6.8. Execution Control

This section describes the execution control functions of the CUDA runtime application programming interface.

Some functions have overloaded C++ API template versions documented separately in the C++ API Routines module.

Functions

__host__ ​ __device__ ​cudaError_t cudaFuncGetAttributes ( cudaFuncAttributes* attr, const void* func )
Find out attributes for a given function.
__host__ ​cudaError_t cudaFuncGetName ( const char** name, const void* func )
Returns the function name for a device entry function pointer.
__host__ ​cudaError_t cudaFuncGetParamInfo ( const void* func, size_t paramIndex, size_t* paramOffset, size_t* paramSize )
Returns the offset and size of a kernel parameter in the device-side parameter layout.
__host__ ​cudaError_t cudaFuncSetAttribute ( const void* func, cudaFuncAttribute attr, int  value )
Set attributes for a given function.
__host__ ​cudaError_t cudaFuncSetCacheConfig ( const void* func, cudaFuncCache cacheConfig )
Sets the preferred cache configuration for a device function.
__device__ ​ void* cudaGetParameterBuffer ( size_t alignment, size_t size )
Obtains a parameter buffer.
__device__ ​ void cudaGridDependencySynchronize ( void )
Programmatic grid dependency synchronization.
__host__ ​cudaError_t cudaLaunchCooperativeKernel ( const void* func, dim3 gridDim, dim3 blockDim, void** args, size_t sharedMem, cudaStream_t stream )
Launches a device function where thread blocks can cooperate and synchronize as they execute.
__host__ ​cudaError_t cudaLaunchCooperativeKernelMultiDevice ( cudaLaunchParams* launchParamsList, unsigned int  numDevices, unsigned int  flags = 0 )
Launches device functions on multiple devices where thread blocks can cooperate and synchronize as they execute.
__device__ ​cudaError_t cudaLaunchDevice ( void* func, void* parameterBuffer, dim3 gridDimension, dim3 blockDimension, unsigned int  sharedMemSize, cudaStream_t stream )
Launches a specified kernel.
__host__ ​cudaError_t cudaLaunchHostFunc ( cudaStream_t stream, cudaHostFn_t fn, void* userData )
Enqueues a host function call in a stream.
__host__ ​cudaError_t cudaLaunchKernel ( const void* func, dim3 gridDim, dim3 blockDim, void** args, size_t sharedMem, cudaStream_t stream )
Launches a device function.
__host__ ​cudaError_t cudaLaunchKernelExC ( const cudaLaunchConfig_t* config, const void* func, void** args )
Launches a CUDA function with launch-time configuration.
__host__ ​cudaError_t cudaSetDoubleForDevice ( double* d )
Converts a double argument to be executed on a device.
__host__ ​cudaError_t cudaSetDoubleForHost ( double* d )
Converts a double argument after execution on a device.
__device__ ​ void cudaTriggerProgrammaticLaunchCompletion ( void )
Programmatic dependency trigger.

Functions

__host__ ​ __device__ ​cudaError_t cudaFuncGetAttributes ( cudaFuncAttributes* attr, const void* func )
Find out attributes for a given function.
Parameters
attr
- Return pointer to function's attributes
func
- Device function symbol
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction

Description

This function obtains the attributes of a function specified via func. func is a device function symbol and must be declared as a __global__ function. The fetched attributes are placed in attr. If the specified function does not exist, then cudaErrorInvalidDeviceFunction is returned. For templated functions, pass the function symbol as follows: func_name<template_arg_0,...,template_arg_N>

Note that some function attributes such as maxThreadsPerBlock may vary based on the device that is currently being used.

Note:

See also:

cudaFuncSetCacheConfig ( C API), cudaFuncGetAttributes ( C++ API), cudaLaunchKernel ( C API), cuFuncGetAttribute

__host__ ​cudaError_t cudaFuncGetName ( const char** name, const void* func )
Returns the function name for a device entry function pointer.
Parameters
name
- The returned name of the function
func
- The function pointer to retrieve name for
Returns

cudaSuccess, cudaErrorInvalidValue, cudaErrorInvalidDeviceFunction

Description

Returns in **name the function name associated with the symbol func . The function name is returned as a null-terminated string. This API may return a mangled name if the function is not declared as having C linkage. If **name is NULL, cudaErrorInvalidValue is returned. If func is not a device entry function, cudaErrorInvalidDeviceFunction is returned.

