quda_internal.h File Reference

#include <quda_cuda_api.h>
#include <sys/time.h>
#include <string>
#include <complex>
#include <enum_quda.h>
#include <quda.h>
#include <util_quda.h>
#include <malloc_quda.h>
#include <object.h>
#include <vector>

Include dependency graph for quda_internal.h:

Go to the source code of this file.

Classes
struct	ParityHw
struct	FullHw
struct	QUDA_DiracField
struct	quda::Timer
class	quda::TimeProfile

Namespaces
	quda

Macros
#define	MAX_SHORT   32767.0f
#define	TEX_ALIGN_REQ   (512*2)
#define	ALIGNMENT_ADJUST(n)   ( (n+TEX_ALIGN_REQ-1)/TEX_ALIGN_REQ*TEX_ALIGN_REQ)
#define	MAX_MULTI_BLAS_N   1
#define	PUSH_RANGE(name, cid)
#define	POP_RANGE
#define	TPSTART(idx)   Start_(__func__, __FILE__, __LINE__, idx)
#define	TPSTOP(idx)   Stop_(__func__, __FILE__, __LINE__, idx)
#define	TPRESET()   Reset_(__func__, __FILE__, __LINE__)

Typedefs
typedef void *	ParityGauge

Enumerations
enum	quda::QudaProfileType {   quda::QUDA_PROFILE_H2D, quda::QUDA_PROFILE_D2H, quda::QUDA_PROFILE_INIT, quda::QUDA_PROFILE_PREAMBLE,   quda::QUDA_PROFILE_COMPUTE, quda::QUDA_PROFILE_COMMS, quda::QUDA_PROFILE_EPILOGUE, quda::QUDA_PROFILE_FREE,   quda::QUDA_PROFILE_IO, quda::QUDA_PROFILE_LOWER_LEVEL, quda::QUDA_PROFILE_PACK_KERNEL, quda::QUDA_PROFILE_DSLASH_KERNEL,   quda::QUDA_PROFILE_GATHER, quda::QUDA_PROFILE_SCATTER, quda::QUDA_PROFILE_LAUNCH_KERNEL, quda::QUDA_PROFILE_EVENT_RECORD,   quda::QUDA_PROFILE_EVENT_QUERY, quda::QUDA_PROFILE_STREAM_WAIT_EVENT, quda::QUDA_PROFILE_FUNC_SET_ATTRIBUTE, quda::QUDA_PROFILE_EVENT_SYNCHRONIZE,   quda::QUDA_PROFILE_STREAM_SYNCHRONIZE, quda::QUDA_PROFILE_DEVICE_SYNCHRONIZE, quda::QUDA_PROFILE_MEMCPY_D2D_ASYNC, quda::QUDA_PROFILE_MEMCPY_D2H_ASYNC,   quda::QUDA_PROFILE_MEMCPY2D_D2H_ASYNC, quda::QUDA_PROFILE_MEMCPY_H2D_ASYNC, quda::QUDA_PROFILE_COMMS_START, quda::QUDA_PROFILE_COMMS_QUERY,   quda::QUDA_PROFILE_CONSTANT, quda::QUDA_PROFILE_TOTAL, quda::QUDA_PROFILE_COUNT }

Functions
bool	quda::canReuseResidentGauge (QudaInvertParam *inv_param)

Variables
cudaDeviceProp	deviceProp
cudaStream_t *	streams
const int	quda::Nstream = 9

Macro Definition Documentation

ALIGNMENT_ADJUST

#define ALIGNMENT_ADJUST

(

n

)

( (n+TEX_ALIGN_REQ-1)/TEX_ALIGN_REQ*TEX_ALIGN_REQ)

Definition at line 32 of file quda_internal.h.

Referenced by quda::ColorSpinorField::create(), quda::cpuColorSpinorField::create(), quda::GaugeField::createGhostZone(), quda::ColorSpinorField::createGhostZone(), quda::GaugeField::GaugeField(), and quda::ColorSpinorField::reset().

MAX_MULTI_BLAS_N

#define MAX_MULTI_BLAS_N   1

Definition at line 49 of file quda_internal.h.

Referenced by quda::blas::axpyBzpcx(), quda::blas::caxpy_recurse(), quda::blas::caxpyBxpz(), quda::blas::caxpyz_recurse(), quda::blas::initReduce(), multiblasCuda(), multiReduceCuda(), quda::blas::multicaxpy_< NXZ, Float2, FloatN >::operator()(), quda::blas::multicaxpyz_< NXZ, Float2, FloatN >::operator()(), and quda::blas::multi_caxpyBxpz_< NXZ, Float2, FloatN >::operator()().

