sdfg::fp_math::QuantizeContraction Class Reference

Computes a floating-point matmul (MatMul or GEMM node) on integers, quantized with a scale per row of A and per column of B. More...

#include <sdfg/fp_math/transformations/quantize_contraction.h>

Inheritance diagram for sdfg::fp_math::QuantizeContraction:

Public Member Functions
	QuantizeContraction (data_flow::LibraryNode &node, int bits=8, Rescaling rescaling=Rescaling::Exponents)
virtual std::string	name () const override
	Get the name of this transformation.
virtual bool	can_be_applied (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) override
	Check if this transformation can be safely applied.
virtual void	apply (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager) override
	Apply this transformation to the SDFG.
virtual void	to_json (nlohmann::json &j) const override
	Serialize this transformation to JSON.
int	bits () const
Rescaling	rescaling () const
Public Member Functions inherited from sdfg::transformations::Transformation
virtual	~Transformation ()=default
virtual void	set_report (PassReportConsumer *report)
	Set the report consumer for transformation feedback.
virtual bool	try_apply (builder::StructuredSDFGBuilder &builder, analysis::AnalysisManager &analysis_manager)
	Attempt to apply the transformation if possible.

Static Public Member Functions
static QuantizeContraction	from_json (builder::StructuredSDFGBuilder &builder, const nlohmann::json &j)

Additional Inherited Members
Protected Attributes inherited from sdfg::transformations::Transformation
PassReportConsumer *	report_ = nullptr

Detailed Description

Computes a floating-point matmul (MatMul or GEMM node) on integers, quantized with a scale per row of A and per column of B.

With q = 2^(bits-1) - 1, s_A[i] = max_k |A[i, k]| and s_B[j] = max_k |B[k, j]|:

A_q[i, k] = rint(q * A[i, k] / s_A[i])          (Int8)
B_q[k, j] = rint(q * B[k, j] / s_B[j])          (Int8)
P         = A_q @ B_q                            (Int8 -> Int32 MatMul, exact)
Y[i, j]   = P[i, j] * (s_A[i] / q) * (s_B[j] / q)

The scales are computed at run time, so the result changes with every rounding of A and B: about 1% relative error for 8 bits. The Int8 matmul stays a library node for the target to map, e.g. onto integer matrix units.

Applies to a Float or Double MatMul, also one repeating over batch dimensions with scales per matrix, or a row-major GEMM (whose alpha and beta it applies to the result), with a constant K small enough for exact Int32 accumulation.

Constructor & Destructor Documentation

QuantizeContraction()

sdfg::fp_math::QuantizeContraction::QuantizeContraction ( data_flow::LibraryNode & node, int bits = 8, Rescaling rescaling = Rescaling::Exponents )

explicit

Member Function Documentation

apply()

virtual void sdfg::fp_math::QuantizeContraction::apply ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager )

overridevirtual

Apply this transformation to the SDFG.

Parameters

builder	The SDFG builder
analysis_manager	The analysis manager (may be invalidated)

Implements sdfg::transformations::Transformation.

bits()

int sdfg::fp_math::QuantizeContraction::bits ( ) const

inline

can_be_applied()

virtual bool sdfg::fp_math::QuantizeContraction::can_be_applied ( builder::StructuredSDFGBuilder & builder, analysis::AnalysisManager & analysis_manager )

overridevirtual

Check if this transformation can be safely applied.

Parameters

builder	The SDFG builder
analysis_manager	The analysis manager providing necessary analyses

Returns

true if all preconditions are met

Implements sdfg::transformations::Transformation.

from_json()

QuantizeContraction sdfg::fp_math::QuantizeContraction::from_json ( builder::StructuredSDFGBuilder & builder, const nlohmann::json & j )

static

name()

virtual std::string sdfg::fp_math::QuantizeContraction::name ( ) const

overridevirtual

Get the name of this transformation.

Returns

The transformation name (e.g., "LoopInterchange")

Implements sdfg::transformations::Transformation.

rescaling()

Rescaling sdfg::fp_math::QuantizeContraction::rescaling ( ) const

inline

to_json()

virtual void sdfg::fp_math::QuantizeContraction::to_json ( nlohmann::json & j ) const

overridevirtual

Serialize this transformation to JSON.

Parameters

j	JSON object to populate with transformation description

Exceptions

std::logic_error

if not implemented

Implements sdfg::transformations::Transformation.

---

The documentation for this class was generated from the following file:

• opt/include/sdfg/fp_math/transformations/quantize_contraction.h