Modification to the Micro Kernel

Content Packing Blocks from Matrix A Possible Solution Packing Blocks from Matrix B Possible Solution GEMM Micro Kernel Possible Solution GEMM Macro Kernel Possible Solution GEMM: Implementation of the Frame Algorithm Possible Solution: ulmBLAS Source File Modification to the Micro Kernel Exercise Possible Solution: ulmBLAS Source File (State after Session 16)

The micro kernel has a huge influence on the overall performance for the GEMM operation. Before we investigate this further in the next session we will do here some simple experiments.

Exercise

In the micro kernel, switch the two most inner loops when computing

\[R \leftarrow A \cdot B\]

Re-run the benchmark!

View document source

© Michael C. Lehn