Apparently it's #MyTwitterAnniversary, so I'll use it as an opportunity to link to the Mastodon account I now use:

Quick reminder that I'm mainly t[oo|wee]ting on Mastodon these days:

Me on Mastodon: I don't plan to disable this account, but I'm generally posting there these days.

Do you know about automatic differentiation and / or Enzyme? Could you help sketch out how AD support for @numba_jit could be implemented? Issue / thread:

I've been on Mastodon a while but I got out of the habit of using it - this is my account that I'm starting to get active with again: (is this usually written as @gmarkall@mastodon.social?). It has an old profile pic, still need to upload my current one.

@anthonypjshaw 6. Finally, maybe this should have been a blog post with some code! Would that be interesting? What else should be answered / elaborated on if I write this up?.

5. "CPython Internals" by @anthonypjshaw is a great intro / reference to CPython, saved me a lot of time, and got me up to speed quickly for this endeavour. Highly recommended if you want to poke about with this sort of stuff! (.

4. Performance is not good for my naive implementation - literally every Python alloc results in a CUDA driver API call, instead of allocating arenas and doing other clever stuff that the Python allocators do. Though, that can be solved, maybe reusing a lot of what CPython does.

3. Context management a bit fiddly - e.g. if the allocator is the first use of CUDA it needs to cuInit and set up context, but then other libraries like Numba can be unhappy. Also other libraries switching the context (such as Numba again) needs to be handled gracefully. .

2. Using the CUDA Memory Management APIs (cuMemAllocManaged etc) was straightforward to implement the required methods (malloc, calloc, realloc, free). cuMemGetAddressRange was convenient for realloc. I used the driver API, to keep things simple (for me):

1. The PyMem_SetAllocator API ( perhaps doesn't support as broad a range of use cases as I'd like - perhaps intended mainly to support tracing and debugging? . I'm not sure I can turn off my allocator once it's on - I see no way to migrate allocations. .

I've been experimenting with using CUDA Unified Memory for the Python heap, towards a general goal of experimenting with a more unified CPU / GPU execution model within CPython, implemented as PyMemAllocatorEx instances. I have a few thoughts so far. .

RT @RAPIDSai: Do more on GPUs with less code - check out @gmarkall's new blog on the @numba_jit high-level API

RT @numba_jit: Public service announcement: Yes, the Numba team is aware of the Python 3.11 release. 💥 Yes, we are working on it. 💪Please….

Lil identify xei9pa2AeQu6.

RT @__mharrison__: Do you use the GPU in Python? If so, for what, and what library do you use? 🤔.

"Use a font that is way too small" - every corporate Powerpoint template for technical talks ever.

With @numba_jit 0.56 you can use the High-level API to extend the CUDA target. Much simpler to use than the low-level API, this notebook shows how to use it for some quick examples: HLA docs: Pic: using it to implement clock64()

Money-saving tip: instead of buying an expensive chef's knife, simply take a PCIe slot blanking plate out of a cheap PC case.