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 module

vnni_intrinsics

Provides wrappers around Intel VNNI (Vector Neural Network Instructions) for dot-product and multiply-accumulate operations on integer data.

Functions

dot_i16_to_i32_AVX2: The dot product of the two words in each int32 element of a and b plus a int32 from src.
dot_i16_to_i32_x86: The dot product of the two words in each int32 element of a and b plus a int32 from src using VNNI or AVX2.
dot_i8_to_i32_AVX2: The dot product of the four bytes in each int32 element of a and b plus a int32 from src.
dot_i8_to_i32_saturated_AVX2: The dot product of the four bytes in each int32 element of a and b plus a int32 from src.
dot_i8_to_i32_saturated_x86: The dot product of the four bytes in each int32 element of a and b plus a int32 from src using VNNI or AVX2.
dot_i8_to_i32_x86: The dot product of the four bytes in each int32 element of a and b plus a int32 from src using VNNI or AVX2.
pmaddubs: Multiplies adjacent unsigned-signed byte pairs and returns the int16 results packed as int32.
pmaddw: Multiplies adjacent signed 16-bit integer pairs and adds the products, returning int32 results.
vpdpbusd: Computes a dot product of four unsigned-signed byte pairs per int32 element using the VPDPBUSD Intel AVX-512 VNNI instruction.
vpdpbusds: Computes a saturating dot product of four unsigned-signed byte pairs per int32 element using the VPDPBUSDS Intel AVX-512 VNNI instruction.
vpdpwssd: Computes a multiply-accumulate of signed 16-bit integers using the VPDPWSSD Intel AVX-512 VNNI instruction.
vpdpwssds: Computes a saturating multiply-accumulate of signed 16-bit integers using the VPDPWSSDS Intel AVX-512 VNNI instruction.

Functions​

Functions