cuda | cuda计时cudaevent C++&python

• C++
  #include "driver_types.h"

float time_elapsed = 0;
cudaEvent_t start, stop;
cudaEventCreate(&start);      // 创建Event
cudaEventCreate(&stop);
cudaEventRecord(start, 0);     // 记录当前时间

m_Context->enqueue(m_BatchcSize, m_Bindings.data(), m_CudaStream, nullptr);

cudaEventRecord(stop, 0);   // 记录当前时间
cudaEventSynchronize(start);     // Watis for an event to complate
cudaEventSynchronize(stop);      // Waits for an event to complate.Record 之前的任务
cudaEventElapsedTime(&time_elapsed, start, stop);    // 计算时间差
cudaEventDestroy(start);         // destory the event
cudaEventDestroy(stop);
std::cout << "time one map for a batch = " << time_elapsed << std::endl;

• Python
  import cupy

start = cupy.cuda.Event()
end = cupy.cuda.Event()
start.record()
m.run()
end.record()
start.synchronize()
end.synchronize()
time_elapsed = cupy.cuda.get_elapsed_time(start, end)
print("running time: ", time_elapsed)