CUDA:Event
lastmod 2024-07-22created 2024-07-206 edits
This section describes the event management functions of the CUDA runtime application programming interface.
About
Jihunlee25.tistory.com_-_cuda_event.png
Leaves a mark at a specific point in the flow of commands inside a stream (a kind of mark → marker)
Because the GPU runs separately from (in parallel with) the CPU, checking the time on the CPU is not synchronized. Therefore, Event synchronization logic that runs in a separate flow for the GPU (CUDA:Stream) is needed.
Functions
Creation and Destruction
-
__host__ cudaError_t cudaEventCreate ( cudaEvent_t* event ) - Creates an event object.
-
__host__ __device__ cudaError_t cudaEventCreateWithFlags ( cudaEvent_t* event, unsigned int flags ) - Creates an event object with the specified flags.
-
__host__ __device__ cudaError_t cudaEventDestroy ( cudaEvent_t event ) - Destroys an event object.
Recording Events
-
__host__ __device__ cudaError_t cudaEventRecord ( cudaEvent_t event, cudaStream_t stream = 0 ) - Records an event.
- Marks the event position in the stream
- Information is recorded in the event when it is taken out of the stream
-
__host__ cudaError_t cudaEventRecordWithFlags ( cudaEvent_t event, cudaStream_t stream = 0, unsigned int flags = 0 ) - Records an event with the specified flags.
Synchronization / Querying
-
__host__ cudaError_t cudaEventSynchronize ( cudaEvent_t event ) - Waits for an event to complete.
- Waits until the event occurs
-
__host__ cudaError_t cudaEventQuery ( cudaEvent_t event ) - Queries an event's status.
- Checks whether the event has occurred
Measure the elasped time
-
__host__ cudaError_t cudaEventElapsedTime ( float* ms, cudaEvent_t start, cudaEvent_t end ) - Computes the elapsed time between events.
- Computes the time elapsed between start and end in milliseconds
Example
Example of measuring time with Events:
//create two events
cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start); //record start event on the default stream
kernel <<< grid, block >>> (arguments); //execute kernel
cudaEventRecord(stop); //record stop event on the default stream
cudaEventSynchronize(stop); //wait until the stop event completes
//calculate the elapsed time between two events
float time;
cudaEventElapsedTime(&time, start, stop);
//clean up the two events
cudaEventDestroy(start);
cudaEventDestroy(stop);