#include "arch_device_api.h"
#include <stdio.h>
#include <mutex>
#include "include/splitvector/splitvec.h"
#include "include/hashinator/hashinator.h"
#include "../definitions.h"
#include "../vlasovsolver/vec.h"
#include "../velocity_mesh_parameters.h"
#include <phiprof.hpp>

Include dependency graph for gpu_base.hpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
struct	ColumnOffsets
struct	GPUMemoryManager

Macros
#define	THREADS_PER_MP 2048
#define	REGISTERS_PER_MP 65536
#define	MAXCPUTHREADS 512
#define	NO_POINTER_DEVICE 0
#define	DEVICE_POINTER 1
#define	HOST_POINTER 2
#define	SINGLE_ARG(...)
#define	DEFINITIONS_HERE
#define	DEFINITIONS_END
#define	CREATE_UNIQUE_POINTER(object, member)
#define	CREATE_SUBPOINTERS(object, member, amount)
#define	ALLOCATE_GPU(object, member, bytes)
#define	ALLOCATE_WITH_BUFFER(object, member, bytes, buffer)
#define	HOST_ALLOCATE_GPU(object, member, bytes)
#define	HOST_ALLOCATE_WITH_BUFFER(object, member, bytes, buffer)
#define	SUBPOINTER_ALLOCATE(object, member, index, bytes)
#define	SUBPOINTER_HOST_ALLOCATE(object, member, index, bytes)
#define	SESSION_ALLOCATE(object, type, member, bytes)
#define	SESSION_HOST_ALLOCATE(object, type, member, bytes)
#define	ALLOCATE_GPU_ASYNC(object, member, bytes, stream)
#define	SUBPOINTER_ALLOCATE_ASYNC(object, member, index, bytes, stream)
#define	GET_POINTER(object, type, member)
#define	GET_SUBPOINTER(object, type, member, index)
#define	GET_SESSION_POINTER(object, type, member)
#define	GET_SESSION_HOST_POINTER(object, type, member)
#define	SET_SUBPOINTER(object, type, member, index, subPointerIndex)

Functions
static const uint	INIT_VMESH_SIZE (32768/WID3)
static const uint	INIT_MAP_SIZE (16 - WID)
void	gpu_init_device ()
void	gpu_clear_device ()
gpuStream_t	gpu_getStream ()
gpuStream_t	gpu_getPriorityStream ()
uint	gpu_getThread ()
uint	gpu_getMaxThreads ()
int	gpu_getDevice ()
uint	gpu_getAllocationCount ()
int	gpu_reportMemory (const size_t local_cap=0, const size_t ghost_cap=0, const size_t local_size=0, const size_t ghost_size=0)
unsigned int	nextPowerOfTwo (unsigned int n)
void	gpu_vlasov_allocate (uint maxBlockCount)
void	gpu_calculateProbeAllocation (uint maxBlockCount)
void	gpu_vlasov_deallocate ()
void	gpu_vlasov_allocate_perthread (uint cpuThreadID, uint maxBlockCount)
uint	gpu_vlasov_getSmallestAllocation ()
void	gpu_batch_allocate (uint nCells=0, uint maxNeighbours=0)
void	gpu_acc_allocate (uint maxBlockCount)
void	gpu_acc_allocate_perthread (uint cpuThreadID, uint firstAllocationCount, uint columnSetAllocationCount=0)
void	gpu_acc_deallocate ()
void	gpu_trans_allocate (cuint nAllCells=0, cuint largestVmesh=0, cuint unionSetSize=0)
void	gpu_trans_deallocate ()

Variables
int	gpuMultiProcessorCount
int	blocksPerMP
int	threadsPerMP
static const uint	VLASOV_BUFFER_MINBLOCKS = 32768/WID3
static const uint	VLASOV_BUFFER_MINCOLUMNS = 2000/WID
static const double	BLOCK_ALLOCATION_PADDING = 1.2
static const double	BLOCK_ALLOCATION_FACTOR = 1.1
static const int	GPU_PROBEFLAT_N = 5
static const int	TRANSLATION_BUFFER_ALLOCATION_FACTOR = 5
gpuStream_t	gpuStreamList []
gpuStream_t	gpuPriorityStreamList []
GPUMemoryManager	gpuMemoryManager
ColumnOffsets *	host_columnOffsetData
uint	gpu_largest_columnCount
size_t	gpu_probeFullSize
size_t	gpu_probeFlattenedSize
size_t	gpu_probeStride
std::vector< uint >	gpu_vlasov_allocatedSize
uint	gpu_acc_allocatedColumns
uint	gpu_acc_foundColumnsCount

Macro Definition Documentation

◆ ALLOCATE_GPU

#define ALLOCATE_GPU	(	object,
		member,
		bytes )

Value:

object.allocate(object.member, bytes)

Definition at line 403 of file gpu_base.hpp.

