Kunal

We've just published "Unsafe C# code best (worst) practices" article - a set of 25 good/bad examples: learn.microsoft.com/en-us/dotnet/s… Let's crash productions in style together! 🧐

@nietras1 @dyaroshev @geofflangdale replied > hope you don't mind of course not. :)

@KuXMLP @dyaroshev @geofflangdale I've mentioned you on a PR on GitHub in Sep repo with details, hope you don't mind. For others, LD4 approach is clearly slower in this specific case probably because here we have to go from u16 to u8 which means 2 x LD4 per "move mask". Which means 2xLD4=16 cyc vs 4xLDP = 8 cyc

New blog post "Sep 0.11.0 - 9.5 GB/s CSV Parsing Using ARM NEON SIMD on Apple M1 🚀" 🛠️ New #ARM #NEON #SIMD parser based on @geofflangdale bulk move mask 📈 Sep #performance up from 7 GB/s on #Apple #M1 🧑‍💻 #csharp and #ARM assembly on #dotnet 9.0 Link below or via profile 👇

@nietras1 @dyaroshev @geofflangdale definitely LD4 is expensive...8 cycles vs. LDP 2 cycles. Might be able to tell the slowness if I see the asm diffs between the two approaches.

@KuXMLP @dyaroshev @geofflangdale Fixed spilling, perf still worse though, asm looks fine, so wondering if latency/tp of new instructions worse than AddPair approach without LD4... LD4 appears to have 8-9 cycle lat. on N2. @KuXMLP sound right?

@nietras1 @dyaroshev @geofflangdale Unfortunately no

@KuXMLP @dyaroshev @geofflangdale @KuXMLP is there no overload taking a ref? Only pointers?

@nietras1 @dyaroshev @geofflangdale the names might be mixed up, but this is what I just verified: So maybe you want the first one i.e. `Load4xVector128`

@KuXMLP @dyaroshev @geofflangdale actually I would think this is not the Unzip version, as we want the interleaving.

@nietras1 @dyaroshev @geofflangdale I believe you are asking about `Load4xVector128AndUnzip`? - learn.microsoft.com/en-us/dotnet/a… If not, which one of the variants are you looking for? x.com/KuXMLP/status/…

@dyaroshev @geofflangdale Dumb brain fart by me ofc 🤦 @KuXMLP it would be nice if #dotnet got LD4 support, as this could unlock several possible optimization in runtime and third party libraries as already discussed on github in other issues.

The highlight of Microsoft Build—thrilled to be part of the .NET team, advancing Arm workloads to new levels!

New blog post "Sep 0.10.0 - 21 GB/s CSV Parsing Using SIMD on AMD 9950X 🚀" 📈 Sep #performance from 7 GB/s to 21 GB/s over last two years 🧑‍💻 #csharp #SIMD and #x64 assembly on #dotnet 9.0 🛠️ Tweaks and new #AVX512-to-256 parser 🔢 Lots of benchmarks 👇 nietras.com/2025/05/09/sep…

Brief presentation of SVE in .NET 9 at Arm's Innovation Coffee youtube.com/watch?v=foTAbw…

Want to see how to use the new .NET SVE APIs? Check out - community.arm.com/arm-community-…

Check out new JIT features added in .NET 9 by our architect Andy! youtu.be/1bsTnaLchi4?si…

Introducing "Scalable Vector Extension" in .NET 9 #Dotnet devblogs.microsoft.com/dotnet/enginee…

@maxx_don It's link time optimization that mostly does interprocedural analysis

@KuXMLP Excuse my ignorance, but what's LTO? Where can I read more about it?

Most accidental and by far, the easiest change to get 15%~20% throughout improvement - Enable LTO for coreclr linux builds. #dotnetcore github.com/dotnet/runtime…

NVIDIA just released FREE online courses in AI. Here are 9 courses you can't afford to miss:

Plan to add 643 new SVE APIs in .NET github.com/dotnet/runtime…

.NET added 1123 new Arm64 SVE instructions. Great work from .NET + Arm engineers. We wrote a tool that read the Arm manual and generated C++ code needed to implement instructions in RyuJIT, cutting the implementation time drastically. github.com/dotnet/runtime… #dotnetcore #sve

@elonmusk - is there a way to give a sequence number to car that wants to use a supercharger? At busy stations it gets chaotic and an altercation about who is in line and whose turn is next to charge.