_mm256_slli_si256 and _mm256_srli_si256

[bronxzv]

the intrinsics _mm256_slli_si256 and _mm256_srli_si256 map to the AVX2 instructions vpslldq and vpsrldq which shift 128-bit values according to the AVXprogramming reference and the behavior ofthe SDE

I'll suggest to rename these intrinsics as _mm256_slli_si128 and _mm256_srli_si128 to be consistent with the rest of the nomenclature

the intrinsics guide explanationalso requires a modification since they are explained as 256-bit shifts at the moment

[Georg_Z_Intel]

Hello,

you're right that shifts are affecting both 128-bit lanes (for _mm256... variants). I've checked our documentation and it's (only) mentioned incorrectly in the Intrinsics Guide for Intel Advanced Vector Extensions 2 (v2.6).
I've filed a defect to fix that: DPD200272778.

Thank you for bringing this up!

Regarding the naming scheme: I don't see how your proposal would fit better, though. Post-fix si256 refers to the 256-bit size source vector register. That's still valid as we do 2x 128-bit value shifts. However, I agree that it's still a bit misleading... esp. for users who did not take a look a deeper look into the INTEL ADVANCED VECTOR EXTENSIONS (Ref. # 319433-011) document.
In the end it's also a question how compact you'd like those intrinsic names to be. Describing (full) semantics in the name would be too much.

Best regards,

Georg Zitzlsberger

[bronxzv]

"Post-fix si256 refers to the 256-bit size source vector register."

Ah yes I missed the "si" vs "epi"thing,but since the "s" in si256 means "scalar"it gives the wrong impressionthat it shifta single value when it's in fact shifting independently two packed value

I you insist keeping it as is, something that will make some sensewill be to introducesome aliases called "_mm256_srli_epi128" and "_mm256_slli_epi128" ? I'm sure it will be quite more clear for newbies

[Mark_S_Intel1]

The compiler development team decided to introduce the new "_si128" names _mm256_slli_si128 and _mm256_srli_si128, and keepthe "_si256" names for compatibility since they were already released.

Thanks,
--mark

[bronxzv]

why not "_epi128" or "_pi128" instead ?, these are shifting two packed 128-bit values not a single scalar value

[Mark_S_Intel1]

Thenew names will be"_mm_bs[rl]li_si128" and "_mm256_bs[rl]li_epi128" .

--mark

[Georg_Z_Intel]

Hello,

the new names will be available with Intel Composer XE 2011 Update 10.
Thank you Mark for taking care!

DPD200272778 is still in the works.

Best regards,

Georg Zitzlsberger