if the goal is to vectorize all of the implementations, i'm going to be here for the rest of my life

MMaxine if the goal is to vectorize all of the implementations, i'm going to be here for...

tannergooding•10/21/23, 12:12 AM

https://github.com/dotnet/runtime/blob/main/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/TensorPrimitives.netcore.cs#L10876-L10926
>.>

tannergooding•10/21/23, 12:13 AM

(doesn't have acosh, asinh, or atanh atm; but has the other 3 + exp + log + a few others)

Aqua•10/21/23, 12:13 AM

what is a tensor primitive

Ttannergooding <https://github.com/dotnet/runtime/blob/main/src/libraries/System.Numerics.Tenso...

MaxineOP•10/21/23, 12:13 AM

holy cow, is this actually fast?

tannergooding•10/21/23, 12:13 AM

it's "tensor operations" but directly over a Span<T>Span<T>

Aqua•10/21/23, 12:14 AM

i see

MaxineOP•10/21/23, 12:14 AM

i feel like for Vector3, using FPU operations in serial is still faster than doing all of this SIMD stuff

tannergooding•10/21/23, 12:14 AM

tensors being just the abstract terminology encompassing scalar + vector + dense matrix + sparse matrix

Aqua•10/21/23, 12:14 AM

In mathematics, a tensor is an algebraic object that describes a multilinear relationship between sets of algebraic objects related to a vector space.

why is math

MMaxine i feel like for Vector3, using FPU operations in serial is still faster than doi...

tannergooding•10/21/23, 12:14 AM

not even close

Ttannergooding not even close 😄

MaxineOP•10/21/23, 12:14 AM

really?

MaxineOP•10/21/23, 12:14 AM

for 3 components?

tannergooding•10/21/23, 12:15 AM

yeah, the x87 FPU (assuming you're referring to fsinfsin and friends) has been effectively deprecated for 20 years
and SSE/SSE2 has been faster for nearly 25

for the scalar implementation, you're doing SIMD operations anyways, just scalar ones (so addssaddss rather than addpsaddps)

MaxineOP•10/21/23, 12:16 AM

i hate computers

tannergooding•10/21/23, 12:16 AM

so the SIMD algorithm is basically the scalar one, just doing 3 at once
so it takes the same amount of time

MaxineOP•10/21/23, 12:16 AM

does this apply for all operators?

MaxineOP•10/21/23, 12:17 AM

i would think serial divide vs SIMD divide would not make a difference

tannergooding•10/21/23, 12:17 AM

the one exception is you replace branching with ConditionalSelect
and so you still relegate the exceptional cases

for example, with LogLog we have the fast path which is the normal scalar algorithm
and only if we have a special case (zero, infinity, nan, subnormal) do we do the extra computation
https://github.com/dotnet/runtime/blob/main/src/libraries/System.Numerics.Tensors/src/System/Numerics/Tensors/TensorPrimitives.netcore.cs#L11004-L11047

MaxineOP•10/21/23, 12:18 AM

hmm interesting

tannergooding•10/21/23, 12:18 AM

the exceptional cases are expected to be exceptional, so the branch is fine for the majority case where we don't encounter them
and no worse than we'd have for scalar anyways

MMaxine i would think serial divide vs SIMD divide would not make a difference

tannergooding•10/21/23, 12:20 AM

think about it this way, a scalar floating-point dividescalar floating-point divide is [11; 12] cycles[11; 12] cycles
3 of them is then [33; 36] cycles[33; 36] cycles

where-as a packed floating-point dividepacked floating-point divide is also [11; 12] cycles[11; 12] cycles
because SIMD is implemented by having countcount parallel units

tannergooding•10/21/23, 12:21 AM

but those units aren't separate ports; and so you can't dispatch count scalar instructions to them at the same time
because they logically execute together, so you can't have two cores trying to compete for them (how hyperthreading works)

MaxineOP•10/21/23, 12:21 AM

hm.

tannergooding•10/21/23, 12:22 AM

so using SIMD is almost always worth it; especially if you don't need to branch or conditional select
if you do need to branch or conditional select, then you often need to tweak it a bit
but it tends to be no worse than the equivalent scalar code

MaxineOP•10/21/23, 12:22 AM

that's wild

MaxineOP•10/21/23, 12:22 AM

thank you!

tannergooding•10/21/23, 12:23 AM

sure thing!

in general most of the C based math functions are available under the MIT license from AMD here: https://github.com/amd/aocl-libm-ose/tree/aocl-4.1
so it's a solid place to port things from

AMD's libm is the current implementation used by Windows and MSVC internally (and has been for years)

tannergooding•10/21/23, 12:23 AM

(although I think we tend to be a couple iterations behind)

tannergooding•10/21/23, 12:24 AM

https://github.com/amd/aocl-libm-ose/tree/aocl-4.1/src/optimized tends to be the most up to date
and the vectorized sometimes isn't as optimal as possible, I took some liberties on improving it for .NET, since we can take advantage of having a JIT and compiling for the current hardware, as an example

MaxineOP•10/21/23, 12:25 AM

inchresting

tannergooding•10/21/23, 12:25 AM

the vectorized is also often slightly faster and allows slightly more error than the scalar version
since the domains it gets used (like games) often allow and like that
but, you can always port the scalar for extra precision if you want that

PPerksey Someone should port miniaudio to C#

dfkeenan🍍🍕•10/21/23, 12:47 AM

https://tenor.com/view/himym-how-i-met-your-mother-neil-patrick-harris-barney-challenge-gif-4820522

Tenor

Challenge Accepted

dfkeenan🍍🍕•10/21/23, 12:48 AM

Ok, maybe not. i have so many things on the go already

MaxineOP•10/21/23, 1:02 AM

i saw it was all one source file, and then closed the tab

MaxineOP•10/21/23, 1:02 AM

given my success porting blender's triangulation alg, i don't trust myself to do a better job here

dfkeenan🍍🍕•10/21/23, 1:06 AM

what we need is a tool that can take c/c++ and convert it to C# using the silk bindings.

Aqua•10/21/23, 1:07 AM

c-clangsharp?
also that is what the new buildtools is being designed to do

AAqua c-clangsharp? also that is what the new buildtools is being designed to do

MaxineOP•10/21/23, 1:07 AM

not what keenan meant

Aqua•10/21/23, 1:09 AM

Aqua•10/21/23, 1:10 AM

oh wait
literally something that transpiles c/c++ into c#?

dfkeenan🍍🍕•10/21/23, 1:10 AM

yeah

Aqua•10/21/23, 1:10 AM

there is something out there that exists

Aqua•10/21/23, 1:10 AM

I think it only works with C though

Aqua•10/21/23, 1:11 AM

it's used in StbImageSharp

MaxineOP•10/21/23, 1:11 AM

not all C is transpilable to C#

tannergooding•10/21/23, 1:16 AM

ClangSharp handles basically everything it can. The stuff it can’t handle is stuff like templates, because there is no equivalent

tannergooding•10/21/23, 1:17 AM

It could try to do generics and type checks in some cases, but it doesn’t translate cleanly without interface constraints

And of course you can’t implement such interfaces on existing types you don’t own

dfkeenan🍍🍕•10/21/23, 1:18 AM

could you not pre "expand" the macros then translate that to C#?

Lyris the Kitori•10/21/23, 1:18 AM

that wont work for the macros themselves

Lyris the Kitori•10/21/23, 1:18 AM

sure you can preprocess it

if the goal is to vectorize all of the implementations, i'm going to be here for the rest of my life

Similar Threads