As the title states. i have discovered an efficiency anomaly in the mojo compiler, where a function is about 100 times slower than expected. How to proceed? I can create a small repro case to illustrate it, so you compiler people can have a closer look at it, and hopefully fix it! Is it alright to…

At a few hundred lines, a git repo link is probably a better idea so we can easily get it onto our own computers and run a profilers.

I’ll figure out how to create one, thanks! And once i have it and can share it, create a new topic?

[image] GitHub - notnot/mojo-25.1-flaw-1: repo case of a suspected Mojo 25.1 compiler flaw repo case of a suspected Mojo 25.1 compiler flaw let me know if it works (it it my first time i share a github link) Have a look at the benchmark timings of dis…

I finally got some time to work on it and making things a little more high level helped a lot: #### RGBA8 ##################################################################### @value @register_passable("trivial") struct RGBA8: var rgba: SIMD[DType.uint8, 4] fn __init__( out self,…

Thanks for this interesting generalization that Mojo makes possible! It is certainly surprising to see that there is no abstraction cost here. The suggested code doesn’t produce correct answers, at a first glance i think it is because the abs_delta() function isn’t doing what it is supposed to be d…

For those interested, the proper abs_delta() implementation is: fn abs_delta[width: Int]( owned a: SIMD[DType.uint8, width], owned b: SIMD[DType.uint8, width]) -> SIMD[DType.uint8, width]: vmin = min(a, b) vmax = max(a, b) return vmax - vmin This runs at a good speed (simi…

You’re correct, I forgot an abs: fn abs_delta[width: Int](owned a: SIMD[DType.uint8, width], owned b: SIMD[DType.uint8, width]) -> SIMD[DType.uint8, width]: return abs(a.cast[DType.int16]() - b.cast[DType.int16]()).cast[DType.uint8]()

I think what was causing the performance issues was the large amount of branching, which meant that a lot of optimization passes couldn’t match on the code you were writing. It’s possible that the opt-pass couldn’t deal with the sparse access pattern for 4 items, but has a pass that deals with (x, y…

Yes i was thinking along similar lines. What worried me was when i compared the Mojo code to equivalent Go code, which didn´t suffer from the slow execution. Go is ~3x slower in this test, but it doesn´t have an anomalous difference in the timing of distance_Env3x3_L1 and distance_Env4x4_L1. Many t…

I have discovered a suspect efficiency anomaly in the mojo compiler, how to proceed?

Mojo

sora March 7, 2025, 3:49pm 19

@notnot Seeing the MLIR/LLVM or ASM code generated by Mojo - #5 by sora

Topic		Replies	Views
Llama2.🔥 performance degradation after updating code to the latest Mojo compiler Mojo performance	1	274	November 21, 2025
Very slow compile times with FFI functions Mojo	5	282	July 2, 2025
Unexpected argument mutation with `mojo test` — not under `mojo run` Mojo discussion , mojo-compiler	2	138	September 19, 2025
Mojo vs max mandelbrot performance Mojo performance	5	343	May 12, 2025
A Benchmark with Files and Bytes (standard benchmark warnings apply) Community Showcase discussion	9	388	January 6, 2025

I have discovered a suspect efficiency anomaly in the mojo compiler, how to proceed?

Related topics