Debugger API :: CUDA Toolkit Documentation

3.10. Events

One of those events will create a CUDBGEvent:

the elf image of the current kernel has been loaded and the addresses within its DWARF sections have been relocated (and can now be used to set breakpoints),
a device breakpoint has been hit,
a CUDA kernel is ready to be launched,
a CUDA kernel has terminated.

When a CUDBGEvent is created, the debugger is notified by calling the callback functions registered with setNotifyNewEventCallback() after the API struct initialization. It is up to the debugger to decide what method is best to be notified. The debugger API routines cannot be called from within the callback function or the routine will return an error.

Upon notification the debugger is responsible for handling the CUDBGEvents in the event queue by using CUDBGAPI_st::getNextEvent(), and for acknowledging the debugger API that the event has been handled by calling CUDBGAPI_st::acknowledgeEvent(). In the case of an event raised by the device itself, such as a breakpoint being hit, the event queue will be empty. It is the responsibility of the debugger to inspect the hardware any time a CUDBGEvent is received.

Example:

‎CUDBGEvent event;
      CUDBGResult res;
      for (res = cudbgAPI->getNextEvent(&event);
           res == CUDBG_SUCCESS && event.kind != CUDBG_EVENT_INVALID;
           res = cudbgAPI->getNextEvent(&event)) {
          switch (event.kind)
              {
              case CUDBG_EVENT_ELF_IMAGE_LOADED:
                  //...
                  break;
              case CUDBG_EVENT_KERNEL_READY:
                  //...
                  break;
              case CUDBG_EVENT_KERNEL_FINISHED:
                  //...
                  break;
              default:
                  error(...);
              }
          }

See cuda-tdep.c and cuda-linux-nat.c files in the cuda-gdb source code for a more detailed example on how to use CUDBGEvent.

Classes

struct CUDBGEvent: Event information container.
struct CUDBGEventCallbackData: Event information passed to callback set with setNotifyNewEventCallback function.
struct CUDBGEventCallbackData40: Event information passed to callback set with setNotifyNewEventCallback function.

Typedefs

typedef void ( *CUDBGNotifyNewEventCallback )( CUDBGEventCallbackData* data ): function type of the function called to notify debugger of the presence of a new event in the event queue.
typedef void ( *CUDBGNotifyNewEventCallback31 )( void* data ): function type of the function called to notify debugger of the presence of a new event in the event queue.

Enumerations

enum CUDBGEventKind: CUDA Kernel Events.

Variables

CUDBGResult ( *CUDBGAPI_st::acknowledgeEvent30 )( CUDBGEvent30* event ): Inform the debugger API that the event has been processed.
CUDBGResult ( *CUDBGAPI_st::acknowledgeEvents42 )( ): Inform the debugger API that synchronous events have been processed.
CUDBGResult ( *CUDBGAPI_st::acknowledgeSyncEvents )( ): Inform the debugger API that synchronous events have been processed.
CUDBGResult ( *CUDBGAPI_st::getNextAsyncEvent50 )( CUDBGEvent50* event ): Copies the next available event in the asynchronous event queue into 'event' and removes it from the queue. The asynchronous event queue is held separate from the normal event queue, and does not require acknowledgement from the debug client.
CUDBGResult ( *CUDBGAPI_st::getNextAsyncEvent55 )( CUDBGEvent55* event ): Copies the next available event in the asynchronous event queue into 'event' and removes it from the queue. The asynchronous event queue is held separate from the normal event queue, and does not require acknowledgement from the debug client.
CUDBGResult ( *CUDBGAPI_st::getNextEvent )( CUDBGEventQueueType type, CUDBGEvent* event ): Copies the next available event into 'event' and removes it from the queue.
CUDBGResult ( *CUDBGAPI_st::getNextEvent30 )( CUDBGEvent30* event ): Copies the next available event in the event queue into 'event' and removes it from the queue.
CUDBGResult ( *CUDBGAPI_st::getNextEvent32 )( CUDBGEvent32* event ): Copies the next available event in the event queue into 'event' and removes it from the queue.
CUDBGResult ( *CUDBGAPI_st::getNextEvent42 )( CUDBGEvent42* event ): Copies the next available event in the event queue into 'event' and removes it from the queue.
CUDBGResult ( *CUDBGAPI_st::getNextSyncEvent50 )( CUDBGEvent50* event )
CUDBGResult ( *CUDBGAPI_st::getNextSyncEvent55 )( CUDBGEvent55* event ): Copies the next available event in the synchronous event queue into 'event' and removes it from the queue.
CUDBGResult ( *CUDBGAPI_st::setNotifyNewEventCallback )( CUDBGNotifyNewEventCallback callback ): Provides the API with the function to call to notify the debugger of a new application or device event.
CUDBGResult ( *CUDBGAPI_st::setNotifyNewEventCallback31 )( CUDBGNotifyNewEventCallback31 callback, void* data ): Provides the API with the function to call to notify the debugger of a new application or device event.
CUDBGResult ( *CUDBGAPI_st::setNotifyNewEventCallback40 )( CUDBGNotifyNewEventCallback40 callback ): Provides the API with the function to call to notify the debugger of a new application or device event.

