is this effectively just viewing complex as float? If yes, it would work for cumsum, however it won't work for cumprod as is, since multiplication for two complex numbers is not defined as a dot product. (a+bi)*(c+di) = ac-bd + (ad+bc)i

Here we are allocating shared memory buffer and not viewing into the actual data.

Since we can't allocate the shared memory for the c10::complex as it requires explicit initialization. So instead we allocate the equivalent memory in terms of the base scalar type and view that as a buffer for c10::complex.

So we are not actually viewing it as independent base type memory in the tensor_kernel_scan_innermost_dim_impl but as actual c10::complex.
So cumprod won't be an issue.

Plus all the tests pass.

Could you explain why we can't allocate the shared memory for the c10::complex as it requires explicit initialization.?
c10::complex has a default constructor, and shouldn't it be initialized to its default value automatically ?

Edit: never mind, I was thinking something differently

I am not quite sure why we can't allocate the shared memory for the c10::complex either. @kshitij12345 can you explain how is it related to explicit initialization requirement of c10::complex?

A minimum repro of this problem could be

struct A {
    int a = 0;
    A() = default;
};

__global__ void f(A *p) {
    __shared__ A s[5];  // error: error: initializer not allowed for __shared__ variable 
    s[0] = *p;
}

offline sync with @zasdfgbnm :
__shared__ variables cannot have an initialization as part of their declaration as explained here http://www.ncsa.illinois.edu/People/kindr/projects/hpca/files/singapore_p2.pdf

updating the complex constructor to not assign real and imag in constructor would resolve the problem.

@anjali411 @zasdfgbnm Thanks. I was actually misguided in understanding the error.

Thanks for the reference as well.