x86-64.net

Instruction

VPMULHUW

Multiplies unsigned words and keeps the high 16 bits of each product.

Extensions
Forms
10 (4 VEX, 6 EVEX)
Data
Intel XED v2026.08.23

VPMULHUW multiplies unsigned words and keeps the high 16 bits of each 32-bit product: the VEX and EVEX form of PMULHUW, whose page shows it multiplying by a fraction and dividing by constants. VPMULHW is the signed twin, with the same encodings: AVX and AVX2 for VEX, AVX512BW or AVX10.1 for EVEX, with a write mask per word but no broadcast. On Zen 5, vpmulhuw ymm1, ymm2, ymm3 gave 7FFFh for 8000h times FFFFh, where VPMULHW’s recorded case gave 0000h for the same pair, 7FFEh for 7FFFh times FFFFh and 4000h for 8000h times 8000h, the one product on which both agree. vpmulhuw ymm7{k7}{z}, ymm12, ymm7 kept FFFEh, the top of FFFFh squared, in word 0 and zeroed the six words K7 leaves out.

In compiled code

In our builds at -O2 over fixed-size uint16_t arrays, ((uint32_t)b[i] * c[i]) >> 16 is one VPMULHUW per vector in GCC 14 and Clang 23, on ymm registers with -march=x86-64-v3 and on zmm registers in GCC with -march=x86-64-v4. Division by constants keeps the sequences of the SSE2 baseline: b[i] / 10 is VPMULHUW by CCCDh and a VPSRLW by 3, and b[i] / 7 adds the correction with VPSUBW, VPSRLW and VPADDW. At -v4 GCC broadcasts the factor straight from a general-purpose register with VPBROADCASTW.

The full products, (uint32_t)b[i] * c[i], no longer need the pair. Clang zero-extends both operands with VPMOVZXWD and multiplies doublewords with VPMULLD at both levels, where for SSE2 it combined PMULLW and PMULHUW. GCC still pairs VPMULLW and VPMULHUW at -v3, interleaving within the 128-bit lanes and repairing the order with two VPERM2I128, and turns to VPMOVZXWD and VPMULLD on zmm registers at -v4.

AMD’s latency spreadsheet for Zen 5 gives the VEX and EVEX rows alike three cycles, two per cycle on pipes FP0 and FP1.

Forms

Every encoding of VPMULHUW that Intel XED knows, one row per XED instruction form (iform). How to read the notation.
InstructionEncodingRequires
VPMULHUW xmm, xmm, m128 VPMULHUW_XMMdq_XMMdq_MEMdqVEX.128.66.0F.WIG E4 /rAVX
VPMULHUW xmm, xmm, xmm VPMULHUW_XMMdq_XMMdq_XMMdqVEX.128.66.0F.WIG E4 /rAVX
VPMULHUW ymm, ymm, m256 VPMULHUW_YMMqq_YMMqq_MEMqqVEX.256.66.0F.WIG E4 /rAVX2
VPMULHUW ymm, ymm, ymm VPMULHUW_YMMqq_YMMqq_YMMqqVEX.256.66.0F.WIG E4 /rAVX2
VPMULHUW xmm {k}{z}, xmm, xmm VPMULHUW_XMMu16_MASKmskw_XMMu16_XMMu16_AVX512EVEX.128.66.0F.WIG E4 /rAVX10.1 or AVX512BW + AVX512VL
VPMULHUW xmm {k}{z}, xmm, m128 VPMULHUW_XMMu16_MASKmskw_XMMu16_MEMu16_AVX512EVEX.128.66.0F.WIG E4 /rAVX10.1 or AVX512BW + AVX512VL
VPMULHUW ymm {k}{z}, ymm, ymm VPMULHUW_YMMu16_MASKmskw_YMMu16_YMMu16_AVX512EVEX.256.66.0F.WIG E4 /rAVX10.1 or AVX512BW + AVX512VL
VPMULHUW ymm {k}{z}, ymm, m256 VPMULHUW_YMMu16_MASKmskw_YMMu16_MEMu16_AVX512EVEX.256.66.0F.WIG E4 /rAVX10.1 or AVX512BW + AVX512VL
VPMULHUW zmm {k}{z}, zmm, zmm VPMULHUW_ZMMu16_MASKmskw_ZMMu16_ZMMu16_AVX512EVEX.512.66.0F.WIG E4 /rAVX10.1 or AVX512BW
VPMULHUW zmm {k}{z}, zmm, m512 VPMULHUW_ZMMu16_MASKmskw_ZMMu16_MEMu16_AVX512EVEX.512.66.0F.WIG E4 /rAVX10.1 or AVX512BW

