std::simd types should be regular Deviating from standard value-semantic types, the proposed == operator for SIMD types returns a mask instead of bool. This inconsistency detrimentally effects user experience. https://t.co/qASewaVCpC Coauthored with @__qualname__

If you want to deep dive on the string concatenation optimization I mentioned a few days ago, Joe has all the details:

Thanks to our lightning talkers yesterday: https://t.co/FhGmV6MD2H @MattBachmann @chrisremmel @nedbat @glindahl @__qualname__ @bhrhugs

I've been working on an experiment to add arena allocators to Python. The goal is to reduce the time spent allocating and destroying objects. The motivating use cases are compilers/optimizers that work with AST objects and web server response handlers.

Maybe someone from @numpy_team knows if this is possible?

I was following these docs: https://t.co/MZnduJN1bG and was able create a dtype with scalar conversions to and from PyBytesObject. This dtype supports various operations like np.empty (using NPY_NEEDS_INIT) and np.sort. Right now each element with size > 8 leaks the data ptr.

Does anyone know how to implement a user-defined numpy dtype for types with non-trivial destructors? I'm trying to make a dtype for: struct sbo_string { size_t size; union { char* ptr; char arr[8]; } data; }; I need to be able to free data.ptr

Having someone open a PR to fix the docs in an open source project you started feels great. Having another contributor review and merge the pull request is even better.

Someone described a feature I wrote as: "... probably the craziest python code I have ever seen in my entire life." Strong endorsement.

C++Now 2019: “test_resource: the pmr detective” by Attila Fehér https://t.co/eqmhzA2yyi #CppNow #cpp #cplusplus

This was also done without making the jlist any larger. I made the tag for homogeneous_ob = 0; and store it in the lower 3 bits of a PyTypeObject*. If the tag == 0, then the pointer holds the PyTypeObject for each element. PyTypeObject is 8 byte aligned so these are always 0

I added support for homogeneous collections of PyObject to jlist. This speeds up elementwise operations (containment, any, sum), as well as sorting for boxed lists. This is similar to the optimizations that list.sort already has, but jlist maintains this type from the start.

Last night I wrote a near drop-in replacement for Python list objects that internally stores unboxed int64s and doubles if possible. Calling append will re-box if needed. It uses the unboxed repr. for faster ops. It even passes the CPython list test suite

Found this gem in the HDF5 frequently asked questions: "Performance-wise, how does HDF5 compare to a relational database? It really depends on your application. HDF5 is tuned to do efficient I/O and storage for "big" data (hundreds of megabytes and more). ..."

I have a ticket for pax tomorrow but can't go, lmk if you are interested.

JLPT N4を合格しましたけどまだ日本語ができません

God bless @scottbsanderson for this wonderful gift! I can't wait to enjoy a delicious meal at my neighborhood Chilli's Grill & Bar!

Super proud of my team at @quantopian today. We announced our first major milestone for international data: pricing and fundamental data for 21 (!) markets: https://t.co/Xme2FghG9F. Excited to see what happens as we expand our (already very global) community around the world!

$ git fetch * [new branch] somebody-once-told-me-the-world-is-gonna-roll-me-i-aint-the-sharpest-tool-in-the-shed -> origin/somebody-once-told-me-the-world-is-gonna-roll-me-i-aint-the-sharpest-tool-in-the-shed @__qualname__ is trolling my bash history now...

First 1e6 integers, represented as binary vectors indicating their prime factors, and laid out using the sparse matrix support in @leland_mcinnes's UMAP dimensionality reduction algorithm. This is from a 1000000x78628 (!) binary matrix. Very pretty structure emerges.