01:21 gfxstrand[d]: mhenning[d]: What's up with that first patch that lets you combine sets with different key types. That seems... Odd.
02:20 mhenning[d]: gfxstrand[d]: oh, it was meant to let you combine sets with the same key type (right now it's uint only) but maybe I didn't constrain the types enough
02:20 gfxstrand[d]: Oh, maybe I misunderstood
02:22 mhenning[d]: no, I think you're right, although it's a problem that exists before that patch
02:23 mhenning[d]: none of the BitSetStream stuff got constrained when the type parameter was added
02:23 gfxstrand[d]: Right. BitSetStream needs a generic type.
02:23 mhenning[d]: so you can do binary ops on different types already
02:23 mhenning[d]: yeah
02:23 gfxstrand[d]: Should probably fix that while you're in the neighborhood
02:24 mhenning[d]: yeah, something to do on monday
02:24 gfxstrand[d]: 👍🏻
10:18 kar1m0[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1378678970668548126/Screenshot_20250601_131726.png?ex=683d7a59&is=683c28d9&hm=d039d55f9f884c18c9cbe5e97b545a0265e2af1898c65aadf7ba7a823bd8122f&
10:18 kar1m0[d]: GOverlay partially works with nouveau!
10:18 kar1m0[d]: the gpu usage and memory speed of the gpu can be seen
11:58 capreplytimeout: I suppose precision qualifiers are for saving memory as well as bandwidth, and those last are used interchangeably, so in essense in that context you configure the memory controller to fetch per bytes or two bytes and this configuration can be per varyings and uniforms as well as framebuffer objects, cause depth only accessible in fragment shader had it long time ago, stencil samewise.
11:58 capreplytimeout: But i actually think it's even there the same packet in underlying hardware its copy engine or hoewever called dma cp. Not too sure about pcie bridges where no dedicated memory on board, then it tends to be pcie async lines or host memory controller altogether. I still think that even earlier intel's atom processor gpus had though vertex shader on cpu, but fragment processors had it's
11:58 capreplytimeout: own dma controller to the ddr3 memory, i.e they were not locking the processor ones. But i never knew how it was implemented it showed out as pci-to-pcie bridge in sw though.
12:22 kar1m0[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1378710358356787230/794a1350-7bcd-47c8-b819-2b02a93f44cd.png?ex=683d9794&is=683c4614&hm=9b5a85409bb063efaf9eec519d6edec8131a7f20b479ec4b327da24aae0ef55c&
12:22 kar1m0[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1378710405580591135/0ac756e4-d3f6-4c71-849a-ecb7363c347e.png?ex=683d979f&is=683c461f&hm=00ff0a2cd587501a8e5042a1364d4cb75c7f63aaedc6bd8218fe4719ed30cbc6&
12:24 dalaimala: It's capable of driving any width of framebuffer regardless, only limited to defacto 24-bit display controller according to google AI. and has framebuffer object as well as registers already 32bit the least, hence you need to put more thought as i put into your programming. So precision qualifiers is not hardware limitation of displays, but just smart specification to save memory, this can be
12:24 dalaimala: overcome with hashing as to way i showed, and is exactly the most thoughtful way to configure the shader internally, but output converted non packed value to the display.
13:05 dalaimala: But i got sober realizing we are equal, you have not been that nice to me, i responded with same, but all of this is some waste of time, i get in with VPNs still, and failed minorly bit more before i got specification together. Since the specification seems to hold in it's entirety all this can be flushed with me leaving, and it's just pointless to argue about your take on mental illness as
13:05 dalaimala: it is pointless to argue about any bits as to what you prefer to have , kar1m0 individually you can still reverse engineer anything in eu or us or wherever, you are not allowed to distribute. Bunch of licenses and stuff on the way, maybe even on directX i was wrong, now i remember the microsoft development kit where directx was part of it, can be distributed, drivers can't be or something,
13:05 dalaimala: API can be translated in between.
13:06 dalaimala: So my approach was all pointless , to get others to grasp what i do, since no one picks to be dick almost no one because one is stupid, but more like cause one likes to be dick.
13:16 dalaimala: So basically in trying to teach you as to what i learned i bit my own tail even more than before, but i try my best not to go there anymore, i think the specifications are solid as of now , theory that computers can be even more than they are today isn't seeming to be that much of surprise to anyone. Maybe even the clown was me, who was just played with.
13:16 dalaimala: that is the world of mindbuggling.
13:17 dalaimala: You are every day behind computers performing some work, i suppose you already knew everything i did tell you.
