So I Tried To Learn Shaders...

i'm just waiting for Prime gamedev arc where he writes his own 3d engine in c++

But Acerola,

Prime’s gamedev arc continues. We’ll get Prime Game before GTA VI

Seriously, how I learn how to do shaders: 1. Watch a Blender tutorial for the effect I'm going for. 2. Dissect the Blender source code for how the nodes worked together to create the effect. 3. It's now OUR code. 4. Success. For anyone new to shaders, I recommend starting with Godot's shader language, it's by far the easiest way to get started. There's a lot of high level language features and parameter handling that makes things much easier starting out.

Shaders continue to break my head every time, and I write them every day.

Good advice, when you get distracted, just dont get distracted :D Got it, ADHD solved. If I was able to not get distracted, I would not be watching this video...

My friends Patricio and Jen made book of shaders - you should have them on your show. They are awesome!

1:30 Gaussian blur, as far as I know, is done with a convolution with a gaussian kernel. That means pretty much what prime was saying, you take the kernel, which is a grid of values you apply to the the center and neighboring pixels, to then add them up, and put them in the new image in each location where the center pixel was before. We had to do make a hardware accelerator to do this in a digital design class using an FPGA. This is very likely not the only way to do this, but its pretty much what you think it is, kind of "averaging" each pixel with its neighbors, but with a higher weight for the center pixel, getting smaller as you go to the edges of the kernal/mask.

GPU has hardware level integration for dealing with floats without need for conversion, stacking or splitting. It's built into the hardware unlike CPUS where it's a software operation in most cases. This is why the float precision actually changes rendering speed unlike CPU bound operations. It's happening in parallel in the hardware rather than being scheduled. GPU cores, they can just pull and push specialized data so damn fast. The 4090 can push 1,000Gb/s through it's compute units in total, when you pair that with the fact that all the native operation for GLSL and Vulkan are directly built into the silicone it makes most GPU single operations process as fast at the speed of electrons in the medium (perceivably instant in copper and aluminium) as they're already complete before the software touches the data. AMD Ryzen 9 7950X3D: Internal data rates up to 512 GB/s and then that's further bottlenecked by the fact the architecture is designed for random compute/broad tasks rather than having actual silicone dedicated to the specific GL calls and matrix math. Some GL calls are on the GPU in silicone so they run extremely fast.

i Love seeing how prime learns something new so many people spamming use shadertoy but it allowed him to test his own variables which really helped me even understand how things really work under the hood.