GPU Class Reference

#include <gpu.h>

Inheritance diagram for GPU:

Collaboration diagram for GPU:

Public Member Functions
	GPU (int device=0, int stream=0)
	~GPU ()
int	device () const
void *	stream () const
int	streamId () const
size_t	lastFreeBytes () const
size_t	totalBytes () const
size_t	lastUsedBytes () const
void	updateMemoryInfo ()
void	peekLastError () const
void	set ()
void	synchAll () const
void	synch () const
std::string	getUUID () const
bool	isSet () const
void	lockMemory (const void *h_mem, size_t bytes) override
void	unlockMemory (const void *h_mem) override
bool	isMemoryLocked (const void *h_mem) override
bool	isGpuPointer (const void *)
int	getCudaVersion ()
Public Member Functions inherited from HW
	HW (unsigned parallelUnits)
virtual	~HW ()
unsigned	noOfParallUnits () const

Static Public Member Functions
static void	pinMemory (const void *h_mem, size_t bytes, unsigned int flags=0)
static void	unpinMemory (const void *h_mem)
static bool	isMemoryPinned (const void *h_mem)
static void	setDevice (int device)
static int	getDeviceCount ()

Additional Inherited Members
Protected Attributes inherited from HW
unsigned	m_parallUnits
size_t	m_totalBytes
size_t	m_lastFreeBytes
std::string	m_uuid

Detailed Description

Definition at line 36 of file gpu.h.

Constructor & Destructor Documentation

GPU()

GPU::GPU ( int  device = 0, int  stream = 0  )

inlineexplicit

Definition at line 38 of file gpu.h.

                                                :
        HW(1),
        m_device(device),
        m_streamId(stream), m_stream(nullptr),
        m_isSet(false) {};

~GPU()

GPU::~GPU

(

)

Definition at line 32 of file gpu.cpp.

          {
    if (m_isSet) {
        synch();
        auto s = (cudaStream_t*)m_stream;
        gpuErrchk(cudaStreamDestroy(*s));
        delete (cudaStream_t*)m_stream;
        m_stream = nullptr;
        m_uuid = std::string();
    }
    m_isSet = false;
}

Member Function Documentation

device()

int GPU::device ( ) const

inline

Definition at line 46 of file gpu.h.

                              {
        return m_device;
    }

getCudaVersion()

int GPU::getCudaVersion

(

)

Definition at line 44 of file gpu.cpp.

                        {
    int version = 0;
    gpuErrchk(cudaRuntimeGetVersion(&version));
    return version;
}

getDeviceCount()

int GPU::getDeviceCount ( )

static

Definition at line 118 of file gpu.cpp.

                        {
    int deviceCount = 0;
    gpuErrchk(cudaGetDeviceCount(&deviceCount));
    return deviceCount;
}

getUUID()

std::string GPU::getUUID ( ) const

inlinevirtual

Reimplemented from HW.

Definition at line 90 of file gpu.h.

                                     {
        check();
        return HW::getUUID();
    }

isGpuPointer()

bool GPU::isGpuPointer

(

const void *

p

)

Definition at line 153 of file gpu.cpp.

                                    {
    cudaPointerAttributes attr;
    if (cudaPointerGetAttributes(&attr, p) == cudaErrorInvalidValue) {
        cudaGetLastError(); // clear out the previous API error
        return false;
    }
#if defined(CUDART_VERSION) && CUDART_VERSION >= 10000
    return (cudaMemoryTypeDevice == attr.type) || (cudaMemoryTypeManaged == attr.type);
#else
    return cudaMemoryTypeDevice == attr.memoryType;
#endif
}

isMemoryLocked()

bool GPU::isMemoryLocked ( const void *  h_mem )

inlineoverridevirtual

Implements HW.

Definition at line 112 of file gpu.h.

                                                    {
        return GPU::isMemoryPinned(h_mem);
    }

isMemoryPinned()

bool GPU::isMemoryPinned ( const void *  h_mem )

static

Definition at line 140 of file gpu.cpp.

                                          {
    cudaPointerAttributes attr;
    if (cudaPointerGetAttributes(&attr, h_mem) != cudaSuccess) {
        cudaGetLastError(); // clear out the previous API error
        return false;
    }
    #if defined(CUDART_VERSION) && CUDART_VERSION >= 10000
        return (cudaMemoryTypeHost == attr.type) || (cudaMemoryTypeManaged == attr.type);
    #else
        return cudaMemoryTypeHost == attr.memoryType;
    #endif
}

isSet()

bool GPU::isSet ( ) const

inline

Definition at line 95 of file gpu.h.

