// CHECK{LITERAL}: %cst_1 = arith.constant dense<[[[[1.000000e+00], [2.000000e+00]], [[2.000000e+00], [4.000000e+00]]]]> : tensor<1x2x2x1xf32>
// CHECK: %3 = tfl.mul %2, %cst_1 {fused_activation_function = "NONE"} : tensor<1x2x2x1xf32>
// CHECK: %cst_2 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

func.func @NoPadStridedSliceNonNewAxisMask(%arg0: tensor<1x2x3x1xf32>) -> tensor<1x2x3x1xf32> {
  %cst = arith.constant dense<0> : tensor<4xi32>
  %cst_0 = arith.constant dense<1> : tensor<4xi32>
  %0 = "tf.StridedSlice"(%arg0, %cst, %cst, %cst_0) {begin_mask = 15 : i64, ellipsis_mask = 0 : i64, end_mask = 15 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 0 : i64} : (tensor<1x2x3x1xf32>, tensor<4xi32>, tensor<4xi32>, tensor<4xi32>) -> tensor<1x2x3x1xf32>

// CustomOpNotWeightOnly-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> tensor<*xf32> attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w = arith.constant dense<127.0> : tensor<1024x1x1x1xf32>

// CHECK-LABEL: testSpaceToDepth
func.func @testSpaceToDepthF32(%arg0: tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32> {
  // CHECK: %[[ARG:.*]]: tensor<1x2x2x1xf32>
  // CHECK: "tfl.space_to_depth"(%[[ARG]]) <{block_size = 2 : i32}> : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  %0 = "tfl.space_to_depth"(%arg0) {block_size = 2: i32} : (tensor<1x2x2x1xf32>) -> tensor<1x1x1x4xf32>
  func.return %0 : tensor<1x1x1x4xf32>
}

// -----

  // We cannot fuse this tfl.mul into the preceding conv op because %cst2 is not broadcast-compatible to %cst0.
  %1 = "tfl.mul"(%0, %cst2) {fused_activation_function = "RELU6"} : (tensor<1x4x4x2xf32>, tensor<4x2xf32>) -> tensor<1x4x4x2xf32>

  func.return %1 : tensor<1x4x4x2xf32>

  %0:2 = "tf.SplitV"(%arg0, %arg1, %arg2) : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
  func.return %0#0 : tensor<1x4x2x3xf32>

  // CHECK-LABEL: splitv
  // CHECK: "tfl.split_v"(%arg0, %arg1, %arg2) <{num_splits = 2 : i32}> : (tensor<1x4x3x3xf32>, tensor<2xi32>, tensor<i32>) -> (tensor<1x4x2x3xf32>, tensor<1x4x1x3xf32>)
}