// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-prepare-quantize=enable-per-channel-quantized-weight=true -verify-diagnostics | FileCheck %s

// -----

module {
  // CHECK-LABEL: conv_with_bias_and_relu
  func.func private @conv_with_bias_and_relu(%arg0: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32> {
    %cst = "tf.Const"() {device = "", value = dense<[7.11401462, 7.05456924]> : tensor<2xf32>} : () -> tensor<2xf32>

  if (dot_general_op == nullptr) return std::nullopt;
  const int64_t filter_rank =
      mlir::dyn_cast<ShapedType>(dot_general_op.getOperand(1).getType())
          .getRank();

  // To quantize rhs per-channel, we currently only consider the case where
  // `stablehlo.dot_general` is legalizable to `tfl.fully_connected`.
  const bool is_per_axis_quantizable =
      IsDotGeneralFullyConnected(dot_general_op).value();

      mlir::TFL::PassConfig(quant_specs), pm);

  if (failed(pm.run(module.get()))) {
    absl::string_view err = statusHandler.ConsumeStatus().message();
    LOG(ERROR) << "Failed to quantize: " << err;
    return kTfLiteError;
  }

  // Export the results to the builder
  std::string result;
  tflite::FlatbufferExportOptions options;
  options.toco_flags.set_force_select_tf_ops(false);

// RUN: stablehlo-quant-opt %s -split-input-file -verify-diagnostics \
// RUN:     -stablehlo-quantize-composite-functions | FileCheck --check-prefix=CHECK %s

// Test that per-tensor weight-only quantized dot_general op is produced when
// empty `weight_only_ptq` is provided.

module attributes {tf_saved_model.semantics} {
  func.func private @quantize_dot_general_per_tensor(%arg0: tensor<1x2xf32>) -> tensor<1x3xf32> attributes {tf._original_func_name = "main_0"} {

// RUN: stablehlo-quant-opt %s -split-input-file -stablehlo-prepare-quantize=enable-per-channel-quantized-weight=false -verify-diagnostics | FileCheck %s

// -----

// CHECK-LABEL: func @dot
// CHECK-SAME: (%[[ARG_0:.*]]: tensor<?x3xf32>) -> tensor<?x2xf32>
func.func @dot(%arg0: tensor<?x3xf32>) -> tensor<?x2xf32> {
  // CHECK: %[[cst:.*]] = stablehlo.constant
  // CHECK: %[[q1:.*]] = "quantfork.qcast"(%[[cst]])

      quantization.}],
      "bool", "GetDynamicRangeQuantKernelSupport",
      (ins), [{}], [{return false;}]>,
    InterfaceMethod<
      [{Returns whether the op requires asymmetric quantize input attribute
      setting.}],
      "bool", "RequireAsymmetricQuantizeInputsAttr",
      (ins), [{}], [{return false;}]>,
  ];
}

// Specify this trait if the op has a fixed output value range.

        returned_type = quant::ConvertSignedQuantizedToUnsigned(
            dequantize_input.getType(), dequantize_op.getLoc());
        // replace the dequantize op by a quantize op
        TypeAttr type_attr = TypeAttr::get(returned_type);
        auto quantize_op = builder.create<QuantizeOp>(
            dequantize_op.getLoc(), returned_type, dequantize_input, type_attr);

// CHECK-NOT: "tfl.batch_matmul"
func.func @Batchmatmul2FullyconnectedQDQ(%arg0: tensor<4x128x2xf32>, %arg1: tensor<2x1xf32>) -> (tensor<4x128x1xf32>) {
  %0 = arith.constant dense<[[1.0], [2.0]]> : tensor<2x1xf32>
  %1 = "tfl.quantize"(%0) {qtype = tensor<2x1x!quant.uniform<i8:f32, 0.024986599940879671:92>>} : (tensor<2x1xf32>) -> tensor<2x1x!quant.uniform<i8:f32, 0.024986599940879671:92>>

func.func @main(%arg0: tensor<1x2xf32>) -> tensor<1x2xf32> {
  %0 = stablehlo.add %arg0, %arg0 : tensor<1x2xf32>
  return %0 : tensor<1x2xf32>
}
// Tests that `composite_add_fn_1` does not quantize when quantizing
// only compute-heavy ops.

// STATIC-RANGE-PTQ-TO-COMPUTE-HEAVY: %[[CONST:.+]] = stablehlo.constant dense<2.000000e+00>