.todo

Beta release tasks:
  🗹 Figure out how to properly emit fast_powi / fast_powf automagically depending
    on data type
  🗹 Get the rest of the write race warnings
  🗹 Get the new ambiguous dependency warnings as well
  ☐ Make platform specification completely independent of pyopencl -- currently, it requires the opencl platform to be installed on the generating machine, but it should only require a platform file.
    🗹 In progress, need to do a full test on a machine w/o pyopencl installed
  ☐ Documentation
Improvements:
  🗹 Make libgen use codepy, since we depend on it already
  ☐ Fusion of similar kernels (e.g., the different Troe / SRI / Lindemann falloff derivative variants)
  ☐ Preload of commonly reused data into __local mem for deep-vectorizations on GPUs
    ☐ Note: the utility of this is limited at the moment, as NVIDIA doesn't yet support atomics
      hence, this will become more pressing when CUDA support is enabled
  ☐ New targets: ISPC, CUDA, Vectorized OpenMP
Add-ons:
  ☐ Add common BLAS / LAPACK implementations (e.g. matrix mul, inversion, etc.), particularly
    for sparse matrices
Codegen:
  ☐ Find a way to unify kernels, such that a single library / pywrapper exposes
    all available code
      🗹 In progress -- we have the ability to generate different classes that can call
         each respective kernel, but we don't (yet) put a class for each possiblty kernel
         in the calling program
  🗹 Reimplement wrapper generation using Cog
    ☐ Make the wrapper a C++ interface, such that we can embed useful info (species ordering, etc.), as well as simplify kernel exection / memory allocation
      🗹 In progress, need to add more information variables / array creators
    🗹 Serialize the pre-initialize memory manager, and load in Cog generation to simplify this
    ☐ Leave stubs for extension to CUDA / ISPC / etc.
    🗹 Unit tests of code-gen should be make to accompany
  ☐ Longer-term: define use fstrings for loopy-code generation
      ☐ Also, implement a compiler pre-define in loopy to make life easier
  🗹 Much more in-depth codegen tests
  ☐ Implement long-running tests on Travis
      ☐ At very least, implement a 'very-long' (only) test on Travis
      ☐ Running of performance / validation tester would be nice as well
misc:
  ☐ Better tests of instruction creator
  ☐ Figure out Adept packaging bug for Conda on older servers
Schemas:
  ☐ Figure out how to directly coerce schemas into optionloops / dicts to save trouble
  ☐ Various improvements
     🗹 Allow memory-limit specification in a variable/codegen-platform to implement overrides of global


Near-term plans:
  🗹 Allow memory-limit specification in a variable/codegen-platform to implement overrides of global
  🗹 Tag pre-print version
  🗹 explict-simd
  🗹 mechanism sorting
  🗹 Change data pre-allocation for non-inputs / outputs to only use vector-size * num_threads * array size
    🗹 Along these lines, allow for fixed-size allocations / fully implement "interface" mode for Hekki
  ☐ Switch to int64 to avoid overflow limiting
    🗹 This is currently possible, but I haven't tested yet.
  ☐ Unify the command-line flags w/ the oploops for the test-matrix to make that more sane
    🗹 This has at been somewhat mitigated (the depth / width options have been converted)
  ☐ In this vein, change the text matrix to an object that can be coerced from yaml input directly via ceberus
  ☐ Simplify split vector size division to right-shift and test
    🗹 This is partially resolved, at least for explicit simd
  ☐ Fix linting errors
  ☐ Interface with Chris' RK / ROS solvers for accelerInt