...because you only have to clip primitives that have both pixels in the viewport AND extend beyond the margin (the guard band). This should be very rare, and in your case you can just assume that it never happens.

Show this thread

Won Chun

@won3d

Mar 22, 2021

The idea of guard bands for rasterization: clipping in pixel space is easy (simply don't visit pixels outside viewport) but your rasterizer has limited range, so you sometimes need to process the geometry. However, if your range has some margin, then you rarely have to do this...

Negabinary! Looks kind of related to ZigZag encoding with some interesting alternation of bits. Also, does this mean there are nega p-adics? twitter.com/pinkhairedcyn/

This Tweet is unavailable.

If you’re in STEM and you support #BlackLivesMatter, please retweet and like this! I’m trying to see something 🚀👩🏾‍🚀🛰

12K

14.5K

Won Chun

@won3d

May 7, 2020

Wow, this thread is super cool! h/t

Quote Tweet

Very excited to share #SIGGRAPH2020 paper w/ @rohansawhney1 on "Monte Carlo Geometry Processing" cs.cmu.edu/~kmcrane/Proje We reimagine geometric algorithms without mesh generation or linear solves. Basically "ray tracing for geometry"—and that analogy goes pretty deep (1/n)

Show this thread

*mem hah

Anyway, I was thinking recently that there is very little language support for heterogeneously structured things packed together. Like VLAs and that's it? Risky things with virtuals and placement new?

Speaking of compacting men, I question the idea that you'll run out of mem bandwidth with a only about one core accessing memory. I mean I guess that could have just been a proxy for cache hotness/utilization, but not bandwidth

Like you could cache and gather the results of your 2 pass decoder to save on indirections/compact working memory, avoid linked structures, etc.

There could be generic traversal apis as well as more cooked, contextual ones (else after then). There can be different representations, either as you experiment, or even maybe different trees get written out differently based on access pattern

Also it seems like you could alleviate a lot of worry about your tree rep by using accessor inlines/macros to do the tree nav. If you have tags or pointers that are elidable in mem, that can all be hidden by the accessors and give you flexibility.

I didn't get through the whole video, but DWARF's on disk representation is similar to your tree layout, although (usually) uses varint section-relative offsets. Also have "sibling" offsets to skip past subtrees.

These vector panels at #SIGGRAPH2019 are magic

Won Chun

@won3d

Jul 22, 2019

Who else is going to #SIGGRAPH2019 ?

Won Chun

@won3d

Mar 3, 2019

Getting a massive en.m.wikipedia.org/wiki/Chromoste effect from this GIF on my phone

Quote Tweet

bmcnett

@bmcnett

Mar 3, 2019

sous vide breakfast meats... QUALITY UP!

GIF

Won Chun

@won3d

Jan 7, 2019

Asus announced a 17" 16:10 narrow-bezel laptop workstation with a centered keyboard. The numpad shares spaces with the trackpad which unfortunately lacks physical buttons.

notebookcheck.net

Slim and light Asus StudioBook S (W700) offers Intel Xeon CPU and Nvidia Quadro P3200 GPU

Asus has revealed the StudioBook S (W700) workstation that comes with either an Intel Xeon E-2176 or an Intel Core i7-8750H processor and an Nvidia Quadro P3200 graphics card. The powerful 2.39...

This NeurIPS talk by

is blowing my mind! I don't think I can summarize this in a tweet except that biology has come a long way. youtube.com/watch?v=RjD1aL

I'd tell you a TCP joke bu

Won Chun

@won3d

Sep 25, 2018

64 entries are for 4K pages, which is 256K. If you use 2M pages, you get 8 entries, or 16M. With 1G pages you get 4 entries or 4G. So the lesson is: use bigger pages if you can!

Won Chun

@won3d

Sep 25, 2018

Had a convo yesterday where someone claimed that L1 TLBs were teeny tiny (single-digits) so I had to look it up. According to wikichip.org, SKL/SKX L1 dTLB is 64 entries, 8-ways, statically shared between hyperthreads, so effectively 32 entries. Never big enough!

That time Fred Rogers shook a theater full of celebrities like Anton Ego tasting ratatouille

youtube.com

Fred Rogers Acceptance Speech - 1997

Fred Rogers Accepts the Lifetime Achievement Award at the 24th Annual Daytime Emmy Awards.Throughout his career Fred Rogers, both on television and off, emph...

Won Chun

@won3d

Aug 30, 2018

I should note that the P52 only "throttles" during turbo (which is to say: it doesn't actually throttle badly at all). It has no problem maintaining base clocks while silent, which is pretty great.

Anyway, I wish they had a 3:2 display option and a centered, 7-row keyboard and all sorts of ponies, but I definitely feel good about getting many good years out of this. P52 thumbs up!

I got the 1080p display, which is fine. Supposedly P51 1080p was abysmal. I was willing to chance it on the P52, because mixed DPI is not really ready. Mixing 4K on the AW17 and my monitor was not a good experience. I ended up running at 1080p anyway.

Fortunately, I kept the original windows partition around so I could boot back into the spindle, activate there (maybe it came activated...?), and hope for the best. It worked, although the activation process is so opaque that I'm not even sure if that's what did it.

The first thing I did was reimage the spindle drive on the the NVMe (Samsung 970 EVO) since ain't nobody got time for rotational latency. But, I almost got locked out of activation. Turns out you should activate and use a MS online account before you migrate!

The machine feels very well built, and the internals are very easy to access, clearly made to be upgraded and serviced. The first thing I did was add an NVMe drive, since the upcharge on NVMe is way too high. I almost got into trouble because of this!

Non-centered keyboard. Well, this is a little annoying, especially since apparently they shrank the keys since the P51, but the keyfeel is excellent. Trackpad is fine with discrete buttons. I love the Trackpoint, though, which goes a long way to forgive these ills.

I haven't done much testing of the Quadro 3200 Max-Q, but it is an interesting part. It does about as well in 3DMark Time Spy as a 1060, but the configuration is pretty different. More cores (still less than a 1070) and less bandwidth (192-bit like the 1060 but lower clocks).

Fortunately, incremental build/link is relatively bursty, so I do get some benefit from the higher turbo clocks of the 8850. Undervolting -0.1V helps so that it stop power thottling and delays the onset of thermal throttling. YMMV of course,

I got the P52 when customization became available. To get the best GPU, Quadro 3200 Max-Q, you also need to get the 8850H CPU. It doesn't come with the 8750H or the Xeon. So the question is: will it throttle? Answer: of course.

The SKL AW17 seems to have bugs with mixed GPU. Doing stuff in explorer sometimes causes weird hitches as the dGPU turns on and off. Sometimes apps mysteriously crash in deep sleep. Forcing the dGPU seems to fix these issues, at the cost of fan noise. Performance is good at least

AW17's downside is obvious. It is heavy and huge and the hinge forward design means it won't fit in most large laptop bags. It isn't comfortable on a lap -- you really need it on a table. You really don't want to lug this thing around, not even to the next room or anything.

I just got a new laptop: a Thinkpad P52. Here are some thoughts since there isn't too much info on it. But first, shoutouts to

@ThaLobsta

for loaning me his Alienware 17 for a few weeks. Some thoughts on that unit as well.

This concludes your Friday procrastineering thread.

Edgerouter-4 looks amazing but had the following downsides in my setup: 1. expensive overkill compared to ER-X (still looks to be a good value) 2. no integrated switch 3. no PoE 2 and 3 are obviously solved with a PoE switch, but that means more $ and power. ER-X is too tidy!

Show this thread