pitrou opened a new issue, #37678: URL: https://github.com/apache/arrow/issues/37678
### Describe the enhancement requested Users should generally prefer using the checked versions of arithmetic functions such as "add", because unchecked arithmetic is inherently unsafe and can silently corrupt data (MySQL is commonly hated for this). However, our checked arithmetic functions are quite slower than unchecked: For example, "add" vs. "add_checked": ```console $ python -m timeit -s "import pyarrow as pa, pyarrow.compute as pc; a = pa.array([42]*1_000_000, type='int32')" "pc.add(a, a)" 2000 loops, best of 5: 128 usec per loop $ python -m timeit -s "import pyarrow as pa, pyarrow.compute as pc; a = pa.array([42]*1_000_000, type='int32')" "pc.add_checked(a, a)" 200 loops, best of 5: 1.92 msec per loop ``` ... that's 7.8 Gitems/sec vs. 0.5 Gitems/sec., or a 15x difference. "sum" vs. "sum_checked" shows a similar problem in PR #37536. Implementation-wise, "add_checked" is currently simple and naïve: it adds and detects overflow on each pair of values, one by one. This adds many conditional branches and probably stifles parallelism on the CPU, as well as auto-vectorization on the compiler. A potentially better implementation would be to operate on K-sized batches (with a K such as 256), in two steps: 1. detect potential overflow over all K pairs of values 2. if no overflow was detected, add all K pairs or values Each step is then unconditional and a reasonable candidate for auto-vectorization. Of course, we can also consider that 0.5 Gitems/sec is good enough and unlikely to be a bottleneck for non-trivial workloads. ### Component(s) C++ -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
