Haoyang Peng

Thank you for the answer! My motivation is to use BN in inference stage, so can i just pass the trained mean and var in tensor C and tensor D...

Well, so that means that i should put the variance in tensor E?

thank you, i'll take a look.

@hwu36 hello,i follow your instructions and make a new epilogue, but i cannot get through the compilation step. Here is the error Log. ``` error: function "cutlass::gemm::warp::MmaTensorOp::operator() [with Shape_=cutlass::gemm::GemmShape, ElementA_=float,...

Here is my definition of the arguments: ``` typename Conv2dFprop::Arguments arguments{{ problem_size, tensor_a, //

thank you, but after I change the type definition as you instructed, it still tells me the implicitGemm is an incomplete type. Here is all my definition: ``` #include #include...

so i can't use NCHW layout data in Conv computation?

thank you, i'll take a look

alright, now i can have conv with normal linearcombination epilogue works. But when i switch to other epilogues like linearcombinationbiasrelu it still pops out errors. The only difference is that...

``` cutlass/cutlass/include/cutlass/epilogue/threadblock/epilogue.h(500): error: no instance of overloaded function "cutlass::epilogue::thread::LinearCombinationBiasRelu::operator() [with ElementC_=float, ElementAccumulator_=float, ElementCompute_=float, ElementZ_=float, ElementsPerAccess=4, StoreT=true]" matches the argument list argument types are: (const cutlass::Array) object type is: const cutlass::epilogue::thread::LinearCombinationBiasRelu...