13:27 dalaimala: AMD GCN as well as intel and those cause it's dma controller i suppose the fbo and uniforms still being one packet, it's the final thing as to how to support decoders and encoders on gpu without bitwise, instead of 32float granularity one can fetch two bytes with stride to one register or however it wants, hoewver those bits are deep in the hardware driver in case of desktop gl, but mesa
13:27 dalaimala: offers all GL's of embedded ones too, so now here we go, we still have everything.
13:29 dalaimala: So i imagine the sign extend and zero extend are in the hardware for most chipsets, and converting between high and whatever precision is just a matter of configuring the controller.
13:39 dalaimala: Yes, it's possible to use both highp and lowp precision qualifiers in the same shader in GLSL ES (OpenGL ES Shading Language). However, the choice of precision is crucial for performance and accuracy, especially in mobile devices.
13:39 dalaimala: Explanation: the AI output of this i discarded the explanation, so technically done deal.
13:41 dalaimala: so you would always start from highp with something then lift to low-precision in some slots , they likely use two dma channels to configure the controller and call it a conversion shader between opengl desktop on opengl es.
13:43 dalaimala: all mistakes that opengl desktop did, were fixed in opengl ES none the less, but mesa had all opengl ES drivers available, so i see no problems there.
13:47 dalaimala: and it is not actually a conversion it would do both execution in new paradigm and converting everything, there is just enough instruction slots to merge that stuff
14:08 vboglessurface: there is a function called glMapBuffer in opengl ES, i either read the thing wrong or it's unspecified, but opengl ES let's in the specification level to map a surface or texture or a thing alike into vram, where as you can share a VBO with one of the FBO attachments, which would allow you to escape the graphics hell so passthrough fragment shader can be written. Now you thought fragment
14:08 vboglessurface: shader runs that much more, there was another hack of doing it on a quad of how many pixels wide needed, maybe it required glviewport i can't remember, but the thing is now you can render everything in fragment shader i tested none of this however, but i expect this to function.
14:09 vboglessurface: VBO can render geometry without shaders too as i remember, so all the transformation you end up doing in fragment shader hashes, but on FBO only 4 colors per run can be output which is kinda lame without texture write mem or buffers extensions.
14:59 convertallcorvette: but how to convert those hashes to be functional on FP24 is where i do not know , that is my final task. It needs a converter from somewhere, my gnome-calculator does not show bits of float where even negative twos complement values are taken there with a function.
17:45 Heisendracula: now the part to lift those packing procedures to floating point isn't expected to take very long time, but i am in the middle of another bully case at my home, my jewish grandmother kicking me out i no longer serve the interest of this leech. However it's that exponent/signiffcand power bits are marked in format like 5 9 13 17 21 25 29, where the sum is below 128 luckily enough, and i
17:45 Heisendracula: suppose mantissas powers could be 0.5 0.9 0.13 upto 32-7 of those from ideal distribution except the ones we already used for exponent. But myself i am too tired to test it at the moment(i would like to see how much memory it takes from following the bits), i need a new setup to do that.
19:54 integerunits99: Ah this design of floating point and saturation on FP24 i.e minimum on OpenGL ES shading language isn't very much fitting stuff, exponent is not carried along to other bits and integer is often 16bits on highp since this meets minimum requirements, it's not fitting more alot more than ten 32bit values, due to the fact. it's not that much worth to deal with such equipment , in my opinion
19:54 integerunits99: i am not fan of such gpus, and i would understand floating point's need if the mantissa on saturated outputs could grow over the account of exponent, like there was a floating on request mantissa, but such stuff does not exist. opengl es3.0 already 32bit integers.
22:43 snowycoder[d]: mhenning[d]: That would be almost perfect, but I need forward data-flow analysis.
22:43 snowycoder[d]: Maybe we can generalize it in a common trait and two sub-traits that implement only the forwards and backwards algorithms(?)
22:49 mhenning[d]: My plan was to just write a second Trait for forward dataflow
22:50 mhenning[d]: Not really sure how to share code between forward/backward in a meaningful way
23:24 snowycoder[d]: We could've shared the initialization + worklist logic (making `transfer` and succ/pred generic).
23:24 snowycoder[d]: I've put together a PoC but I forgot about specialization and now the compiler complains that one could implement both forward and backward dataflow -.-
23:25 snowycoder[d]: (*for the same struct)
23:25 mangodev[d]: i can't tell who's real and who's not anymore 😭
23:26 mangodev[d]: hdkr is the only one i know that's not one of the many alts in irc
23:26 HdkR:dances
23:26 mangodev[d]: see
23:26 mangodev[d]: human response
23:28 snowycoder[d]: mhenning[d]: Yeah 10 lines of duplicated code aren't worth the abstraction + type confusion, sorry