CODEWITHSUNDEEP
c#simdintrinsics
user12722843
user12722843

Reputation:

Are there macros for SIMD instruction sets?

Are there any macros that we can use for instruction set detection?

I know that we have the runtime ones:

if (Avx.IsSupported)
{
//...
}

but do we have defined symboles for that e.g.

#if AVX
//...
#endif

And if there are none, can we create ones or do we have a workaround for that?

NOTE

.method public hidebysig specialname static 
    bool get_IsSupported () cil managed 
{
    // Method begins at RVA 0x142fa4
    // Code size 6 (0x6)
    .maxstack 8

    // return IsSupported;
    IL_0000: call bool System.Runtime.Intrinsics.X86.Avx::get_IsSupported()
    IL_0005: ret
} // end of method Avx::get_IsSupported

public int F() 
{
    if (Avx.IsSupported)
        return 0;
    if (Avx2.IsSupported)
        return 1;

    return 2;
}

asm

Program.F()
    L0000: xor eax, eax
    L0002: ret

As you can see it somehow figured out that Avx is supported and didn't include the branches. Is this a normal and well defined behaviour?

Upvotes: 2

Views: 358

Answers (2)

Eamon
Eamon

Reputation: 1909

if (Avx.IsSupported)
{
}
else
{
}

is the correct way to achieve what you want:

The IsSupported checks are treated as runtime constants by the JIT (when optimizations are enabled) and so you do not need to cross-compile to support multiple different ISAs, platforms, or architectures. Instead, you just write your code using if-statements and the unused code paths (any code path which is not executed, due to the condition for the branch being false or an earlier branch being taken instead) are dropped from the generated code (the native >assembly code generated by the JIT at runtime).

https://devblogs.microsoft.com/dotnet/hardware-intrinsics-in-net-core/#:~:text=The%C2%A0IsSupported%C2%A0checks%20are,by%20the%20JIT%20at%20runtime).

Upvotes: 1

Ondrej Tucny
Ondrej Tucny

Reputation: 27962

Avx.IsSupported is a run-time feature because it needs to dynamically detect the capabilities of the target CPU where your code is actually being run.

On the other hand, #if is a compile-time feature, conditional compilation. It's not really useful to have an AVX conditional compilation symbol and then just presume for the target CPU to do have support for AVX. It is simply better to check on run-time and behave accordingly.

However, should you really need to use conditional compilation, you can always declare your own symbol.

Upvotes: 2

Related Questions