2 ms·

> help the GPU scheduler assign tasks to specialized blocks Source? I can't find anything for '"cublas" scheduler'…

• rubatuga 5y ago
  The GEMM instructions are sent to Tensor Cores by the scheduler. The cutlass library is the open source version of cublas: https://github.com/NVIDIA/cutlass https://github.com/NVIDIA/cutlass
  • floatboth 5y ago
    Seems like you can leverage those via VK_NV_cooperative_matrix…