ox-arrays/cbits, branch repro-9.14-branchNested, compositional struct-of-arrays orthotope arrays
https://git.tomsmeding.com/ox-arrays/atom?h=repro-9.14-branch2025-04-15T15:18:34Zarith: Better typing of enum stats_binary_id2025-04-15T15:18:34ZTom Smedingtom@tomsmeding.com2025-04-15T15:18:34Zurn:sha1:9106993eeb2036f1dc5165535e1f2be77c273f59arith: Don't use C23 features2025-04-15T15:18:18ZTom Smedingtom@tomsmeding.com2025-04-15T15:18:18Zurn:sha1:8210378510b92f8ec224c6adcda3ecc77625f1a0Dotprod: Optimise reversed and replicated dimensions2025-03-25T16:09:20ZTom Smedingtom@tomsmeding.com2025-03-25T16:09:20Zurn:sha1:a78ddeaa5d34fa8b6fa52eee42977cc46e8c36a5Separate arith routines into a library2025-03-20T12:01:24ZTom Smedingtom@tomsmeding.com2025-03-20T12:01:24Zurn:sha1:55036a5ea4a6e590d0404638b2823c6a4aec3fba
The point is that this separate library does not depend on orthotope.
arith stats: Print timings with 3 digits precision2025-03-18T22:51:51ZTom Smedingtom@tomsmeding.com2025-03-18T22:51:51Zurn:sha1:5414434df62b2b196354b9748b265093c168601b
If you render microseconds timings as milliseconds, you _have_ only 3
digits behind the decimal point.
arith stats: Improve output format2025-03-18T22:37:56ZTom Smedingtom@tomsmeding.com2025-03-18T22:37:56Zurn:sha1:7883bed5997f430219077202c84af7bf80ada2b7
This makes it nicer to process using unix tools. Try:
$ sed -n '/ox-arrays-arith-stats start/,/ox-arrays-arith-stats end/ !d; /===/ !p' | sort -n -k4,4 -k6,6
Arith statistics collection from C2025-03-18T21:32:16ZTom Smedingtom@tomsmeding.com2025-03-18T21:32:16Zurn:sha1:cb758277b3fa2d74551c45340b8ff0539713078cOptimise reductions and dotprod with more vectorisation2025-03-14T20:58:51ZTom Smedingtom@tomsmeding.com2025-03-14T20:57:56Zurn:sha1:6276ed3c7bcd20c8b860e1275386ecd068671bcc
Turns out that if you don't supply -ffast-math, the C compiler will
faithfully reproduce your linear reduction order, which is rather
disastrous for parallelisation with vector units.
This changes the summation order, so numerical results might differ
slightly. To wit: the test suite needed adjustment.
arith: Remove CASE1, add restrict2025-03-14T13:40:02ZTom Smedingtom@tomsmeding.com2025-03-14T13:40:02Zurn:sha1:639acb0abed995400d984203684e178a11d91fa1
Turns out that GCC already splits generates separate code for an inner
stride of 1 automatically, so no need to do fancy stuff in C.
Also, GCC generated a whole bunch of superfluous code to correctly
handle the case where output and input arrays overlap; since this never
happens in our case, let's add `restrict` and save some binary size.
Add atan22025-03-13T08:28:24ZTom Smedingtom@tomsmeding.com2025-03-13T08:26:20Zurn:sha1:a87c80b1fbaa826142605d0846479c94d6ee2bcc