#include <blas_quda.h>
#include <tune_quda.h>
#include <float_vector.h>
#include <color_spinor_field_order.h>
#include <cub_helper.cuh>
#include <algorithm>
#include <texture.h>
#include <reduce_core.cuh>
#include <reduce_core.h>
#include <reduce_mixed_core.h>

Include dependency graph for reduce_quda.cu:

Classes
struct	ScalarType< typename >

struct	ScalarType< double >

struct	ScalarType< double2 >

struct	ScalarType< double3 >

struct	ScalarType< double4 >

struct	Vec2Type< typename >

struct	Vec2Type< double >

struct	quda::blas::ReduceFunctor< ReduceType, Float2, FloatN >

struct	quda::blas::Norm1< ReduceType, Float2, FloatN >

struct	quda::blas::Norm2< ReduceType, Float2, FloatN >

struct	quda::blas::Dot< NXZ, ReduceType, Float2, FloatN >

struct	quda::blas::DotNormA< ReduceType, Float2, FloatN >

struct	quda::blas::axpyNorm2< ReduceType, Float2, FloatN >

struct	quda::blas::AxpyReDot< ReduceType, Float2, FloatN >

struct	quda::blas::xmyNorm2< ReduceType, Float2, FloatN >

struct	quda::blas::caxpyNorm2< ReduceType, Float2, FloatN >

struct	quda::blas::caxpyxmaznormx< ReduceType, Float2, FloatN >

struct	quda::blas::cabxpyaxnorm< ReduceType, Float2, FloatN >

struct	quda::blas::Cdot< NXZ, ReduceType, Float2, FloatN >

struct	quda::blas::xpaycdotzy< ReduceType, Float2, FloatN >

struct	quda::blas::caxpydotzy< ReduceType, Float2, FloatN >

struct	quda::blas::CdotNormA< ReduceType, Float2, FloatN >

struct	quda::blas::CdotNormB< ReduceType, Float2, FloatN >

struct	quda::blas::caxpbypzYmbwcDotProductUYNormY_< ReduceType, Float2, FloatN >

struct	quda::blas::axpyCGNorm2< ReduceType, Float2, FloatN >

struct	quda::blas::HeavyQuarkResidualNorm_< ReduceType, Float2, FloatN >

struct	quda::blas::xpyHeavyQuarkResidualNorm_< ReduceType, Float2, FloatN >

struct	quda::blas::tripleCGReduction_< ReduceType, Float2, FloatN >

Namespaces
	quda

	quda::blas

	quda::blas::reduce

Macros
#define	QudaSumFloat double

#define	QudaSumFloat2 double2

#define	QudaSumFloat3 double3

#define	QudaSumFloat4 double4

Functions
void	checkSpinor (const ColorSpinorField &a, const ColorSpinorField &b)

void	checkLength (const ColorSpinorField &a, ColorSpinorField &b)

cudaStream_t *	quda::blas::getStream ()

void *	quda::blas::getDeviceReduceBuffer ()

void *	quda::blas::getMappedHostReduceBuffer ()

void *	quda::blas::getHostReduceBuffer ()

cudaEvent_t *	quda::blas::getReduceEvent ()

void	quda::blas::initReduce ()

void	quda::blas::endReduce ()

template<typename ReduceType >
__device__ __host__ ReduceType	quda::blas::norm1_ (const double2 &a)

template<typename ReduceType >
__device__ __host__ ReduceType	quda::blas::norm1_ (const float2 &a)

template<typename ReduceType >
__device__ __host__ ReduceType	quda::blas::norm1_ (const float4 &a)

double	quda::blas::norm1 (const ColorSpinorField &b)

template<typename ReduceType >
__device__ __host__ void	quda::blas::norm2_ (ReduceType &sum, const double2 &a)

template<typename ReduceType >
__device__ __host__ void	quda::blas::norm2_ (ReduceType &sum, const float2 &a)

template<typename ReduceType >
__device__ __host__ void	quda::blas::norm2_ (ReduceType &sum, const float4 &a)

double	quda::blas::norm2 (const ColorSpinorField &a)

template<typename ReduceType >
__device__ __host__ void	quda::blas::dot_ (ReduceType &sum, const double2 &a, const double2 &b)

template<typename ReduceType >
__device__ __host__ void	quda::blas::dot_ (ReduceType &sum, const float2 &a, const float2 &b)

template<typename ReduceType >
__device__ __host__ void	quda::blas::dot_ (ReduceType &sum, const float4 &a, const float4 &b)

double	quda::blas::reDotProduct (ColorSpinorField &x, ColorSpinorField &y)

template<typename ReduceType , typename InputType >
__device__ __host__ ReduceType	quda::blas::dotNormA_ (const InputType &a, const InputType &b)

double2	quda::blas::reDotProductNormA (ColorSpinorField &a, ColorSpinorField &b)

double	quda::blas::axpyNorm (const double &a, ColorSpinorField &x, ColorSpinorField &y)

double	quda::blas::axpyReDot (const double &a, ColorSpinorField &x, ColorSpinorField &y)

double	quda::blas::xmyNorm (ColorSpinorField &x, ColorSpinorField &y)

