// WHOLE-PASSES-NEXT:  %[[q2:.*]] = "tfl.quantize"(%[[cc1]]) <{qtype = tensor<1x4x3x!quant.uniform<u8:f32, {{.*}}>>}> : (tensor<1x4x3x!quant.uniform<u8:f32, {{.*}}>>) -> tensor<1x4x3x!quant.uniform<u8:f32, {{.*}}>>

      func.return %%1 : tensor<1x4x4xf32>
    }
  })";
  std::string mat_mul_method =
      absl::StrFormat(kMatMulModuleStr, test_case.mat_mul_method);
  TF_ASSERT_OK_AND_ASSIGN(
      XlaCompiler::CompilationResult result,
      CompileMlirModule(
          mat_mul_method.c_str(),

    %1 = "quantfork.dcast"(%0) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 6.000000e-03:-128>>) -> tensor<2x3x3x2xf32>

  %2 = "quantfork.qcast"(%arg0) : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3x!quant.uniform<i8:f32, 0.0011764706057660721:-43>>
  %3 = "quantfork.dcast"(%2) : (tensor<1x3x4x3x!quant.uniform<i8:f32, 0.0011764706057660721:-43>>) -> tensor<1x3x4x3xf32>

    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "tf.XlaCallModule"(%arg0, %cst) {
        Sout = [#tf_type.shape<1x3x4x2>], dim_args_spec = [], disabled_checks = [],
        has_token_input_output = false, module = "", platforms = [], version = 5 : i64,
        _entry_function = @composite_conv_fn, _original_entry_function = "composite_conv_fn",

  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<1xi32>, tensor<1x2xi32>) -> tensor<1x2xi32>
  func.return %0: tensor<1x2xi32>
}

// CHECK-LABEL: func @broadcast_multi_dim_add
// TODO(laurenzo): Change this to a (4x1x1 + 1x4x4x4) shaped add once upstream
// broadcastable bug is fixed (helps make the CHECK matching unambiguous)
func.func @broadcast_multi_dim_add(%arg0: tensor<4x1x1xi32>, %arg1: tensor<4x4x4x4xi32>) -> tensor<4x4x4x4xi32> {

    // CHECK-NEXT: %[[CMP:.*]] = mhlo.compare GT, %[[INP:.*]], %[[ZERO]], NOTYPE : (tensor<1x4x4xf32>, tensor<1x4x4xf32>) -> tensor<1x4x4xi1>
    // CHECK-NEXT: %[[RES:.*]] = mhlo.select %[[CMP]], %[[GRADIENT]], %[[LEAKYGRAD]] : tensor<1x4x4xi1>, tensor<1x4x4xf32>
    // CHECK-NEXT: return %[[RES]] : tensor<1x4x4xf32>

  %0, %1, %2 = "tfl.split_v"(%arg0, %size_splits, %split_dim) {num_splits = 3 : i32} : (tensor<16x4x4xf32>, tensor<3xi32>, tensor<i32>) -> (tensor<1x4x4xf32>, tensor<1x4x4xf32>, tensor<14x4x4xf32>)
  func.return %0, %1, %2 : tensor<1x4x4xf32>, tensor<1x4x4xf32>, tensor<14x4x4xf32>
}

// -----

func.func @testSplitVOpWithBadSizeSplitsConstantSum(%arg0: tensor<16x4x4xf32>) -> (tensor<0x4x4xf32>, tensor<16x4x4xf32>) {

func.func @partitioned_input_rank_mismatch(%arg0: tensor<!tf_type.resource<tensor<1x4x4xf32>>>) {
  // expected-error @+1 {{rank}}
  %0 = "tf.TPUPartitionedInputV2"(%arg0) {_XlaSharding = "\08\03\1A\05\04\01\01\01\02\22\08\00\01\02\03\04\05\06\070\01", partition_dims = [4, 1, 1, 2], is_packed = true} : (tensor<!tf_type.resource<tensor<1x4x4xf32>>>) -> tensor<!tf_type.resource<tensor<4x4x4xf32>>>