// CHECK: {
// CHECK-NEXT:  version: 3,
// CHECK-NEXT:  operator_codes: [ {
// CHECK-NEXT:    deprecated_builtin_code: 67,
// CHECK-NEXT:    version: 1,
// CHECK-NEXT:    builtin_code: TRANSPOSE_CONV
// CHECK-NEXT:  } ],
// CHECK-NEXT:  subgraphs: [ {
// CHECK-NEXT:    tensors: [ {
// CHECK-NEXT:      shape: [ 4 ],
// CHECK-NEXT:      type: INT32,
// CHECK-NEXT:      buffer: 1,

  %b = arith.constant dense<[1.0e-2, 2.1473647e1, -2.1473647e2]> : tensor<3xf32>
  %transpose_conv = "tfl.transpose_conv"(%arg1, %w, %0, %b) {
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, fused_activation_function = "NONE"
  } : (tensor<4xi32>, tensor<3x1x1x2xf32>, tensor<1x5x5x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %transpose_conv : tensor<1x5x5x3xf32>

  %w = arith.constant dense<127.0> : tensor<1x32x42x128xf32>
  %b = arith.constant dense<0.0> : tensor<1x32x42x128xf32>
  %tconv = "tfl.transpose_conv"(%arg1, %w, %arg0, %b) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<1x32x42x128xf32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>) -> tensor<1x32x42x128xf32>

  %3 = "tfl.transpose"(%1, %cst_0) : (tensor<3x3x16x3x!quant.uniform<i8<-127:127>:f32, 0.047244094488188976>>, tensor<4xi32>) -> tensor<16x3x3x3x!quant.uniform<i8<-127:127>:f32, 0.047244094488188976>>

  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]])
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]])
  // CHECK-DAG: %[[VAL4:.+]] = "tfl.transpose_conv"(%[[SHAPE]], %[[VAL2]], %arg0, %[[VAL3]]) <{fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}>
  // CHECK: return %[[VAL4]]

// tfl.relu / tfl.relu6 / tfl.rsqrt / tfl.sin / tfl.slice / tfl.softmax /
// tfl.space_to_depth / tfl.sqrt / tfl.square / tfl.squared_difference /
// tfl.strided_slice / tfl.tanh / tfl.transpose / tfl.transpose_conv
class GpuBasicSupportedOpNoCost : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override { return 0; }

  bool IsOpSupported(mlir::Operation* op) const override {

  %cst = "tfl.no_value"() {value = unit} : () -> none
  // CHECK: _arithmetic_count = 176160768 : i64
  %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2, %cst) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x32xf32>

  %w = arith.constant dense<127.0> : tensor<1x32x42x128xf32>
  %b = arith.constant dense<0.0> : tensor<1x32x42x128xf32>
  %tconv = "tfl.transpose_conv"(%arg1, %w, %0, %b) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<1x32x42x128xf32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>) -> tensor<1x32x42x128xf32>

  %out = "tfl.transpose_conv"(%output_shape, %dq_weights, %arg0, %bias) {fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<4xi32>, tensor<4x2x2x2048xf32>, tensor<2x2x3x2048xf32>, none) -> tensor<2x3x2x2048xf32>
  func.return %out : tensor<2x3x2x2048xf32>

  // CHECK-NOT: "tfl.dequantize"

  %cst_3 = arith.constant dense<[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0]> : tensor<32xf32>
  %0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>