Expressing strided output tensors

Questions
#1

Hi all,
I’m trying to write a tensor computation with a strided output but I’m not sure how to express the output dimensions in the lambda computation for tvm.compute.
My output is of the form O[a][b+c].

Thanks,

[EDIT]
Following is a more concrete example, the backward computation of a convolutional layer.

def bp(A,W,B):
   for n in range(0,N):
     for k in range(0,K) :
       for h in range(0, H):
         for w in range(0, W):
           for c in range(0, C):
             for r in range(0,R):
               for s in range(0, S):
                 B[n][h+r][w+s][c] += A[n][h][w][k]*W[k][c][r][s]

As @eqy mentioned, it is possible that different values of h,w,r,s map to the same h+r and w+s indices.

#2

Can you describe your computation in more detail? Naively, just writing O[a][b+c] seems ambiguous from the perspective of the tensor expression language e.g., a=0, b=1, c=0 and a=0, b=0, c=1 will index to the same output.

#3

Thanks, I updated the original question.

#4

UPDATE:
I tried passing the Python function bp as the fcompute argument to tvm.compute where A,B,W are tvm.placeholder objects and I get the following error.

TypeError: 'TensorSlice' object does not support item assignment

Any idea how to solve this issue?
Thanks,

Dose TVM support training models？