public static void cuEventRecord(CUevent hEvent, CUstream hStream)
{
Wrap(() =>
{
try
{
var error = nativeEventRecord(hEvent, hStream);
if (error != CUresult.CUDA_SUCCESS)
{
throw new CudaException(error);
}
}
catch (CudaException)
{
throw;
}
catch (DllNotFoundException dnfe)
{
throw new CudaException(CudaError.NoDriver, dnfe);
}
catch (Exception e)
{
throw new CudaException(CudaError.Unknown, e);
}
});
}
/// <summary>
/// Records an event. If <c>stream</c> is non-zero, the event is recorded after all preceding operations in the stream have been
/// completed; otherwise, it is recorded after all preceding operations in the CUDA context have been completed. Since
/// operation is asynchronous, <see cref="Query"/> and/or <see cref="Synchronize"/> must be used to determine when the event
/// has actually been recorded. <para/>
/// If <see cref="Record()"/> has previously been called and the event has not been recorded yet, this function throws
/// <see cref="CUResult.ErrorInvalidValue"/>.
/// </summary>
public void Record()
{
CUstream _stream = new CUstream();
res = DriverAPINativeMethods.Events.cuEventRecord(_event, _stream);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuEventRecord", res));
if (res != CUResult.Success) throw new CudaException(res);
}
/// <summary>
/// Creates a new 1D FFT plan (new API)
/// </summary>
/// <param name="handle">cufftHandle object</param>
/// <param name="nx">The transform size (e.g., 256 for a 256-point FFT)</param>
/// <param name="type">The transform data type (e.g., C2C for complex to complex)</param>
/// <param name="batch">Number of transforms of size nx</param>
/// <param name="stream">A valid CUDA stream created with cudaStreamCreate() (or 0 for the default stream)</param>
/// <param name="mode">The <see cref="Compatibility"/> option to be used</param>
/// <param name="size"></param>
/// <param name="autoAllocate">indicates that the caller intends to allocate and manage
/// work areas for plans that have been generated.</param>
public CudaFFTPlan1D(cufftHandle handle, int nx, cufftType type, int batch, CUstream stream, Compatibility mode, ref SizeT size, bool autoAllocate)
: this(handle, nx, type, batch, ref size)
{
SetStream(stream);
SetCompatibilityMode(mode);
SetAutoAllocation(autoAllocate);
}
public CUstream CreateStream(StreamFlags flags)
{
CUstream phStream = new CUstream();
this.LastError = CUDADriver.cuStreamCreate(ref phStream, (uint)flags);
return(phStream);
}
public static extern nvjpegStatus nvjpegDecodeBatched(
nvjpegHandle handle,
nvjpegJpegState jpeg_handle,
IntPtr[] data,
SizeT[] lengths,
nvjpegImage[] destinations,
CUstream stream);
/// <summary>
/// Creates a new 2D FFT plan (new API)
/// </summary>
/// <param name="handle">cufftHandle object</param>
/// <param name="nx">The transform size in the X dimension (number of rows)</param>
/// <param name="ny">The transform size in the Y dimension (number of columns)</param>
/// <param name="type">The transform data type (e.g., C2R for complex to real)</param>
/// <param name="stream">A valid CUDA stream created with cudaStreamCreate() (or 0 for the default stream)</param>
/// <param name="mode">The <see cref="Compatibility"/> option to be used</param>
/// <param name="autoAllocate">indicates that the caller intends to allocate and manage
/// work areas for plans that have been generated.</param>
public CudaFFTPlan2D(cufftHandle handle, int nx, int ny, cufftType type, CUstream stream, Compatibility mode, bool autoAllocate)
: this(handle, nx, ny, type)
{
SetStream(stream);
SetCompatibilityMode(mode);
SetAutoAllocation(autoAllocate);
}
public CUdeviceptr CopyHostToDeviceAsync(IntPtr buffer, uint size, CUstream stream)
{
CUdeviceptr devPtr = this.Allocate(size);
this.CopyHostToDeviceAsync(devPtr, buffer, size, stream);
return(devPtr);
}
public static extern nvjpegStatus nvjpegDecodeJpeg(
nvjpegHandle handle,
nvjpegJpegDecoder decoder,
nvjpegJpegState decoder_state,
nvjpegJpegStream jpeg_bitstream,
ref nvjpegImage destination,
nvjpegDecodeParams decode_params,
CUstream stream);
/// <summary>
/// Set the current stream for CURAND kernel launches. All library functions
/// will use this stream until set again.
