Rust wgpu now has a
-
@MaddieM4 @ireneista *is* that done on on-board controller silicon? I would have assumed it's done in a CPU-driven userspace thing installed in to the regular operating system alongside the driver.
Like I'm certain they *could* do it with on-board controller silicon, if they wanted, but even then I suspect that on-board controller silicon would look a lot like a small ARM running a CPU-based compiler in a tiny Linux
-
@aeva @mcc if you want to start from historic, fatmap.txt is a pretty fun read (and you can then make your renderer look like it’s genuinely old and needs small polygons and has ordering issues and no perspective divide, which is also fun, but at the same time don’t actually have to optimize *that* much)
-
@ireneista @MaddieM4 @mcc @decay @amarioguy anyways, i _really_ struggle to actually *talk with* people who don't do well with this approach, and with social reach in general, so i haven't invested too much into this, but some patterns i've noticed nonetheless:
* some people just.... don't.... like.... symbols??? (i've been realizing how difficult a concept "abstraction" really is, and how much of the "cabal of Elite
️ universities" pressure (and toxicity) is in part an attempt to force you to learn their particular style of doing abstraction. but, otoh, the part i _really_ don't get is that there seem to exist people who don't like symbols in "equations" but are perfectly happy with symbols in computer code or on a graph)* a lot of this is insufficiently physical (even a project-based curriculum *run by the EE department* doesn't usually have as many fiddly doohickeys as "musicians" seem to like)
* everything is connected to everything else. this is either spelled out in *too much* detail or *too little* detail, never "just right". at the same time, everything is either too much "you are smart, you can do the computer-y bits by yourself" or else too much "here is something to plug into the computer", never "just right" either. (in case examples help: using pencil-and-paper to do the derivation from FT->DTFT->DFT is *really cool*, but unless you're already in the "right" mindset it just looks like shuffling a bunch of very similar very confusing symbols. afterwards, you either _never_ mention e.g. the python-scientific-stack or else you hand everyone a pile of python/matlab code for homework (where you need to fill some part in) )
@ireneista @MaddieM4 @mcc @decay @amarioguy oh also
* digital modeling of analog systems is *fiendishly* hard (or at least interdisciplinary), which makes it hard to fit into an EE-department-run curriculum. but musicians seem to have a *particular* interest in doing this
* ime the EE department (or at least ones that have, uh "communications" (and "war" of course) heritage) really focuses on various forms of accuracy and optimality. we have been discovering that, apparently, musicians *do not*. this implies that "optimizations and hacks" are not covered
-
@MaddieM4 @ireneista *is* that done on on-board controller silicon? I would have assumed it's done in a CPU-driven userspace thing installed in to the regular operating system alongside the driver.
Like I'm certain they *could* do it with on-board controller silicon, if they wanted, but even then I suspect that on-board controller silicon would look a lot like a small ARM running a CPU-based compiler in a tiny Linux
@mcc @ireneista it's not currently done with on-board silicon, it's just that there's capacity. Conventional GPUs work more like you're describing, doing that kind of work within the driver on the CPU.
I think it *should* be translated on the GPU controller, and that the reasons not to are more about proprietary economics than technology. I also think Linux is overkill for this, but Linux-compatible microcontrollers are so cheap at this point that it wouldn't be an insane way to implement it. To me, part of the point is establishing a stable (but extensible) binary standard for both programs and high level control commands, so that GPUs are a standardized micro-appliance. To me it's not so different from microcode allowing many generations of AMD64 with different microarchitecture to all speak AMD64.
-
@mcc @decay @amarioguy @MaddieM4 Fourier and wavelet transforms are really important though, even though most programs don't do them. but you raise a good point.
@ireneista @mcc @decay @amarioguy @MaddieM4 Fourier transforms will also happily run at GB/s rates on any modern GPU.
VkFFT is awesome.