Note:

cudaFuncGetName (C++ API)

__host__ ​cudaError_t cudaFuncGetParamInfo ( const void* func, size_t paramIndex, size_t* paramOffset, size_t* paramSize )
Returns the offset and size of a kernel parameter in the device-side parameter layout.
Parameters
func
- The function to query
paramIndex
- The parameter index to query
paramOffset
- The offset into the device-side parameter layout at which the parameter resides
paramSize
- The size of the parameter in the device-side parameter layout
Returns

CUDA_SUCCESS, CUDA_ERROR_INVALID_VALUE,

Description

Queries the kernel parameter at paramIndex in func's list of parameters and returns parameter information via paramOffset and paramSize. paramOffset returns the offset of the parameter in the device-side parameter layout. paramSize returns the size in bytes of the parameter. This information can be used to update kernel node parameters from the device via cudaGraphKernelNodeSetParam() and cudaGraphKernelNodeUpdatesApply(). paramIndex must be less than the number of parameters that func takes.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

__host__ ​cudaError_t cudaFuncSetAttribute ( const void* func, cudaFuncAttribute attr, int  value )
Set attributes for a given function.
Parameters
func
- Function to get attributes of
attr
- Attribute to set
value
- Value to set
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction, cudaErrorInvalidValue

Description

This function sets the attributes of a function specified via func. The parameter func must be a pointer to a function that executes on the device. The parameter specified by func must be declared as a __global__ function. The enumeration defined by attr is set to the value defined by value. If the specified function does not exist, then cudaErrorInvalidDeviceFunction is returned. If the specified attribute cannot be written, or if the value is incorrect, then cudaErrorInvalidValue is returned.

Valid values for attr are:

  • cudaFuncAttributeMaxDynamicSharedMemorySize - The requested maximum size in bytes of dynamically-allocated shared memory. The sum of this value and the function attribute sharedSizeBytes cannot exceed the device attribute cudaDevAttrMaxSharedMemoryPerBlockOptin. The maximal size of requestable dynamic shared memory may differ by GPU architecture.

  • cudaFuncAttributePreferredSharedMemoryCarveout - On devices where the L1 cache and shared memory use the same hardware resources, this sets the shared memory carveout preference, in percent of the total shared memory. See cudaDevAttrMaxSharedMemoryPerMultiprocessor. This is only a hint, and the driver can choose a different ratio if required to execute the function.

  • cudaFuncAttributeRequiredClusterWidth: The required cluster width in blocks. The width, height, and depth values must either all be 0 or all be positive. The validity of the cluster dimensions is checked at launch time. If the value is set during compile time, it cannot be set at runtime. Setting it at runtime will return cudaErrorNotPermitted.

  • cudaFuncAttributeRequiredClusterHeight: The required cluster height in blocks. The width, height, and depth values must either all be 0 or all be positive. The validity of the cluster dimensions is checked at launch time. If the value is set during compile time, it cannot be set at runtime. Setting it at runtime will return cudaErrorNotPermitted.

  • cudaFuncAttributeRequiredClusterDepth: The required cluster depth in blocks. The width, height, and depth values must either all be 0 or all be positive. The validity of the cluster dimensions is checked at launch time. If the value is set during compile time, it cannot be set at runtime. Setting it at runtime will return cudaErrorNotPermitted.

  • cudaFuncAttributeNonPortableClusterSizeAllowed: Indicates whether the function can be launched with non-portable cluster size. 1 is allowed, 0 is disallowed.

  • cudaFuncAttributeClusterSchedulingPolicyPreference: The block scheduling policy of a function. The value type is cudaClusterSchedulingPolicy.

