// CHECK-LABEL: matmulBiasAdd
func.func @matmulBiasAdd(%arg0: tensor<64xf32>, %arg1: tensor<8x32xf32>, %arg2: tensor<32x64xf32>) -> (tensor<*xf32>) {
  // CHECK: %[[VAL_3:.*]] = "tf._FusedMatMul"(%arg1, %arg2, %arg0) <{epsilon = 0.000000e+00 : f32, fused_ops = ["BiasAdd"], transpose_a = false, transpose_b = false}> : (tensor<8x32xf32>, tensor<32x64xf32>, tensor<64xf32>) -> tensor<*xf32>

      func.func @main() -> (tensor<32x64xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}) {
        %cst = "tf.Const"() {value = dense<[524170, 523952]> : tensor<2xi32>} : () -> tensor<2xi32>
        %cst_0 = "tf.Const"() {value = dense<[32, 64]> : tensor<2xi32>} : () -> tensor<2xi32>
        %0 = "tf.StatelessRandomNormal"(%cst_0, %cst) : (tensor<2xi32>, tensor<2xi32>) -> tensor<32x64xf32>
        return %0 : tensor<32x64xf32>
    }

  %cst = arith.constant dense<[3, 4]> : tensor<2xi32>
  %cst_0 = arith.constant dense<1.000000e+00> : tensor<3x4xf32>
  %0 = "tfl.reshape"(%arg0, %cst) : (tensor<*xf32>, tensor<2xi32>) -> tensor<3x4xf32>
  %1 = "tfl.add"(%0, %cst_0) {fused_activation_function = "NONE"} : (tensor<3x4xf32>, tensor<3x4xf32>) -> tensor<3x4xf32>
  func.return %1 : tensor<3x4xf32>
// CHECK-LABEL: ReshapeAddUnknownShape

func.func @move_transpose_handle_broadcast(%arg0:tensor<8x64xf32>, %arg1:tensor<8x64x64xf32>) -> tensor<512x64xf32> {
  %cst = "tf.Const"() {value = dense<3> : tensor<i32>} : () -> tensor<i32>
  %cst_1 = "tf.Const"() {value = dense<[2, 0, 1]> : tensor<3xi32>} : () -> tensor<3xi32>
  %cst_2 = "tf.Const"() {value = dense<[512, 64]> : tensor<2xi32>} : () -> tensor<2xi32>
  %0 = "tf.ExpandDims"(%arg0, %cst) {device = ""} : (tensor<8x64xf32>, tensor<i32>) -> tensor<8x64x1xf32>

func.func @batchmatmulv2_basic(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
// CHECK-LABEL:   func @batchmatmulv2_basic
// CHECK-SAME:        ([[LHS:%.*]]: tensor<1x4x2xf32>, [[RHS:%.*]]: tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
// CHECK:           [[LHSSHAPE:%.*]] = shape.shape_of [[LHS]] : tensor<1x4x2xf32>
// CHECK:           [[RHSSHAPE:%.*]] = shape.shape_of [[RHS]] : tensor<3x2x4xf32>
// CHECK:           [[CM2:%.*]] = arith.constant -2 : index

// fallback lowering is preferred for static shaped operands when available.

// CHECK-LABEL: batchmatmulv2
func.func @batchmatmulv2(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
  // NO_FALLBACK: mhlo.dynamic_broadcast_in_dim
  // NO_FALLBACK: mhlo.dot_general

  // SUPPORTED_FALLBACK_DEVICE: mhlo.reduce
  // SUPPORTED_FALLBACK_DEVICE: mhlo.dot_general

    %0 = stablehlo.constant dense<0.000000e+00> : tensor<1x64xf32>
    %1 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x3xf32>, tensor<3x64xf32>) -> tensor<1x64xf32>
    %2 = stablehlo.add %1, %arg2 : tensor<1x64xf32>
    %3 = stablehlo.maximum %2, %0 : tensor<1x64xf32>

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>

func.func @tfl_wrapped_jax_random_normal(%arg0: tensor<2xui32>) -> tuple<tensor<3x4xf32>> {
  // This is a fake jax random normal body.
  %0 = stablehlo.constant dense<0.0> : tensor<12xf32>
  %1 = "stablehlo.reshape"(%0) : (tensor<12xf32>) -> tensor<3x4xf32>
  %2 = "stablehlo.tuple"(%1) : (tensor<3x4xf32>) -> tuple<tensor<3x4xf32>>
  func.return %2 : tuple<tensor<3x4xf32>>
}


// CHECK-LABEL:   func @tfl_wrapped_jax_random_uniform(

func.func @test_batch_matmul_to_einsum(%arg0: tensor<1x2x3xf32>, %arg1: tensor<3x4xf32>) -> tensor<1x2x4xf32> {
  // CHECK-LABEL: test_batch_matmul_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>
  func.return %0: tensor<1x2x4xf32>
}