I have exactly this problem. It is reasonably easy to convert to SoA using
something like this:
struct tComplexFloat
{
float Real;
float Imag;
};
static inline tComplexFloat ReadFromArrayOfComplexFloat(
uniform const tComplexFloat Input[], int32 Index
)
{
const uint64 Raw = ((uniform uint64 *)Input )[Index];
float Imag = floatbits((uint32)(Raw >> 32));
float Real = floatbits((uint32)(Raw & 0xFFFFFFFF));
const tComplexFloat Value = {Real, Imag};
return Value;
}
...
foreach(i = 0...N)
{
tComplexFloat Value = ReadFromArrayOfComplexFloat(InputData, i);
// Operate on Value
}
However, while this might be more efficient than using a gather, it is
still significantly less efficient than hand-coded (with the FMADDSUB, in
our case) C+intrinsics.
This is the main reason why we are still using intrinsics, instead of ISPC,
for a lot of our code. However, I don't see how ISPC could make use of the
ADDSUB instructions, due to the programming model. You could, of course,
manually swap values between program instances and work on them that way
but this would be complicated and likely to result in code just as
difficult to write and understand as using intrinsics.
I would also be very interested to learn if there is a way around this.
Regards,
S.
On Wednesday, 12 January 2022 at 09:33:43 UTC [email protected] wrote:
> Dear ISPC Community,
>
> Is there any support for arrays of complex numbers in ISPC?
>
> The inadequacy of std::complex and complex to SIMDization,
> due to the AoS layout, can be partially relaxed by the ISA (see vaddsubps
> for example).
>
> Having seamless support of ISPC for std::complex would be best.
> My guess is that there would be some incompatibility in terms of ISPC
> masking?
> (For complex<float>, I would mask i32x8 or i64x4 depending on the
> operation at hand.)
>
> It would be great to have at least some builtin functions dealing with
> such arrays,
> to "unzip" them into real and imaginary component arrays (SoA), zip them
> again, and builtin functions z = f(x,y), hence consuming 4 input scalars
> and producing 2 output scalars.
>
> What are your thoughts on this?
>
> Best wishes,
>
> Diego
>
>
--
You received this message because you are subscribed to the Google Groups
"Intel SPMD Program Compiler Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion on the web visit
https://groups.google.com/d/msgid/ispc-users/eac895ef-3541-43fa-85c5-c81534e6c52dn%40googlegroups.com.
