| Tweetovi |
|
damageboy
@damageboy
|
40 min |
|
Would be great if MSFT could either share official JIT/codge-gen plans, or alternatively at least say when things would cement...
|
||
|
|
||
|
damageboy
@damageboy
|
41 min |
|
That would be so weird. I used to rely on Mono/LLVM for a living, having so many seemingly unrelated code-generation methods depending on what uninformed users copy-paste from the internet is slightly scary.
I assumed everything would be joined with tiered compilation...
|
||
|
|
||
| damageboy proslijedio/la je tweet | ||
|
Andrew Clay Shafer 雷启理
@littleidea
|
1. velj |
|
OH: weeks of coding can save you hours of planning
|
||
|
|
||
|
damageboy
@damageboy
|
7 h |
|
And sponsoring it, also it appears!
|
||
|
|
||
| damageboy proslijedio/la je tweet | ||
|
Internet of Shit
@internetofshit
|
14 h |
|
99 smart phones in a pull-cart = instant traffic jam simonweckert.com/googlemapshack… pic.twitter.com/gNDa05OeVi
|
||
|
|
||
|
damageboy
@damageboy
|
24 h |
|
I've been informed by my blog review team (population count: 1) in no uncertain terms that I should either budget adult diapers for my readership, or split pt. 4 into 3 more parts.
Given Damage Open Source Industries' dire financial circumstances, I'm choosing the latter :)
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
It's funny how I started, a least in my mind, thinking about the problem as a quick-sort with vectors, and ended up acknowledging that even for huge array sorting, it's mostly about partitioning down to small enough groups for small-array sorting to take over.
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
I think his is as vague without being misleading as I need it to be? pic.twitter.com/xdCJ363PDn
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
Yeap. This is a reasonable outcome.
In some ways it should be slightly easier given that sorting is a relatively demanding task, so everyone's expectations should be dialed in into a more accepting mood ;)
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
You and me both.
There will be a draft PR.
I expect (rightful) headwind w.r.t code-size.
Let's hope I can swing that 11x carrot convincingly.
Seriously, there's a case to be made for cutting down 3/4 of the code for 8x.
We should add speed-up per machine-code metric to BDN :)
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
Yeah, no pressure here. None at all :) twitter.com/trav_downs/sta… pic.twitter.com/LjZW3Pxv0M
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
I'll rephrase and ping you for some validation :)
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
4 is mostly done. I have one last epic failure to document...
5 is about dealing with small sorting, where I document both the stable+unstable small sort approaches...
Pt6 will have a little sorting "a-ha" moment, not sure if I can convert it to better perf yet though.
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
I'll clarify right then and there. I wanted to leave a small cliff hanger at the end there...
Maybe I cut it too short.
My fight with the predictor is actually mostly resolved with unrolling, but I'm still trying to squeeze a little more in pt6...
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
I think (still researching) that in pt6 I can really show something I have not seen before about quick sorting and branch prediction...
On that note, I also think I have a minor novelty in the middle on pt4 re. alignment...
Tune in, it will hopefully be entertaining...
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
Well, it's in pt4, but the loop unrolling pays in spades here since that branch is sneaking in a big mis-prediction penalty since it's essentially depending on the random input data...
But unrolling cuts that rate since we can change the amount of mis-prediction per work...
|
||
|
|
||
|
damageboy
@damageboy
|
1. velj |
|
What I've "published" for now is the relatively "easy" stuff. I think of this, at best, as a very well explained+animated SIMD series based on the paper I linked in the first post.
I hope that some of the stuff I have in pt. 4+5 might constitute as "new" ideas, BTW
|
||
|
|
||
|
damageboy
@damageboy
|
31. sij |
|
I'm still left wondering though how many people got the 11 reference.
I thought it had a nice self deprecating touch, but am fearful people just missed it...
|
||
|
|
||
|
damageboy
@damageboy
|
31. sij |
|
This sort of things gets my blood pressure up instantly:
gisanddata.maps.arcgis.com/apps/opsdashbo… pic.twitter.com/AX3LWkhj3S
|
||
|
|
||
|
damageboy
@damageboy
|
31. sij |
|
It's all screwed only in so far as much you can lie to yourself by calling it a _compression scheme" to keep cache and RAM pressure down :)
|
||
|
|
||