VPCMPB

VPCMPB compares 16, 32 or 64 pairs of signed bytes and writes one bit per pair to a mask register. The immediate picks the test, of which only bits 2:0 count, as on VPCMPD’s page: 0 equal, 1 less, 2 less or equal, 4 not equal, 5 greater or equal, 6 greater, 3 false and 7 true. VPCMPUB is the unsigned twin. Both need AVX512BW (or AVX10.1) and take no broadcast operand. SSE and AVX compare bytes only with PCMPEQB and PCMPGTB, which write a vector of all-ones and zero bytes.

Recorded on Zen 5#

• Signed order. In vpcmpb k6{k1}, zmm17, zmm30, 0x1, less than, the pair 80h and 01h set its bit, -128 being below 1, and 7Fh against FFh left it clear, 127 not being below -1; an unsigned compare would answer both the other way.
• The write mask clears. Only the 32 elements that K1 selected could be set, and every bit outside K1 came out 0, whatever K6 held.
• Predicate 7 copies the mask. vpcmpb k1{k5}, xmm24, xmm3, 0x7 gave D5E1h, the low 16 bits of K5, and cleared bits 63:16.
• Bits 7:3 are ignored. vpcmpb k3, ymm9, ymm9, 0xa tested less or equal, as 2 does, and gave FFFFFFFFh for 32 equal pairs.

In compiled code#

With -march=x86-64-v4, GCC 14 at -O2 compiles every byte compare intrinsic on registers to VPCMPB with its predicate, _mm512_cmpeq_epi8_mask and _mm512_cmpgt_epi8_mask included, while Clang 23 keeps VPCMPEQB and VPCMPGTB for those and uses VPCMPB, printed vpcmpneqb or vpcmpnltb, for not equal and greater or equal. Less than swaps the operands into VPCMPGTB, unless the second source is in memory: for o[i] = a[i] < b[i] ? a[i] : 0 over int8_t arrays, Clang emits vpcmpltb with b[i] as the memory operand on YMM registers, and GCC VPCMPB with predicate 1 on ZMM, each followed by a zero-masked VMOVDQU8.

AMD’s Zen 5 latency spreadsheet lists VPCMPB as one macro-op on pipes FP1 or FP2, two for a memory source, with 3, 4 and 6 cycles of latency for XMM, YMM and ZMM, as VPCMPUB.

In the optimization guide

What the optimization guide says about VPCMPB, a line per claim. Each links to the claim in the guide, with its sources and where they differ.

Zen 4

• Vectors and floating point: An AVX-512 compare that writes a mask register, such as VPCMPD, takes 3, 4 or 5 cycles for 128, 256 or 512 bits, where the older compares into a vector register take one cycle for integers and two for floats. sources differ

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 // 8):
    sa = zmm2.sb[i]
    sb = src3.sb[i]
    k1[i] = k2[i] and compare(imm8, sa, sb)
# imm8 & 7: 0 ==, 1 <, 2 <=, 3 false, 4 !=, 5 >=, 6 >, 7 true
# the other bits of k1 are cleared

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

write masking

Memory forms

memory fault suppression

Intrinsics

The C intrinsics of Clang 23.1.2 that compile to VPCMPB, each with the form it compiles to and the Clang options it needs. The forms come from compiling every intrinsic and decoding the result with Intel XED; all intrinsics A–Z explains the method.

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

• VPADDUSW
• VPADDW
• VPALIGNR
• VPANDD
• VPANDND
• VPANDNQ
• VPANDQ
• VPAVGB
• VPAVGW
• VPBLENDMB
• VPBLENDMD
• VPBLENDMQ
• VPBLENDMW
• VPBROADCASTB
• VPBROADCASTD
• VPBROADCASTMB2Q
• VPBROADCASTMW2D
• VPBROADCASTQ
• VPBROADCASTW
• VPCLMULQDQ
• VPCMPD
• VPCMPEQB
• VPCMPEQD
• VPCMPEQQ
• VPCMPEQW
• VPCMPGTB
• VPCMPGTD
• VPCMPGTQ
• VPCMPGTW
• VPCMPQ
• VPCMPUB
• VPCMPUD
• VPCMPUQ
• VPCMPUW
• VPCMPW
• VPCOMPRESSB
• VPCOMPRESSD
• VPCOMPRESSQ
• VPCOMPRESSW
• VPCONFLICTD

All 686 AVX10.1 instructions