Note:

cudaLaunchKernel (C++ API), cudaFuncSetCacheConfig ( C++ API), cudaFuncGetAttributes ( C API),

__host__ ​cudaError_t cudaFuncSetCacheConfig ( const void* func, cudaFuncCache cacheConfig )
Sets the preferred cache configuration for a device function.
Parameters
func
- Device function symbol
cacheConfig
- Requested cache configuration
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction

Description

On devices where the L1 cache and shared memory use the same hardware resources, this sets through cacheConfig the preferred cache configuration for the function specified via func. This is only a preference. The runtime will use the requested configuration if possible, but it is free to choose a different configuration if required to execute func.

func is a device function symbol and must be declared as a __global__ function. If the specified function does not exist, then cudaErrorInvalidDeviceFunction is returned. For templated functions, pass the function symbol as follows: func_name<template_arg_0,...,template_arg_N>

This setting does nothing on devices where the size of the L1 cache and shared memory are fixed.

Launching a kernel with a different preference than the most recent preference setting may insert a device-side synchronization point.

The supported cache configurations are:

Note:

See also:

cudaFuncSetCacheConfig ( C++ API), cudaFuncGetAttributes ( C API), cudaLaunchKernel ( C API), cuFuncSetCacheConfig

__device__ ​ void* cudaGetParameterBuffer ( size_t alignment, size_t size )
Obtains a parameter buffer.
Parameters
alignment
- Specifies alignment requirement of the parameter buffer
size
- Specifies size requirement in bytes
Returns

Returns pointer to the allocated parameterBuffer

Description

Obtains a parameter buffer which can be filled with parameters for a kernel launch. Parameters passed to cudaLaunchDevice must be allocated via this function.

This is a low level API and can only be accessed from Parallel Thread Execution (PTX). CUDA user code should use <<< >>> to launch kernels.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

See also:

cudaLaunchDevice

__device__ ​ void cudaGridDependencySynchronize ( void ) [inline]
Programmatic grid dependency synchronization.
Description

This device function will block the thread until all direct grid dependencies have completed. This API is intended to use in conjuncture with programmatic / launch event / dependency. See cudaLaunchAttributeID::cudaLaunchAttributeProgrammaticStreamSerialization and cudaLaunchAttributeID::cudaLaunchAttributeProgrammaticEvent for more information.

__host__ ​cudaError_t cudaLaunchCooperativeKernel ( const void* func, dim3 gridDim, dim3 blockDim, void** args, size_t sharedMem, cudaStream_t stream )
Launches a device function where thread blocks can cooperate and synchronize as they execute.
Parameters
func
- Device function symbol
gridDim
- Grid dimentions
blockDim
- Block dimentions
args
- Arguments
sharedMem
- Shared memory
stream
- Stream identifier
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction, cudaErrorInvalidConfiguration, cudaErrorLaunchFailure, cudaErrorLaunchTimeout, cudaErrorLaunchOutOfResources, cudaErrorCooperativeLaunchTooLarge, cudaErrorSharedObjectInitFailed

Description

The function invokes kernel func on gridDim (gridDim.xgridDim.ygridDim.z) grid of blocks. Each block contains blockDim (blockDim.xblockDim.yblockDim.z) threads.

The device on which this kernel is invoked must have a non-zero value for the device attribute cudaDevAttrCooperativeLaunch.

The total number of blocks launched cannot exceed the maximum number of blocks per multiprocessor as returned by cudaOccupancyMaxActiveBlocksPerMultiprocessor (or cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors as specified by the device attribute cudaDevAttrMultiProcessorCount.

The kernel cannot make use of CUDA dynamic parallelism.

If the kernel has N parameters the args should point to array of N pointers. Each pointer, from args[0] to args[N - 1], point to the region of memory from which the actual parameter will be copied.

For templated functions, pass the function symbol as follows: func_name<template_arg_0,...,template_arg_N>

sharedMem sets the amount of dynamic shared memory that will be available to each thread block.

stream specifies a stream the invocation is associated to.

