[Golden] TTRT segfault when trying to copy tensor data

[mtopalovicTT]

I'm working on bug which has this as input IR:

module @jit_convolution attributes {} {
  func.func @forward(%arg0: tensor<1x2048x7x7xf32> {ttir.name = "x"}) -> (tensor<1x2048x1x1xf32> {ttir.name = "AvgPool2d.output_avg_pool2d_0"}) {
    %0 = tensor.empty() : tensor<1x1x2048x49xf32>
    %1 = "ttir.reshape"(%arg0, %0) <{shape = [1 : i32, 1 : i32, 2048 : i32, 49 : i32]}> : (tensor<1x2048x7x7xf32>, tensor<1x1x2048x49xf32>) -> tensor<1x1x2048x49xf32>
    %2 = tensor.empty() : tensor<1x1x49x2048xf32>
    %3 = "ttir.transpose"(%1, %2) <{dim0 = -2 : si32, dim1 = -1 : si32}> : (tensor<1x1x2048x49xf32>, tensor<1x1x49x2048xf32>) -> tensor<1x1x49x2048xf32>
    %4 = tensor.empty() : tensor<1x1x1x2048xf32>
    %5 = "ttir.mean"(%3, %4) <{keep_dim = true}> {dim = -2 : si32} : (tensor<1x1x49x2048xf32>, tensor<1x1x1x2048xf32>) -> tensor<1x1x1x2048xf32>
    %6 = tensor.empty() : tensor<1x2048x1x1xf32>
    %7 = "ttir.reshape"(%5, %6) <{shape = [1 : i32, 2048 : i32, 1 : i32, 1 : i32]}> : (tensor<1x1x1x2048xf32>, tensor<1x2048x1x1xf32>) -> tensor<1x2048x1x1xf32>
    return %7 : tensor<1x2048x1x1xf32>
  }
}

When I run ttrt on silicon I get segfault when trying to get output tensor of to_layout operation. Below is log tail:

2024-12-25 21:59:48,810 - DEBUG - evaluating program=0 for binary=avg.ttnn
                 Always |  WARNING | Golden information not found
2024-12-25 21:59:48,811 - DEBUG - generating inputs/outputs for loop=1/1 for binary=avg.ttnn
2024-12-25 21:59:48,812 - DEBUG - starting loop=1/1 for binary=avg.ttnn
            RuntimeTTNN |    DEBUG | Executing operation: %0 = "ttnn.get_device"() <{mesh_shape = #ttnn<mesh_shape 1x1>}> : () -> !tt.device<<workerGrid = #tt.grid<8x8, (d0, d1) -> (0, d0, d1)>, l1Map = (d0, d1)[s0, s1] -> (0, d0 floordiv s0, d1 floordiv s1, (d0 mod s0) * s1 + d1 mod s1), dramMap = (d0, d1)[s0, s1] -> (0, 0, ((((d0 floordiv s0) * 8 + d1 floordiv s1) * (s1 * s0) + (d0 mod s0) * s1 + d1 mod s1) floordiv 8192) mod 12, (((d0 floordiv s0) * 8 + d1 floordiv s1) * (s1 * s0) + (d0 mod s0) * s1 + d1 mod s1) floordiv 98304 + (((d0 floordiv s0) * 8 + d1 floordiv s1) * (s1 * s0) + (d0 mod s0) * s1 + d1 mod s1) mod 8192), meshShape = , chipIds = [0]>> loc("avg.ttnn.mlir":14:10)
2024-12-25 21:59:48,812 - DEBUG - executing golden comparison
                 Always |  WARNING | Golden information not found
                 Always |  WARNING | Output tensor not found in tensor pool
2024-12-25 21:59:48,812 - DEBUG - Golden tensor is None - skipping golden comparison
            RuntimeTTNN |    DEBUG | Executing operation: %1 = "ttnn.to_layout"(%arg0) <{layout = #ttnn.layout<tile>}> : (tensor<1x2048x7x7xf32, #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 14336 + d1 * 7 + d2, d3), <1x1>, memref<14336x7xf32, #ttnn.buffer_type<system_memory>>>>) -> tensor<1x2048x7x7xf32, #ttnn.ttnn_layout<(d0, d1, d2, d3) -> (d0 * 14336 + d1 * 7 + d2, d3), <1x1>, memref<448x1x!tt.tile<32x32, f32>, #ttnn.buffer_type<dram>>, <interleaved>>> loc("avg.ttnn.mlir":15:10)
2024-12-25 21:59:48,864 - DEBUG - executing golden comparison
                 Always |  WARNING | Golden information not found
Segmentation fault

I'm running mlir in debug build on commit 109d917. If any information is missing please reach out.