Silk.NET•3y ago

trying to reuse code for vectorized equals for all of the primitives... time to see if the JIT can l

trying to reuse code for vectorized equals for all of the primitives... time to see if the JIT can lower this

MMaxine trying to reuse code for vectorized equals for all of the primitives... time to ...

MaxineOP•10/22/23, 1:29 AM

*ignore the obvious errors

MaxineOP•10/22/23, 1:54 AM

i got it to work by removing the loop

MaxineOP•10/22/23, 1:54 AM

shame it can't be unrolled but, good enough

Perksey•10/22/23, 1:55 AM

what code can’t it unroll

MaxineOP•10/22/23, 1:55 AM

sequenceequal at the bottom

MaxineOP•10/22/23, 1:56 AM

basically, the JIT can't inline the length in SpanHelpers.SequenceEqual because it's a mutated variable

Perksey•10/22/23, 1:57 AM

you can probably make an unrolling impl by recursion + inlining

MaxineOP•10/22/23, 1:57 AM

my solution, since this method is only meant to handle scalar lists of length 1 to 4, is to hardcode the equality checks

MaxineOP•10/22/23, 1:57 AM

god help me

dfkeenan🍍🍕•10/22/23, 2:00 AM

congrats on getting v 2.18 out

dfkeenan🍍🍕•10/22/23, 2:01 AM

time to update "all" my things

PPerksey you can probably make an unrolling impl by recursion + inlining

MaxineOP•10/22/23, 2:15 AM

hard to tell, but i think this impl is just recursing

MaxineOP•10/22/23, 2:15 AM

maybe i'm just doing the pattern slightly wrong and the JIT doesn't get it

MaxineOP•10/22/23, 2:26 AM

should we have interfaces for the static methods on primitives? i guess i can implement them explicitly to discourage direct use

dfkeenan🍍🍕•10/22/23, 2:50 AM

Had no dramas updating. (Though nuget restore went wonky and I had to update twice.

)

MaxineOP•10/22/23, 3:12 AM

another optimization dead-end:

static unsafe bool IsCanonical<T>(Vector3D<T> value) where T : INumberBase<T> => FastAndDelegateForScalarEntries<Vector3D<T>, T>(value, &T.IsCanonical);

internal static unsafe bool FastAndDelegateForScalarEntries<T, TElem>(in T self, delegate*<TElem, bool> toCall)
{
  ... etc
}

static unsafe bool IsCanonical<T>(Vector3D<T> value) where T : INumberBase<T> => FastAndDelegateForScalarEntries<Vector3D<T>, T>(value, &T.IsCanonical);

internal static unsafe bool FastAndDelegateForScalarEntries<T, TElem>(in T self, delegate*<TElem, bool> toCall)
{
  ... etc
}

for T=intint, JIT attemps to lower it, but doesn't realize that int.IsCanonical always returns true, so the entire method should become mov eax, 1; ret 0xcmov eax, 1; ret 0xc

MaxineOP•10/22/23, 3:13 AM

tanner pls fix /j

MaxineOP•10/22/23, 3:27 AM

are static virtual interface methods meant to not be callable from inside the derived type?

khitiara•10/22/23, 3:32 AM

i think they should work?

khitiara•10/22/23, 3:32 AM

not 100% sure

Perksey•10/22/23, 3:35 AM

@grifon.petre how much bandwidth do you have atm

Perksey•10/22/23, 3:36 AM

i want to nerdsnipe someone else into writing a competing solution to the overload problem

Perksey•10/22/23, 3:36 AM

so that we have multiple prototypes to try out in a short timeframe

Kkhitiara i think they should work?

MaxineOP•10/22/23, 3:37 AM

it don't work!

khitiara•10/22/23, 3:37 AM

huh

Perksey•10/22/23, 3:37 AM

you need a TSelf

khitiara•10/22/23, 3:37 AM

that almost feels like a bug lol

MaxineOP•10/22/23, 3:37 AM

yeah its such a strange design decision

khitiara•10/22/23, 3:37 AM

the static abstracts can be called from the member type but also i guess those have a separate declaration on that type

MaxineOP•10/22/23, 3:37 AM

yeah

MaxineOP•10/22/23, 3:38 AM

i wanted to use interfaces to reuse code

MaxineOP•10/22/23, 3:38 AM

that ain't happening now

MaxineOP•10/22/23, 3:38 AM

if c# had templates this shit would be done in minutes

PPerksey i want to nerdsnipe someone else into writing a competing solution to the overlo...

Perksey•10/22/23, 3:39 AM

so the glass in my mind has been shattered a little bit

Perksey•10/22/23, 3:39 AM

and... it's workable but it's making me realise just how ridiculous it is what we're doing

Perksey•10/22/23, 3:39 AM

i still abstain that this is the only way without excessive overloads to achieve the level of flexibility we want

Perksey•10/22/23, 3:39 AM

but okay here it goes

Perksey•10/22/23, 3:39 AM

the Ptr type

Perksey•10/22/23, 3:39 AM

how do you store one?

Perksey•10/22/23, 3:40 AM

like in a class

MaxineOP•10/22/23, 3:40 AM

not meant to

Perksey•10/22/23, 3:40 AM

but

Perksey•10/22/23, 3:40 AM

then we have no safe storage medium

MaxineOP•10/22/23, 3:40 AM

you store either a T* or a T[], and that gets cast to Ptr<T> implicitly

Perksey•10/22/23, 3:40 AM

ok but consider glfwCreateWindow

Perksey•10/22/23, 3:40 AM

it returns a pointer

MaxineOP•10/22/23, 3:41 AM

it should instead return some sort of handle to GLFWwindow

MaxineOP•10/22/23, 3:41 AM

that can be cast to ptr<T> as well

Perksey•10/22/23, 3:42 AM

yeah that's where i'm at. making a StoredPtrStoredPtr type (or add a UntrackedUntracked variant of each type so you can just do PtrPtr<T>.UntrackedPtrPtr<T>.Untracked)

Perksey•10/22/23, 3:42 AM

but that's kinda crap

trying to reuse code for vectorized equals for all of the primitives... time to see if the JIT can l

Similar Threads

Similar Threads

Similar Threads