limitations under the License.
==============================================================================*/
//
// This file defines support utilities for interoperating with FakeQuant* based
// QAT (Quantized Aware Training) computations, as implemented by TFLite. Note
// that FakeQuant* operators mix multiple concerns specific to how TFLite
// originally implemented quantization. As such, utilities here enforce

//              tfl.dequantize
//                   |
//
//
// Warns if the (most likely unwanted, currently not quite correctly handled)
// case of back-to-back tf.FakeQuant occurs
//
//             tf.FakeQuant*
//                   |
//             tf.FakeQuant*
//
template <typename TFFakeQuantOp, bool PerAxis, class FetchMinMax>
class InsertTFLQuantOpsAfterTFFakeQuantOp {
 public:

        TF::FakeQuantWithMinMaxArgsOp, /*PerAxis=*/false,
        FetchMinMaxAttrs<TF::FakeQuantWithMinMaxArgsOp>>;

// Removes the wrapper of the tf.FakeQuant* ops and creates the tfl.quantize
// and tfl.dequantize pairs before tf.FakeQuant* being foled.
LogicalResult ConvertFakeQuantOps(func::FuncOp func, MLIRContext* ctx,
                                  bool use_fake_quant_num_bits) {
  OpBuilder builder(func);

// Removes the wrapper of the tf.FakeQuant* ops and creates the quant.qcast
// and quant.dcast pairs before tf.FakeQuant* ops are being foled.
LogicalResult ConvertFakeQuantOps(func::FuncOp func, MLIRContext* ctx,
                                  bool use_fake_quant_num_bits) {
  OpBuilder builder(func);

  // Insert the quant.qcast/quant.dcast ops in place of the tf.FakeQuant* ops to
  // preserve the quantization parameters.

//                   |                          |
//
// Warns if the (most likely unwanted, currently not quite correctly handled)
// case of back-to-back tf.FakeQuant occurs
//
//             tf.FakeQuant*
//                   |
//             tf.FakeQuant*
//
template <typename TFFakeQuantOp, bool PerAxis, class FetchMinMax>
class ConvertFakeQuantOpToQuantOps {
 public:

  if (failed(
          ConvertFakeQuantOps(func, ctx, /*use_fake_quant_num_bits=*/false))) {
    func.emitError() << "quant-convert-fake-quant-to-qdq pass failed.";
    signalPassFailure();
  }

  // For removing dead FakeQuant* ops
  RewritePatternSet patterns(ctx);
  if (failed(applyPatternsAndFoldGreedily(func, std::move(patterns)))) {
    signalPassFailure();
  }
}

}  // namespace

          diag << "Skipping reordering between FakeQuant and "
               << (*target_ops.begin())->getName()
               << ", since there are other ops using the FakeQuant result.";
        });
      }
    }
  }
  return ::mlir::success();
}

// Reorder the FakeQuant operation for specific ops (ReorderOp).
// The transformation pattern looks like below:
//

namespace mlir {
namespace func {
class FuncOp;
}  // namespace func

namespace quantfork {

/// Creates a pass that converts quantization simulation operations (i.e.
/// FakeQuant and those like it) to casts into/out of supported QuantizedTypes.
std::unique_ptr<OperationPass<func::FuncOp>> createConvertSimulatedQuantPass();

/// Creates a pass that converts constants followed by a qbarrier to a

  Type storageType;
  int64_t qmin;
  int64_t qmax;
  if (getDefaultStorageParams(numBits, narrowRange, isSigned, ctx, storageType,
                              qmin, qmax)) {
    return (emitError(loc, "unsupported FakeQuant number of bits: ") << numBits,
            nullptr);
  }

  // Special case where min/max is close enough. The tensor contents are all
  // 0.0s, so the scale is set to 1.0 and the tensor can be quantized to zero

// Checks if a value can be symetrically quantized.
def CanBeSymmetricallyQuantized : Constraint<CPred<"CanBeSymmetricallyQuantized($0)">>;

// Multiplies the value followed by a FakeQuant op and adjusts its params.
def MultiplyFakeQuantValue : NativeCodeCall<
  "MultiplyFakeQuantValue($_builder, $_loc, $0...)">;

// Convert AddV2Op following an AffineOp to BiasAddOp.