Jack Kosaian

>Theoretically, in the next call to a convolution, should it be saved already, as I understood? I don't fully understand what you're thinking of here, unfortunately. > Could it be...

Thanks for the additional details. > its filters and save them in GPU shared memory Any modifications done in shared memory will not persist through multiple kernel calls (e.g., for...

Thanks for the additional details. I think I understand your intention a bit better. I agree that it makes sense to modify the filters once and reuse them across remaining...

The approach that you prototyped is slightly different than what I had in mind. I was suggesting that you write one kernel that does: ```c++ for (int i = start_idx;...

Ok. I personally think it might still be easier to have two kernels, but, if you want to fuse this to a CUTLASS kernel, you'll want to perform the scaling...

If you just want to access the global memory, you can do so via the `ptr_A` and `ptr_B` members of `Params` [here](https://github.com/NVIDIA/cutlass/blob/24f991e87930e1159f1f5a47e329d43bcfbd76b9/include/cutlass/conv/kernel/implicit_gemm_convolution.h#L205-L208). These are available in the `operator()` method, which...

> Moreover, params.first_call = false; gives me an error: error: expression must be a modifiable lvalue. `params` are declared as `const` [here](https://github.com/NVIDIA/cutlass/blob/24f991e87930e1159f1f5a47e329d43bcfbd76b9/include/cutlass/conv/kernel/implicit_gemm_convolution.h#L281). They cannot be modified. > how you would...

>Besides that, I find there is a bigger issue. If params are const and they cannot be modified, how then can I performed a modification of the filters (like i...

Sorry for the confusion. The sole point of introducing static assert was to determine which iterator type was being used. Now that you've determined this, you can remove the static...

Thanks for pointing this out. You're right: the logic seems to have been corrupted during the CUTLASS 3.4 release. We'll fix the logic in the next release. In the meantime,...