www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - auto vectorization observations

reply Bruce Carneal <bcarneal gmail.com> writes:
Auto vectorization (autovec) can yield significant performance 
improvements but this back end technology may struggle with some 
very simple forms, leaving you with lackluster performance.  When 
that happens in performance critical code D's __vector types will 
come in handy.

ldc and gdc differ in their autovec capabilities with gdc coming 
out ahead in at least one important area: dealing with 
conditionals.

As an example, gdc is able to vectorize the following for both 
ARM SVE and x86-SIMD architectures while ldc, per my godbolt 
testing at least, can not.

```d
alias T = ubyte; // 16 wide (128 bit HW) to 64 wide (512 bit HW)
alias CT = const(T);

void choose(size_t n, CT* src, T threshold, CT* a, CT* b, T* dst)
{
     foreach(i; 0 .. n)
         dst[i] = src[i] < threshold ? a[i] : b[i];
}
```

You can handle conditionals manually in the __vector world but 
it's tedious and error prone so kudos Iain and the gcc crew.

Additional observations wrt D and auto vectorization, good and 
bad, are welcome.
Jun 08 2022
next sibling parent Guillaume Piolat <first.last gmail.com> writes:
On Wednesday, 8 June 2022 at 18:41:44 UTC, Bruce Carneal wrote:

 Additional observations wrt D and auto vectorization, good and 
 bad, are welcome.


Also failed to get it autovectorized in LDC.  restrict, going 
short, or else doesn't seem to work - strange.
Jun 08 2022
prev sibling parent reply Siarhei Siamashka <siarhei.siamashka gmail.com> writes:
On Wednesday, 8 June 2022 at 18:41:44 UTC, Bruce Carneal wrote:

 As an example, gdc is able to vectorize the following for both 
 ARM SVE and x86-SIMD architectures while ldc, per my godbolt 
 testing at least, can not.


I was unable to confirm this: https://d.godbolt.org/z/Y9fEvn83e 
(neither GDC nor LDC can vectorize it). Could you please post a 
link to your godbolt results with the right compiler versions and 
optimization options?
Jun 09 2022
parent reply Bruce Carneal <bcarneal gmail.com> writes:
On Thursday, 9 June 2022 at 14:28:31 UTC, Siarhei Siamashka wrote:

 On Wednesday, 8 June 2022 at 18:41:44 UTC, Bruce Carneal wrote:

 As an example, gdc is able to vectorize the following for both 
 ARM SVE and x86-SIMD architectures while ldc, per my godbolt 
 testing at least, can not.


 I was unable to confirm this: https://d.godbolt.org/z/Y9fEvn83e 
 (neither GDC nor LDC can vectorize it). Could you please post a 
 link to your godbolt results with the right compiler versions 
 and optimization options?


https://godbolt.org/z/1exqWT49c

The above is a link to gdc/ldc godbolt comparison with x86-64-v4 
targets. I'm not sure which subset of the v4 capabilities are 
required for gdc to vectorize the code.  It does not vectorize 
v3, so *very* recent x86s in any event.

The ARM story is similar, SVE is very recent. Note: I think I had 
to go to C equivalent code to get a godbolt visible ARM SVE 
compiler.

Auto vectorization could be the future but if you're targetting 
non-cutting-edge HW SIMT is a better bet if available and 
__vector if not.
Jun 09 2022
parent reply user1234 <user1234 12.de> writes:
On Thursday, 9 June 2022 at 20:32:44 UTC, Bruce Carneal wrote:

 On Thursday, 9 June 2022 at 14:28:31 UTC, Siarhei Siamashka 
 wrote:

 On Wednesday, 8 June 2022 at 18:41:44 UTC, Bruce Carneal wrote:

 As an example, gdc is able to vectorize the following for 
 both ARM SVE and x86-SIMD architectures while ldc, per my 
 godbolt testing at least, can not.


 I was unable to confirm this: 
 https://d.godbolt.org/z/Y9fEvn83e (neither GDC nor LDC can 
 vectorize it). Could you please post a link to your godbolt 
 results with the right compiler versions and optimization 
 options?


 https://godbolt.org/z/1exqWT49c

 The above is a link to gdc/ldc godbolt comparison with 
 x86-64-v4 targets. I'm not sure which subset of the v4 
 capabilities are required for gdc to vectorize the code.  It 
 does not vectorize v3, so *very* recent x86s in any event.


Thanks for the precision, I wondered the same as Siarhei.
gdc options "-O3 -mavx512bw" produceq the same output as your.
With no surprise LDC does not do better with -mattr=+avx512bw.
Jun 09 2022
parent Bruce Carneal <bcarneal gmail.com> writes:
On Thursday, 9 June 2022 at 23:35:11 UTC, user1234 wrote:

 On Thursday, 9 June 2022 at 20:32:44 UTC, Bruce Carneal wrote:

 Thanks for the precision, I wondered the same as Siarhei.
 gdc options "-O3 -mavx512bw" produceq the same output as your.
 With no surprise LDC does not do better with -mattr=+avx512bw.


Turns out gdc can vectorize it with the much more widely 
available x86-64-v3 target if your data elements are 4 bytes or 
larger.  Unfortunately ldc still can not.

Your avx512bw discovery prompted me to give the larger types a 
try.  Thanks for digging into it.

Here's the link:
https://godbolt.org/z/ajjjj8vfY
Jun 09 2022