◆ ALLOCATE_GPU_ASYNC

#define ALLOCATE_GPU_ASYNC	(	object,
		member,
		bytes,
		stream )

Value:

object.allocateAsync(object.member, bytes, stream)

Definition at line 643 of file gpu_base.hpp.

◆ ALLOCATE_WITH_BUFFER

#define ALLOCATE_WITH_BUFFER	(	object,
		member,
		bytes,
		buffer )

Value:

object.allocateWithBuffer(object.member, bytes, buffer)

Definition at line 429 of file gpu_base.hpp.

◆ CREATE_SUBPOINTERS

#define CREATE_SUBPOINTERS	(	object,
		member,
		amount )

Value:

object.createSubPointers(object.member, amount)

Definition at line 322 of file gpu_base.hpp.

◆ CREATE_UNIQUE_POINTER

#define CREATE_UNIQUE_POINTER	(		object,
			member )

Value:

object.createPointer(object.member)

Definition at line 296 of file gpu_base.hpp.

◆ DEFINITIONS_END

#define DEFINITIONS_END

Definition at line 290 of file gpu_base.hpp.

◆ DEFINITIONS_HERE

#define DEFINITIONS_HERE

Definition at line 288 of file gpu_base.hpp.

◆ DEVICE_POINTER

#define DEVICE_POINTER 1

Definition at line 276 of file gpu_base.hpp.

◆ GET_POINTER

#define GET_POINTER	(	object,
		type,
		member )

Value:

object.getPointer<type>(object.member)

Definition at line 809 of file gpu_base.hpp.

◆ GET_SESSION_HOST_POINTER

#define GET_SESSION_HOST_POINTER	(	object,
		type,
		member )

Value:

object.getSessionHostPointer<type>(object.member)

Definition at line 853 of file gpu_base.hpp.

◆ GET_SESSION_POINTER

#define GET_SESSION_POINTER	(	object,
		type,
		member )

Value:

object.getSessionPointer<type>(object.member)

Definition at line 833 of file gpu_base.hpp.

◆ GET_SUBPOINTER

#define GET_SUBPOINTER	(	object,
		type,
		member,
		index )

Value:

object.getSubPointer<type>(object.member, index)

index

#define index(i, j, k)

Definition map_test_3d.cpp:5

Definition at line 819 of file gpu_base.hpp.

◆ HOST_ALLOCATE_GPU

#define HOST_ALLOCATE_GPU	(	object,
		member,
		bytes )

Value:

object.hostAllocate(object.member, bytes)

Definition at line 455 of file gpu_base.hpp.

◆ HOST_ALLOCATE_WITH_BUFFER

#define HOST_ALLOCATE_WITH_BUFFER	(	object,
		member,
		bytes,
		buffer )

Value:

object.hostAllocateWithBuffer(object.member, bytes, buffer)

Definition at line 481 of file gpu_base.hpp.

◆ HOST_POINTER

#define HOST_POINTER 2

Definition at line 277 of file gpu_base.hpp.

◆ MAXCPUTHREADS

#define MAXCPUTHREADS 512

Definition at line 73 of file gpu_base.hpp.

◆ NO_POINTER_DEVICE

#define NO_POINTER_DEVICE 0

Definition at line 275 of file gpu_base.hpp.

◆ REGISTERS_PER_MP

#define REGISTERS_PER_MP 65536

Definition at line 45 of file gpu_base.hpp.

◆ SESSION_ALLOCATE

#define SESSION_ALLOCATE	(	object,
		type,
		member,
		bytes )

Value:

object.sessionAllocate<type>(object.member, bytes)

Definition at line 557 of file gpu_base.hpp.

◆ SESSION_HOST_ALLOCATE

#define SESSION_HOST_ALLOCATE	(	object,
		type,
		member,
		bytes )

Value:

object.sessionHostAllocate<type>(object.member, bytes)

Definition at line 600 of file gpu_base.hpp.

◆ SET_SUBPOINTER

#define SET_SUBPOINTER	(	object,
		type,
		member,
		index,
		subPointerIndex )

Value:

object.setSubPointer<type>(object.member, index, subPointerIndex)

Definition at line 879 of file gpu_base.hpp.

◆ SINGLE_ARG

#define SINGLE_ARG ( ... )

Value:

__VA_ARGS__

Definition at line 280 of file gpu_base.hpp.

◆ SUBPOINTER_ALLOCATE

#define SUBPOINTER_ALLOCATE	(	object,
		member,
		index,
		bytes )

Value:

object.subPointerAllocate(object.member, index, bytes)

Definition at line 507 of file gpu_base.hpp.

◆ SUBPOINTER_ALLOCATE_ASYNC

#define SUBPOINTER_ALLOCATE_ASYNC	(	object,
		member,
		index,
		bytes,
		stream )

Value:

object.subPointerAllocateAsync(object.member, index, bytes, stream)

Definition at line 669 of file gpu_base.hpp.

