Emulating AVX512 vpcompressd

[Mysticial]

I can't say why anyone would want to do this, but I'm looking for a 2nd set of eyes.

Can this instruction:

    vpcompressd [mem]{k}, zmm

be emulated as follows:

    vpcompressd zmm0{k1}{z}, zmm0
    kmovd       eax, k1
    pext        eax, -1, eax     (the -1 in a register of course)
    kmovd       k1, eax
    vmovdqu32   [mem]{k1}, zmm0

I haven't tested this sequence yet, but I feel like this is too simple to be correct.

[bsquared]

Originally Posted by Mysticial

I can't say why anyone would want to do this, but I'm looking for a 2nd set of eyes.

Can this instruction:

    vpcompressd [mem]{k}, zmm

be emulated as follows:

    vpcompressd zmm0{k1}{z}, zmm0
    kmovd       eax, k1
    pext        eax, -1, eax     (the -1 in a register of course)
    kmovd       k1, eax
    vmovdqu32   [mem]{k1}, zmm0

I haven't tested this sequence yet, but I feel like this is too simple to be correct.

Compress epi32's into zmm (packing selected words into lower-order locations of zmm).
Pack the selected mask-bit locations into the lower-order bits of eax(k1).
Mask store them into unaligned memory.

Looks good to me!

Why do you want to do this, again?

[Mysticial]

Originally Posted by bsquared

Compress epi32's into zmm (packing selected words into lower-order locations of zmm).
Pack the selected mask-bit locations into the lower-order bits of eax(k1).
Mask store them into unaligned memory.

Looks good to me!

Why do you want to do this, again?

Thanks! I think it could also be done as:

    vpcompressd zmm0{k1}{z}, zmm0
    vpcompressd zmm1{k1}{z}, [-1]      (constant of all 1s)
    vpcmpd k1, zmm1, [-1], 0           (constant of all 1s), compare for equality
    vmovdqu32   [mem]{k1}, zmm0

There may be a reason to do this soon.

[janwas]

Ah :) Now the reason becomes clear (Zen4).
FWIW users of compressstoreu will typically want to know the popcount of the mask, so they can increment pointers. Seems like that would work better with the PEXT approach because popcount also requires the mask in a GPR.

[bsquared]

Originally Posted by janwas

Ah :) Now the reason becomes clear (Zen4).
FWIW users of compressstoreu will typically want to know the popcount of the mask, so they can increment pointers. Seems like that would work better with the PEXT approach because popcount also requires the mask in a GPR.

Agreed. Yafu uses compressstoreu along with popcount in its bucket sieve.