__device__ __host__ void	quda::blas::Caxpy_ (const double2 &a, const double2 &x, double2 &y)

__device__ __host__ void	quda::blas::Caxpy_ (const float2 &a, const float2 &x, float2 &y)

__device__ __host__ void	quda::blas::Caxpy_ (const float2 &a, const float4 &x, float4 &y)

double	quda::blas::caxpyNorm (const Complex &a, ColorSpinorField &x, ColorSpinorField &y)

double	quda::blas::caxpyXmazNormX (const Complex &a, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

double	quda::blas::cabxpyAxNorm (const double &a, const Complex &b, ColorSpinorField &x, ColorSpinorField &y)

template<typename ReduceType >
__device__ __host__ void	quda::blas::cdot_ (ReduceType &sum, const double2 &a, const double2 &b)

template<typename ReduceType >
__device__ __host__ void	quda::blas::cdot_ (ReduceType &sum, const float2 &a, const float2 &b)

template<typename ReduceType >
__device__ __host__ void	quda::blas::cdot_ (ReduceType &sum, const float4 &a, const float4 &b)

Complex	quda::blas::cDotProduct (ColorSpinorField &, ColorSpinorField &)

Complex	quda::blas::xpaycDotzy (ColorSpinorField &x, const double &a, ColorSpinorField &y, ColorSpinorField &z)

Complex	quda::blas::caxpyDotzy (const Complex &a, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

template<typename ReduceType , typename InputType >
__device__ __host__ void	quda::blas::cdotNormA_ (ReduceType &sum, const InputType &a, const InputType &b)

double3	quda::blas::cDotProductNormA (ColorSpinorField &a, ColorSpinorField &b)

template<typename ReduceType , typename InputType >
__device__ __host__ void	quda::blas::cdotNormB_ (ReduceType &sum, const InputType &a, const InputType &b)

double3	quda::blas::cDotProductNormB (ColorSpinorField &a, ColorSpinorField &b)

double3	quda::blas::caxpbypzYmbwcDotProductUYNormY (const Complex &a, ColorSpinorField &x, const Complex &b, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &u)

Complex	quda::blas::axpyCGNorm (const double &a, ColorSpinorField &x, ColorSpinorField &y)

double3	quda::blas::HeavyQuarkResidualNorm (ColorSpinorField &x, ColorSpinorField &r)

double3	quda::blas::xpyHeavyQuarkResidualNorm (ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &r)

double3	quda::blas::tripleCGReduction (ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

Variables
struct {
const char * vol_str

const char * aux_str

char aux_tmp [quda::TuneKey::aux_n]

}	blasStrings

static QudaSumFloat *	d_reduce =0

static QudaSumFloat *	h_reduce =0

static QudaSumFloat *	hd_reduce =0

static cudaEvent_t	reduceEnd

Macro Definition Documentation

◆ QudaSumFloat

#define QudaSumFloat double

Definition at line 33 of file reduce_quda.cu.

Referenced by quda::blas::initReduce().

◆ QudaSumFloat2

#define QudaSumFloat2 double2

Definition at line 34 of file reduce_quda.cu.

◆ QudaSumFloat3

#define QudaSumFloat3 double3

Definition at line 35 of file reduce_quda.cu.

◆ QudaSumFloat4

#define QudaSumFloat4 double4

Definition at line 36 of file reduce_quda.cu.

Function Documentation

◆ checkLength()

void checkLength	(	const ColorSpinorField &	a,
		ColorSpinorField &	b
	)

Definition at line 49 of file reduce_quda.cu.

References a, b, and errorQuda.

Referenced by blasCuda(), and reduceCuda().

Here is the caller graph for this function:

◆ checkSpinor()

void checkSpinor	(	const ColorSpinorField &	a,
		const ColorSpinorField &	b
	)

Definition at line 40 of file reduce_quda.cu.

References a, b, and errorQuda.

Referenced by multiReduceCuda().

Here is the caller graph for this function:

Variable Documentation

◆ aux_str

const char* aux_str

Definition at line 58 of file reduce_quda.cu.

◆ aux_tmp

char aux_tmp[quda::TuneKey::aux_n]

Definition at line 59 of file reduce_quda.cu.

◆ blasStrings

struct { ... } blasStrings

◆ d_reduce

QudaSumFloat* d_reduce =0

static

Definition at line 63 of file reduce_quda.cu.

Referenced by quda::blas::endReduce(), quda::blas::getDeviceReduceBuffer(), and quda::blas::initReduce().

◆ h_reduce

QudaSumFloat* h_reduce =0

static

Definition at line 64 of file reduce_quda.cu.

Referenced by quda::blas::endReduce(), quda::blas::getHostReduceBuffer(), quda::blas::initReduce(), and reduceLaunch().

◆ hd_reduce

QudaSumFloat* hd_reduce =0

static

Definition at line 65 of file reduce_quda.cu.

Referenced by quda::blas::endReduce(), quda::blas::getMappedHostReduceBuffer(), quda::blas::initReduce(), and reduceLaunch().

◆ reduceEnd

cudaEvent_t reduceEnd

static

Definition at line 66 of file reduce_quda.cu.

Referenced by quda::blas::endReduce(), quda::blas::getReduceEvent(), quda::blas::initReduce(), and reduceLaunch().

◆ vol_str

const char* vol_str

Definition at line 57 of file reduce_quda.cu.

Classes

Namespaces

Macros

Functions

Variables

Macro Definition Documentation

◆ QudaSumFloat

◆ QudaSumFloat2

◆ QudaSumFloat3

◆ QudaSumFloat4

Function Documentation

◆ checkLength()

◆ checkSpinor()

Variable Documentation

◆ aux_str

◆ aux_tmp

◆ blasStrings

◆ d_reduce

◆ h_reduce

◆ hd_reduce

◆ reduceEnd

◆ vol_str