%computation = "tf_device.cluster_func"(%read) {func = @computation, use_spmd_for_xla_partitioning = true} : (tensor<i32>) -> tensor<i32>
      "tf.AssignVariableOp"(%partitioned_variable, %computation) : (tensor<!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
      return
    }

    func @computation(%arg0: tensor<i32>) -> tensor<i32> {
      return %arg0: tensor<i32>

        [](Operation* user) { return isa<TF::AssignVariableOp>(user); });

    // TODO(b/186717563): Eliminate all pass through results once XLA correctly
    // handles empty computations. Another approach could be to drop empty
    // clusters within MLIR but that seems to trigger other failures but can be
    // considered again.
    // Old bridge only removes unsupported TPU types (only string for now)

  // Internal tensors are those that do not flow in or out of the operation,
  // but instead are part of internal computation. As such, the operation's
  // implementation may manage its memory more efficiently. They are needed
  // however (i.e. not just an implementation detail) since they are part of the
  // computation, which may require relevant metadata such as quantization
  // parameters.
  intermediates:[int];

//
// (Note that this can be viewed equivalently as:
//
//	actionID(binary)/buildID(main.a)/contentID(binary)
//
// Storing the buildID(main.a) in the middle lets the computations that care
// about the prefix or suffix halves ignore the middle and preserves the
// original build ID as a contiguous string.)
//
// During the build, when it's time to build main.a, the gofmt binary has the

	physPageSize = getPageSize()

	// Windows dynamic priority boosting assumes that a process has different types
	// of dedicated threads -- GUI, IO, computational, etc. Go processes use
	// equivalent threads that all do a mix of GUI, IO, computations, etc.
	// In such context dynamic priority boosting does nothing but harm, so we turn it off.
	stdcall2(_SetProcessPriorityBoost, currentProcess, 1)
}

//go:nosplit

			if _, acc := x.val.Int(i); acc == big.Exact {
				return makeInt(i)
			}

			// If we can get an integer by rounding up or down,
			// assume x is not an integer because of rounding
			// errors in prior computations.

			const delta = 4 // a small number of bits > 0
			var t big.Float
			t.SetPrec(prec - delta)

			// try rounding down a little
			t.SetMode(big.ToZero)
			t.Set(x.val)

   * usually have much noticeable impact. A value of one permits only one thread to modify the cache
   * at a time, but since read operations and cache loading computations can proceed concurrently,
   * this still yields higher concurrency than full synchronization.
   *
   * <p>Defaults to 4. <b>Note:</b>The default may change in the future. If you care about this

   * usually have much noticeable impact. A value of one permits only one thread to modify the cache
   * at a time, but since read operations and cache loading computations can proceed concurrently,
   * this still yields higher concurrency than full synchronization.
   *
   * <p>Defaults to 4. <b>Note:</b>The default may change in the future. If you care about this

	//         == n * (n-1) * ... * (n-k+1) / 1 * (1+1) * ... * k
	//
	// Using the multiplicative formula produces smaller values
	// at each step, requiring fewer allocations and computations:
	//
	// z = 1
	// for i := 0; i < k; i = i+1 {
	//     z *= n-i
	//     z /= i+1
	// }
	//
	// finally to avoid computing i+1 twice per loop:
	//
	// z = 1
	// i := 0

      result = op->getResult(0);
    }
    rewriter.replaceOp(op, {result});
    return success();
  }
};

// Lowers ResizeNearestNeighbor to an indices computations with a gather along
// the combined spatial dimensions. Generating the indices along the
// width/height index could be used to gather along each of W and H dimension