Latency and throughput

What the scheduling models of LLVM 23.1.2 give for each form, not measurements: latency / reciprocal throughput in cycles, the ports under them. A row is a model, or models that agree; a dash, a form that its processors lack. Memory forms: latency from the address, ports without the load and store units. Which processors each model covers.
Modelxmm, xmm, m128
xmm {k}{z}, xmm, m128no mask		xmm, xmm, xmm
xmm {k}{z}, xmm, xmmno mask
ymm {k}{z}, ymm, ymmno mask		ymm, ymm, m256
ymm {k}{z}, ymm, m256no mask		ymm, ymm, ymmxmm {k}{z}, xmm, xmmwith a mask, zeroing
ymm {k}{z}, ymm, ymmwith a mask, zeroing		xmm {k}{z}, xmm, m128with a mask, zeroingymm {k}{z}, ymm, m256with a mask, zeroingzmm {k}{z}, zmm, zmmno maskzmm {k}{z}, zmm, zmmwith a mask, zeroingzmm {k}{z}, zmm, m512no maskzmm {k}{z}, zmm, m512with a mask, zeroing
Sandy Bridge11 / 1p05 / 1p0
Haswell11 / 1p05 / 1p012 / 1p05 / 1p0
Broadwell10 / 1p05 / 1p011 / 1p05 / 1p0
Skylake11 / 0.5p015 / 0.5p0112 / 0.5p015 / 0.5p01
Skylake-SP, Ice Lake11 / 0.5p015 / 0.5p0112 / 0.5p015 / 0.5p015 / 0.5p0111 / 0.5p0112 / 0.5p015 / 0.5p055 / 0.5p0512 / 0.5p0512 / 0.5p05
Alder Lake13 / 0.5p015 / 0.5p0113 / 0.5p015 / 0.5p01
Sapphire Rapids13 / 0.5p015 / 0.5p0113 / 0.5p015 / 0.5p017 / 0.5p0114 / 0.5p0114 / 0.5p015 / 1p07 / 1p013 / 1p014 / 1p0
Lunar Lake10 / 2V0 V14 / 2V0 V111 / 2.33V0 V14 / 2V0 V1
Piledriver9 / 1.5FPMMA FPU04 / 1FPMMA FPU0
Zen11 / 1FPU04 / 1FPU011 / 2FPU04 / 2FPU0
Zen 211 / 1FPU04 / 1FPU011 / 1FPU04 / 1FPU0
Zen 310 / 0.5FP0 FP33 / 0.5FP0 FP310 / 0.5FP0 FP33 / 0.5FP0 FP3
Zen 410 / 0.5FP0 FP33 / 0.5FP0 FP310 / 0.5FP0 FP33 / 0.5FP0 FP33 / 0.5FP0 FP310 / 0.5FP0 FP310 / 0.5FP0 FP33 / 1FP0 FP33 / 1FP0 FP310 / 1FP0 FP310 / 1FP0 FP3
Jaguar7 / 1FPU0 VIMUL2 / 1FPU0 VIMUL

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.

Element-wise: each word of the result comes from the word in the same place in each source.
VPMULHUW: element by elementResult word i comes from word i of xmm2 and xmm3. Result words 7 to 0: r7, unchanged (bit 6 of k1 is 0), r5, r4, r3, r2, r1, r0.01234567xmm2a11111110k1r0r1r2r3r4r5r7xmm101234567xmm3or m128b

Every word works alike: 8, 16 and 32 words in the xmm, ymm and zmm forms. The EVEX forms write under a mask: where its bit is 0, the element keeps its old value, or zeroes it with {z}.

