// CHECK: return %[[v4]] : tensor<?x36x32xf32>
}

func.func @einsum_with_runtime_shape2(%arg0 : tensor<?x?x8x64xf32>, %arg1 : tensor<8x8x64xf32>) -> tensor<?x?x8xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {device = "", equation = "ABNH,DNH->ABD"} : (tensor<?x?x8x64xf32>, tensor<8x8x64xf32>) -> tensor<?x?x8xf32>
  func.return %0 : tensor<?x?x8xf32>
// CHECK-LABEL: einsum_with_runtime_shape2

  %2 = "tf.Transpose"(%1, %cst_1) {device = ""} : (tensor<8x64x64xf32>, tensor<3xi32>) -> tensor<64x8x64xf32>
  %3 = "tf.Reshape"(%2, %cst_2) {device = ""} : (tensor<64x8x64xf32>, tensor<2xi32>) -> tensor<512x64xf32>

  func.return %3 : tensor<512x64xf32>

  // CHECK-DAG: %[[CST_0:.*]] = "tf.Const"() <{value = dense<[2, 0, 1]> : tensor<3xi32>}> : () -> tensor<3xi32>

  %1 = "tf.Add"(%0, %arg5) : (tensor<8x8x40xf32>, tensor<40xf32>) -> tensor<8x8x40xf32>
  %2 = "tf.BatchMatMulV2"(%1, %arg4) {adj_x = false, adj_y = true} : (tensor<8x8x40xf32>, tensor<10x40xf32>) -> tensor<8x8x10xf32>
  %3 = "tf.Add"(%2, %arg1) : (tensor<8x8x10xf32>, tensor<8x10xf32>) -> tensor<8x8x10xf32>

// DISABLE-ALL-DOT-GENERAL: @main
func.func @main(%arg0: tensor<1x1x167xf32>) -> tensor<1x1x64xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<167x64xf32>
  %1 = stablehlo.dot_general %arg0, %0, contracting_dims = [2] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  return %1 : tensor<1x1x64xf32>
}

// DISABLE-ALL-DOT-GENERAL: %[[CONST:.+]] = stablehlo.constant dense<2.000000e+00>

    %19 = stablehlo.convert %18 : (tensor<8x16x4xi32>) -> tensor<8x16x4xf32>
    %20 = stablehlo.broadcast_in_dim %9, dims = [0, 1, 2] : (tensor<1x1x1xf32>) -> tensor<8x16x4xf32>
    %21 = stablehlo.multiply %19, %20 : tensor<8x16x4xf32>  // * s1 s2
    %22 = call @uniform_quantize_1(%21, %7, %8) : (tensor<8x16x4xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<8x16x4xi8>

  %5 = stablehlo.broadcast_in_dim %1, dims = [2] : (tensor<64xf32>) -> tensor<1x1x64xf32>
  %6 = stablehlo.add %4, %5 : tensor<1x1x64xf32>
  %7 = stablehlo.clamp %2, %6, %3 : tensor<1x1x64xf32>
  func.return %7: tensor<1x1x64xf32>
}
// CHECK: %[[CONST_0:.*]] = stablehlo.constant dense<2.000000e+00>
// CHECK: %[[CONST_1:.*]] = stablehlo.constant dense<2.000000e+00>

    precision_config = #vhlo.array_v1<[#vhlo<precision_v1 DEFAULT>, #vhlo<precision_v1 DEFAULT>]>}> : (tensor<1x1x167xf32>, tensor<167x64xf32>) -> tensor<1x1x64xf32>
  return %0 : tensor<1x1x64xf32>
}

//CHECK:func.func private @dot_general(%arg0: tensor<1x1x167xf32>, %arg1: tensor<167x64xf32>) -> tensor<1x1x64xf32> {
//CHECK-NEXT: %0 = "vhlo.dot_general_v1"(%arg0, %arg1) <{
//CHECK-SAME:    lhs_batching_dimensions = #vhlo.tensor_v1<dense<0> : tensor<1xi64>>,

// CHECK:           return %[[VAL_5]] : tensor<8x8x8x64xf32>
// CHECK:         }
func.func @convert_conv2d_back_prop_input(%arg0: tensor<8x4x4x32xf32>, %arg1: tensor<3x3x64x32xf32>) -> tensor<8x8x8x64xf32> {
  %0 = "mhlo.convolution"(%arg0, %arg1) {batch_group_count = 1 : i64,
    dimension_numbers = #mhlo.conv<raw

  %cst_0 = arith.constant dense<[1, 1, 64]> : tensor<3xi32>
  %0 = "tfl.reshape"(%arg0, %cst_0) : (tensor<1x64xf32>, tensor<3xi32>) -> tensor<1x1x64xf32>
  %1 = "tfl.batch_matmul"(%0, %arg1) {adj_x = false, adj_y = false} : (tensor<1x1x64xf32>, tensor<1x64x1024xf32>) -> tensor<1x1x1024xf32>
  %2 = "tfl.reshape"(%1, %cst) : (tensor<1x1x1024xf32>, tensor<2xi32>) -> tensor<1x1024xf32>
  return %2 : tensor<1x1024xf32>

    // CHECK: %[[CONST3:.*]] = "tf.Const"() <{value = dense<
    // CHECK-SAME: [7, 7, 3, 64]
    // CHECK: %[[CONST4:.*]] = "tf.Const"() <{value = dense<
    // CHECK-SAME: 0
    // CHECK: %[[SLICE0:.*]] = "tf.Slice"(%[[RESHAPE1:.*]], %[[CONST4:.*]], %[[CONST3:.*]]) : (tensor<8x8x3x64xf32>, tensor<4xi64>, tensor<4xi32>) -> tensor<7x7x3x64xf32>