AVX2 broadcast from register?

[acaly]

If I understand correctly, all broadcast intrinsics go in the Avx2 class. There are a bunch of overloads, but for parameters smaller than or equal to 64-bit, there are only overloads that accept memory pointers. The x86 vpbroadcastx (where x = b/w/d/q) instruction supports both memory form and register form. See, for example, https://www.felixcloutier.com/x86/vpbroadcast . The Avx2 class is not providing these methods now.

There are workarounds of using the memory form, taking the address of the local variable, or using the Create method, but in either case, .NET cannot optimize it into the desired vpbroadcastq (sharplab).

So what should we do here if we want to broadcast from a double (or a Vector64<double>) to a Vector256<double>?

[acaly]

I found an answer myself. It seems that we can use Avx2.BroadcastScalarToVector256(Vector128.CreateScalarUnsafe(tmp)).

But I am not sure this is good, because we must use the "Unsafe" form in order for the optimization to occur.

[EgorBo]

Unsafe here means that it is not guaranteed that all components of Vector128 except the first one will be 0, there can be some garbage values. Which is perfectly fine for your case since you're broadcasting the first component only

[tfenise]

There is an overload of Vector256.Create that takes only one parameter, which should be optimized properly.