Kirill Dmitrenko

To celebrate the release of Dungeons & Dragons Neverwinter Nights 2: Enhanced Edition on Nintendo Switch from @AspyrMedia, we have a couple of codes to give away. Repost and follow to enter. Ends July 18. Open to all regions, but we are giving out US codes.

@s1ddok That's an unfortunate turn:( If you don't mind, how long did it take to get an answer? I saw processing times in the UK and decided to now even try

Canada rejected my visa, so I won’t be able to attend our SIGGRAPH booth in Vancouver this year. I'm pretty heartbroken, but I still encourage everyone to come and experience our demos firsthand!

@s1ddok Generally speaking, not necessarily: fgiesen.wordpress.com/2015/02/05/a-s…

modern GPUs being scalar is bit of different matter for my confusion. let's assume for a second that we still have one-instruction 4 component vector multiplication. then shouldn't all samples insist on multiplying vector from the left? otherwise it makes stuff very inefficient (additional shuffles are required) and almost every sample from every shader language by default contains matrix on the left and vector on the right

I'm confused. All modern GPU APIs default to column major matrices, with a common wisdom of them being "faster bc SIMD". Presumably that means that it should be true if matrix/vertex multiplication can be expressed in 4 dot products, each working on 4-component vector. But then all API follows up with sample where vector is multiplied on the right, meaning that you first have to scatter elements from each column to form a row and then multiply. Am I missing something or it makes no sense? Shouldn't we all be multiplying vectors from the left?

@s1ddok E.g. if you squint hard enough, you can see here that matrix-vector mul is compiled to a flurry of FMAs: godbolt.org/z/d1sEE9Wsq.

@s1ddok AFAIK, almost all current GPU architectures are scalar, so it doesn't matter. There's no "dot product" instruction that operates on four floats at once. There are specific "dot product" instructions for packed 16-bit or 8-bit values, though.

@SebAaltonen I think that's the answer to a different question :) Mine was about what you think about a possibility of WebGPU being the only GAPI used in the HypeHype.

@dmikis I had to add validation disable also to the WebGPU instance. Device wasn't enough.

WebGPU performance: WebGPU (Dawn native): 0.7 ms MoltenVK: 0.7 ms Metal: 0.6 ms 791 draw calls in 10 render passes. Our Metal 2.0 backend uses manual resource tracking, unmanaged command buffers, placement heaps and argument buffers. This is a very good result for WebGPU!

@SebAaltonen Would you be open to publishing the code as a gist or something? :)

CLARIFICATION: I am not intending to ship a graphics API. I am only writing mock code to show how much simpler it would be to write GPU code without a complex graphics API.

I am again writing my "No graphics API" blog post. Wrote the first C/C++ example that creates a shader pipeline, texture, RT, buffer, does a dispatch with sampler, output texture, buffer and root data struct containing all of them. Total 84 lines. Much better than Vulkan :)

@kenpex @maxliani @yotam_erel Blender runs CPython inside. You can write native addons for it. For example, we (and other people, AFAIK) have successfully run PyTorch within an add-on to the stock Blender.

@maxliani @yotam_erel And AFAIK you don't even really develop plugins for blender, it's either python or you patch the source? But I dunno how limiting it is, couldn't you patch blender & release the source of the interconnect you need, then use that to develop your tech privately?

Autodesk does not seem to be reacting much to the Blender phenomenon? Now, ofc I have no data about this, but it seems that Blender keeps growing - the moment it will be a clear danger for Maya it will be too late for Maya to steer. They probably need to do something radical now.

@StefanPijnacker What is the test model?

Started working again on my renderer that uses DirectX Ray Tracing. Making good progress right now. Gonna do some more modern techniques soon together with supporting quality settings.

@SebAaltonen Shared pointers in Metal are hidden in the bowels of objective c runtime. There're still very much there.

Why is the C++ wrapper for Vulkan riddled with shared pointers. C++ wrapper for Metal seems much more clean.

I just found out that there's a C++ wrapper for Vulkan when I was reading Flutter's Impeller source code. Of course there's some template magic there. Turns out that I can get underlying Vulkan handles with overloaded type cast operator. Obviously...

@FreyaHolmer @tom_forsyth Same :)

@dmikis thanks! also hey its @tom_forsyth i know that nerd! (as in, from twitter)

ok I need a low down on vertex streams! is there a good resource to learn these things? some questions I have right now: 1. are streams a separate concept from the actual data layout, and are there performance reasons you would/wouldn't put each vertex attribute on a separate stream? the only thing I'm aware of here is that tiled renderers prefer having the position buffer in its own stream for better performance when doing,, one of the passes, I forgot the name of it 2. do meshes actually use an array of structs, or is it a struct of arrays, or, do you get to pick, and if so, what are the performance tradeoffs? 3. do index buffers prefer having referenced vertices close to each other in the vertex buffer, or does it not matter?

@meuns D3D is supported via windows crate in manner similar to vulkan and ash.

I'm interested using rust in some tiny graphics project but it seems the only usable/supported/mature API is vulkan. Maybe I need to use wgpu... 😩

@self_shadow @kenpex Thanks for doing this! Invaluable effort.

And it begins

@pixnblox I'm waiting for my Delta flight to denver and this twit had me worried for a second that they've changed the gate 😅

Ready for Denver! 💪 #SIGGRAPH2024

I will be in Denver next week to attend @siggraph 2024. Feel free to get in touch if you're interested in talking about our work at @cinemersivelabs and graphics overall.