#include <atomic>
#include <blas_quda.h>
#include <tune_quda.h>
#include <float_vector.h>
#include <color_spinor_field_order.h>
#include <launch_kernel.cuh>
#include <jitify_helper.cuh>
#include <kernels/reduce_core.cuh>
#include <generic_reduce.cuh>

Include dependency graph for reduce_quda.cu:

Classes
class	quda::blas::ReduceCuda< doubleN, ReduceType, FloatN, M, SpinorX, SpinorY, SpinorZ, SpinorW, SpinorV, Reducer >

Namespaces
	quda

	quda::blas

Functions
cudaStream_t *	quda::blas::getStream ()

void *	quda::blas::getDeviceReduceBuffer ()

void *	quda::blas::getMappedHostReduceBuffer ()

void *	quda::blas::getHostReduceBuffer ()

cudaEvent_t *	quda::blas::getReduceEvent ()

bool	quda::blas::getFastReduce ()

void	quda::blas::initFastReduce (int32_t words)

void	quda::blas::completeFastReduce (int32_t words)

void	quda::blas::initReduce ()

void	quda::blas::endReduce ()

template<typename doubleN , typename ReduceType , typename FloatN , int M, typename Arg >
doubleN	quda::blas::reduceLaunch (Arg &arg, const TuneParam &tp, const cudaStream_t &stream, Tunable &tunable)

template<typename doubleN , typename ReduceType , typename RegType , typename StoreType , typename zType , int M, template< typename ReducerType, typename Float, typename FloatN > class Reducer, int writeX, int writeY, int writeZ, int writeW, int writeV>
doubleN	quda::blas::nativeReduce (const double2 &a, const double2 &b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &v, int length)

template<typename doubleN , typename ReduceType , template< typename ReducerType, typename Float, typename FloatN > class Reducer, int writeX, int writeY, int writeZ, int writeW, int writeV, bool siteUnroll>
doubleN	quda::blas::uni_reduce (const double2 &a, const double2 &b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &v)

template<typename doubleN , typename ReduceType , template< typename ReducerType, typename Float, typename FloatN > class Reducer, int writeX, int writeY, int writeZ, int writeW, int writeV, bool siteUnroll>
doubleN	quda::blas::mixed_reduce (const double2 &a, const double2 &b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &v)

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

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

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

double	quda::blas::axpbyzNorm (double a, ColorSpinorField &x, double b, ColorSpinorField &y, ColorSpinorField &z)

double	quda::blas::axpyReDot (double a, ColorSpinorField &x, ColorSpinorField &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::cabxpyzAxNorm (double a, const Complex &b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

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

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

double3	quda::blas::cDotProductNormA (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 (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)

double4	quda::blas::quadrupleCGReduction (ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

double	quda::blas::quadrupleCG3InitNorm (double a, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &v)

double	quda::blas::quadrupleCG3UpdateNorm (double a, double b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z, ColorSpinorField &w, ColorSpinorField &v)

double	quda::blas::doubleCG3InitNorm (double a, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

double	quda::blas::doubleCG3UpdateNorm (double a, double b, ColorSpinorField &x, ColorSpinorField &y, ColorSpinorField &z)

Variables
static QudaSumFloat *	d_reduce =0

static QudaSumFloat *	h_reduce =0

static QudaSumFloat *	hd_reduce =0

static cudaEvent_t	reduceEnd

static bool	fast_reduce_enabled = false