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

build_default_matmul_schedule

def build_default_matmul_schedule[num_k_tiles: Int, num_m_mmas: Int, num_n_mmas: Int, num_k_mmas: Int, MMA_M: Int, MMA_N: Int, a_loads_per_thread: Int, b_loads_per_thread: Int]() -> ScheduleCompiler

Build the complete software pipeline schedule for the default matmul.

Uses ScheduleCompiler with SingleBufferSchedule trait implementation. All phases derived from the logical body; no magic numbers in derivation.

Returns a ScheduleCompiler with all phases as Lists. Use as a comptime value; the kernel reads entries via comptime for.

Parameters:

num_k_tiles (Int): Number of K-tiles along the contraction dimension, each producing one fragment load and one MMA compute op.
num_m_mmas (Int): Number of MMA tiles along the M dimension of the output tile.
num_n_mmas (Int): Number of MMA tiles along the N dimension of the output tile.
num_k_mmas (Int): Number of MMA tiles along the K dimension per COMPUTE body op.
MMA_M (Int): M dimension of a single MMA instruction (for example 16 or 32).
MMA_N (Int): N dimension of a single MMA instruction (for example 16 or 32).
a_loads_per_thread (Int): Number of DS_WRITE ops per STORE_SMEM for the A operand tile.
b_loads_per_thread (Int): Number of DS_WRITE ops per STORE_SMEM for the B operand tile.

Returns:

ScheduleCompiler