Concurrent Kernel Execution

This example will demonstrate how to use multiple and out of order command queues to simultaneously execute multiple kernels on an FPGA.

KEY CONCEPTS: Concurrent execution, Out of Order Command Queues, Multiple Command Queues

KEYWORDS: CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, setCallback

This example illustrates two ways to implement concurrent kernel execution. 1. Multiple command enqueues 2. Single command queue with out of order execution

Design contains 3 kernels mscale, madd and mmult. madd needs to wait for mscale to complete but mmult can run independently (no dependency).

multiple_command_queues function creates two sequential command queues, one for mscale and madd and other for mmult.

cl::CommandQueue ordered_queue1(context, device, CL_QUEUE_PROFILING_ENABLE, &err);
cl::CommandQueue ordered_queue2(context, device, CL_QUEUE_PROFILING_ENABLE, &err);

Both Command queues are created as in-order execution. msacle and madd are enqueued in first command queue one after another. As Command queue is in-order, madd will only start when mscale will finish. So the order of execution is handled by command queue.

There is another approach of achieve the same using single out-of-order command queue. out_of_order_queue function creates a single command queue with out of order execution enabled which means that enqueues in the command queue can go out of order (without waiting for a previous execution to finish).

cl::CommandQueue ooo_queue(context, device,
          CL_QUEUE_PROFILING_ENABLE | CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE);

However, since we need that madd kernel only starts executing after mscale has finished, cl::event is used to wait for mscale to finish execution.

cl::Event event;
vector<cl::Event> kernel_wait_events
kernel_wait_events.push_back(event);
err = ooo_queue.enqueueNDRangeKernel(kernel_mscale, offset, global, local, nullptr, &event));

err = ooo_queue.enqueueNDRangeKernel(kernel_madd, offset, global, local,
                                          &kernel_wait_events, // Event from previous call
                                          nullptr);

EXCLUDED PLATFORMS:

• All NoDMA Platforms, i.e u50 nodma etc

DESIGN FILES

Application code is located in the src directory. Accelerator binary files will be compiled to the xclbin directory. The xclbin directory is required by the Makefile and its contents will be filled during compilation. A listing of all the files in this example is shown below

src/host.cpp
src/madd.cpp
src/mmult.cpp
src/mscale.cpp

Access these files in the github repo by clicking here.

COMMAND LINE ARGUMENTS

Once the environment has been configured, the application can be executed by

./concurrent_kernel_execution <matrix_ops XCLBIN>