Representation of L1 Interleaved layout

[fbajraktariTT]

Method getLayoutSizeBytes returns incorrect memory usage for L1 Interleaved layouts.

Let's say we have a LayoutAttr showed in the image below that was created as a default DRAM layout for tensor<1x1024xbf16>:

Now, in LegalGridAnalysis we construct L1 Interleaved layout for this tensor:

It seems that we get an incorrect representation for this layout. It is expected to get memref<1x1x!tt.tile<32x32, bf16> instead of memref<1x4x!tt.tile<32x32, bf16>.

The problem appears to be in withGrid method that relies on calculateLogicalShardShape. Furthermore, we use the same withGrid method for both sharding layouts and L1 Interleaved layout. It seems that in the current implementation we are handling L1 Interleaved layouts as Block sharded layouts.

Therefore, method getLayoutSizeBytes used for calculating L1 usage doesn't work properly for L1 Interleaved layouts. In the mentioned case, it returns 8192 bytes while the actual L1 usage is 2048 bytes.