TLDR Tech / Practical
Optimizing FP16 GEMM Kernels on AMD RDNA 3/4 GPUs
This technical guide outlines an eight-stage process for developing and optimizing an FP16 GEMM kernel on AMD RDNA 3 and RDNA 4 graphics processors. The author details methods such as expanding block tiling, implementing double buffering, utilizing LDS padding and permutation, employing buffer resource descriptors, and interleaving prefetch operations to enhance arithmetic intensity and improve cache utilization.
Read the full story