// RUN: tf-quant-opt %s -split-input-file -quant-propagate-quantize-type | FileCheck %s

module {
  func.func @not_propagate_matmul(%arg0: tensor<1x2x2x2xf32>) -> tensor<*xf32> {
    %cst = "tf.Const"() {value = dense<127> : tensor<2x1024xi8>} : () -> tensor<2x1024xi8>
    %cst_0 = "tf.Const"() {value = dense<0.0157480314> : tensor<f32>} : () -> tensor<f32>
    %0 = "tf.Identity"(%cst) : (tensor<2x1024xi8>) -> tensor<2x1024xi8>

// CHECK-LABEL: func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>)
func.func @testCase(%arg0: tensor<i32>, %arg1: tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32> {
  %0 = "tf.Case"(%arg0, %arg1) {branches = [@branch_0, @branch_1], is_stateless = false} : (tensor<i32>, tensor<!tf_type.resource<tensor<1x2x3xf32>>>) -> tensor<1x2x3xf32>

  // CHECK: [[Result0:%.*]] = "tf.CaseRegion"

    %7 = call @uniform_quantize_0(%arg0, %0, %1) : (tensor<1x4x2xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x2xi8>
    %8 = stablehlo.convert %7 : (tensor<1x4x2xi8>) -> tensor<1x4x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [2] x [0] : (tensor<1x4x2xf32>, tensor<2x3xf32>) -> tensor<1x4x3xf32>

func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
// CHECK: return %[[CUSTOM_0]] : tensor<1x2x2x1xf32>
}

// support dynamic shaped operands like the native lowering. Verify that
// 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

  %1 = "tfl.add"(%cst_2, %cst_1) {fused_activation_function = "NONE"} : (tensor<2x2x2xf32>, tensor<  2x2xf32>) -> tensor<2x2x2xf32>
  %2 = "tfl.add"(%cst_0, %cst_2) {fused_activation_function = "NONE"} : (tensor<    2xf32>, tensor<2x2x2xf32>) -> tensor<2x2x2xf32>

  func.return %0, %1, %2 : tensor<2x2xf32>, tensor<2x2x2xf32>, tensor<2x2x2xf32>

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x2x2x3xf32>

    %0 = "tf.Div"(%arg1, %cst_0) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %1 = "tf.AddV2"(%0, %cst_1) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %2 = "tf.Floor"(%1) : (tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
    %3 = "tf.ClipByValue"(%2, %cst_5, %cst_6) : (tensor<2x2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2x2xf32>
    %4 = "tf.Cast"(%3) {Truncate = false} : (tensor<2x2x2xf32>) -> tensor<2x2x2xi8>

    %1 = "tf.PartitionedCall"(%arg0, %cst, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_with_bias_and_relu6_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>, tensor<2xf32>) -> tensor<*xf32> loc(callsit...

  func.return %2, %3 : tensor<128x16xf32>, tensor<128xi32>
}

// CHECK-LABEL: PruneUnusedLstm
func.func @PruneUnusedLstm(%arg0: tensor<1x28x28xf32>) -> (tensor<1x28x28xf32>) {
    %input = "tfl.quantize"(%arg0) {qtype = tensor<1x28x28x!quant.uniform<i8:f32, 0.003:-128>>} : (tensor<1x28x28xf32>) -> tensor<1x28x28x!quant.uniform<i8:f32, 0.003:-128>>