◆ SUBPOINTER_HOST_ALLOCATE

#define SUBPOINTER_HOST_ALLOCATE	(	object,
		member,
		index,
		bytes )

Value:

object.subPointerHostAllocate(object.member, index, bytes)

Definition at line 534 of file gpu_base.hpp.

◆ THREADS_PER_MP

#define THREADS_PER_MP 2048

Definition at line 42 of file gpu_base.hpp.

Function Documentation

◆ gpu_acc_allocate()

void gpu_acc_allocate ( uint maxBlockCount )

Definition at line 537 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_acc_allocate_perthread()

void gpu_acc_allocate_perthread	(	uint	cpuThreadID,
		uint	firstAllocationCount,
		uint	columnSetAllocationCount = 0 )

Definition at line 567 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_acc_deallocate()

void gpu_acc_deallocate ( )

Definition at line 556 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_batch_allocate()

void gpu_batch_allocate	(	uint	nCells,
		uint	maxNeighbours )

Allocation and deallocation for pointers used by batch operations in block adjustment

Definition at line 462 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_calculateProbeAllocation()

void gpu_calculateProbeAllocation ( uint maxBlockCount )

Definition at line 359 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_clear_device()

void gpu_clear_device ( )

Definition at line 229 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_getAllocationCount()

uint gpu_getAllocationCount ( )

Definition at line 259 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_getDevice()

int gpu_getDevice ( )

Definition at line 253 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_getMaxThreads()

uint gpu_getMaxThreads ( )

Definition at line 84 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_getPriorityStream()

gpuStream_t gpu_getPriorityStream ( )

Definition at line 248 of file gpu_base.cpp.

Here is the call graph for this function:

◆ gpu_getStream()

gpuStream_t gpu_getStream ( )

Definition at line 244 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_getThread()

uint gpu_getThread ( )

Definition at line 77 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_init_device()

void gpu_init_device ( )

Definition at line 103 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_reportMemory()

int gpu_reportMemory	(	const size_t	local_cap = 0,
		const size_t	ghost_cap = 0,
		const size_t	local_size = 0,
		const size_t	ghost_size = 0 )

Definition at line 266 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_trans_allocate()

void gpu_trans_allocate	(	cuint	nAllCells = 0,
		cuint	largestVmesh = 0,
		cuint	unionSetSize = 0 )

Definition at line 607 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_trans_deallocate()

void gpu_trans_deallocate ( )

Definition at line 663 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_vlasov_allocate()

void gpu_vlasov_allocate ( uint maxBlockCount )

Definition at line 335 of file gpu_base.cpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ gpu_vlasov_allocate_perthread()

void gpu_vlasov_allocate_perthread	(	uint	cpuThreadID,
		uint	maxBlockCount )

Definition at line 421 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_vlasov_deallocate()

void gpu_vlasov_deallocate ( )

Definition at line 401 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ gpu_vlasov_getSmallestAllocation()

uint gpu_vlasov_getSmallestAllocation ( )

Definition at line 410 of file gpu_base.cpp.

Here is the caller graph for this function:

◆ INIT_MAP_SIZE()

const uint INIT_MAP_SIZE ( 16 - WID )

static

◆ INIT_VMESH_SIZE()

const uint INIT_VMESH_SIZE ( 32768/ WID3 )

static

Here is the caller graph for this function:

◆ nextPowerOfTwo()

unsigned int nextPowerOfTwo ( unsigned int n )

Definition at line 92 of file gpu_base.cpp.

Here is the caller graph for this function:

Variable Documentation

◆ BLOCK_ALLOCATION_FACTOR

const double BLOCK_ALLOCATION_FACTOR = 1.1

static

Definition at line 61 of file gpu_base.hpp.

◆ BLOCK_ALLOCATION_PADDING

const double BLOCK_ALLOCATION_PADDING = 1.2

static

Definition at line 60 of file gpu_base.hpp.

◆ blocksPerMP

int blocksPerMP

extern

Definition at line 43 of file gpu_base.cpp.

◆ gpu_acc_allocatedColumns

uint gpu_acc_allocatedColumns

extern

◆ gpu_acc_foundColumnsCount

uint gpu_acc_foundColumnsCount

extern

◆ gpu_largest_columnCount

uint gpu_largest_columnCount

extern

Definition at line 75 of file gpu_base.cpp.

◆ GPU_PROBEFLAT_N

const int GPU_PROBEFLAT_N = 5

static

Definition at line 65 of file gpu_base.hpp.

◆ gpu_probeFlattenedSize

size_t gpu_probeFlattenedSize

Definition at line 897 of file gpu_base.hpp.

