LLVM Clang 15 Enables Faster Square Root Instructions For AMD Zen

[phoronix]

Phoronix: LLVM Clang 15 Enables Faster Square Root Instructions For AMD Zen

As part of an effort to update LLVM Clang's "-mtune" handling to cater to newer processors, AMD Zen processors with LLVM/Clang 15 later this year will be able to enjoy faster and more accurate square root calculations with tuning to use SQRTSS/SQRTPS instructions...

https://www.phoronix.com/scan.php?page=news_item&px=LLVM-15-Faster-Zen-Sqrt

[birdie]

Michael, would be great if you saved and posted screenshots as PNG. Thank you.

[JEBjames]

Michael

Typo: "generuc" should be "generic".

[discordian]

Wonder why they (AMD) can't get their shit together when it comes to software. Being behind the curve compared to Intel, completely outclassed by Nvidia.
Was acceptable 3-4 years ago, but nowadays their financials are very healthy and I just don't get it.

[jabl]

Wonder why they (AMD) can't get their shit together when it comes to software.

I'd imagine having a couple of compiler engineers on staff making sure that the pipeline descriptions and tuning heuristics are in a good shape in upstream GCC and LLVM for their current processors would be a much cheaper way to increase the performance of their products by a few % than making the actual raw chip performance better by the same amount.

[brucethemoose]

Michael, would be great if you saved and posted screenshots as PNG. Thank you.

PNGs take up much more bandwidth, even in very compressible content like these screenshots, even when heavily optimized.

There are some programs that try to quantize them, but it looks awful.


Cue the discussion on webp, jpeg-xl, avif and so on, and how they will never be adopted...

[birdie]

PNGs take up much more bandwidth, even in very compressible content like these screenshots, even when heavily optimized.

There are some programs that try to quantize them, but it looks awful.


Cue the discussion on webp, jpeg-xl, avif and so on, and how they will never be adopted...

PNG compresses screenshots made mostly of text a lot better than JPEG unless JPEG at 70% with a ton of fringing is OK for you. Modern web stacks weigh megabytes, the poor screenshot in the article would barely take more than 100KB in PNG while being picture perfect. We are not in 1995 while using 33.6Kbit modems.

Let's check the next article,

JPEG: 230,595 bytes (looks like absolute crap)
PNG: 183,786 bytes (100% pristine and bit perfect)

Take up more bandwidth you say? I'm getting bloody angry from a ton of things I've been reading in the comments section lately. It's like each new person wants to embarrass themselves even more than the previous one already did.

[QuImUfu]

PNG compresses screenshots made mostly of text a lot better than JPEG unless JPEG at 70% with a ton of fringing is OK for you. Modern web stacks weigh megabytes, the poor screenshot in the article would barely take more than 100KB in PNG while being picture perfect. We are not in 1995 while using 33.6Kbit modems.

Let's check the next article,

JPEG: 230,595 bytes (looks like absolute crap)
PNG: 183,786 bytes (100% pristine and bit perfect)

Take up more bandwidth you say? I'm getting bloody angry from a ton of things I've been reading in the comments section lately. It's like each new person wants to embarrass themselves even more than the previous one already did.

These pictures look identical to me at 100%. As does the image as WebP (that takes only 94,566 bytes). As does a PNG with a limited palette (that takes only 86,577 bytes).

So the images are neither as small as possible without significant losses in visual quality nor lossless.
The software used to write this site should be improved, but probably no one cares enough.

[skeevy420]

PNG compresses screenshots made mostly of text a lot better than JPEG unless JPEG at 70% with a ton of fringing is OK for you. Modern web stacks weigh megabytes, the poor screenshot in the article would barely take more than 100KB in PNG while being picture perfect. We are not in 1995 while using 33.6Kbit modems.

Let's check the next article,

JPEG: 230,595 bytes (looks like absolute crap)
PNG: 183,786 bytes (100% pristine and bit perfect)

Take up more bandwidth you say? I'm getting bloody angry from a ton of things I've been reading in the comments section lately. It's like each new person wants to embarrass themselves even more than the previous one already did.

Your PNG was cropped. It doesn't have the Phoronix logo or Gitlab Sidebar. Of course a cropped image with things removed will have a smaller file size.

Anyhoo, my PC reports the article JPG at 226kb, GIMP with the default settings converted that to a PNG at 601kb, GIMP with the TinyPNG trick netted it down to 212kb, Gimp to WebP goes down to 134kb, AVIF was 123kb.

JPG	226kb
PNG	610kb
TinyPNG	212kb
WebP	134kb
AVIF	123kb

Note: DO NOT use the photo viewer in Windows 10 to view an AVIF image. It looks like crap with their decoder. The same image looks great in Firefox.

[sdack]

This will probably just have made the news only so Phoronix can whinge again about the supposedly missing Zen optimizations. Anyhow, I bet there are a lot of programmers who use the square root function without needing it, where they could just square their formula to get rid of the square root. And why is this instruction on Zen so much better when it is most likely implemented as microcode anyway? An inlined, discrete algorithm of sqrt() could possibly see further optimizations in the scope it was used in.