The Weird Concept of Branchless Programming

(sanixdk.xyz)

1. Legend2440 ◴[28 Sep 25 18:04 UTC] No.45406467[source]▶

Article doesn’t mention this, but I’d consider neural networks a form of branchless programming. It’s all a bunch of multiply and threshold operations.

replies(2): >>45406503 #>>45407289 #

2. abyesilyurt ◴[28 Sep 25 18:09 UTC] No.45406503[source]▶

>>45406467 (TP) #

Thresholding requires branching no?

replies(3): >>45406688 #>>45406694 #>>45407482 #

3. WJW ◴[28 Sep 25 18:32 UTC] No.45406688[source]▶

>>45406503 #

No, you can do that without branching. See https://en.algorithmica.org/hpc/pipelining/branchless/#predi... for example.

4. Legend2440 ◴[28 Sep 25 18:33 UTC] No.45406694[source]▶

>>45406503 #

No, it’s just a math operation and can be applied to any amount of data in parallel.

5. david-gpu ◴[28 Sep 25 19:43 UTC] No.45407289[source]▶

>>45406467 (TP) #

> It’s all a bunch of multiply and threshold operations.

Real-world high-performance matrix multiplication functions do contain branches internally, even on GPUs. If you are ever curious about what that looks like, NVidia maintains an open-source library called CUTLASS.

6. bqmjjx0kac ◴[28 Sep 25 20:05 UTC] No.45407482[source]▶

>>45406503 #

As demonstrated in the article, you can compute clamp(x, min, max) with straight-line code.

↑