CPU kernel. More...

#include <kernel.h>

Inheritance diagram for Coal::CPUKernel:

Collaboration diagram for Coal::CPUKernel:

Public Member Functions
	CPUKernel (CPUDevice device, Kernel kernel, llvm::Function *function)
	Constructor.
	~CPUKernel ()
size_t	workGroupSize () const
	Maximum work-group size of a kernel.
cl_ulong	localMemSize () const
	Local memory used by the kernel.
cl_ulong	privateMemSize () const
	Private memory used by the kernel.
size_t	preferredWorkGroupSizeMultiple () const
	Preferred work-group size multiple.
size_t	guessWorkGroupSize (cl_uint num_dims, cl_uint dim, size_t global_work_size) const
	Optimal work-group size.
Kernel *	kernel () const
	`Coal::Kernel` object this kernel will run
CPUDevice *	device () const
	device on which the kernel will be run
llvm::Function *	function () const
	`llvm::Function` representing the kernel but not to be run
llvm::Function *	callFunction ()
	stub function used to run the kernel, see Using Clang and LLVM to Launch Kernels
Static Public Member Functions
static size_t	typeOffset (size_t &offset, size_t type_len)
	Calculate where to place a value in an array.
Private Attributes
CPUDevice *	p_device
Kernel *	p_kernel
llvm::Function *	p_function
llvm::Function *	p_call_function
pthread_mutex_t	p_call_function_mutex

Detailed Description

CPU kernel.

This class holds passive information about a kernel (Coal::Kernel object and device on which it is run) and provides the callFunction() function.

This function is described at the end of Using Clang and LLVM to Launch Kernels .

See also:: Coal::CPUKernelWorkGroup

Definition at line 71 of file cpu/kernel.h.

Constructor & Destructor Documentation

CPUKernel::CPUKernel	(	CPUDevice *	device,
		Kernel *	kernel,
		llvm::Function *	function
	)

Constructor.

Parameters:

device	device on which the kernel will be run
kernel	`Coal::Kernel` object holding information about this kernel
function	`llvm::Function` to run

Definition at line 58 of file cpu/kernel.cpp.

References p_call_function_mutex.

CPUKernel::~CPUKernel ( )

Definition at line 65 of file cpu/kernel.cpp.

References p_call_function, and p_call_function_mutex.

Member Function Documentation

llvm::Function * CPUKernel::callFunction ( )

stub function used to run the kernel, see Using Clang and LLVM to Launch Kernels

Definition at line 179 of file cpu/kernel.cpp.

References Coal::Kernel::arg(), p_call_function, p_call_function_mutex, p_function, p_kernel, typeOffset(), Coal::Kernel::Arg::valueSize(), and Coal::Kernel::Arg::vecDim().

Referenced by Coal::CPUKernelWorkGroup::run().

CPUDevice * CPUKernel::device ( ) const

device on which the kernel will be run

Definition at line 146 of file cpu/kernel.cpp.

References p_device.

Referenced by Coal::CPUKernelWorkGroup::callArgs(), Coal::CPUKernelWorkGroup::getImageData(), and Coal::CPUKernelWorkGroup::run().

llvm::Function * CPUKernel::function ( ) const

llvm::Function representing the kernel but not to be run

Definition at line 136 of file cpu/kernel.cpp.

References p_function.

Referenced by Coal::CPUKernelWorkGroup::barrier(), and Coal::CPUKernelWorkGroup::builtinNotFound().

size_t CPUKernel::guessWorkGroupSize	(	cl_uint	num_dims,
		cl_uint	dim,
		size_t	global_work_size
	)		const `[virtual]`

Optimal work-group size.

This function allows a device to calculate the optimal work-group size for this kernel, using it's memory usage, SIMD dimension, etc.

Coal::CPUDevice tries to split the kernel into a number of work-groups the closest possible to the number of CPU cores.

Parameters:

num_dims	Number of working dimensions
dim	Dimension for which the multiple is being calculated
global_work_size	Total number of work-items to split into work-groups

Returns:: optimal size of a work-group, for the dim dimension.

Implements Coal::DeviceKernel.

Definition at line 106 of file cpu/kernel.cpp.

References k_exp(), Coal::CPUDevice::numCPUs(), and p_device.

Kernel * CPUKernel::kernel ( ) const

Coal::Kernel object this kernel will run

Definition at line 141 of file cpu/kernel.cpp.

References p_kernel.

Referenced by Coal::CPUKernelWorkGroup::callArgs(), and Coal::CPUKernelWorkGroup::run().

cl_ulong CPUKernel::localMemSize ( ) const [virtual]

Local memory used by the kernel.

Returns:: Local memory used by the kernel, in bytes

Implements Coal::DeviceKernel.

Definition at line 78 of file cpu/kernel.cpp.

size_t CPUKernel::preferredWorkGroupSizeMultiple ( ) const [virtual]

Preferred work-group size multiple.

Returns:: The size multiple a work-group can have to work the best and the fastest on the device

Implements Coal::DeviceKernel.

Definition at line 88 of file cpu/kernel.cpp.

cl_ulong CPUKernel::privateMemSize ( ) const [virtual]

Private memory used by the kernel.

Returns:: Private memory used by the kernel, in bytes

Implements Coal::DeviceKernel.

Definition at line 83 of file cpu/kernel.cpp.

size_t CPUKernel::typeOffset	(	size_t &	offset,
		size_t	type_len
	)		`[static]`

Calculate where to place a value in an array.

This function is used to calculate where to place a value in an array given its size, properly aligning it.

This function is called repeatedly to obtain the aligned position of each value that must be place in the array

 size_t array_len = 0, array_offset = 0;
 void *array;

 // First, get the array size given alignment constraints
 typeOffset(array_len, sizeof(int));
 typeOffset(array_len, sizeof(float));
 typeOffset(array_len, sizeof(void *));

 // Then, allocate memory
 array = malloc(array_len)

 // Finally, place the arguments
 *(int *)((char *)array + typeOffset(array_offset, sizeof(int))) = 1337;
 *(float *)((char *)array + typeOffset(array_offset, sizeof(int))) = 3.1415f;
 *(void **)((char *)array + typeOffset(array_offset, sizeof(int))) = array;

Parameters:

offset	offset at which the value will be placed. This variable gets incremented by `type_len + padding`.
type_len	size in bytes of the value that will be stored