-
@ireneista @MaddieM4 @mcc @decay @amarioguy oh also
* digital modeling of analog systems is *fiendishly* hard (or at least interdisciplinary), which makes it hard to fit into an EE-department-run curriculum. but musicians seem to have a *particular* interest in doing this
* ime the EE department (or at least ones that have, uh "communications" (and "war" of course) heritage) really focuses on various forms of accuracy and optimality. we have been discovering that, apparently, musicians *do not*. this implies that "optimizations and hacks" are not covered
@ireneista @MaddieM4 @mcc @decay @amarioguy oh, and, the "rest of the fucking owl" *software engineering* of computer music is essentially *not culturally connected* to this cluster of EE at all. this is before we get to the part where a bunch of this is """just""" convention and so the "cabal of Elite
️ universities" don't value itthink things like:
* "what is VST?" (what is an "API" more generallly?)
* "what is a real-time system?"
* "what is MIDI?"
* "what is DMX?"
* "how do i get signals from the real world into and out of my computer?"
* "what is a protocol, more generally?" -
@ireneista @mcc @decay @amarioguy @MaddieM4 Fourier transforms will also happily run at GB/s rates on any modern GPU.
VkFFT is awesome.
@ireneista @mcc @decay @amarioguy @MaddieM4 the big advantage of simd on modern cpu/gpu systems is very small inner loops and things like memcpy where the overhead of going to GPU and back overcomes the speed benefits
-
@ireneista @MaddieM4 @mcc @decay @amarioguy oh, and, the "rest of the fucking owl" *software engineering* of computer music is essentially *not culturally connected* to this cluster of EE at all. this is before we get to the part where a bunch of this is """just""" convention and so the "cabal of Elite
️ universities" don't value itthink things like:
* "what is VST?" (what is an "API" more generallly?)
* "what is a real-time system?"
* "what is MIDI?"
* "what is DMX?"
* "how do i get signals from the real world into and out of my computer?"
* "what is a protocol, more generally?"@r @ireneista @MaddieM4 @mcc @amarioguy me, perking up from behind my behringer x32
-
@ireneista @mcc @decay @amarioguy That's a good particular example. A part of computing freedom that I think is pretty important is "being able to make music with computers, ideally in realtime." I have that on my list of vital user stories for my someday project microkernel OS. The arts are the kind of thing that make computers worth doing.
A lot of realtime music plugin stuff depends of Fourier transforms and SIMD in general.
@MaddieM4 @ireneista @mcc @decay @amarioguy Music, as well as speech synthesis!
-
@MaddieM4 @decay @riley @ireneista @amarioguy *rubbing eyes* is SPIR-V an ISA? I thought it was literally a fork of LLVM IR
That is a very different abstraction level with very different forward compatibility commitment risks
@mcc @MaddieM4 The original SPIR was a fork of LLVM IR. SPIR-V is more or less "GLSL encoded as an SSA-based VM bytecode". @decay @riley @ireneista @amarioguy
-
@ireneista @mcc @decay @amarioguy @MaddieM4 the big advantage of simd on modern cpu/gpu systems is very small inner loops and things like memcpy where the overhead of going to GPU and back overcomes the speed benefits
@ireneista @mcc @decay @amarioguy @MaddieM4 like in ngscopeclient we ended up deleting a significant amount of the AVX code because it was slower than doing it on the GPU and was just one more implementation to keep in sync with everything else when we made changes, test, etc.
-
@r @ireneista @MaddieM4 @mcc @amarioguy me, perking up from behind my behringer x32
@decay @ireneista @MaddieM4 @mcc @amarioguy it's *still* wild to me that there exists audio gear that costs more than something like an eNodeB
-
@mcc @decay @riley @ireneista @amarioguy Architecture revision is a very reasonable good faith assumption to make, which GPU companies IMHO do not even slightly deserve. It's just an IP protection thing.
SPIR-V would not exist, in an extremely successful way, if GPU ISAs were a bad idea for technical reasons. It exists nestled into Vulkan as a compromise with IP protection.
