We develop a fused matrix multiplication kernel that unifies sampled\ndense-dense matrix multiplication and sparse-dense matrix multiplication under\na single operation called FusedMM. By using user-defined functions, FusedMM can\ncapture almost all computational patterns needed by popular graph embedding and\nGNN approaches. FusedMM is an order of magnitude faster than its equivalent\nkernels in Deep Graph Library. The superior performance of FusedMM comes from\nthe low-level vectorized kernels, a suitable load balancing scheme and an\nefficient utilization of the memory bandwidth. FusedMM can tune its performance\nusing a code generator and perform equally well on Intel, AMD and ARM\nprocessors. FusedMM speeds up an end-to-end graph embedding algorithm by up to\n28x on different processors.\n