Module layered 

The layered matmul combines multiple component-based algorithm implementations with selection logic to pick the optimal kernel for a set of parameters.

Modules

double_buffering

double_unit

ordered_double_buffering

simple

simple_barrier

simple_tma

simple_unit

vecmat

Enums

Selection

TileSizeSelection

Constants

NUM_SM_APPROX

NUM_TENSOR_CORES_APPROX

Traits

Algorithm

Specifications for a matmul algorithm

Functions

find_instruction_size

A heuristic to choose the instruction to use, based on input shape

launch

Launch a matrix multiplication kernel.

launch_kernel_concrete

Select which kernel to launch for the given Algorithm.

launch_kernel_virtual

Select which kernel to launch for the given Algorithm.

launch_ref

Launch a matrix multiplication kernel.

launch_with_config

matmul_cmma_tma_ref_no_check