Why clEnqueueNDRangeKernel does not return to the host immediately? - opencl

Question

I checked out some answers, which indicates an immediate return after clEnqueueNDRangeKernel starts (for example in this link the 1st question). But in my case, the kernel executes serially (one by one).
A simple example of my code:
cl_int err;
cl_platform_id* platforms;
cl_device_id* devices;
cl_context context;
cl_command_queue queue;
cl_int gpu_platform_id;
SetupPlatform(&platforms, &gpu_platform_id);
SetupDevice(platforms, &devices, gpu_platform_id);
SetupContext(platforms, devices, &context, 1, gpu_platform_id);
SetupQueue(&queue, context, devices);
cl_program programs[2];
cl_event events[3];
...
programs[0] = SetupBuildProgramWithBinary(context, devices, "Kernel1_Opt.bin");
programs[1] = SetupBuildProgramWithBinary(context, devices, "Kernel2_Opt.bin");
cl_kernel kernels[2];
char kernel_name[][50] = { "kernel1","kernel2"};
SetupKernel(kernels, programs, 2, kernel_name);
...(some host data preparation)
err = clEnqueueNDRangeKernel(queue, kernels[0], 3, 0, kernel1_global_size, kernel1_local_size,
0, NULL, &events[0]); checkErr(err);
clWaitForEvents(1, &events[0]);
for(int i=0;i<2;i++){
...(some host data preparation)
err = clEnqueueNDRangeKernel(queue, kernels[1], 2, 0, kernel2_global_size,
kernel2_local_size, 0, NULL, &events[i+1]); checkErr(err);
}
I want the kernel execution in the for loop to be asynchronousso that the host data preparation in the second loop can be executed while the kernel in the first loop is executing. But the fact is that clEnqueueNDRangeKernel does not return immediately to the host, so I can not process the data preparation in the second loop immediately.
My queue setup is:
cl_int err = 0;
cl_command_queue_properties props[] = { CL_QUEUE_PROFILING_ENABLE |
CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE };
//choose 0th device
*queue = clCreateCommandQueue(context, devices[0], *props, &err); checkErr(err);
I am using Windows OS, NVIDIA RTX 3090 GPU card, kernels are compiled with OpenCL 3.0.