Typedefs

void ( *CUDBGNotifyNewEventCallback )( CUDBGEventCallbackData* data ): function type of the function called to notify debugger of the presence of a new event in the event queue.
void ( *CUDBGNotifyNewEventCallback31 )( void* data ): function type of the function called to notify debugger of the presence of a new event in the event queue.

Enumerations

enum CUDBGEventKind

Values

CUDBG_EVENT_INVALID = 0x000: Invalid event.
CUDBG_EVENT_ELF_IMAGE_LOADED = 0x001: The ELF image for a CUDA source module is available.
CUDBG_EVENT_KERNEL_READY = 0x002: A CUDA kernel is about to be launched.
CUDBG_EVENT_KERNEL_FINISHED = 0x003: A CUDA kernel has terminated.
CUDBG_EVENT_INTERNAL_ERROR = 0x004: An internal error occur. The debugging framework may be unstable.
CUDBG_EVENT_CTX_PUSH = 0x005: A CUDA context was pushed.
CUDBG_EVENT_CTX_POP = 0x006: A CUDA CTX was popped.
CUDBG_EVENT_CTX_CREATE = 0x007: A CUDA CTX was created.
CUDBG_EVENT_CTX_DESTROY = 0x008: A CUDA context was destroyed.
CUDBG_EVENT_TIMEOUT = 0x009: An timeout event is sent at regular interval. This event can safely ge ignored.
CUDBG_EVENT_ATTACH_COMPLETE = 0x00a: The attach process has completed and debugging of device code may start.
CUDBG_EVENT_DETACH_COMPLETE = 0x00b
CUDBG_EVENT_ELF_IMAGE_UNLOADED = 0x00c
CUDBG_EVENT_FUNCTIONS_LOADED = 0x00d

Variables

CUDBGResult ( *CUDBGAPI_st::acknowledgeEvent30 )( CUDBGEvent30* event )

Inform the debugger API that the event has been processed. Since CUDA 3.0.

Parameters

event: - pointer to the event that has been processed

Returns

CUDBG_SUCCESS

CUDBGResult ( *CUDBGAPI_st::acknowledgeEvents42 )( )

Inform the debugger API that synchronous events have been processed. Since CUDA 3.1.

Returns

CUDBG_SUCCESS

CUDBGResult ( *CUDBGAPI_st::acknowledgeSyncEvents )( )

Inform the debugger API that synchronous events have been processed. Since CUDA 5.0.

Returns

CUDBG_SUCCESS

CUDBGResult ( *CUDBGAPI_st::getNextAsyncEvent50 )( CUDBGEvent50* event )

Copies the next available event in the asynchronous event queue into 'event' and removes it from the queue. The asynchronous event queue is held separate from the normal event queue, and does not require acknowledgement from the debug client. Since CUDA 5.0.