Open yzh119 opened 2 years ago
Currently, we only use m16n16k16 wmma instructions in RGCN/Sparse Conv which might not be optimal, we should refactor the code to enable different WMMA shapes (m8n32k16/m32n8k16/m16n16k16).
Currently, we only use m16n16k16 wmma instructions in RGCN/Sparse Conv which might not be optimal, we should refactor the code to enable different WMMA shapes (m8n32k16/m32n8k16/m16n16k16).