Making my debug build run 100x faster so that it is finally usable (opens in new tab)

(gaultier.github.io)

87 pointsbroken_broken_1y ago20 comments

20 comments

12 comments · 5 top-level

bean-weevil1y ago· 5 in thread

Why not just compile that particular object with optimizations on and the rest of the file with optimizations off?

Yes, that's the obvious (and boring!) answer, that I mention in the introduction and that's in a way the implicit conclusion. But that does not teach us SIMD then :)

saurik1y ago

Your article isn't really about, though, how to speed up a debug build, and I thereby think you're likely not going to find the right audience. Like, to be honest, I gave up on your article, because while I found the premise of speeding up a debug build really interesting, I (currently) have no interest in hand-optimizing SIMD... but, in another time, or if I were someone else, I might find that really interesting, but then would not have thought to look at this article. "Hand-optimizing SHA-1 using SIMD intrinsics and assembly" is just a very different mental space than "making my debug build run 100x faster", even if they are two ways to describe the same activity. "Using SIMD and assembly to avoid relying on compiler optimizations for performance" also feels better? I would at least get it if your title was a pun or a joke or was in some way fun--at which point I would blame Hacker News for pulling articles out of their context and not having a good policy surrounding publicly facing titles or subtitles--but it feels like, in this case, the title is merely a poor way to describe the content.

1 more reply

corysama1y ago

You would appreciate https://aras-p.info/blog/2024/09/14/Vector-math-library-code...

TLDR: SIMD intrinsics are great. But, their performance in debug builds is surprisingly bad.

1 more reply

mperham1y ago

Yep, I was hoping to learn how to do this. Seems like a much better long term lesson.

senkora1y ago

For gcc: #pragma GCC optimize ("O0")

For clang: #pragma clang optimize off

For MSVC: #pragma optimize("", off)

Put one of these at the top of your source file.

2 more replies

molenzwiebel1y ago· 2 in thread

For this use-case, you can squeeze out even more performance by using the SHA-1 implementation in Intel ISA-L Crypto [1]. The SHA-1 implementation there allows for multi-buffer hashes, giving you the ability to calculate the hashes for multiple chunks in parallel on a single core. Given that that is basically your usecase, it might be worth considering. I doubt it'll provide much speedup if you're already I/O bound here though.

[1]: https://github.com/intel/isa-l_crypto

broken_broken_OP1y ago

Thank you, I will definitely have a look, and update the article if there’s any interesting finding

secondcoming1y ago

I came across this repo recently and it looks great. It's a pity there doesn't seem to be an official Ubuntu package for it though. There is one for the Intelligent Storage Acceleration Library though.

ack_complete1y ago

SHA1 is difficult to vectorize due to a tight loop-carried dependency in the main operation. In an optimized build, I've only seen about a 15% speedup over the scalar version with x64 SSSE3 without hardware SHA1 support. Debug builds of course can benefit more from the reduction in operations since the inefficient code generation is a bigger issue there than the dependency chains. I think the performance delta is bigger for ARM64 CPUs, but it's pretty rare to not have the Crypto extension (except notably some Raspberry Pi models).

The comments in the SSE2 version are a bit odd as it references MMX, and the Pentium M and Efficeon CPUs. Those CPUs are ancient -- 2003/2004 era. The vectorized code you have also uses SSE2 and not MMX, which is important since SSE2 is double the width and has different performance characteristics from MMX. IIRC, Intel CPUs didn't start supporting SHA until ~2019 with Ice Lake, so the target for non-hardware-accelerated vectorized SHA1 for Intel CPUs would be mostly Skylake-based.

tvbusy1y ago

I understand the post is about learning to speed up SHA1 calculation, that I have no comment. However, the state file is a solved problem for me. It's a rare case where state files are corrupted and it's simple to just re-check the file. I cannot imagine a torrent client checking the hash of TBs of files for every single start. It's not a coincidence that many torrent clients have a feature to skip hash checking and just immediately assume the file is correct and start seeding immediately.

Neywiny1y ago

If I were a betting person I'd bet that the sha1 instructions and the openssl instruction map to similar enough uops. Unsure if there's a way to check, but that's my understanding of the thousands of instructions in modern processors - mostly just assigning names to common patterns.

j / k navigate · click thread line to collapse

20 comments

12 comments · 5 top-level

bean-weevil1y ago· 5 in thread

Why not just compile that particular object with optimizations on and the rest of the file with optimizations off?

broken_broken_OP1y ago

Yes, that's the obvious (and boring!) answer, that I mention in the introduction and that's in a way the implicit conclusion. But that does not teach us SIMD then :)

saurik1y ago

1 more reply

corysama1y ago

You would appreciate https://aras-p.info/blog/2024/09/14/Vector-math-library-code...

TLDR: SIMD intrinsics are great. But, their performance in debug builds is surprisingly bad.

1 more reply

mperham1y ago

Yep, I was hoping to learn how to do this. Seems like a much better long term lesson.

senkora1y ago

For gcc: #pragma GCC optimize ("O0")

For clang: #pragma clang optimize off

For MSVC: #pragma optimize("", off)