For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).

Mojo function

vpdpbusd

def vpdpbusd[width: SIMDLength, a_type: DType, b_type: DType, c_type: DType](src: SIMD[c_type, width], a: SIMD[a_type, width], b: SIMD[b_type, width]) -> SIMD[c_type, width]

Computes a dot product of four unsigned-signed byte pairs per int32 element using the VPDPBUSD Intel AVX-512 VNNI instruction.

For each int32 lane, treats the corresponding four bytes of a as uint8 and four bytes of b as int8, multiplies them element-wise, and adds the four products to the accumulator src.

Parameters:

width (SIMDLength): Number of int32 output elements (4, 8, or 16).
a_type (DType): DType of the a operand.
b_type (DType): DType of the b operand.
c_type (DType): DType of the accumulator; must be int32.

Args:

src (SIMD[c_type, width]): Int32 accumulator vector.
a (SIMD[a_type, width]): Uint8 input packed as int32-wide vector.
b (SIMD[b_type, width]): Int8 input packed as int32-wide vector.

Returns:

SIMD[c_type, width]: Updated int32 accumulator after byte dot-product accumulation.