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

gemm_kernel_apple_8x8

def gemm_kernel_apple_8x8[c_type: DType, a_type: DType, b_type: DType, c_layout: TensorLayout, a_layout: TensorLayout, b_layout: TensorLayout, c_storage: TensorStorage, a_storage: TensorStorage, b_storage: TensorStorage, transpose_b: Bool = False, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None, s_type: DType = get_accum_type[c_type](), BLOCK_M: Int = Int(64), BLOCK_N: Int = Int(64), BLOCK_K: Int = Int(16), NUM_SIMDGROUPS: Int = Int(4)](c: TileTensor[c_type, c_layout, MutAnyOrigin, Storage=c_storage], a: TileTensor[a_type, a_layout, ImmutAnyOrigin, Storage=a_storage], b: TileTensor[b_type, b_layout, ImmutAnyOrigin, Storage=b_storage], m: Int, n: Int, k: Int)

Launchable wrapper for the 8x8 simdgroup-matrix GEMM (bench/test).

Parameters:

c_type (DType): Element type of the output C tile.
a_type (DType): Element type of the input A tile.
b_type (DType): Element type of the input B tile.
c_layout (TensorLayout): Memory layout of the output C tile.
a_layout (TensorLayout): Memory layout of the input A tile.
b_layout (TensorLayout): Memory layout of the input B tile.
c_storage (TensorStorage): Storage kind of the output C tile.
a_storage (TensorStorage): Storage kind of the input A tile.
b_storage (TensorStorage): Storage kind of the input B tile.
transpose_b (Bool): Whether B is stored as (N, K) rather than (K, N) (defaults to False).
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue applied to each output element before store (defaults to None).
s_type (DType): Accumulator element type (defaults to the accumulator type of c_type).
BLOCK_M (Int): M-dimension block tile size (defaults to 64).
BLOCK_N (Int): N-dimension block tile size (defaults to 64).
BLOCK_K (Int): K-dimension block tile size (defaults to 16).
NUM_SIMDGROUPS (Int): Simdgroups per threadgroup, each owning a 32x32 subtile (defaults to 4).

Args:

c (TileTensor[c_type, c_layout, MutAnyOrigin, Storage=c_storage]): Output TileTensor of shape (m, n).
a (TileTensor[a_type, a_layout, ImmutAnyOrigin, Storage=a_storage]): Input TileTensor of shape (m, k).
b (TileTensor[b_type, b_layout, ImmutAnyOrigin, Storage=b_storage]): Input TileTensor of shape (k, n), or (n, k) when transpose_b is set.
m (Int): Number of rows in C and A.
n (Int): Number of columns in C and B.
k (Int): Contraction dimension shared by A and B.