                              {
        return m_isSet;
    }

lastFreeBytes()

size_t GPU::lastFreeBytes ( ) const

inlinevirtual

Reimplemented from HW.

Definition at line 59 of file gpu.h.

                                        {
        check();
        return HW::lastFreeBytes();
    }

lastUsedBytes()

size_t GPU::lastUsedBytes ( ) const

inlinevirtual

Reimplemented from HW.

Definition at line 69 of file gpu.h.

                                        {
        check();
        return HW::lastUsedBytes();
    }

lockMemory()

void GPU::lockMemory ( const void *  h_mem, size_t  bytes  )

inlineoverridevirtual

Implements HW.

Definition at line 104 of file gpu.h.

                                                              {
        GPU::pinMemory(h_mem, bytes, 0);
    }

peekLastError()

void GPU::peekLastError

(

)

const

Definition at line 87 of file gpu.cpp.

                              {
    check();
    gpuErrchk(cudaPeekAtLastError());
}

pinMemory()

void GPU::pinMemory ( const void *  h_mem, size_t  bytes, unsigned int  flags = 0  )

static

Definition at line 92 of file gpu.cpp.

                            {
    if (isMemoryPinned(h_mem)
            && (isMemoryPinned((char*)h_mem + bytes - 1))) {
        return;
    }
    assert(0 == cudaHostRegisterDefault); // default value should be 0
    // check that it's aligned properly to the beginning of the page
    if (0 != ((size_t)h_mem % 4096)) {
        // otherwise the cuda-memcheck and cuda-gdb tends to randomly crash (confirmed on cuda 8 - cuda 10)
        REPORT_ERROR(ERR_PARAM_INCORRECT, "Only pointer aligned to the page size can be registered");
    }
    // we remove const, but we don't change the data
    gpuErrchk(cudaHostRegister(const_cast<void*>(h_mem), bytes, flags));
}

set()

void GPU::set ( )

virtual

Reimplemented from HW.

Definition at line 50 of file gpu.cpp.

              {
    // set device (for current context / thread)
    setDevice(m_device);
    if ( ! m_isSet) {
        // create stream
        m_stream = new cudaStream_t;
        gpuErrchk(cudaStreamCreate((cudaStream_t*)m_stream));
        // remember the state
        m_isSet = true;
        // get additional info
        HW::set();
    }
    peekLastError();
}

setDevice()

void GPU::setDevice ( int  device )

static

Definition at line 135 of file gpu.cpp.

                              {
    gpuErrchk(cudaSetDevice(device));
    gpuErrchk(cudaPeekAtLastError());
}

stream()

void* GPU::stream ( ) const

inline

Definition at line 50 of file gpu.h.

                                {
        check();
        return m_stream;
    }

streamId()

int GPU::streamId ( ) const

inline

Definition at line 55 of file gpu.h.

                                {
        return m_streamId;
    }

synch()

void GPU::synch ( ) const

virtual

Implements HW.

Definition at line 129 of file gpu.cpp.

                      {
    check();
    auto stream = (cudaStream_t*)m_stream;
    gpuErrchk(cudaStreamSynchronize(*stream));
}

synchAll()

void GPU::synchAll ( ) const

virtual

Implements HW.

Definition at line 124 of file gpu.cpp.

                         {
    check();
    gpuErrchk(cudaDeviceSynchronize());
}

totalBytes()

size_t GPU::totalBytes ( ) const

inlinevirtual

Reimplemented from HW.

Definition at line 64 of file gpu.h.

                                     {
        check();
        return HW::totalBytes();
    }

unlockMemory()

void GPU::unlockMemory ( const void *  h_mem )

inlineoverridevirtual

Implements HW.

Definition at line 108 of file gpu.h.

                                                  {
        GPU::unpinMemory(h_mem);
    }

unpinMemory()

void GPU::unpinMemory ( const void *  h_mem )

static

Definition at line 108 of file gpu.cpp.

                                       {
    // we remove const, but we don't change the data
    auto err = cudaHostUnregister(const_cast<void*>(h_mem));
    if (cudaErrorHostMemoryNotRegistered == err) {
        cudaGetLastError(); // clear out the previous API error
    } else {
        gpuErrchk(err);
    }
}

updateMemoryInfo()

void GPU::updateMemoryInfo ( )

virtual

Implements HW.

Definition at line 82 of file gpu.cpp.

                           {
    check();
    gpuErrchk(cudaMemGetInfo(&m_lastFreeBytes, &m_totalBytes));
}

---

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

• xmipp/libraries/reconstruction_cuda/gpu.h
• xmipp/libraries/reconstruction_cuda/gpu.cpp