For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Mojo struct
MatmulKernelConfig
struct MatmulKernelConfig
Block/warp/MMA shape configuration for 4-wave kernels.
Shared by AMD4WaveMatmul's matmul and conv2d entry points: both
use the same 4-warp 2×2 quadrant layout and the same MFMA shape
selection, so the per-call surface only needs the workgroup /
warp / MMA tile shapes.
Fields
- block_shape (
IndexList[Int(3)]): Workgroup tile shape(BM, BN, BK). - warp_shape (
IndexList[Int(3)]): Per-warp tile shape(WM, WN, WK). - mma_shape (
IndexList[Int(3)]): Single-MMA instruction shape(MMA_M, MMA_N, MMA_K).
Implemented traits
AnyType,
Copyable,
Deinitable,
ImplicitlyCopyable,
Movable,
Writable
Methods
__init__
def __init__(out self, *, block_shape: IndexList[Int(3)], warp_shape: IndexList[Int(3)], mma_shape: IndexList[Int(3)])
Constructs a MatmulKernelConfig from the three tile shapes.
Args:
- block_shape (
IndexList[Int(3)]): Workgroup tile shape(BM, BN, BK). - warp_shape (
IndexList[Int(3)]): Per-warp tile shape(WM, WN, WK). - mma_shape (
IndexList[Int(3)]): Single-MMA instruction shape(MMA_M, MMA_N, MMA_K).
num_threads
def num_threads(self) -> Int
Returns the total threads per workgroup (warps x WARP_SIZE).
Returns:
Int: The number of threads in one workgroup.
write_to
def write_to(self, mut writer: T)
Writes a human-readable shape tag to writer.
Args:
- writer (
T): Sink for the rendered config tag.
write_repr_to
def write_repr_to(self, mut writer: T)
Writes a debug representation of this config to writer.
Args:
- writer (
T): Sink for the rendered config tag.