QuantUtils.cpp File Reference

#include "mlir/Dialect/Tosa/Utils/QuantUtils.h"

Go to the source code of this file.

Macros
#define	GET_UQTYPE(inputType)    (llvm::dyn_cast<quant::UniformQuantizedType>((inputType).getElementType()))
#define	GET_QTYPE(inputType)    (llvm::dyn_cast<quant::QuantizedType>((inputType).getElementType()))

Functions
static void	computeMultiplierAndShiftTosaScale16 (double scale, int32_t &multiplier, int32_t &shift)
	From a scale value, generates multiplier and shift values where mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that multiplier = mantissa*2^shift for 16-bit scaling. More...
static void	computeMultiplierAndShiftTosaScale32 (double scale, int32_t &multiplier, int32_t &shift)
	From a scale value, generates multiplier and shift values where mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that multiplier = mantissa*2^shift for 32-bit scaling. More...
static std::optional< std::pair< std::int64_t, std::int64_t > >	getConvZeroPoints (Value input, Value weight)

Macro Definition Documentation

GET_QTYPE

#define GET_QTYPE

(

inputType

)

(llvm::dyn_cast<quant::QuantizedType>((inputType).getElementType()))

Definition at line 119 of file QuantUtils.cpp.

GET_UQTYPE

#define GET_UQTYPE

(

inputType

)

(llvm::dyn_cast<quant::UniformQuantizedType>((inputType).getElementType()))

Definition at line 117 of file QuantUtils.cpp.

Function Documentation

computeMultiplierAndShiftTosaScale16()

static void computeMultiplierAndShiftTosaScale16 ( double  scale, int32_t &  multiplier, int32_t &  shift  )

static

From a scale value, generates multiplier and shift values where mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that multiplier = mantissa*2^shift for 16-bit scaling.

Definition at line 22 of file QuantUtils.cpp.

References frexp(), max(), and mlir::presburger::round().

Referenced by mlir::tosa::computeMultiplierAndShift().

computeMultiplierAndShiftTosaScale32()

static void computeMultiplierAndShiftTosaScale32 ( double  scale, int32_t &  multiplier, int32_t &  shift  )

static

From a scale value, generates multiplier and shift values where mantissa is in [-1.0,-0.5] or [0.5, 1.0] such that multiplier = mantissa*2^shift for 32-bit scaling.

Definition at line 60 of file QuantUtils.cpp.

References frexp(), max(), and mlir::presburger::round().

Referenced by mlir::tosa::computeMultiplierAndShift().

getConvZeroPoints()

static std::optional<std::pair<std::int64_t, std::int64_t> > getConvZeroPoints ( Value  input, Value  weight  )

static

Definition at line 123 of file QuantUtils.cpp.

References GET_UQTYPE, and mlir::Value::getType().

Referenced by mlir::tosa::buildConvOpQuantizationAttr(), and mlir::tosa::createZPsAsConst().