Allocate temp buffer for tensorize

As the example below:

produce B {
  for (i, 0, 12) {
    B[i] = 0.000000f
    for (k.outer, 0, 16) {
      for (k.inner, 0, 4) {
        B[i] = (B[i] + A[((((i*16) + k.outer)*4) + k.inner)])
      }
    }
  }
}

The number of elements in vector instruction is fixed, like 4. If we want to tensorize the k.inner, we need to allocate a extra buffer to store the intermediate result, or else it would be overwritten. Here may be what we expect:

produce B {
  for (i, 0, 12) {
    B[i] = 0.000000f
    for (k.outer, 0, 16) {
      tmp = intrinsic("summation", A)
      B[i] = B[i] + tmp
    }
  }
}

So is there any way to allocate buffer for intermediate result?

[tqchen]

I think we need an factor axis argument which indicate which axis we want to insert the factor dimension

[tqchen]

move to https://github.com/dmlc/tvm/issues/867