Why use it

Byte dot products appear in quantized linear algebra and integer filters. DotProd multiplies four pairs of bytes per accumulator lane and adds their sum directly to a 32-bit accumulator. The indexed forms let several dot products share the same four coefficients.

Operations

Import native.arm.dotprod, or use the native.arm or native hub. Operands and results use native::simd<T,N,Arch> with a common Arch.

Operation	Accumulator/result	Byte operands	Right-hand selection
sdot<Arch>(acc,a,b)	simd<std::int32_t,2,Arch> or simd<std::int32_t,4,Arch>	Matching-width signed 8-bit vectors: 8 or 16 lanes	Corresponding groups of four bytes
udot<Arch>(acc,a,b)	simd<std::uint32_t,2,Arch> or simd<std::uint32_t,4,Arch>	Matching-width unsigned 8-bit vectors: 8 or 16 lanes	Corresponding groups of four bytes
sdot_lane<Arch,Lane>(acc,a,b)	Either signed accumulator	Matching-width a; 8 or 16 signed bytes in b	One four-byte group of b for every result lane
udot_lane<Arch,Lane>(acc,a,b)	Either unsigned accumulator	Matching-width a; 8 or 16 unsigned bytes in b	One four-byte group of b for every result lane

For the ordinary forms, output lane j accumulates sum(a[4*j+k] * b[4*j+k], k=0..3). The indexed forms replace the second index with 4*Lane+k. Lane is a compile-time immediate: 0–1 for an eight-byte source, 0–3 for a sixteen-byte source.

#include <cstdint>
#include <native/targets.h>
import native.arm.dotprod;
 
constexpr auto required = native::feature_closure(native::arm_feature::dotprod);
 
__attribute__((target("dotprod")))
native::simd<std::int32_t,4,required> accumulate(
    native::simd<std::int32_t,4,required> acc,
    native::simd<std::int8_t,16,required> a,
    native::simd<std::int8_t,16,required> b) {
  return native::sdot<required>(acc, a, b);
}
 
bool available() {
  return native::classify_isa(native::observe_arm_capabilities(),
    required, NATIVE_TARGET_MINIMUM).admitted();
}

Caveats

Accumulation wraps modulo 2³². Signed overflow has the instruction's wrapping result; it does not saturate or invoke C++ signed-overflow undefined behavior.

Runtime calls require arm_feature::dotprod and a caller compiled for "dotprod". Check admission before entering the target function. DotProd adds NEON as a prerequisite and can be requested independently of RDM, FP16 or I8MM. An import does not enable the target or dispatch to a supported implementation.

All forms support constant evaluation with the same modular arithmetic. A tag without DotProd permits only consteval calls, provided its SIMD storage shapes exist. It supplies no runtime software fallback.

See the Arm Advanced SIMD intrinsic reference.