Which AT_DISPATCH... file should be used?
unholdmx opened this issue · comments
I have a very hard time understanding, which AT_DISPATCH from here needs to be used in order to write tensors of different dtypes to a kernel.
While I understand, how #define AT_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...)
or #define AT_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...)
can be used when all tensors have the same dtype,
(here in a torch environment)
AT_DISPATCH_FLOATING_TYPES(at::ScalarType::Float,"Some floating error message", [&] {
some_cuda_kernel<scalar_t><<<blocks, threads>>>(
some_float_tensor.packed_accessor<scalar_t,2,torch::RestrictPtrTraits,size_t>(),
....
);
}
);
all attempts to dispatch tensors when both tensors have different dtypes fail.