Note:

See also:

cudaLaunchCooperativeKernel (C++ API), cudaLaunchCooperativeKernelMultiDevice, cuLaunchCooperativeKernel

__host__ ​cudaError_t cudaLaunchCooperativeKernelMultiDevice ( cudaLaunchParams* launchParamsList, unsigned int  numDevices, unsigned int  flags = 0 )
Launches device functions on multiple devices where thread blocks can cooperate and synchronize as they execute.
Parameters
launchParamsList
- List of launch parameters, one per device
numDevices
- Size of the launchParamsList array
flags
- Flags to control launch behavior
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction, cudaErrorInvalidConfiguration, cudaErrorLaunchFailure, cudaErrorLaunchTimeout, cudaErrorLaunchOutOfResources, cudaErrorCooperativeLaunchTooLarge, cudaErrorSharedObjectInitFailed

Deprecated

This function is deprecated as of CUDA 11.3.

Description

Invokes kernels as specified in the launchParamsList array where each element of the array specifies all the parameters required to perform a single kernel launch. These kernels can cooperate and synchronize as they execute. The size of the array is specified by numDevices.

No two kernels can be launched on the same device. All the devices targeted by this multi-device launch must be identical. All devices must have a non-zero value for the device attribute cudaDevAttrCooperativeMultiDeviceLaunch.

The same kernel must be launched on all devices. Note that any __device__ or __constant__ variables are independently instantiated on every device. It is the application's responsiblity to ensure these variables are initialized and used appropriately.

The size of the grids as specified in blocks, the size of the blocks themselves and the amount of shared memory used by each thread block must also match across all launched kernels.

The streams used to launch these kernels must have been created via either cudaStreamCreate or cudaStreamCreateWithPriority or cudaStreamCreateWithPriority. The NULL stream or cudaStreamLegacy or cudaStreamPerThread cannot be used.

The total number of blocks launched per kernel cannot exceed the maximum number of blocks per multiprocessor as returned by cudaOccupancyMaxActiveBlocksPerMultiprocessor (or cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors as specified by the device attribute cudaDevAttrMultiProcessorCount. Since the total number of blocks launched per device has to match across all devices, the maximum number of blocks that can be launched per device will be limited by the device with the least number of multiprocessors.

The kernel cannot make use of CUDA dynamic parallelism.

The cudaLaunchParams structure is defined as: 

‎        struct cudaLaunchParams
              {
                  void *func;
                  dim3 gridDim;
                  dim3 blockDim;
                  void **args;
                  size_t sharedMem;
                  cudaStream_t 
                  stream;
              };
where:

By default, the kernel won't begin execution on any GPU until all prior work in all the specified streams has completed. This behavior can be overridden by specifying the flag cudaCooperativeLaunchMultiDeviceNoPreSync. When this flag is specified, each kernel will only wait for prior work in the stream corresponding to that GPU to complete before it begins execution.

Similarly, by default, any subsequent work pushed in any of the specified streams will not begin execution until the kernels on all GPUs have completed. This behavior can be overridden by specifying the flag cudaCooperativeLaunchMultiDeviceNoPostSync. When this flag is specified, any subsequent work pushed in any of the specified streams will only wait for the kernel launched on the GPU corresponding to that stream to complete before it begins execution.

Note:

See also:

cudaLaunchCooperativeKernel (C++ API), cudaLaunchCooperativeKernel, cuLaunchCooperativeKernelMultiDevice

__device__ ​cudaError_t cudaLaunchDevice ( void* func, void* parameterBuffer, dim3 gridDimension, dim3 blockDimension, unsigned int  sharedMemSize, cudaStream_t stream )
Launches a specified kernel.
Parameters
func
- Pointer to the kernel to be launched
parameterBuffer
- Holds the parameters to the launched kernel. parameterBuffer can be NULL. (Optional)
gridDimension
- Specifies grid dimensions
blockDimension
- Specifies block dimensions
sharedMemSize
- Specifies size of shared memory
stream
- Specifies the stream to be used
Returns

cudaSuccess, cudaErrorInvalidDevice, cudaErrorLaunchMaxDepthExceeded, cudaErrorInvalidConfiguration, cudaErrorStartupFailure, cudaErrorLaunchPendingCountExceeded, cudaErrorLaunchOutOfResources

Description

Launches a specified kernel with the specified parameter buffer. A parameter buffer can be obtained by calling cudaGetParameterBuffer().

