EvalContextMoECache Class Reference

Thread-local evaluator context utilizing the MoE network. More...

#include <EvalContextMoECache.h>

Inheritance diagram for EvalContextMoECache:

Collaboration diagram for EvalContextMoECache:

Public Member Functions
	EvalContextMoECache (const std::string &weightsPath)
	EvalContextMoECache (std::shared_ptr< const EvalContextMoECacheSharedModel > sharedModel)
float	evaluate (const Board &board, int ply=0) override
	Evaluates the board from the perspective of the side to move.
WDLConverter::WDL	evaluateWDL (const Board &board, int ply=0) override
	Evaluates the board and returns Win/Draw/Loss probabilities.
void	setAggression (float aggression) override
	Sets the contempt or aggression factor for the evaluator.
void	setEvalScale (int base, int weight) override
	Sets the scaling parameters for the evaluation score.
void	setEvalNormalization (bool enable) override
	Enables or disables dynamic evaluation normalization.
void	setIncrementalRebuildInterval (int interval) override
	Sets the interval for forcing full feature rebuilds (to correct accumulation errors).
uint64_t	getFullRebuilds () const override
	Retrieves the number of full feature rebuilds performed (for profiling).
Public Member Functions inherited from IEvaluator
virtual	~IEvaluator ()=default
virtual float	evaluate (const ChessInput &input)
	Evaluates the position using pre-computed features.

Static Public Member Functions
static std::shared_ptr< const EvalContextMoECacheSharedModel >	loadSharedModel (const std::string &weightsPath)
	Loads the shared model weights from disk.

Private Member Functions
void	init_from_shared_model (std::shared_ptr< const EvalContextMoECacheSharedModel > shared)
	Initializes the thread-local context from the pre-loaded shared weights.
void	load_weights_into_model (const std::string &weightsPath)
	Loads neural network weights directly into the thread-local model (legacy usage).

Static Private Member Functions
static uint32_t	read_u32 (std::ifstream &in)
	Reads a 32-bit unsigned integer from the binary weights stream.
template<typename FloatContainer>
static void	read_floats (std::ifstream &in, FloatContainer &dst, size_t n)
	Reads a block of float values into a resizing container from the binary stream.
static void	read_floats_raw (std::ifstream &in, float *dst, size_t n)
	Reads a raw block of floats into a contiguous C-style array.
static ExpertPoolMode	pool_mode_from_code (int code)
	Converts an integer ID stored in the weights file to an ExpertPoolMode enum.
template<typename SrcContainer, typename DstContainer>
static void	transpose_copy (const SrcContainer &src, DstContainer &dst, int rows, int cols)
	Copies and transposes a matrix tensor (used to align weights for faster cache hits during inference).
static void	load_weights_into_target (const std::string &weightsPath, BenchConfig &cfg, SharedMoEWeights &weights)
	Loads neural network weights from a binary file into a shared weights container.

Private Attributes
BenchConfig	cfg_ {}
std::shared_ptr< const EvalContextMoECacheSharedModel >	sharedModel_ {}
std::array< MoEDoubleAccumulator, 2 >	models_ {}
FactorizedInput	scratchInput_ {}
std::array< FactorizedInput, 2 >	prevInputByStm_ {}
std::array< bool, 2 >	hasPrevByStm_ {{false, false}}
WDLConverter	wdlConverter_
int	evalScaleBase_ = 750
int	evalScaleWeight_ = 25
bool	enableEvalNormalization_ = true
int	rebuildEveryNEvals_ = 0
std::array< uint32_t, 2 >	evalsSinceFullByStm_ {{0, 0}}
std::atomic< uint64_t >	totalRebuilds_ {0}

Static Private Attributes
static constexpr uint32_t	kMagicWeights = 0x32454F4D
static constexpr uint32_t	kVersion = 1

Detailed Description

Thread-local evaluator context utilizing the MoE network.

Maintains incremental accumulator state and evaluates positions rapidly.

Constructor & Destructor Documentation

EvalContextMoECache() [1/2]

EvalContextMoECache::EvalContextMoECache ( const std::string & weightsPath )

explicit

Here is the call graph for this function:

EvalContextMoECache() [2/2]

EvalContextMoECache::EvalContextMoECache ( std::shared_ptr< const EvalContextMoECacheSharedModel > sharedModel )

explicit

Here is the call graph for this function:

Member Function Documentation

evaluate()

float EvalContextMoECache::evaluate ( const Board & board, int ply = 0 )

overridevirtual

Evaluates the board from the perspective of the side to move.

Parameters

board	The current chess board state.
ply	The current depth from the root of the search (used for scaling).

Returns

Evaluation score in centipawns (e.g., 150 = +1.5 pawns advantage).

Implements IEvaluator.

Here is the call graph for this function:

evaluateWDL()

WDLConverter::WDL EvalContextMoECache::evaluateWDL ( const Board & board, int ply = 0 )

overridevirtual

Evaluates the board and returns Win/Draw/Loss probabilities.

Parameters

board	The current chess board state.
ply	The current depth from the root of the search.

Returns

WDL probability distribution.

Reimplemented from IEvaluator.

Here is the call graph for this function:

Here is the caller graph for this function:

getFullRebuilds()

uint64_t EvalContextMoECache::getFullRebuilds ( ) const

inlineoverridevirtual

Retrieves the number of full feature rebuilds performed (for profiling).

