%2 = "mhlo.slice"(%arg0) <{limit_indices = dense<[3, 1, 512]> : tensor<3xi64>, start_indices = dense<[2, 0, 0]> : tensor<3xi64>, strides = dense<1> : tensor<3xi64>}> : (tensor<3x1x512xf32>) -> tensor<1x1x512xf32>
  %r = "mhlo.concatenate"(%0, %1, %2) <{dimension = 0 : i64}> : (tensor<1x1x512xf32>, tensor<1x1x512xf32>, tensor<1x1x512xf32>) -> tensor<3x1x512xf32>

  func.return %result : tensor<1x1x512xf32>

func.func @quantized_dot_general_not_converted(%arg0: tensor<1x1x512xf32>, %arg1: tensor<512x512x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x1x512xf32> {
  %0 = "mhlo.dot_general"(%arg0, %arg1) {
    dot_dimension_numbers = #mhlo.dot<
      lhs_contracting_dimensions = [2],
      rhs_contracting_dimensions = [0]
    >} : (tensor<1x1x512xf32>, tensor<512x512x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x1x512xf32>
  func.return %0 : tensor<1x1x512xf32>
}

func.func @mean_4d_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = true} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x1x1x512xf32>
  func.return %0 : tensor<1x1x1x512xf32>
}

// CHECK:       func @mean_4d_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {

  return %0 : tensor<1x1x2x2xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_4(%arg0: tensor<1x1x3x3xf32>) -> tensor<1x1x2x2xf32>

// CHECK-LABEL: avg_pool2d_5
// CHECK: %cst = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x1x2xbf16>) -> tensor<1x1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x1x2xf32>) -> tensor<1x1x2xf32>
  return %3 : tensor<1x1x2xf32>
}

// CHECK: func @cast_bf16_batch_matmul_v2_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>

// CHECK:           %[[VAL_8:.*]] = "tfl.reshape"(%[[VAL_7]], %[[VAL_3]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x2xf32>, tensor<1xi32>) -> tensor<2xf32>

  %1 = "tf.Maximum"(%0, %cst_0) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  %2 = "tf.Minimum"(%1, %cst_1) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  func.return %2 : tensor<1x3x4x2xf32>
// CHECK-DAG: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<1x1x3x2xf32>}> : () -> tensor<1x1x3x2xf32>

    %7 = stablehlo.constant dense<5.000000e-01> : tensor<1x1x1xf32>  // Output inverse scale (1 / s3).
    %8 = stablehlo.constant dense<-5> : tensor<1x1x1xi8>  // Output zero point (z3).
    %9 = stablehlo.constant dense<1.250000e+01> : tensor<1x1x1xf32>  // Merged scale (s1 * s2).
    %10 = call @uniform_quantize(%arg0, %1, %2) : (tensor<8x16x16xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<8x16x16xi8>  // q1

  // CHECK-DAG: %[[CST:.+]] = mhlo.constant dense<[1.000000e-01, 2.000000e-01]> : tensor<2xf32>
  // CHECK-DAG: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>
  // CHECK-DAG: %[[NEW_FILTER:.+]] =  mhlo.multiply %[[CST_BCAST]], %[[FILTER]] : tensor<1x1x3x2xf32>