Why use it

Fixed-point filters need the product and accumulator to meet before rounding and saturation. RDM performs that combined operation for signed 16-bit and 32-bit samples. Keeping the full intermediate matters when a large product cancels part of the accumulator.

Operations

Import native.arm.rdm, or use the native.arm or native hub. sqrdmlah<Arch>(acc,a,b) adds the doubled product; sqrdmlsh<Arch>(acc,a,b) subtracts it. The corresponding _lane<Arch,Lane> forms broadcast b[Lane].

Accumulator, a and result	Ordinary b	Indexed b
std::int16_t	std::int16_t	simd<std::int16_t,4,Arch> or simd<std::int16_t,8,Arch>
simd<std::int16_t,4,Arch> or simd<std::int16_t,8,Arch>	Same vector type	Either 16-bit vector width
std::int32_t	std::int32_t	simd<std::int32_t,2,Arch> or simd<std::int32_t,4,Arch>
simd<std::int32_t,2,Arch> or simd<std::int32_t,4,Arch>	Same vector type	Either 32-bit vector width

All vector operands share Arch. Lane is a compile-time immediate within the right operand's range. For element width w, each result follows:

sqrdmlah(acc,a,b) = signed_saturate_w(floor((acc*2^w + 2*a*b + 2^(w-1)) / 2^w))

sqrdmlsh(acc,a,b) = signed_saturate_w(floor((acc*2^w - 2*a*b + 2^(w-1)) / 2^w))

Caveats

Ties round toward positive infinity. Saturation occurs after combining the full product and accumulator. In particular, minimum-times-minimum can cancel a negative accumulator without intermediate saturation. A saturating multiply followed by a saturating add has different results.

Runtime saturation sets sticky FPSR.QC; a nonsaturating operation does not clear it. An unused result still executes the instruction. The wrappers leave FPCR unchanged and neither save nor restore FPSR. Constant evaluation computes only the value, with no access to the calling thread's QC state.

Runtime calls require arm_feature::rdm and an "rdm" caller target. feature_closure(arm_feature::rdm) includes NEON. Admit that requirement and NATIVE_TARGET_MINIMUM before entering the target function. RDM does not require DotProd, FP16 or the rest of Armv8.1-A. The wrappers use exact inline instructions to avoid Clang 23's broader v8.1a builtin requirement.

Scalar calls may omit Arch, using the owning module's captured NATIVE_BASELINE. An importer's target attribute cannot change that default. Clang 23 defines __ARM_FEATURE_QRDMX for Armv8.1-A, but not for armv8-a+rdm alone; use an explicit ISA for the latter with the current macro-based baseline snapshot. Vector calls deduce Arch from their operands.

All forms support constant evaluation. A tag without RDM permits only consteval calls, provided its SIMD storage shapes exist. There is no runtime software fallback or dispatch.

See the Arm Advanced SIMD intrinsic reference and LLVM's big-endian NEON representation notes.