◆ gpu_probeFullSize

size_t gpu_probeFullSize

extern

Definition at line 57 of file gpu_base.cpp.

◆ gpu_probeStride

size_t gpu_probeStride

Definition at line 897 of file gpu_base.hpp.

◆ gpu_vlasov_allocatedSize

std::vector<uint> gpu_vlasov_allocatedSize

extern

Definition at line 69 of file gpu_base.cpp.

◆ gpuMemoryManager

GPUMemoryManager gpuMemoryManager

extern

Definition at line 64 of file gpu_base.cpp.

◆ gpuMultiProcessorCount

int gpuMultiProcessorCount

extern

Definition at line 42 of file gpu_base.cpp.

◆ gpuPriorityStreamList

gpuStream_t gpuPriorityStreamList[]

extern

Definition at line 52 of file gpu_base.cpp.

◆ gpuStreamList

gpuStream_t gpuStreamList[]

extern

Definition at line 51 of file gpu_base.cpp.

◆ host_columnOffsetData

ColumnOffsets* host_columnOffsetData

extern

Definition at line 55 of file gpu_base.cpp.

◆ threadsPerMP

int threadsPerMP

extern

Definition at line 44 of file gpu_base.cpp.

◆ TRANSLATION_BUFFER_ALLOCATION_FACTOR

const int TRANSLATION_BUFFER_ALLOCATION_FACTOR = 5

static

Definition at line 71 of file gpu_base.hpp.

◆ VLASOV_BUFFER_MINBLOCKS

const uint VLASOV_BUFFER_MINBLOCKS = 32768/WID3

static

Definition at line 56 of file gpu_base.hpp.

◆ VLASOV_BUFFER_MINCOLUMNS

const uint VLASOV_BUFFER_MINCOLUMNS = 2000/WID

static

Definition at line 57 of file gpu_base.hpp.

Classes

Macros

Functions

Variables

Macro Definition Documentation

◆ ALLOCATE_GPU

◆ ALLOCATE_GPU_ASYNC

◆ ALLOCATE_WITH_BUFFER

◆ CREATE_SUBPOINTERS

◆ CREATE_UNIQUE_POINTER

◆ DEFINITIONS_END

◆ DEFINITIONS_HERE

◆ DEVICE_POINTER

◆ GET_POINTER

◆ GET_SESSION_HOST_POINTER

◆ GET_SESSION_POINTER

◆ GET_SUBPOINTER

◆ HOST_ALLOCATE_GPU

◆ HOST_ALLOCATE_WITH_BUFFER

◆ HOST_POINTER

◆ MAXCPUTHREADS

◆ NO_POINTER_DEVICE

◆ REGISTERS_PER_MP

◆ SESSION_ALLOCATE

◆ SESSION_HOST_ALLOCATE

◆ SET_SUBPOINTER

◆ SINGLE_ARG

◆ SUBPOINTER_ALLOCATE

◆ SUBPOINTER_ALLOCATE_ASYNC

◆ SUBPOINTER_HOST_ALLOCATE

◆ THREADS_PER_MP

Function Documentation

◆ gpu_acc_allocate()

◆ gpu_acc_allocate_perthread()

◆ gpu_acc_deallocate()

◆ gpu_batch_allocate()

◆ gpu_calculateProbeAllocation()

◆ gpu_clear_device()

◆ gpu_getAllocationCount()

◆ gpu_getDevice()

◆ gpu_getMaxThreads()

◆ gpu_getPriorityStream()

◆ gpu_getStream()

◆ gpu_getThread()

◆ gpu_init_device()

◆ gpu_reportMemory()

◆ gpu_trans_allocate()

◆ gpu_trans_deallocate()

◆ gpu_vlasov_allocate()

◆ gpu_vlasov_allocate_perthread()

◆ gpu_vlasov_deallocate()

◆ gpu_vlasov_getSmallestAllocation()

◆ INIT_MAP_SIZE()

◆ INIT_VMESH_SIZE()

◆ nextPowerOfTwo()

Variable Documentation

◆ BLOCK_ALLOCATION_FACTOR

◆ BLOCK_ALLOCATION_PADDING

◆ blocksPerMP

◆ gpu_acc_allocatedColumns

◆ gpu_acc_foundColumnsCount

◆ gpu_largest_columnCount

◆ GPU_PROBEFLAT_N

◆ gpu_probeFlattenedSize

◆ gpu_probeFullSize

◆ gpu_probeStride

◆ gpu_vlasov_allocatedSize

◆ gpuMemoryManager

◆ gpuMultiProcessorCount

◆ gpuPriorityStreamList

◆ gpuStreamList

◆ host_columnOffsetData

◆ threadsPerMP

◆ TRANSLATION_BUFFER_ALLOCATION_FACTOR

◆ VLASOV_BUFFER_MINBLOCKS

◆ VLASOV_BUFFER_MINCOLUMNS