VPBLENDMQ

VPBLENDMQ builds each quadword of its destination from one of two sources, chosen by a mask register: the second source where the mask bit is set, the first where it is clear, or zero there with {z}. It is the quadword form of VPBLENDMD, whose page sets it against masked moves and the older blends. With at most eight quadwords it reads only the low 8 bits of the mask.

Recorded on Zen 5#

• The low byte of the mask. In vpblendmq zmm0{k7}, zmm31, zmm16 K7 held EBB9F4D13D1DF2E3h; its low byte, E3h, took quadwords 0, 1 and 5 to 7 from ZMM16 and 2 to 4 from ZMM31, and none of the old ZMM0 survived.
• Broadcast. vpblendmq zmm17{k3}, zmm30, qword bcst [mem] wrote the memory quadword into the six elements where EBh, the low byte of K3, has a bit set, and ZMM30’s quadwords 2 and 4 into the others.
• Narrow forms use fewer bits: vpblendmq ymm17{k5}, ymm30, ymmword ptr [mem] took only element 0 from memory, by bits 3:0 of K5, and cleared bits 256 and up.

In compiled code#

_mm512_mask_blend_epi64(k, a, b) takes b where k is set. GCC 14 and Clang 23 at -O2 with -march=x86-64-v4 compile it as the word form on VPBLENDMW’s page: a masked move or load when the result may replace a or b is in memory, VPBLENDMQ when the operands come the other way round. Neither folds a broadcast: with _mm512_set1_epi64(*p) as an operand both broadcast first with VPBROADCASTQ and then move under the mask.

A select with a constant fallback is where both emit it. For o[i] = a[i] > b[i] ? c[i] : 5 over int64_t arrays, both compare into a mask register, GCC with VPCMPQ and Clang with the EVEX VPCMPGTQ, and blend a vector of fives with c[i] read straight from memory, GCC on ZMM registers and Clang on YMM. For o[i] = a[i] ? 7 : o[i] they part as for words: GCC blends with VPBLENDMQ, Clang marks the nonzero elements with VPTESTMQ and stores sevens with a masked VMOVDQU64.

AMD’s Zen 5 latency spreadsheet gives VPBLENDMQ one cycle and four per cycle on any of the four vector pipes, as VPBLENDMD. The 6 recorded cases cover the register, memory and broadcast forms at all three widths.

Forms

Latency and throughput

Data flow

Where the elements of the sources go, element 0 on the right; results that depend on a control value are drawn for an example. How to read.

Operation

What the forms do, in pseudocode that our tests run against results recorded on real processors. How to read.

for i in range(VL // 64):
    if k1[i]:
        zmm1.q[i] = src3.q[i]
    elif zeroing:
        zmm1.q[i] = 0
    else:
        zmm1.q[i] = zmm2.q[i]
# m64bcst: src3.q[i] is the m64 for every i
# zmm1 is zeroed from bit VL up

Operands

Flags

None: XED records no flag that the forms test or change.

Exceptions

The exception class is XED's name for the set of exception conditions that a form follows. How to read.

All forms

AVX512-E4

Attributes

What else XED records about the forms. How to read.

All forms

opmask selects the source

Memory forms

embedded broadcast (tuple Full); memory fault suppression

CPUID and processors

Each form belongs to an XED ISA set. A form can be used when the processor reports every CPUID bit of one of its ISA set's alternatives.

More AVX10.1 instructions

• VPACKSSWB
• VPACKUSDW
• VPACKUSWB
• VPADDB
• VPADDD
• VPADDQ
• VPADDSB
• VPADDSW
• VPADDUSB
• VPADDUSW
• VPADDW
• VPALIGNR
• VPANDD
• VPANDND
• VPANDNQ
• VPANDQ
• VPAVGB
• VPAVGW
• VPBLENDMB
• VPBLENDMD
• VPBLENDMW
• VPBROADCASTB
• VPBROADCASTD
• VPBROADCASTMB2Q
• VPBROADCASTMW2D
• VPBROADCASTQ
• VPBROADCASTW
• VPCLMULQDQ
• VPCMPB
• VPCMPD
• VPCMPEQB
• VPCMPEQD
• VPCMPEQQ
• VPCMPEQW
• VPCMPGTB
• VPCMPGTD
• VPCMPGTQ
• VPCMPGTW
• VPCMPQ
• VPCMPUB

All 686 AVX10.1 instructions