Result word i comes from word i of xmm2 and xmm3. Result words 7 to 0: r7, unchanged (bit 6 of k1 is 0), r5, r4, r3, r2, r1, r0.

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 // 16):
    if k1[i]:  # 1 without a write mask
        a = zmm2.w[i]
        b = src3.w[i]
        zmm1.w[i] = a * b >> 16
    elif zeroing:
        zmm1.w[i] = 0
    # else: zmm1.w[i] keeps its value
# zmm1 is zeroed from bit VL up

Run on AMD Zen 5 (EPYC 9645): 10 cases.

Operands

What each form reads and writes, as Intel XED records it: the operands of its syntax, in order, and the implicit ones that the syntax does not name. How to read.
OperandKindAction
VPMULHUW xmm, xmm, xmm/m128 VPMULHUW ymm, ymm, ymm/m256
xmm, ymmexplicitwrite
xmm, ymmexplicitread
xmm/m128, ymm/m256explicitread
VPMULHUW xmm {k}{z}, xmm, xmm/m128 VPMULHUW ymm {k}{z}, ymm, ymm/m256 VPMULHUW zmm {k}{z}, zmm, zmm/m512
xmm, ymm, zmmexplicitwrite (masked)
{k}explicitread (write mask)
xmm, ymm, zmmexplicitread
xmm/m128, ymm/m256, zmm/m512explicitread

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.

VEX forms
avx-type-4
EVEX forms
AVX512-E4

Attributes

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

EVEX forms
write masking
EVEX memory forms
memory fault suppression

Intrinsics

The C intrinsics of Clang 23.1.2 that compile to VPMULHUW, 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.

IntrinsicCompiles to
_mm_mulhi_epu16 __m128i _mm_mulhi_epu16(__m128i __a, __m128i __b)

VPMULHUW xmm, xmm, xmm VEX.128 -mavxwith AVX enabled

_mm_mask_mulhi_epu16 __m128i _mm_mask_mulhi_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)

VPMULHUW xmm {k}, xmm, xmm EVEX.128 -mavx512vl -mavx512bw

_mm_maskz_mulhi_epu16 __m128i _mm_maskz_mulhi_epu16(__mmask8 __U, __m128i __A, __m128i __B)

VPMULHUW xmm {k}{z}, xmm, xmm EVEX.128 -mavx512vl -mavx512bw

_mm256_mulhi_epu16 __m256i _mm256_mulhi_epu16(__m256i __a, __m256i __b)

VPMULHUW ymm, ymm, ymm VEX.256 -mavx2

_mm256_mask_mulhi_epu16 __m256i _mm256_mask_mulhi_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B)

VPMULHUW ymm {k}, ymm, ymm EVEX.256 -mavx512vl -mavx512bw

_mm256_maskz_mulhi_epu16 __m256i _mm256_maskz_mulhi_epu16(__mmask16 __U, __m256i __A, __m256i __B)

VPMULHUW ymm {k}{z}, ymm, ymm EVEX.256 -mavx512vl -mavx512bw

_mm512_mulhi_epu16 __m512i _mm512_mulhi_epu16(__m512i __A, __m512i __B)

VPMULHUW zmm, zmm, zmm EVEX.512 -mavx512bw

_mm512_mask_mulhi_epu16 __m512i _mm512_mask_mulhi_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)

VPMULHUW zmm {k}, zmm, zmm EVEX.512 -mavx512bw

_mm512_maskz_mulhi_epu16 __m512i _mm512_maskz_mulhi_epu16(__mmask32 __U, __m512i __A, __m512i __B)

VPMULHUW zmm {k}{z}, zmm, zmm EVEX.512 -mavx512bw

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.