This is a low level API and can only be accessed from Parallel Thread Execution (PTX). CUDA user code should use <<< >>> to launch the kernels.

Note:

Note that this function may also return error codes from previous, asynchronous launches.

Please refer to Execution Configuration and Parameter Buffer Layout from the CUDA Programming Guide for the detailed descriptions of launch configuration and parameter layout respectively.

See also:

cudaGetParameterBuffer

__host__ ​cudaError_t cudaLaunchHostFunc ( cudaStream_t stream, cudaHostFn_t fn, void* userData )
Enqueues a host function call in a stream.
Parameters
stream
fn
- The function to call once preceding stream operations are complete
userData
- User-specified data to be passed to the function
Returns

cudaSuccess, cudaErrorInvalidResourceHandle, cudaErrorInvalidValue, cudaErrorNotSupported

Description

Enqueues a host function to run in a stream. The function will be called after currently enqueued work and will block work added after it.

The host function must not make any CUDA API calls. Attempting to use a CUDA API may result in cudaErrorNotPermitted, but this is not required. The host function must not perform any synchronization that may depend on outstanding CUDA work not mandated to run earlier. Host functions without a mandated order (such as in independent streams) execute in undefined order and may be serialized.

For the purposes of Unified Memory, execution makes a number of guarantees:

  • The stream is considered idle for the duration of the function's execution. Thus, for example, the function may always use memory attached to the stream it was enqueued in.

  • The start of execution of the function has the same effect as synchronizing an event recorded in the same stream immediately prior to the function. It thus synchronizes streams which have been "joined" prior to the function.

  • Adding device work to any stream does not have the effect of making the stream active until all preceding host functions and stream callbacks have executed. Thus, for example, a function might use global attached memory even if work has been added to another stream, if the work has been ordered behind the function call with an event.

  • Completion of the function does not cause a stream to become active except as described above. The stream will remain idle if no device work follows the function, and will remain idle across consecutive host functions or stream callbacks without device work in between. Thus, for example, stream synchronization can be done by signaling from a host function at the end of the stream.

Note that, in constrast to cuStreamAddCallback, the function will not be called in the event of an error in the CUDA context.

Note:

See also:

cudaStreamCreate, cudaStreamQuery, cudaStreamSynchronize, cudaStreamWaitEvent, cudaStreamDestroy, cudaMallocManaged, cudaStreamAttachMemAsync, cudaStreamAddCallback, cuLaunchHostFunc

__host__ ​cudaError_t cudaLaunchKernel ( const void* func, dim3 gridDim, dim3 blockDim, void** args, size_t sharedMem, cudaStream_t stream )
Launches a device function.
Parameters
func
- Device function symbol
gridDim
- Grid dimentions
blockDim
- Block dimentions
args
- Arguments
sharedMem
- Shared memory
stream
- Stream identifier
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction, cudaErrorInvalidConfiguration, cudaErrorLaunchFailure, cudaErrorLaunchTimeout, cudaErrorLaunchOutOfResources, cudaErrorSharedObjectInitFailed, cudaErrorInvalidPtx, cudaErrorUnsupportedPtxVersion, cudaErrorNoKernelImageForDevice, cudaErrorJitCompilerNotFound, cudaErrorJitCompilationDisabled

Description

The function invokes kernel func on gridDim (gridDim.xgridDim.ygridDim.z) grid of blocks. Each block contains blockDim (blockDim.xblockDim.yblockDim.z) threads.

If the kernel has N parameters the args should point to array of N pointers. Each pointer, from args[0] to args[N - 1], point to the region of memory from which the actual parameter will be copied.

For templated functions, pass the function symbol as follows: func_name<template_arg_0,...,template_arg_N>

sharedMem sets the amount of dynamic shared memory that will be available to each thread block.

stream specifies a stream the invocation is associated to.

