Manual kernels (A5)¶

This directory contains manual, performance-oriented kernel examples targeting Ascend A5.

Examples¶

• Flash-Attention kernel: flash_atten
• MXFP4 matrix multiplication performance kernel: matmul_mxfp4_performance
• MXFP8 matrix multiplication performance kernel: matmul_mxfp8_performance

Common setup¶

These examples typically require a CANN environment to be sourced before building/running. For example:

source ${ASCEND_INSTALL_PATH}/bin/setenv.bash

Then follow the run.sh usage documented in each example directory.