@MaddieM4 @mcc @decay @riley @ireneista @amarioguy GPU shader processor architectures did go through a few different evolutions so there was something to that but I think they've more or less stabilized on what is and isn't a good idea (access memory through a plain pointer, do SIMD between shader invocations, not within one, that sort of thing). But even so there are a lot of subtly different approaches to things like flow control or sampler state to give just two examples.
-
@decay @ireneista @MaddieM4 @mcc @amarioguy it's *still* wild to me that there exists audio gear that costs more than something like an eNodeB
@r @ireneista @MaddieM4 @mcc @amarioguy our production kit is built around a Behringer x32 Producer which is not cheap but not crazy expensive, but you're basically talking a mildly chunky computer attached to a shitload of power amps that has to be able to bounce around in the back of a truck for a substantial amount of its life and work perfectly as soon as power is applied, which is really what you're paying for. Not quite aerospace but close
-
@r @ireneista @MaddieM4 @mcc @amarioguy our production kit is built around a Behringer x32 Producer which is not cheap but not crazy expensive, but you're basically talking a mildly chunky computer attached to a shitload of power amps that has to be able to bounce around in the back of a truck for a substantial amount of its life and work perfectly as soon as power is applied, which is really what you're paying for. Not quite aerospace but close
-
-
@r @MaddieM4 @mcc @decay @amarioguy thank you!
very helpful list
sigh, yes. maybe we can beat some swords into paintbrushes though...
@ireneista @MaddieM4 @mcc @decay @amarioguy my initial sketch for a potential completely different approach to this material (which has *not* been trialed at all, because social stuff fell through, so grains of salt etc.):
1. functions, with a focus on function composition. if you have "USA K-12" brainworms, concepts from the mess that is "pre-calculus" but taught with through a very "computer science" lens, probably with code samples
2. linear maps, with a focus on how this is _an abstraction_
3. matrices and vectors, with a focus on *computation*
4. FIR and IIR filters. the general goal is to get to a working "live demo" with as little "boring" theory on the critical path as possible
5. frequency domain and spectra
6. short time fourier transforms
????
7. < i haven't actually planned it this far yet >
idk if this would be viable or worthless
-
@MaddieM4 @ireneista *is* that done on on-board controller silicon? I would have assumed it's done in a CPU-driven userspace thing installed in to the regular operating system alongside the driver.
Like I'm certain they *could* do it with on-board controller silicon, if they wanted, but even then I suspect that on-board controller silicon would look a lot like a small ARM running a CPU-based compiler in a tiny Linux
@mcc @MaddieM4 @ireneista My suspicion is that it's just much easier to develop / maintain / manage the ever-growing software stack (compiler and friends) as a library that runs on the host rather than as a more "firmware"-y environment deployed on the GPU itself. Combined with the desire to be able to do offline/AOT shader compilation and the like.
-
@r @ireneista @MaddieM4 @mcc @amarioguy me, perking up from behind my behringer x32
@r @ireneista @MaddieM4 @mcc @amarioguy but yeah jumping back up here a lot of this is technical theater. You can go to school for lighting design, stage electrics, audio engineering and like such as and learn this stuff but there's no real connection back to the computer-touching or EE side
-
What I mean is, my daydream GPU was going to be designed to be good for the kinds of 2D acceleration you need to like, run a web browser, scroll a window, move a window around. Want to do anything 3D? It's slow or it kicks you back out to MESA software rendering. I just wanted a computer I can use in a day to day way that felt like it was mine and I owned it.
There are a few rudimentary open source GPUs, but none with this focus, and critically, *none I know that support Vulkan*.
@mcc if my aim were to implement the most minimal hardware 2D accel I'd probably target XRender so I could shim it, from my current understanding
Haiku is maybe getting hardware 3D ... eventually(?), and has forbidden slop, if you hadn't seen