Note:

See also:

cudaLaunchKernel (C++ API), cuLaunchKernel

__host__ ​cudaError_t cudaLaunchKernelExC ( const cudaLaunchConfig_t* config, const void* func, void** args )
Launches a CUDA function with launch-time configuration.
Parameters
config
- Launch configuration
func
- Kernel to launch
args
- Array of pointers to kernel parameters
Returns

cudaSuccess, cudaErrorInvalidDeviceFunction, cudaErrorInvalidConfiguration, cudaErrorLaunchFailure, cudaErrorLaunchTimeout, cudaErrorLaunchOutOfResources, cudaErrorSharedObjectInitFailed, cudaErrorInvalidPtx, cudaErrorUnsupportedPtxVersion, cudaErrorNoKernelImageForDevice, cudaErrorJitCompilerNotFound, cudaErrorJitCompilationDisabled

Description

Note that the functionally equivalent variadic template cudaLaunchKernelEx is available for C++11 and newer.

Invokes the kernel func on config->gridDim (config->gridDim.xconfig->gridDim.yconfig->gridDim.z) grid of blocks. Each block contains config->blockDim (config->blockDim.xconfig->blockDim.yconfig->blockDim.z) threads.

config->dynamicSmemBytes sets the amount of dynamic shared memory that will be available to each thread block.

config->stream specifies a stream the invocation is associated to.

Configuration beyond grid and block dimensions, dynamic shared memory size, and stream can be provided with the following two fields of config:

config->attrs is an array of config->numAttrs contiguous cudaLaunchAttribute elements. The value of this pointer is not considered if config->numAttrs is zero. However, in that case, it is recommended to set the pointer to NULL. config->numAttrs is the number of attributes populating the first config->numAttrs positions of the config->attrs array.

If the kernel has N parameters the args should point to array of N pointers. Each pointer, from args[0] to args[N - 1], point to the region of memory from which the actual parameter will be copied.

N.B. This function is so named to avoid unintentionally invoking the templated version, cudaLaunchKernelEx, for kernels taking a single void** or void* parameter.

Note:

See also:

cudaLaunchKernelEx(const cudaLaunchConfig_t *config, void (*kernel)(ExpTypes...), ActTypes &&... args) "cudaLaunchKernelEx (C++ API)", cuLaunchKernelEx

__host__ ​cudaError_t cudaSetDoubleForDevice ( double* d )
Converts a double argument to be executed on a device.
Parameters
d
- Double to convert
Returns

cudaSuccess

Deprecated

This function is deprecated as of CUDA 7.5

Description

Converts the double value of d to an internal float representation if the device does not support double arithmetic. If the device does natively support doubles, then this function does nothing.

Note:

See also:

cudaFuncSetCacheConfig ( C API), cudaFuncGetAttributes ( C API), cudaSetDoubleForHost

__host__ ​cudaError_t cudaSetDoubleForHost ( double* d )
Converts a double argument after execution on a device.
Parameters
d
- Double to convert
Returns

cudaSuccess

Deprecated

This function is deprecated as of CUDA 7.5

Description

Converts the double value of d from a potentially internal float representation if the device does not support double arithmetic. If the device does natively support doubles, then this function does nothing.

Note:

See also:

cudaFuncSetCacheConfig ( C API), cudaFuncGetAttributes ( C API), cudaSetDoubleForDevice

__device__ ​ void cudaTriggerProgrammaticLaunchCompletion ( void ) [inline]
Programmatic dependency trigger.
Description

This device function ensures the programmatic launch completion edges / events are fulfilled. See cudaLaunchAttributeID::cudaLaunchAttributeProgrammaticStreamSerialization and cudaLaunchAttributeID::cudaLaunchAttributeProgrammaticEvent for more information. The event / edge kick off only happens when every CTAs in the grid has either exited or called this function at least once, otherwise the kick off happens automatically after all warps finishes execution but before the grid completes. The kick off only enables scheduling of the secondary kernel. It provides no memory visibility guarantee itself. The user could enforce memory visibility by inserting a memory fence of the correct scope.