CPUID requirements and processors of the XED ISA sets above. Alternatives are separated by “or”; processors follow XED’s chip model, whose AMD processors end with Zen 2, and the CPU data of this site for Zen 3 to Zen 6.
ISA setCPUIDProcessors
AVX

AVX
CPUID.01H:ECX[28] AVX

Sandy Bridge, Ivy Bridge, Haswell, Broadwell, Skylake, Comet Lake, Cannon Lake, Ice Lake (client), Tiger Lake, Alder Lake, Arrow Lake, Panther Lake, Nova Lake, Skylake-SP, Cascade Lake, Cooper Lake, Ice Lake (server), Sapphire Rapids, Emerald Rapids, Granite Rapids, Diamond Rapids, Sierra Forest, Clearwater Forest, Knights Landing, Knights Mill, Bulldozer, Piledriver, Zen, Zen+, Zen 2, Zen 3, Zen 4, Zen 5, Zen 6
AVX2

AVX2
CPUID.(EAX=07H,ECX=0):EBX[5] AVX2

Haswell, Broadwell, Skylake, Comet Lake, Cannon Lake, Ice Lake (client), Tiger Lake, Alder Lake, Arrow Lake, Panther Lake, Nova Lake, Skylake-SP, Cascade Lake, Cooper Lake, Ice Lake (server), Sapphire Rapids, Emerald Rapids, Granite Rapids, Diamond Rapids, Sierra Forest, Clearwater Forest, Knights Landing, Knights Mill, Zen, Zen+, Zen 2, Zen 3, Zen 4, Zen 5, Zen 6
AVX512BW_128

AVX10.1
CPUID.(EAX=07H,ECX=1):EDX[19] AVX10
CPUID.(EAX=24H,ECX=0):EBX[7:0] ≥ 1 AVX10 version

or

AVX512BW + AVX512VL
CPUID.(EAX=07H,ECX=0):EBX[30] AVX512BW
CPUID.(EAX=07H,ECX=0):EBX[31] AVX512VL

Cannon Lake, Ice Lake (client), Tiger Lake, Nova Lake, Skylake-SP, Cascade Lake, Cooper Lake, Ice Lake (server), Sapphire Rapids, Emerald Rapids, Granite Rapids, Diamond Rapids, Zen 4, Zen 5, Zen 6
AVX512BW_256
Same as AVX512BW_128Same as AVX512BW_128
AVX512BW_512

AVX10.1
CPUID.(EAX=07H,ECX=1):EDX[19] AVX10
CPUID.(EAX=24H,ECX=0):EBX[7:0] ≥ 1 AVX10 version

or

AVX512BW
CPUID.(EAX=07H,ECX=0):EBX[30] AVX512BW

Same as AVX512BW_128

More AVX instructions

All 253 AVX instructions

More AVX2 instructions

All 138 AVX2 instructions

More AVX10.1 instructions

All 686 AVX10.1 instructions

Sources

  1. Intel XED v2026.08.23 (commit 0bcb6237345c): forms, encodings, operands, flags, exception classes, attributes, ISA sets, CPUID bits and chip model.
  2. The CPU data of x86-64.net: Zen 3 to Zen 6, from GCC’s -march feature sets and AMD’s Processor Programming References, which the page of each processor cites.
  3. Clang's x86 intrinsic headers, LLVM 23.1.2 (commit 85ac56026243): the intrinsics, their signatures and doc comments, and the Clang of that release, which compiled each of them.
  4. LLVM 23.1.2's x86 scheduling models (commit 85ac56026243): the latency, reciprocal throughput and ports of each form, as llvm-mca reports them.
  5. Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 2 (325383-093, September 2026): Volume 2B, page 4-379 (PMULHUW)
  6. Intel 64 and IA-32 Architectures Software Developer's Manual, Volume 1 (253665-093, September 2026)
  7. Intel Advanced Vector Extensions 10.2 Architecture Specification, revision 7.0 (361050-007, April 2026)
  8. Software Optimization Guide for the AMD Zen5 Microarchitecture (58455, rev. 1.00, August 2024), Wayback Machine copy of AMD's ZIP

The tables are derived from Intel XED, Copyright Intel Corporation, licensed under the Apache License 2.0; x86-64.net converted and reformatted them. The intrinsics and their signatures come from Clang and the timings from LLVM's scheduling models, both part of the LLVM Project, licensed under the Apache License 2.0 with LLVM Exceptions. The text is our own.