Metadata.h

Go to the documentation of this file.

//===- Metadata.h - Top level types and metadata ----------------*- C++ -*-===//
//
// Part of the MLIR Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file contains top level types needed to construct constraint graphs,
// including context/allocator support and concrete metadata structs for
// different quantization schemes (which must be attached to anchor nodes).
//
//===----------------------------------------------------------------------===//

#ifndef MLIR_QUANTIZER_SUPPORT_METADATA_H
#define MLIR_QUANTIZER_SUPPORT_METADATA_H

#include <limits>

#include "mlir/Dialect/QuantOps/QuantTypes.h"
#include "mlir/IR/MLIRContext.h"
#include "mlir/Quantizer/Support/Rules.h"
#include "llvm/ADT/SmallBitVector.h"

namespace mlir {
namespace quantizer {

class SolverContext {
public:
  SolverContext(MLIRContext &mlirContext) : mlirContext(mlirContext) {}

  MLIRContext &getMlirContext() { return mlirContext; }

  llvm::BumpPtrAllocator &getAllocator() { return allocator; }

  // Optional path to write a debug DOT file for the CAG.
  StringRef getDebugCAGDotPath() const { return debugCAGDotPath; }
  void setDebugCAGDotPath(StringRef p) { debugCAGDotPath = p; }

private:
  MLIRContext &mlirContext;
  llvm::BumpPtrAllocator allocator;
  std::string debugCAGDotPath;
};

struct CandidateQuantizedType {
  // Note that scheme encodes more than just the target type: it also encodes
  // additional constraints.
  enum class Scheme {
    // Uses aggregate range information for all nodes in the cluster to
    // solve for uniform scale and zero point.
    UniformPerLayer,
    // Uses aggregate per-axis range information for all nodes in the cluster
    // to solve for per-axis uniform scale and zero point.
    UniformPerAxisFixedPoint,
    // Uses the |explicitScaleZeroPoint| to set the scale (and zero point = 0)
    // for the uniform type. This typically overrides all other constraints
    // and is used for wide accumulator types (i.e. i32 bias vectors).
    UniformExplicitFixedPointScale,
  };
  unsigned ordinal;
  quant::AnyQuantizedType quantizedType;
  Scheme scheme;
};

struct CAGUniformMetadata {
  static constexpr int SalienceDefault = 0;

  static constexpr int SalienceForced = 100;

  static constexpr int SalienceRequired = 200;

  ExpandingMinMaxFact requiredRange;

  llvm::SmallBitVector disabledCandidateTypes;

  quant::QuantizedType selectedType;

  DiscreteScaleZeroPointFact explicitScaleZeroPoint;

  void printSummary(raw_ostream &os) const;
};

} // end namespace quantizer
} // end namespace mlir

#endif // MLIR_QUANTIZER_SUPPORT_METADATA_H