Is the implementation of Cutlass’s basic gemm benchmark open sourced? It seems the code leads to a kernel called GemmKernel, but where can I find the actual implementation of the matrix multiply? My understanding is that this kernel’s launch configurations are computed based on problem size. Is it feasible to change its launch configurations, if so where (i.e. when calling GemmKernel)?
cutlass::gemm::device::Gemm::GemmKernel is a type alias that refers to cutlass::gemm::kernel::DefaultGemm. The definition of the kernel, including partial specializations for some architectures, can be found in the latter link.