Josh Bainbridge

I’m thrilled to share that OpenQMC, a key part of the technology stack behind Framestore’s feature film VFX, is now open source and part of the Academy Software Foundation! aswf.io/blog/openqmc-b… OpenQMC is a library for sampling high quality Quasi-Monte Carlo (QMC) points...

If you’re interested, come join us and help make these tools even better for everyone! openqmc.org

A huge thank you to everyone who made this possible - from the researchers and paper authors, to the ASWF members, and the technology teams here at Framestore.

I’m thrilled to share that OpenQMC, a key part of the technology stack behind Framestore’s feature film VFX, is now open source and part of the Academy Software Foundation! aswf.io/blog/openqmc-b… OpenQMC is a library for sampling high quality Quasi-Monte Carlo (QMC) points...

@trierman @m_schuetz I believe HIP compiles directly for AMD GPUs using HIP-Clang, analogous to CUDA and NVCC. Although I've heard you can go to OpenCL via SPIR-V if need be.

CUDA is absolutely amazing. Almost full featured C++17 for GPUs. So many C++ features that make development and debugging easier, like strings in this case. Are there any attempts in making C++/CUDA work in Vulkan? Because I never want to use GLSL ever again.

@zenopelgrims @arnoldrenderer Cool! Sounded similar. Results look fantastic.

@joshbainbridge @arnoldrenderer yes, it’s similar. I read the ERPT paper but it’s not a direct derivative

Both renders are at 4AA samples. The first is using Lentil (Arnold plugin) to bidirectionally trace camera rays, the second is the @arnoldrenderer default. All this at a much lower rendertime than if we had to trace e.g 40AA samples to clear up this image. Pretty neat no?

@yiningkarlli So true. Beyond my knowledge. Would be interested to know. And looking forward to this upcoming blog post. It sounds like a good one. They are always a joy to read. I’m sure to learn a thing or two there as well.

@joshbainbridge Turns out that while most real lenses can be modeled using a thick lens model, actually figuring out the parameters of the thick lens that corresponds to a real lens is.... also nontrivial. ;)

I have a dumb question: how is the effective focal length of a real photo lens calculated? I know the definition of focal length, but a real complex lens has many lens elements, each with their own nodal point. From what point in the whole assembly is dist to the image measured?

Before & Afters spoke to VFX Supervisor Adrien Saint Girons (MTL) about the mirror dimension sequences in #SpiderManNoWayHome , the mathematics involved, and how the team mastered astral projection. beforesandafters.com/2022/01/14/her…

@yiningkarlli These are so good, I’ve been using one as a self-hosted runner on GitHub. Does a great job as a GPU config in a CI matrix. Building and unit testing in the background all day.

Uh oh, this arrived today, but I already have too many personal projects that need to be finished or written up or started. Buying one of these might have been a mistake, but I found one for around MSRP on EBay and as a lover of weird SBCs, I couldn’t pass up the chance…

Today we hear from some of our top creatives and technologists about how tech shapes our future. Find out how skills in game design, programming and engineering are becoming increasingly important to the biggest films and TV shows #NextLevelFramestore youtu.be/jshnIJkWCgQ

@attila_afra Amazing, looking forward to testing it out! Any update on support for temporally stable denoising?

Intel #OpenImageDenoise v1.4.2 has been released, which adds support for FP16 input/output images, and fixes a performance issue on Alder Lake CPUs: github.com/OpenImageDenoi…

Showcased at the @COP26 World Leaders Summit Opening Ceremony, Framestore worked with filmmaker Matt Houghton on the post-production and delivery of various pieces of content to bring to life 'Earth to COP' a vivid and emotional film. @earthtocop #Cop26 #ActionThisDay

@anderslanglands @Atrix256 Ah yes, the perfect sequence! Maybe it's out there somewhere, hiding behind an obscure series of xor operations.

@joshbainbridge @Atrix256 Yeah that’s the holy grail - a high-dimensional sequence with good projections on all lower dimensions (and combinations thereof) which can be constructed progressively and cheaply.

@anderslanglands @Atrix256 That's pretty much it. Just looked it up, in the TOG paper (Section 7.1.2) they said the PMJ sequences are shuffled to obtain stratification into 3D, 4D, etc. Although this is just stratification. Maybe there is still advantages to be found with natively high dimensional LDS.

@joshbainbridge @Atrix256 I thought that was just shuffling to get good projections in the first few dimensions?

@anderslanglands @Atrix256 Totally agree! It would be great to see more analysis done on higher dimensions. If I'm not mistaken, in practice the Renderman team were reordering the scrambled 2D sequences to numerically find the most optimal high dimensional sequence. So there must be benefits to be found.

@joshbainbridge @Atrix256 Yeah there was a paper at Siggraph a few years ago investigating different distributions against different functions. Think the result was best dist was dependent on function shape. We need to see more analyses like in the PMJ paper but over more dimensions.

@anderslanglands @Atrix256 If I'm not mistaken, LDS is only optimal for low frequency parts of the integral. Geometric discontinuities often found in high dimensional path tracing result in limited benefits of high dimensional sequences. This might be why padding independent 2D sequences are popular.

@Atrix256 Received wisdom states that the benefits of LDS drop off somewhere after 10(ish?) dimensions and convergence deteriorates to white noise (1/sqrt(N)). Don’t think I’ve ever seen this analysed though.

@zenopelgrims Enoki github.com/mitsuba-render… by @wenzeljakob is another option. It's a great library, I've been using it for some vectorisation work. Or there is NVC++ which can compile Standard Library algorithms natively for Nvidia GPUs.

Graphics twitter, what's the "best" way to write code that runs both on cpu and gpu using cuda? Through the polymorphic allocator, like pbrt v4 does? Or a library like Hemi (8 years since last update..)? Any other options I'm missing?