Meaning of `sizes` for tensor.insert_slice

MLIR
1

The doc says sizes: tensor-rank number of sizes which specify the sizes of the result tensor type. which sounds like the sizes of the dest tensor. However, my understanding is that this is wrong and it should be the source tensor.

Even for source tensor, it’s not clear to me, if it means the slice shape or the source tensor shape. There is a difference between these two if the strides are not equal to 1. Like if the source tensor is 1x4 and stride is 1x3, should the sizes be 1x2 or it should still be 1x4?

2

I got the answer from another channel. Summarizing here:
The sizes is the shape of the source tensor and the entire source tensor space is inserted into the dest at the given offsets. We only consider stride == 1 at the moment. The source tensor is of sizes and it would extract/insert from/into a size * stride subregion and it is the user reponsibility that offset + size * stride of the small tensor fits within the large tensor.

IR like this would be invalid because %arg0: tensor<1x1x3xf32> has to match the [1,1,2] in %init[0, 0, 0] [1, 1, 2] as the entire source tensor space is inserted

builtin.func @f(%arg0: tensor<1x1x3xf32>) ->  tensor<1x1x2xf32> {
     %init = linalg.init_tensor [1, 1, 2] : tensor<1x1x2xf32>
     %ret = tensor.insert_slice %arg0 into %init[0, 0, 0] [1, 1, 2] [1, 1, 1] : tensor<1x1x3xf32> into tensor<1x1x2xf32>
     return %ret : tensor<1x1x2xf32>
}

The size is indeed redundant information in this particular case but there is also a notion of rank-reducing version of these ops where you can drop 1’s from the list. If you have rank-reduced tensors the non-1 dimensions match the destination tensor and missing dimensions should all have sizes 1. Being explicit in all the cases avoids more surprises.

3 Likes