mschuetz

It's kinda like building Legos vs building actual Skyscrapers. The gap between compute shaders and CUDA is massive. At least it feels massive because CUDA has some key features that compute shaders lack, and which make it so much easier to build complex, powerful and fast applications.

One of the features that would get compute shaders far ahead compared to now would be pointers and pointer casting - Just let me have a byte buffer and easily cast the bytes to whatever I want. Another would be function pointers. These two are pretty much the main reason I had to stop doing a project in OpenGL/Vulkan, and start using CUDA. There are so many more, however, that make life easier like cooperative groups with device-wide sync, being able to allocate a single buffer with all the GPU memory, recursion, etc.

Khronos should start supporting C++20 for shaders (basically what CUDA is) and stop the glsl or spirv nonsense.

That ship has sailed. Choices are stdout or stderr. Given those names, I'll of course send errors to stderr, leaving me with stdout for messages/infos/logs.

Most applications nowadays use stdout for exactly what the name implies - any form of output. Complaining about that is yelling at the clouds.

A board member's husband?

How so? Seems like a fairly simple license.

The problem with "unlimited detail" wasn't that it was static (the majority of any game's environment are), it's that it was using voxels which can't really compete with triangles when it comes to a quality-perf trade-off. They could render massive data sets, but not with a quality that is needed for games. Voxel-based data sets tend to require a whole lot of memory, whereas triangle-based data sets can cheaply "fake" higher details with textures. The blockiness of voxels is also a huge issue for anything that's not an axis-aligned-bounding-box, and to fix that you have to invest so much GPU resources, you might as well go back to textured triangles.

I wouldn't be surprised if gaussian splats make it into AAA games, though. Not as the main rendering primitive, but for specific things like vegetation where they really kick ass.

How are they preventing the competition to create something better than CUDA? And how does it hurt the consumers that they are providing a fantastic product that others refuse to provide?

Those are some poor arguments, imho, because there literally is no other option than CUDA. The alternatives are so bad, it's far better to be vendor-locked and being able to get stuff done, than not being able to get stuff done at all.

As I said, I avoided it for years because of the reasons you mentioned. Turns out I could not avoid it any longer because it's the only (meaningful) option that could do what I needed, has serious support, and great UX. And NVIDIA is hardly to blame because they simply made sure to build a good product. It can't stop AMD, Intel or Khronos from creating a competitive alternative, but so far they haven't.

And regarding support, so far NVIDIA has shown excellent continuous support for CUDA, whereas OpenCL and OpenGL are the ones that went down. And I've chosen CUDA over rocm precisely due to support reasons, because AMD has always treated it as some kind of side gig with uncertain future.

It's a great cover...what do you mean with "get away with"?

Why would I blame NVIDIA? If it wasn't for them, we'd still only have needlessly cumbersome APIs and ecosystems. They did what Khronos always failed to do: They created something that is both easy, powerful and fast. Khronos always heavily neglects the easy part.

Not happening. WGSL wants to support the lowest common denominator, so it'll always mainly be a 5-year old mobile-phone API. Also if you want to beat CUDA, you'll need some functionality that's completely missing in compute shaders, especially WGSL. Like pointers and pointer casting (and that glsl buffer reference extension is the worst emulation of that feature I've every seen).

Not just feature parity, but proper UX. Things need to just work, without spending hours or days to make them work.

What's wrong with CUDA? I avoided it for years because it's proprietory but about one year ago I started using it because all the alternatives (OpenGL/Vulkan compute, OpenCL, WebGPU, ...) couldn't quite do what I wanted, and it turned out to be a game changer. Nothing comes close to it. Now I'm hooked because there simply isn't an alternative that's as easy to use, yet powerfull and fast.

I wish there was an open alternative, but NVIDIA did several things right that others, especially Khronos, do not: The UX is top-notch. It makes the common cases easy yet still fast, and from there you can optimize to your hearts content. Khronos, however, usually completely over-engineers things and makes the common case hard and cumbersome with massive entry barriers.

It's used in one of the fastest sorting approaches - counting sort / binning - to compute the location of where to store the sorted/binned items. First you count the number of items per bin, then you use prefix-sums to compute the memory location of each bin, then you insert the items into the respective bins. Some radix-sort implementations also utilize counting sort under the hood, and therefore prefix-sums. (Not sure if all radix-sort implementations need it)

"Turns out we can beat the hardware with triangles much bigger than expected, far past micropoly. We software rasterize any clusters whos triangles are less than 32 pixels long." - https://advances.realtimerendering.com/s2021/Karis_Nanite_SI...

It's not just micro-triangles, it works for triangles that span multiple pixels. And that's not a special case, that's the standard nowadays, except for games targeting very low-end devices.

The dedicated hardware is nice for general-purpose support for arbitrary triangle-soups. But if you structure triangles a certain way and have a certain amount of density (which you want for modern games), you can specifically optimize for that and beat the general-purpose hardware rasterizer.

Huh? It had other options longer than it had async/await. Async/await is a fairly recent addition. E.g. before fetch with async/await, there was XmlHTTPRequest with callbacks. It also had Web Workers as a means for parallel&concurrent processing for way longer than it had async/await.

The standard rendering pipeline is still fairly fixed and mandates shaders for vertices and fragments. With software rasterization, there are no more vertex or fragment shader, only compute. And you don't use the hardware rasterization units - you rasterize the triangles yourself and write the results to the framebuffers with atomic-min/max operations. You decide for yourself how and when you compute the shading in your compute shader. This can be multiple times faster for small triangles, and 10-100 times faster for points. And once you do things that way, there isn't much point for graphics APIs anymore - everything is just buffers and functions that process them.

