Tracking issue for migrating to use mlx-c

[minghuaw]

Initial implementation steps

• stream
• array
• ops
  • logical - feat(api): Arithmetic & Logical ops #39, feat(api): Add remaining logical ops #43
  • arithmetic (partial) - feat(api): Arithmetic & Logical ops #39, Add remaining arith ops #78
  • shapes ( feat(api): indexing + shapes #52 )
  • reduction - feat(api): Arithmetic & Logical ops #39, feat(api): remaining reduction ops #53
  • sorting (Impl sorting related ops #66)
  • convolution - feat(api): convolution ops #47 (missing convolve, depends on shapes)
  • cumulative - feat(api): cumulative ops #46
  • conversion (feat(api): impl ops/conversion #44)
  • indexing ( feat(api): indexing + shapes #52)
  • index mut (impl IndexMutOp #65)
• fft - feat(api): add fft #37
• linalg (feat(api): linalg #51)
• random (feat(api): random #63)
• transforms (feat(api): adding transforms #26, Added async_eval #83)
• quantization (Add quantization ops #81 )

Problems found

• Document that mlx is lazy
• Perform type conversion for item()
• Perform type conversion for as_slice (Unlike item(), we cannot use mlx_astype on the inner pointer here because it creates a new mlx_array which should be freed at the end of this fn, making it the returned slice pointing at a temporary value.)
• Split as_slice into three versions
  • as_slice<T>(&self) -> &[T] - checks if the inner dtype is the same as the output dtype, panics if different. Also panics if pointer is null
  • as_slice_unchecked<T>(&self) -> &[T] - unsafe and does not perform any check
  • try_as_slice<T>(&self) -> Result<&[T], Error> - checks the type and returns an error if the types don't match or ptr is null
• Indexing array (read). mlx allows out of bound indexing (Possible mx.array bug?  ml-explore/mlx#206)
• Indexing array (write). Check if out of bound indexing is still allowed
• Revisit all error types (when everything is more or less ready) not needed with the new error handling
• Revisit whether *_unchecked functions should be unsafe not needed with the new error handling
• Revisit eval() (https://ml-explore.github.io/mlx/build/html/usage/lazy_evaluation.html)
  • Calling eval() on mlx_array or mlx_vector_array multiple times is allowed (is &mut really required)
  • Check if implicit eval() takes place on the c++ side
• (Related to eval()) Whether Array should be Send and/or Sync (impl Send for Array & fix broken links in docs #64)
• Revisit Clone for Array
• Implement PartialEq for Array using array_eq
• Add unit test to cover this case with indexing ( feat(api): indexing + shapes #52 (comment))
• Avoid direct use of the Add or Sub operators (ie. +, +=, -, -=), use alternatives like saturating_add etc instead. This is probably not needed anymore with the new error handling
  • Which should be used then? saturating_*, checked_*, overflowing_*, unchecked_*, wrapping_*
• The ops macros could probably be generated automatically. However, declarative macros cannot be defined in impl or trait blocks.
• Review args that take ownership of Array to check if they really need ownership (if all they need is a ref, change them to .as_ref() in the corresponding macros)
• Review default dtype (with python) for factory ops macros
• arange is missing from the constructors for Array (feat(api): add arange #72)
• Fix bug with as_slice when the underlying array is not contiguous #77
• gpu support for fft ops should be available in mlx 0.15
• What functions should take impl ScalarOrArray as argument? (see ScalarOrArray in the swift binidng)
• Some exceptions cannot be displayed properly with the new error handling (Fixed problem with exception not displayed properly #86)
• Revisit whether the deep clone in index and index_mut is necessary
• Add From<complex64> for Array