func.func private @composite_dot_general_fn_1(%arg0: tensor<3x3xf32>, %arg1: tensor<3x11xf32>) -> tensor<3x11xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<3x3xf32>, tensor<3x11xf32>) -> tensor<3x11xf32>
    return %0 : tensor<3x11xf32>
  }
}

// -----

    %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

  %5 = "tfl.pseudo_const"() {value = dense<[[0.5]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %6 = "tfl.pseudo_const"() {value = dense<[[0.6]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %7 = "tfl.pseudo_const"() {value = dense<[[0.7]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %8 = "tfl.pseudo_const"() {value = dense<[[0.8]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %9 = "tfl.no_value"() {value} : () -> none

    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<3x4xf32>) -> tensor<3x4x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<3x4x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<3x4xf32>
    %7 = stablehlo.slice %6 [1:3, 2:4] : (tensor<3x4xf32>) -> tensor<2x2xf32>

func.func @softmax(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Softmax"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexSoftmax", custom_option = #tfl<const_bytes : "0x07536F66746D617800161207536F66746D61781A002A070A0154120230013200000221191414042801">}> : (tensor<8x16xf32>) -> tensor<8x16xf32>

    func.func @main() -> (tensor<5x11xf32>) {
      %arg0 = "tf.Const"() {value = dense<-3.0> : tensor<5x7xf32>} : () -> tensor<5x7xf32>
      %arg1 = "tf.Const"() {value = dense<-3.0> : tensor<11x7xf32>} : () -> tensor<11x7xf32>

      %1 = "tf.MatMul"(%arg0, %arg1) {transpose_a = false, transpose_b = true} : (tensor<5x7xf32>, tensor<11x7xf32>) -> tensor<5x11xf32>

      func.return %1 : tensor<5x11xf32>
    }

// -----

// CHECK-LABEL: float_matmul
func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

  "tfl.custom"(%0, %2) {custom_code = "FlexAssignVariableOp", custom_option = #tfl<const_bytes : "0x1041737369676E5661726961626C654F70003B121041737369676E5661726961626C654F701A001A002A0B0A056474797065120230012A140A0E76616C69646174655F736861706512022800320000024F3E1414042801">} : (tensor<!tf_type.resource<tensor<1x10xf32>>>, tensor<1x10xf32>) -> ()

  %w = arith.constant dense<127.0> : tensor<4x12xf32>
  %b = arith.constant dense<0.0> : tensor<4xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<4x12xf32>, tensor<4xf32>) -> tensor<1x112x112x4xf32>
  func.return %fc : tensor<1x112x112x4xf32>

  func.return %0: tensor<3x2xf32>
}

func.func @sparse_f16() -> tensor<3x2xf16> {
  // CHECK-LABEL: @sparse_f16