}

// -----

// CHECK-LABEL: testReverseV2
func.func @testReverseV2(%arg0: tensor<1x2x3x4xf32>, %arg1 : tensor<2xi32>) -> tensor<1x2x3x4xf32> {
  // CHECK: "tfl.reverse_v2"(%arg0, %arg1)
  %0 = "tfl.reverse_v2"(%arg0, %arg1): (tensor<1x2x3x4xf32>, tensor<2xi32>) -> tensor<1x2x3x4xf32>
  func.return %0 : tensor<1x2x3x4xf32>
}

// -----

// test select
// CHECK-LABEL: testSelect

  %cst_0 = arith.constant dense<1.000000e+00> : tensor<f32>
  %cst_1 = arith.constant dense<0.000000e+00> : tensor<8xf32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<8x3x3x3xf32>
  %0 = "tfl.sub"(%arg0, %cst_0) {fused_activation_function = "NONE"} : (tensor<3x15x14x3xf32>, tensor<f32>) -> tensor<3x15x14x3xf32>

func.func @fake_quant_conv(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.FakeQuantWithMinMaxArgs"(%arg1) {device = "", max = 2.000000e+00 : f32, min = -1.000000e+00 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<2x3x3x2xf32>) -> tensor<*xf32>

// CHECK: return %[[RES0]] : tensor<256x32x32x16xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "CONV_2D"
    }
  ],
  "subgraphs": [
    {

// RUN: tf-quant-opt %s -split-input-file -quant-insert-quantized-functions -quant-quantize-composite-functions | FileCheck %s

module {
  func.func @conv(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

  %cst2 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<1x112x112x3xf32>, tensor<1x3x3x2xf32>) -> tensor<1x28x23x2xf32>

func.func @QuantizeGatherWeightOnly(%arg0: tensor<3xi32>) -> tensor<3x3x3x3xf32> {
  %w = arith.constant dense<1.270000e+02> : tensor<64x3x3x3xf32>
  %emb = "tfl.gather"(%w, %arg0) {axis = 0 : i32, batch_dims = 0 : i32} : (tensor<64x3x3x3xf32>, tensor<3xi32>) -> tensor<3x3x3x3xf32>
  %emb_s = "quantfork.stats"(%emb) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3xf32>
  func.return %emb_s : tensor<3x3x3x3xf32>

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x1x1x3xf32>) -> tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x1x1x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x1x1x3xf32>
  %5 = "tfl.add"(%1, %4) {fused_activation_function = "NONE"} : (tensor<1x5x5x3xf32>, tensor<1x1x1x3xf32>) -> tensor<1x5x5x3xf32>

// CHECK: return %[[RES0]] : tensor<256x32x32x16xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "CONV_2D"
    }
  ],
  "subgraphs": [

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>) -> tensor<256x8x7x4xf32>
  func.return %rst : tensor<256x8x7x4xf32>