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

apple_amx_intrinsics

Provides low-level wrappers around Apple's AMX assembly instruction set for matrix operations.

Functions

dot_at_b: Performs a matrix multiply C = A^T * B using Apple AMX instructions.
extrx: Extracts a row or moves it to x, result in amx0.
extry: Extracts a row or moves it to y, result in amx0.
fma: Issues an AMX fused multiply-add of X and Y register rows accumulated into the Z accumulator.
fma16: Float16 matrix multiply and subtract.
fma32: Float32 matrix multiply and add.
fma64: Float64 matrix multiply and add.
fms16: Float16 matrix multiply and add.
fsm32: Float32 matrix multiply and subtract.
fsm64: Float64 matrix multiply and subtract.
genlut: Issues the Apple AMX genlut instruction using the address encoded in gpr.
ldx: Loads data from the memory address encoded in gpr into an AMX X register row.
ldy: Loads data from the memory address encoded in gpr into an AMX Y register row.
ldz: Loads data from the memory address encoded in gpr into an AMX Z accumulator row.
ldzi: Loads data from the memory address encoded in gpr into an AMX Z accumulator row using interleaved layout.
load_z: Stores row_count rows from AMX Z accumulator registers starting at start_index to memory at src.
mac16: SI16 matrix multiply and add.
matfp: Float16 matrix multiply.
max_int__: UI16 matrix multiply.
read_x: Stores row_count rows from AMX X registers starting at start_index to memory at src.
read_y: Stores row_count rows from AMX Y registers starting at start_index to memory at src.
store_x: Loads row_count rows from memory at src into AMX X registers starting at start_index.
store_y: Loads row_count rows from memory at src into AMX Y registers starting at start_index.
store_z: Loads row_count rows from memory at src into AMX Z accumulator registers starting at start_index.
stx: Stores an AMX X register row to the memory address encoded in gpr.
sty: Stores an AMX Y register row to the memory address encoded in gpr.
stz: Stores an AMX Z accumulator row to the memory address encoded in gpr.
stzi: Stores an AMX Z accumulator row to the memory address encoded in gpr using interleaved layout.
transpose_z_to_x_or_y: Transposes a downsampled column of the AMX Z register into a row of the X or Y register.
vec_int__: Horizontal ui16 multiply z0[i] += x0[i] + y0[i].
vecfp: Horizontal float16 multiply z0[i] += x0[i] + y0[i].

Functions​

Functions