Returns

The number of full rebuilds.

Reimplemented from IEvaluator.

init_from_shared_model()

void EvalContextMoECache::init_from_shared_model ( std::shared_ptr< const EvalContextMoECacheSharedModel > shared )

private

Initializes the thread-local context from the pre-loaded shared weights.

Here is the caller graph for this function:

load_weights_into_model()

void EvalContextMoECache::load_weights_into_model ( const std::string & weightsPath )

private

Loads neural network weights directly into the thread-local model (legacy usage).

Here is the call graph for this function:

Here is the caller graph for this function:

load_weights_into_target()

void EvalContextMoECache::load_weights_into_target ( const std::string & weightsPath, BenchConfig & cfg, SharedMoEWeights & weights )

staticprivate

Loads neural network weights from a binary file into a shared weights container.

Here is the call graph for this function:

Here is the caller graph for this function:

loadSharedModel()

std::shared_ptr< const EvalContextMoECacheSharedModel > EvalContextMoECache::loadSharedModel ( const std::string & weightsPath )

static

Loads the shared model weights from disk.

Here is the call graph for this function:

Here is the caller graph for this function:

pool_mode_from_code()

ExpertPoolMode EvalContextMoECache::pool_mode_from_code ( int code )

staticprivate

Converts an integer ID stored in the weights file to an ExpertPoolMode enum.

Here is the caller graph for this function:

read_floats()

template<typename FloatContainer>

void EvalContextMoECache::read_floats ( std::ifstream & in, FloatContainer & dst, size_t n )

inlinestaticprivate

Reads a block of float values into a resizing container from the binary stream.

Here is the caller graph for this function:

read_floats_raw()

void EvalContextMoECache::read_floats_raw ( std::ifstream & in, float * dst, size_t n )

staticprivate

Reads a raw block of floats into a contiguous C-style array.

read_u32()

uint32_t EvalContextMoECache::read_u32 ( std::ifstream & in )

staticprivate

Reads a 32-bit unsigned integer from the binary weights stream.

Here is the caller graph for this function:

setAggression()

void EvalContextMoECache::setAggression ( float aggression )

inlineoverridevirtual

Sets the contempt or aggression factor for the evaluator.

Parameters

aggression

The aggression factor (positive encourages risk, negative encourages draws).

Reimplemented from IEvaluator.

setEvalNormalization()

void EvalContextMoECache::setEvalNormalization ( bool enable )

inlineoverridevirtual

Enables or disables dynamic evaluation normalization.

Parameters

enable

True to enable normalization, false to use raw scores.

Reimplemented from IEvaluator.

setEvalScale()

void EvalContextMoECache::setEvalScale ( int base, int weight )

inlineoverridevirtual

Sets the scaling parameters for the evaluation score.

Parameters

base	The base scaling divisor.
weight	The material weight scaling factor.

Reimplemented from IEvaluator.

setIncrementalRebuildInterval()

void EvalContextMoECache::setIncrementalRebuildInterval ( int interval )

inlineoverridevirtual

Sets the interval for forcing full feature rebuilds (to correct accumulation errors).

Parameters

interval

The number of incremental updates before forcing a rebuild.

Reimplemented from IEvaluator.

transpose_copy()

template<typename SrcContainer, typename DstContainer>

void EvalContextMoECache::transpose_copy ( const SrcContainer & src, DstContainer & dst, int rows, int cols )

inlinestaticprivate

Copies and transposes a matrix tensor (used to align weights for faster cache hits during inference).

Here is the caller graph for this function:

Member Data Documentation

cfg_

BenchConfig EvalContextMoECache::cfg_ {}

private

enableEvalNormalization_

bool EvalContextMoECache::enableEvalNormalization_ = true

private

evalScaleBase_

int EvalContextMoECache::evalScaleBase_ = 750

private

evalScaleWeight_

int EvalContextMoECache::evalScaleWeight_ = 25

private

evalsSinceFullByStm_

std::array<uint32_t, 2> EvalContextMoECache::evalsSinceFullByStm_ {{0, 0}}

private

hasPrevByStm_

std::array<bool, 2> EvalContextMoECache::hasPrevByStm_ {{false, false}}

private

kMagicWeights

uint32_t EvalContextMoECache::kMagicWeights = 0x32454F4D

staticconstexprprivate

kVersion

uint32_t EvalContextMoECache::kVersion = 1

staticconstexprprivate

models_

std::array<MoEDoubleAccumulator, 2> EvalContextMoECache::models_ {}

private

prevInputByStm_

std::array<FactorizedInput, 2> EvalContextMoECache::prevInputByStm_ {}

private

rebuildEveryNEvals_

int EvalContextMoECache::rebuildEveryNEvals_ = 0

private

scratchInput_

FactorizedInput EvalContextMoECache::scratchInput_ {}

private

sharedModel_

std::shared_ptr<const EvalContextMoECacheSharedModel> EvalContextMoECache::sharedModel_ {}

private

totalRebuilds_

std::atomic<uint64_t> EvalContextMoECache::totalRebuilds_ {0}

private

wdlConverter_

WDLConverter EvalContextMoECache::wdlConverter_

private

---

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

• engine/src/eval/EvalContextMoECache.h
• engine/src/eval/EvalContextMoECache.cpp