Nanite renders small triangles faster by avoiding the dedicated hardware and using compute, instead.

You're missing the point. CUDA is easy but also powerful, and it shows there is no reason that other APIs need to be hard and cumbersome.

Not really trivial if you have to support any input set of triangles and don't know much about them. Software rasterizion exploits things like localized chunks of triangles, but the hardware rasterizer does not know about that in advance. Also, these kinds of software rasterization algorithms add limitations, which aren't much of an issue with your own rendering pipeline that specifically knows how to deal with these limitations and works with them.

It's the exact reason why I've avoided CUDA for years, but I hit a dead end with OpenGL and Vulkan, and CUDA happened to be a fantastic, easy and fast solution. Of course I don't want graphics programming to be NVIDIA-only, but I want it to be like CUDA, just for all platforms.

Exactly. Things like that already work with a workaround: You can use Cuda-OpenGL interop to expose an OpenGL framebuffer in CUDA, then you can simply write into that framebuffer from your CUDA kernel, and afterwards you get back to OpenGL to display it on screen. Just directly integrate that functionality in CUDA by providing a CUDA native framebuffer and a present(buffer) or buffer swap functionality.

Yeah, but OpenGL doesn't get updates anymore. My timeline goes like: I needed pointers(and pointer casting) for my compute shaders so I checked the corresponding GLSL extension, which was only available in Vulkan so I tried switching from OpenGL to Vulkan. After a week I gave up - the pointer/Buffer reference extension did not look promising anyway - and I tried out CUDA instead. That's when I found out that CUDA is the greatest shit ever. That's what I want graphics programming to be like. Since then I just render all the triangles and lines in CUDA, because it easily handles hundreds of thousands of them in real-time with a naive, unoptimized software-rasterizer, and that's all I need. In addition to the billion points you can also render in real-time in CUDA with atomics.

I meant the development/learning overhead. With Vulkan you can do incredible low-level optimizations to squeeze every last bit of performance out of your 3D application, but because you are basically mandated to do it that way, you have a very harsh learning curve and need lots of code for the simplest tasks. I'd rather prefer approaches that make the common things that everyone wants to do easy (draw your first simple scenes), and then optionally gives you all the features to sqeeze out performance where you really need to. Because at least for me, I don't work with massive scenes with millions of instances of thousands of different objects. I do real-time graphics research, mostly with compute, and I'd just like to present the triangles I've created or the framebuffers I created via compute (like shadertoy).

I'll readily admit, I'm neither smart nor patient enough for Vulkan so I quickly gave up and learned CUDA instead, because it was way easier to write a naive software-rasterizer for triangles in CUDA, than it was to combine a compute shader and a vertex+fragment shader in Vulkan. I'm just rendering a single buffer with ~100k compute-generated triangles, and learning Vulkan for that just wasn't worth it.

Yes, sorry for the confusion. I'd just rather have an API where the common things are easy, and the super powerful low-level optimizations are optional.

> Vulkan and OpenGL both already support mesh shaders, which is a compute-oriented alternative to the traditional rasterization pipeline.

Mesh shaders are a step in the right direction, but they are still embedded in all that unnecessary Vulkan fluff. I would want these things in CUDA because it does the opposite approach of Vulkan - it makes the common things easy, and the hard/powerful things optional. Just let me draw things directly in CUDA, and maybe give access to the hardware rasterizer via a CUDA call.

I know it won't happen overnight, but since it's already possible to do software rasterization for small triangles faster than hardware, having a graphics API framework starts losing its purpose. After all, we want to have the detail of small triangles anyway. Just let us draw to the screen in CUDA without the need for OpenGL/Vulkan interop, and I believe we'll soon see a shift to serious compute-based real-time rendering.

Basically, instead of graphics being a framework, I want graphics to be a straightforward library you include and use in your CUDA/HIP/SYCL/OpenCL code.

Part of Nanite is software rasterization by rendering triangles with 64 bit atomics. You can simply draw the closest fragment of a triangle to screen via atomicMin(framebuffer[pixelID], (depth << 32) | triangleData).

Glad to see more support for OpenGL, but I really hope we'll soon move to a compute-only way of handling graphics. The overhead of vulkan is absolutely insane (and not warranted, in my opinion), and OpenGL is on its last legs.

Things like Nanite spark a little hope, since they've shown that software-rasterization via compute can be faster than the standard graphics pipeline with the hardware rasterizer for small, dense triangles. Seems like a matter of time until everything goes compute, even large triangles. Maybe the recent addition of work-graphs in DirectX is one step in that direction?

I strongly disagree. Async/Await is one of the nicest, cleanest ways to deal with asynchronous tasks, and asynchronous tasks are everywhere. Loading data from disk without blocking and doing something once this is done -> async. Memcpying data from CPU to GPU without blocking -> async. Memcpying data from GPU back to CPU without blocking -> async.

Sure, there are other ways to handle these things like polling state, but async makes it trivial and readable.

But this mostly applies to JS which has a kickass implementation of async/await. Whatever C++ tried to do, it's an awful mess so there I still use threads with busy loops, polling, etc., whatever makes sense for the task at hand.

From an expert's standpoint, it's still quite unusable and complex.