Simple Instructions, Weird Algorithms

1위@CoreDumpped1 day ago

👉Get Rust training from Let’s Get Rusty: https://letsgetrusty.com/start-with-jorge번역 보기

2위@lordcraigdragneel1361 day ago

A video on gpus could expand really well on this vid.번역 보기

3위@CoreDumpped1 day ago (edited)

A few things I would've liked to mention in the video: 1. If a processor supports 128-, 256-, and 512-bit vector operations, that doesn't mean it has three completely separate sets of vector registers. Architecturally, these registers overlap. An XMM register refers to the lower 128 bits of the corresponding YMM/ZMM register, while a YMM register refers to the lower 256 bits of the corresponding ZMM register. So the same register can be accessed at different widths depending on the instruction being used. I learned this the hard way once when I was trying to use vector registers as a sort of cache. 2. Because the algorithm is still O(N²), every version of the program will eventually struggle if the initial particle count gets high enough. However, particles merge when they collide, so the number of particles gradually decreases. Eventually, there are fewer interactions to calculate and the simulation starts running smoothly again—even in the non-SIMD version. Because of this, *how quickly each version can escape that huge initial workload* can actually be used as another benchmark to compare them. 3. I'm not particularly good at explaining physics, so that part of the video is intentionally a little oversimplified. If you want to explore the math in more detail, check out the repo. 4. I usually avoid showing hexadecimal numbers in my videos because they tend to scare or confuse beginners, but these registers are so ridiculously wide that I had to. 5. Editing mistake. As someone else commented bellow: Assembly at 1:10 is wrong. Should be `MUL` instead of `ADD`, or change the `*` to `+`. Have a great weekend.번역 보기