"Zero copy" is slower in my OpenCL/Cloo(C#) program than non-zero copy

Question

This may simply be an issue with the memory objects being allocated by the .NET framework not being properly page-aligned, but I cannot see why zero-copy is slower for me than non-zero copy.

I'll include code inline in this question, but the complete source can be seen here: https://github.com/kwende/ClooMatrixMultiply/blob/master/GiantMatrixOnGPU/GPUMatrixMultiplier.cs.

Since this is my first attempt at getting zero-copy working, I wrote up a simple matrix multiplication example. I first initialize my OpenCL objects:

    private void Initialize()
    {
        // get the intel integrated GPU
        _integratedIntelGPUPlatform = ComputePlatform.Platforms.Where(n => n.Name.Contains("Intel")).First();

        // create the compute context. 
        _context = new ComputeContext(
            ComputeDeviceTypes.Gpu, // use the gpu
            new ComputeContextPropertyList(_integratedIntelGPUPlatform), // use the intel openCL platform
            null,
            IntPtr.Zero);

        // the command queue is the, well, queue of commands sent to the "device" (GPU)
        _commandQueue = new ComputeCommandQueue(
            _context, // the compute context
            _context.Devices[0], // first device matching the context specifications
            ComputeCommandQueueFlags.None); // no special flags

        string kernelSource = null;
        using (StreamReader sr = new StreamReader("kernel.cl"))
        {
            kernelSource = sr.ReadToEnd();
        }

        // create the "program"
        _program = new ComputeProgram(_context, new string[] { kernelSource });

        // compile. 
        _program.Build(null, null, null, IntPtr.Zero);
        _kernel = _program.CreateKernel("ComputeMatrix");
    }

...this is only executed once if my code hasn't been initialized. Then I get into the main body. For non-zero copy, I do the following:

  public float[] MultiplyMatrices(float[] matrix1, float[] matrix2,
  int matrix1Height, int matrix1WidthMatrix2Height, int matrix2Width)
  {
        if (!_initialized)
        {
            Initialize();
            _initialized = true;
        }

        ComputeBuffer<float> matrix1Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix1);
        _kernel.SetMemoryArgument(0, matrix1Buffer);

        ComputeBuffer<float> matrix2Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix2);
        _kernel.SetMemoryArgument(1, matrix2Buffer);

        float[] ret = new float[matrix1Height * matrix2Width];
        ComputeBuffer<float> retBuffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.CopyHostPointer,
            ret);
        _kernel.SetMemoryArgument(2, retBuffer);

        _kernel.SetValueArgument<int>(3, matrix1WidthMatrix2Height);
        _kernel.SetValueArgument<int>(4, matrix2Width);

        _commandQueue.Execute(_kernel,
            new long[] { 0 },
            new long[] { matrix2Width, matrix1Height },
            null, null);

        unsafe
        {
            fixed (float* retPtr = ret)
            {
                _commandQueue.Read(retBuffer,
                    false, 0,
                    ret.Length,
                    new IntPtr(retPtr),
                    null);

                _commandQueue.Finish();
            }
        }

        matrix1Buffer.Dispose();
        matrix2Buffer.Dispose();
        retBuffer.Dispose();

        return ret;
    }

You can see how I'm explicitly setting CopyHostPointer for all of my ComputeBuffer allocations. This executes fine.

I then do the following adjustment to (which includes setting "UseHostPointer" and calling Map/Unmap instead of Read):

    public float[] MultiplyMatricesZeroCopy(float[] matrix1, float[] matrix2,
        int matrix1Height, int matrix1WidthMatrix2Height, int matrix2Width)
    {
        if (!_initialized)
        {
            Initialize();
            _initialized = true;
        }

        ComputeBuffer<float> matrix1Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix1);
        _kernel.SetMemoryArgument(0, matrix1Buffer);

        ComputeBuffer<float> matrix2Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix2);
        _kernel.SetMemoryArgument(1, matrix2Buffer);

        float[] ret = new float[matrix1Height * matrix2Width];
        ComputeBuffer<float> retBuffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadWrite | ComputeMemoryFlags.UseHostPointer,
            ret);
        _kernel.SetMemoryArgument(2, retBuffer);

        _kernel.SetValueArgument<int>(3, matrix1WidthMatrix2Height);
        _kernel.SetValueArgument<int>(4, matrix2Width);

        _commandQueue.Execute(_kernel,
            new long[] { 0 },
            new long[] { matrix2Width, matrix1Height },
            null, null);

        IntPtr retPtr = _commandQueue.Map(
            retBuffer,
            false,
            ComputeMemoryMappingFlags.Read,
            0,
            ret.Length, null);

        _commandQueue.Unmap(retBuffer, ref retPtr, null);
        _commandQueue.Finish();

        matrix1Buffer.Dispose();
        matrix2Buffer.Dispose();
        retBuffer.Dispose();

        return ret;
    }

The timing says it all, however. My program spits this out:

CPU Matrix multiplication: 1178.5ms

GPU Matrix multiplication (copy): 115.1ms

GPU Matrix multiplication (zero copy): 174.1ms

GPU (w/ copy) is 10.23892x faster.

GPU (zero copy) is 6.769098x faster.

...so zero copy is slower.

Answer 1

Thanks to huseyin tugrul buyukisik I was able to figure out what was going on.

I needed to update my Intel drivers. Once I did this, then the zero-copy was much, much faster.

For the sake of posterity, here is the final version of the zero-copy code:

    public float[] MultiplyMatricesZeroCopy(float[] matrix1, float[] matrix2,
        int matrix1Height, int matrix1WidthMatrix2Height, int matrix2Width)
    {
        if (!_initialized)
        {
            Initialize();
            _initialized = true;
        }

        ComputeBuffer<float> matrix1Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix1);
        _kernel.SetMemoryArgument(0, matrix1Buffer);

        ComputeBuffer<float> matrix2Buffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.ReadOnly | ComputeMemoryFlags.CopyHostPointer,
            matrix2);
        _kernel.SetMemoryArgument(1, matrix2Buffer);

        float[] ret = new float[matrix1Height * matrix2Width];
        GCHandle handle = GCHandle.Alloc(ret, GCHandleType.Pinned); 
        ComputeBuffer<float> retBuffer = new ComputeBuffer<float>(_context,
            ComputeMemoryFlags.UseHostPointer,
            ret);
        _kernel.SetMemoryArgument(2, retBuffer);

        _kernel.SetValueArgument<int>(3, matrix1WidthMatrix2Height);
        _kernel.SetValueArgument<int>(4, matrix2Width);

        _commandQueue.Execute(_kernel,
            new long[] { 0 },
            new long[] { matrix2Width, matrix1Height },
            null, null);

        IntPtr retPtr = _commandQueue.Map(
            retBuffer,
            true,
            ComputeMemoryMappingFlags.Read,
            0,
            ret.Length, null);

        _commandQueue.Unmap(retBuffer, ref retPtr, null);
        //_commandQueue.Finish();

        matrix1Buffer.Dispose();
        matrix2Buffer.Dispose();
        retBuffer.Dispose();
        handle.Free(); 

        return ret;
    }