MAX_SHORT

#define MAX_SHORT   32767.0f

Definition at line 29 of file quda_internal.h.

Referenced by quda::copy(), and store_norm().

POP_RANGE

#define POP_RANGE

Definition at line 229 of file quda_internal.h.

Referenced by quda::TimeProfile::Stop_().

PUSH_RANGE

#define PUSH_RANGE	(		name,
			cid
	)

Definition at line 228 of file quda_internal.h.

Referenced by quda::TimeProfile::Start_().

TEX_ALIGN_REQ

#define TEX_ALIGN_REQ   (512*2)

Definition at line 31 of file quda_internal.h.

TPRESET

#define TPRESET

(

)

Reset_(__func__, __FILE__, __LINE__)

Definition at line 322 of file quda_internal.h.

TPSTART

#define TPSTART

(

idx

)

Start_(__func__, __FILE__, __LINE__, idx)

Definition at line 320 of file quda_internal.h.

TPSTOP

#define TPSTOP

(

idx

)

Stop_(__func__, __FILE__, __LINE__, idx)

Definition at line 321 of file quda_internal.h.

Typedef Documentation

ParityGauge

typedef void* ParityGauge

Definition at line 62 of file quda_internal.h.

Variable Documentation

deviceProp

cudaDeviceProp deviceProp

Definition at line 152 of file interface_quda.cpp.

Referenced by quda::Tunable::advanceBlockDim(), quda::ShiftColorSpinorField< Output, Input >::advanceBlockDim(), quda::TunableVectorY::advanceBlockDim(), quda::TunableVectorYZ::advanceBlockDim(), MultiReduceCuda< NXZ, doubleN, ReduceType, FloatN, M, SpinorX, SpinorY, SpinorZ, SpinorW, Reducer >::advanceGridDim(), quda::Tunable::advanceSharedBytes(), quda::Tunable::blockMin(), quda::Tunable::blockStep(), quda::Tunable::checkLaunchParam(), initQudaDevice(), quda::blas::initReduce(), quda::ShiftColorSpinorField< Output, Input >::initTuneParam(), quda::Tunable::initTuneParam(), quda::Tunable::maxBlockSize(), quda::WuppertalSmearing< Float, Ns, Nc, Arg >::maxBlockSize(), quda::Laplace< Float, nDim, nColor, Arg >::maxBlockSize(), MultiReduceCuda< NXZ, doubleN, ReduceType, FloatN, M, SpinorX, SpinorY, SpinorZ, SpinorW, Reducer >::maxBlockSize(), quda::TunableLocalParity::maxBlockSize(), quda::Tunable::maxBlocksPerSM(), quda::Tunable::maxGridSize(), multiReduceLaunch(), and reduceLaunch().

streams

cudaStream_t* streams

Definition at line 153 of file interface_quda.cpp.

Referenced by quda::ApplyClover(), quda::ApplyGamma(), quda::ApplyTwistClover(), quda::ApplyTwistGamma(), anonymous_namespace{dslash_policy.cuh}::commsComplete(), anonymous_namespace{dslash_policy.cuh}::completeDslash(), quda::contractCuda(), anonymous_namespace{dslash_policy.cuh}::DslashPolicyTune::DslashPolicyTune(), endQuda(), quda::cudaGaugeField::exchangeExtendedGhost(), quda::exchangeExtendedGhost(), quda::cudaGaugeField::exchangeGhost(), quda::cudaColorSpinorField::exchangeGhost(), quda::blas::init(), initQudaMemory(), quda::cudaGaugeField::injectGhost(), anonymous_namespace{dslash_policy.cuh}::issueGather(), anonymous_namespace{dslash_policy.cuh}::issuePack(), anonymous_namespace{dslash_policy.cuh}::DslashBasic::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashPthreads::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedExterior::operator()(), quda::XSD::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashGDR::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedGDR::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashGDRRecv::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedGDRRecv::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashZeroCopyPack::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedZeroCopyPack::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashZeroCopyPackGDRRecv::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedZeroCopyPackGDRRecv::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashZeroCopy::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashFusedZeroCopy::operator()(), anonymous_namespace{dslash_policy.cuh}::DslashNC::operator()(), mixed::reduceCuda(), quda::shiftColorSpinorField(), and quda::cudaColorSpinorField::zero().