A newer version of this document is available. Customers should click here to go to the newest version.
Visible to Intel only — GUID: GUID-A9245E47-A740-4337-B926-E1724844B497
Visible to Intel only — GUID: GUID-A9245E47-A740-4337-B926-E1724844B497
Targeting Multiple Homogeneous FPGA Devices
The Intel® oneAPI DPC++/C++ Compiler supports targeting multiple homogeneous FPGA devices from a single host CPU. This allows to improve your design’s throughput by parallelizing the execution of your program on multiple FPGAs.
Intel® recommends creating a single context with multiple device queues because, with multi-context, buffers at OpenCL layer must be copied between contexts, which introduces overhead and impacts overall performance. However, you can use multi-context if your design is simple and the overhead does not affect the overall performance.
Follow one of the following methods to target multiple FPGA devices:
Create a Single Context with Multiple Device Queues
Perform the following steps to target multiple FPGA devices with a single context:
Create a single SYCL* context to encapsulate a collection of FPGA devices of the same platform.
context ctxt(deviceList, &m_exception_handler);
Create a SYCL queue for each FPGA device.
std::vector<queue> queueList; for (unsigned int i = 0; i < ctxt.get_devices().size(); i++) { queue newQueue(ctxt, ctxt.get_devices()[i], &m_exception_handler); queueList.push_back(newQueue); }Submit either the same or different device codes to all available FPGA devices. If you want to target a subset of all available devices, then you must first perform device selection to filter out unwanted devices.
for (unsigned int i = 0; i < queueList.size(); i++) { queueList[i].submit([&](handler& cgh) {...}); }
Create a Context For Each Device Queue (Multi-Context)
Perform the following steps to target multiple FPGA devices with multiple contexts:
Obtain a list of all available FPGA devices. Optionally, you can select a device based on the device member or device property. For device properties such as device name, use the member function get_info()const with the desired device property.
std::vector<device> deviceList = device::get_devices();
Create a SYCL queue for each FPGA device.
std::vector<queue> queueList; for (unsigned int i = 0; i < deviceList.size(); i++) { queue newQueue(deviceList[i], &m_exception_handler); queueList.push_back(newQueue); }Submit either the same or different device codes to all available FPGA devices. If you want to target a subset of all available devices, then you must first perform device selection to filter out unwanted devices.
for (unsigned int i = 0; i < queueList.size(); i++) { queueList[i].submit([&](handler& cgh) {...}); }
Limitations
Consider the following limitations when targeting multiple FPGA devices:
All FPGA devices use the same FPGA bitstream.
All FPGA devices used must be of the same FPGA card (same -Xstarget target)