aware MPI with device<>()-pointer needs af::sync() bug CUDA-
• Errors when CUDA-MPS is active bug CUDA
• OpenCL-OpenGL broken on Windows AMD GPUs bug OpenCL Windows
• gfor example doesn't work bug
• Test failure on AMD APU with Open Source drivers bug

• Internals: jit, memory manager, af::proxy
• Better API documentation
• New examples