• Guillaume Piolat (17/17) Dec 06 DUB package intel-intrinsics is out in version v1.13.0.
• Serg Gini (19/32) Dec 08 Congratulations! Very nice to have this package.
• Guillaume Piolat (10/15) Dec 08 (reminder that intel-intrinsics is translated to Neon)

Guillaume Piolat <first.nam_e gmail.com> writes:

DUB package intel-intrinsics is out in version v1.13.0.
Now with support for all AVX and AVX2 intrinsics.

- Supported: SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, BMI2, AVX, 
AVX2
- Target x86, x86_64, arm64, arm32. All intrinsics have 
emulation/alt paths.
- DMD, LDC, or GDC

DUB: https://code.dlang.org/packages/intel-intrinsics

Finishing the AVX2 set took a surprisingly long time.

Now: should you use AVX for consumer software in D?
- The reach isn't very good, very nascent support on Microsoft 
Prism and Apple Rosetta.
- In arm, neon vectors are 128-bit anyway.
- You will encounter: https://github.com/dlang/dub/issues/3080 
(unless using redub or --combined)

Typical safe targets for consumer software range imho from SSE3 
to SSE4.2 top.

Serg Gini <kornburn yandex.ru> writes:

On Sunday, 7 December 2025 at 00:42:11 UTC, Guillaume Piolat 
wrote:
 DUB package intel-intrinsics is out in version v1.13.0.
 Now with support for all AVX and AVX2 intrinsics.

 - Supported: SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, BMI2, AVX, 
 AVX2
 - Target x86, x86_64, arm64, arm32. All intrinsics have 
 emulation/alt paths.
 - DMD, LDC, or GDC

Congratulations! Very nice to have this package.

 Now: should you use AVX for consumer software in D?

I would say - yes as much as you can! If it is not AVX512, but 
AVX and AVX2 could be applied.
Maybe the flow could be started with auto-vec for simpler code 
maintenance.
But nowadays, AVX(2) is supported pretty widely I think.

Projects for inspiration could be taken from Lemire and Ash.
Their projects SIMD-powered used in other consumer products.
simdutf, simdjson, usearch and others

https://github.com/lemire (blog https://lemire.me/blog/)
https://github.com/ashvardanian (blog 
https://ashvardanian.com/archives/)

 - The reach isn't very good, very nascent support on Microsoft 
 Prism and Apple Rosetta.
 - In arm, neon vectors are 128-bit anyway.
 - You will encounter: https://github.com/dlang/dub/issues/3080 
 (unless using redub or --combined)

Probably separate ARM targeted arm-intrinsics package could be 
created with targeting:
- NEON
- SVE
- SVE2

Guillaume Piolat <first.nam_e gmail.com> writes:

On Monday, 8 December 2025 at 08:41:55 UTC, Serg Gini wrote:
 Probably separate ARM targeted arm-intrinsics package could be 
 created with targeting:
 - NEON
 - SVE
 - SVE2

(reminder that intel-intrinsics is translated to Neon)

Other than for the GATHER instructions, the mismatch isn't large. 
And lot's of code (eg: stb) uses these intrinsics for anything 
SIMD.

But the next intel-intrinsics will probably need to be 
arm-intrinsics as x86 fades out (if it fades out, people disagree 
on that) and maybe do the reverse, translate to x86 intrinsics.

It doesn't even need to be in a separate project, as 
simd-everywhere (the inspiration) translate all <=> all.