Wrong values when using CL_USE_HOST_PTR on a Cyclone V SoC

[JButt5]

For research purposes regarding portability of OpenCL applications I set up a test application that measures the latency of memory enqueueing/dequeueing operations for the various possible API calls OpenCL offers. While generating correct results over a broad variety of platforms, operating systems, etc. it fails in case of constructing the buffer via CL_USE_HOST_PTR when built for a Cyclone V SoC where the host application runs on the ARM. The error that can be observed is that there are simply invalid values.

I read through some docs, hover I found no mention regarding that this should not work when targeting Intel FPGA SoCs. Did I overlook something, is the code below in some ways incorrect or have I run into a bug?

This is the relevant part of the code where the error happens (note that CL_ALLOC_HOST_PTR works fine).

I hope the code snippet below is understandable without the context

nlohmann::json performTestUseHostPtr () { nlohmann::json results;   for (auto size : testSizes) { void* inputBufferHostPtr; void* outputBufferHostPtr;   // Read somewhere in some docs that 64 Bit alignment is required, not sure if this applies to SoCs but just in case... #ifdef OPEN_CL_INTEL_FPGA posix_memalign (&inputBufferHostPtr, 64, size * sizeof (float)); posix_memalign (&outputBufferHostPtr, 64, size * sizeof (float)); #else std::vector<float> inputBufferHost (size); std::vector<float> outputBufferHost (size);   inputBufferHostPtr = inputBufferHost.data(); outputBufferHostPtr = outputBufferHost.data(); #endif   cl_int err;   cl::Buffer inputBuffer (context, CL_MEM_USE_HOST_PTR | CL_MEM_READ_ONLY, size * sizeof (float), inputBufferHostPtr, &err); if (err != CL_SUCCESS) std::cerr << "Error creating input buffer: " << ntlab::OpenCLHelpers::getErrorString (err) << std::endl;   cl::Buffer outputBuffer (context, CL_MEM_USE_HOST_PTR | CL_MEM_WRITE_ONLY, size * sizeof (float), outputBufferHostPtr, &err); if (err != CL_SUCCESS) std::cerr << "Error creating output buffer: " << ntlab::OpenCLHelpers::getErrorString (err) << std::endl;   mapUnmapTestsInternal (size, inputBuffer, outputBuffer);   #ifdef OPEN_CL_INTEL_FPGA free (inputBufferHostPtr); free (outputBufferHostPtr); #endif   results[std::to_string (size)] = calculateTimes(); }   return results; }   nlohmann::json performTestAllocHostPtr() { nlohmann::json results;   for (auto size : testSizes) { cl::Buffer inputBuffer (context, CL_MEM_ALLOC_HOST_PTR | CL_MEM_READ_ONLY, size * sizeof (float)); cl::Buffer outputBuffer (context, CL_MEM_ALLOC_HOST_PTR | CL_MEM_WRITE_ONLY, size * sizeof (float));   mapUnmapTestsInternal (size, inputBuffer, outputBuffer);   results[std::to_string (size)] = calculateTimes(); }   return results; }   void mapUnmapTestsInternal (size_t size, cl::Buffer& inputBuffer, cl::Buffer& outputBuffer) { std::vector<cl::Event> inputBufferEvent (1); std::vector<cl::Event> outputBufferEvent (1); std::vector<cl::Event> kernelEvent (1);   #ifdef OPEN_CL_INTEL_FPGA void* inputBufferMapped = queue.enqueueMapBuffer (inputBuffer, CL_TRUE, CL_MAP_WRITE, 0, size * sizeof (float)); #else void* inputBufferMapped = queue.enqueueMapBuffer (inputBuffer, CL_TRUE, CL_MAP_WRITE_INVALIDATE_REGION, 0, size * sizeof (float)); #endif std::memcpy (inputBufferMapped, testData.data(), size * sizeof (float));   kernel.setArg (0, inputBuffer); kernel.setArg (1, outputBuffer);   for (int i = 0; i < numTests; ++i) { startTimes[i] = ntlab::HighResolutionTimer::now();   queue.enqueueUnmapMemObject (inputBuffer, inputBufferMapped, nullptr, &inputBufferEvent[0]); inputBufferEvent[0].setCallback (CL_COMPLETE, setTimestampCallback, &writeBufferCompletedTimes[i]);   queue.enqueueNDRangeKernel (kernel, cl::NullRange, cl::NDRange (size), cl::NullRange, &inputBufferEvent, &kernelEvent[0]); kernelEvent[0].setCallback (CL_COMPLETE, setTimestampCallback, &kernelCompletedTimes[i]);   void* outputBufferMapped = queue.enqueueMapBuffer (outputBuffer, CL_FALSE, CL_MAP_READ, 0, size * sizeof (float), &kernelEvent, &outputBufferEvent[0]); outputBufferEvent[0].setCallback (CL_COMPLETE, setTimestampCallback, &readBufferCompletedTimes[i]);   #ifdef OPEN_CL_INTEL_FPGA inputBufferMapped = queue.enqueueMapBuffer (inputBuffer, CL_TRUE, CL_MAP_WRITE, 0, size * sizeof (float), &kernelEvent); #else inputBufferMapped = queue.enqueueMapBuffer (inputBuffer, CL_TRUE, CL_MAP_WRITE_INVALIDATE_REGION, 0, size * sizeof (float), &kernelEvent); #endif queue.finish(); std::memcpy (inputBufferMapped, outputBufferMapped, size * sizeof (float));   queue.enqueueUnmapMemObject (outputBuffer, outputBufferMapped); queue.finish(); }   // make sure the results are valid auto* outputBufferMapped = static_cast<float*> (queue.enqueueMapBuffer (outputBuffer, CL_TRUE, CL_MAP_READ, 0, size * sizeof (float))); for (int i = 0; i < size; ++i) if (testData[i] != outputBufferMapped[i]) { std::cerr << "Error: Unexpected value at position " << i << ", expected " << testData[i] << ", got " << outputBufferMapped[i] << std::endl; break; }   queue.enqueueUnmapMemObject (outputBuffer, outputBufferMapped); queue.enqueueUnmapMemObject (inputBuffer, inputBufferMapped); queue.finish (); }

[FawazJ_Altera]

Hello sir, I will look into your code and let you know the feedback soon, Thanks

[FawazJ_Altera]

Hello, The Allocating Shared Memory for OpenCL Kernels Targeting SoCs section of the programming guide mentions that CL_MEM_USE_HOST_PTR doesn’t work with shared memory configurations on SoCs. This appears to be the scenario that you are using now, so it’s expected not to work. In that section, the programming guide describes how to write host code to share data between the ARM CPU and FPGA. I suggest the customer review that example. https://www.intel.com/content/www/us/en/programmable/documentation/iee1516128926240.html Hope this might help. Thanks