// previous quantized layer (Please note that this part is different part
    // from DEBUGGER_TYPE_FLOAT_PER_LAYER). Each layer in the debugging model
    // has a DumpTensor, and it is used to save the entire value of outputs from
    // both the quantized and unquantized layer.
    DEBUGGER_TYPE_INT_PER_LAYER = 2;
    // DEBUGGER_TYPE_FLOAT_PER_LAYER creates a debugging model with both

    return result;
  } else {
    return std::nullopt;
  }
}

// Populates quantized ops from `module_op` to `results`. After going through
// the quantization passes, quantized ops are represented as `func::CallOp` with
// a callee's prefix of `quantized_`.
void PopulateQuantizedResults(ModuleOp module_op,
                              QuantizationResults& results) {

// CHECK: -------- Quantization Summary --------
// CHECK: Number of quantized layers in the model
// CHECK: --------------------------------
// CHECK: Name    Count/Total
// CHECK: ================================
// CHECK: Conv2D  1/2

// CHECK: Number of quantized layers with quantized outputs: 1/1
// CHECK: Number of quantize layers added: 1
// CHECK: Number of dequantize layers added: 1
}

// -----

//   * Input tensors are per-tensor uniform quantized (i8->f32)
//     tensors (full integer) with shape [..., r_x, c_x] or [..., c_x, r_x].
//   * The filter tensor is a per-tensor uniform quantized (i8->f32) tensor
//     (constant or activation) with shape [..., r_y, c_y] or [..., c_y, r_y].
//   * Output tensors are per-tensor uniform quantized (i8->f32) or
//     per-channel uniform quantized (i32->f32) tensors.
//

namespace mlir::quant::stablehlo {

// Checks whether an op is connected with a quantized composite function. If
// not, the same-scale op will not be quantized. This decision is based on the
// current assumption that the performance gain of the same-scale op itself
// could not beat the overhead of the quantize and dequantize routines need to
// be added around that op. When the assumption changes, this policy might

def ComposeUniformQuantizedTypePass : Pass<"compose-uniform-quantized-type", "ModuleOp"> {
  let summary = "Compose uniform quantized types in StableHLO.";
  let constructor = "mlir::odml::CreateComposeUniformQuantizedTypePass()";
  let description = [{
    Identifies uniform quantization patterns and composes them to uniform
    quantized types. This pass targets a specific set of models that are
    quantized from the framework level, which produces "decomposed"

      llvm::SmallVector<Value, 4> quantized;
      for (auto user : returned.getUsers()) {
        if (auto q = Quantized(user)) {
          quantized.push_back(q);
        }
      }
      if (quantized.size() == 1) {
        ret.setOperand(i, quantized.front());
      }
      i++;
    }
  });

  // We prefer to placing quantization emulation ops on the results of the

// RUN: stablehlo-quant-opt %s -stablehlo-test-post-calibration-component='unpack-quantized-types=false' \
// RUN:   -split-input-file | FileCheck %s --check-prefix=CHECK-NO-UNPACK

// Tests that a simple dot_general (lifted as a function) with CustomAggregators
// around it is quantized. The resulting graph has quantized types unpacked into
// int ops.
func.func @main(%arg0: tensor<1x1024xf32>) -> tensor<1x3xf32> {

//      quantized representation may be acceptable.
//
// Especially early in transformation, it is common to have pairs of
// qcast/dcast at points where a transition to a quantized type is
// required. In addition, it is also common to have an identity qcast
// (where the operand and result type are not quantized) at all points where
// it is legal to use a quantized representation (but is not known to be

  // The quantized call op without the _quantization_method attribute is not
  // captured as a `QuantizationResult`.
  ASSERT_THAT(results.results(), IsEmpty());
}

TEST_F(QuantizationReportTest, InitializeWithModuleOpWithInvalidCalleeName) {
  // A quantized dot_general op but the callee function has an invalid name. It
  // is expected to start with `quantized_`.