llvm/mlir/lib/Dialect/Quant/Utils/FakeQuantSupport.cpp

//===- FakeQuantSupport.cpp - Support utilities for FakeQuant ops ---------===//
//
// Part of the LLVM 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
//
//===----------------------------------------------------------------------===//

#include "mlir/Dialect/Quant/IR/QuantTypes.h"
#include "mlir/Dialect/Quant/Utils/FakeQuantSupport.h"

usingnamespacemlir;
usingnamespacemlir::quant;

static bool getDefaultStorageParams(unsigned numBits, bool narrowRange,
                                    bool isSigned, MLIRContext *ctx,
                                    Type &storageType, int64_t &qmin,
                                    int64_t &qmax) { … }

// This is a specific implementation of nudging:
// If 0.0 < rmin < rmax or rmin < rmax < 0.0, the range will be shifted
// to include 0.0, but the range width size (rmax-rmin) isn't changed. The zero
// point is derived from the shifted range, and the scale isn't changed. As
// a consequence some values, which are supposed in the original [rmin, rmax]
// range will be outside the shifted range and be clamped during quantization.
// TODO: we should nudge the scale as well, but that requires the
// fake quant op used in the training to use the nudged scale as well.
static void getNudgedScaleAndZeroPoint(int64_t qmin, int64_t qmax, double rmin,
                                       double rmax, double &scale,
                                       int64_t &nudgedZeroPoint) { … }

UniformQuantizedType
mlir::quant::fakeQuantAttrsToType(Location loc, unsigned numBits, double rmin,
                                  double rmax, bool narrowRange,
                                  Type expressedType, bool isSigned) { … }

UniformQuantizedPerAxisType mlir::quant::fakeQuantAttrsToType(
    Location loc, unsigned numBits, int32_t quantizedDimension,
    ArrayRef<double> rmins, ArrayRef<double> rmaxs, bool narrowRange,
    Type expressedType, bool isSigned) { … }