%flow_1 = "tf.TensorArrayWriteV3"(%handle, %index, %value, %flow) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<2x!tf_type.resource<tensor<1x512xf32>>>, tensor<i32>, tensor<1x512xf32>, tensor<f32>) -> tensor<f32>

  // CHECK: return %[[RESULT]] : tensor<2x3x4x6xf32>
}

// -----

func.func @batchMatMulOneDim(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.BatchMatMul"(%arg0, %arg1) : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>

  // CHECK-LABEL: batchMatMulOneDim

// CHECK:           [[LHSBCAST:%.*]] = "mhlo.dynamic_broadcast_in_dim"([[LHS]], [[LHSSHAPEEXTENTS]]) <{broadcast_dimensions = dense<[0, 1, 2]> : tensor<3xi64>}> : (tensor<1x4x2xf32>, tensor<3xindex>) -> tensor<3x4x2xf32>
// CHECK:           [[RHSBCASTSHAPE:%.*]] = shape.concat [[BCASTHEAD]], [[RHSTAIL]]
// CHECK:           [[RESULT:%.*]] = "mhlo.dot_general"([[LHSBCAST]], [[RHS]])
// CHECK:           return [[RESULT]] : tensor<3x4x4xf32>

  %w = arith.constant dense<[[[[-1.0, 1.0]]], [[[1.0, 2.0]]], [[[-2.0, 1.0]]]]> : tensor<3x1x1x2xf32>
  %b = arith.constant dense<[1.0e-2, 2.1473647e1, -2.1473647e2]> : tensor<3xf32>
  %transpose_conv = "tfl.transpose_conv"(%arg1, %w, %0, %b) {
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, fused_activation_function = "NONE"
  } : (tensor<4xi32>, tensor<3x1x1x2xf32>, tensor<1x5x5x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>

  %cst_add = arith.constant dense<2.0> : tensor<2x512xf32>  // Not 1D
  %cst_weights = arith.constant dense<3.0> : tensor<1024x512xf32>
  %cst_bias = arith.constant dense<5.0> : tensor<1024xf32>

  %0 = "tfl.add"(%arg, %cst_add) {fused_activation_function = "NONE"} : (tensor<2x512xf32>, tensor<2x512xf32>) -> tensor<2x512xf32>

  func.func private @not_quantize_singular_op_without_static_range_ptq(%arg: tensor<3x4x2xf32>) -> tensor<2x3x2x2xf32> attributes {tf._original_func_name = "main_0"} {
    %cst = "tf.Const"() {value = dense<1> : tensor<2x3x2xi32>} : () -> tensor<2x3x2xi32>
    %0 = "quantfork.stats"(%arg) {layerStats = dense<[4.00000000e-6, 9.80000000e-1]> : tensor<2xf32>} : (tensor<3x4x2xf32>) -> tensor<3x4x2xf32>

  constexpr absl::string_view kModuleGather = R"mlir(
    module {
      func.func @main() -> (tensor<2x3x2x2xf32>) {
        %0 = stablehlo.constant dense<1.0> : tensor<3x4x2xf32>
        %1 = stablehlo.constant dense<2> : tensor<2x3x2xi64>
        %2 = "stablehlo.gather"(%0, %1) {
          dimension_numbers = #stablehlo.gather<
            offset_dims = [2, 3],

func.func @no_qdq_except_conv_and_dot_general(%arg0: tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32> attributes {tf._original_func_name = "main_0"} {
  %cst = "tf.Const"() {value = dense<3.000000e-01> : tensor<3x4x2xf32>} : () -> tensor<3x4x2xf32>
  %0 = "tf.XlaCallModule"(%cst, %arg0) {
    Sout = [#tf_type.shape<1x3>], _entry_function = @composite_gather_fn,
    _original_entry_function = "composite_gather_fn", _quantization_method = "weight_only_ptq { }",

  %2 = "tf.SpaceToBatchND"(%arg0, %0, %1) : (tensor<1x8x2xf32>, tensor<1xi32>, tensor<1x2xi32>) -> tensor<3x5x2xf32>

  // CHECK: return [[VAL8]]
  func.return %2 : tensor<3x5x2xf32>
}

// CHECK-LABEL: avoid_lowering_space_to_batch_nd
func.func @avoid_lowering_space_to_batch_nd(%arg0: tensor<1x8x2xf32>, %arg1: tensor<*xi32>) -> (tensor<3x5x2xf32>) {
  %0 = "tf.Const"() {value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>

func.func @lower_einsum(%arg0: tensor<3x4x5xf32>, %arg1: tensor<3x5x6xf32>) -> tensor<3x4x6xf32> {
  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,ikm->ijm"}: (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>
  func.return %0 : tensor<3x4x6xf32>
}
// CHECK-LABEL: lower_einsum
// CHECK: "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = false}> : (tensor<3x4x5xf32>, tensor<3x5x6xf32>) -> tensor<3x4x6xf32>

// -----