/// </summary>
/// <param name="stream"></param>
public void SetStream(CUstream stream)
{
_status = CudaRandNativeMethods.curandSetStream(_generator, stream);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "curandSetStream", _status));
if (_status != CurandStatus.Success)
{
throw new CudaRandException(_status);
}
}
/// <summary>
/// Associates a CUDA stream with a CUFFT plan. All kernel launches
/// made during plan execution are now done through the associated
/// stream, enabling overlap with activity in other streams (for example,
/// data copying). The association remains until the plan is destroyed or
/// the stream is changed with another call to SetStream().
/// </summary>
/// <param name="stream">A valid CUDA stream created with cudaStreamCreate() (or 0 for the default stream)</param>
public void SetStream(CUstream stream)
{
res = CudaFFTNativeMethods.cufftSetStream(_handle, stream);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cufftSetStream", res));
if (res != cufftResult.Success)
{
throw new CudaFFTException(res);
}
}
public static extern nvjpegStatus nvjpegEncodeImage(
nvjpegHandle handle,
nvjpegEncoderState encoder_state,
nvjpegEncoderParams encoder_params,
ref nvjpegImage source,
nvjpegInputFormat input_format,
int image_width,
int image_height,
CUstream stream);
public void launchCopyKernel(physx.PxGpuCopyDescPtr desc, uint count, CUstream stream)
{
physx.PxGpuDispatcherPtr pvk_in_this = this;
physx.PxGpuCopyDescPtr pvk_in_desc = desc;
uint pvk_in_count = count;
CUstream pvk_in_stream = stream;
void_PxGpuDispatcherPtr_launchCopyKernelPtr_PxGpuCopyDescPtr_uint_CUstream_(pvk_in_this, pvk_in_desc, pvk_in_count, pvk_in_stream);
}
public override void Init()
{
cuda = new CUDA(0, true);
var cuCtx = cuda.CreateContext(0, CUCtxFlags.MapHost);
cuda.SetCurrentContext(cuCtx);
cuModule = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, cudaModuleName));
cuFunc = cuda.GetModuleFunction(cudaEvaluatorKernelName);
cuFuncSign = cuda.GetModuleFunction(cudaSignKernelName);
//reserved memory based on dimension of support vector
//svVector = new float[TrainedModel.SupportElements[0].Count];
stream = cuda.CreateStream();
//memSvSize = (uint)(TrainedModel.SupportElements[0].Count * sizeof(float));
memSvSize = (uint)(TrainedModel.SupportElements[0].Dim * sizeof(float));
//allocates memory for buffers
svVecIntPtrs[0] = cuda.AllocateHost(memSvSize);
svVecIntPtrs[1] = cuda.AllocateHost(memSvSize);
mainVecPtr = cuda.CopyHostToDeviceAsync(svVecIntPtrs[0], memSvSize, stream);
cuSVTexRef = cuda.GetModuleTexture(cuModule, "svTexRef");
cuda.SetTextureFlags(cuSVTexRef, 0);
cuda.SetTextureAddress(cuSVTexRef, mainVecPtr, memSvSize);
//todo: copy labels and alphas
float[] svLabels = new float[TrainedModel.SupportElements.Length];
float[] svAlphas = new float[TrainedModel.SupportElements.Length];
Parallel.For(0, TrainedModel.SupportElementsIndexes.Length,
i => {
int idx = TrainedModel.SupportElementsIndexes[i];
svLabels[i] = TrainedModel.Y[i];
//svLabels[i] = TrainningProblem.Labels[idx];
svAlphas[i] = TrainedModel.Alpha[idx];
});
//for (int i = 0; i < TrainedModel.SupportElementsIndexes.Length; i++)
//{
// int idx = TrainedModel.SupportElementsIndexes[i];
// svLabels[i]= TrainningProblem.Labels[idx];
// svAlphas[i] = TrainedModel.Alpha[idx];
//}
labelsPtr = cuda.CopyHostToDevice(svLabels);
alphasPtr = cuda.CopyHostToDevice(svAlphas);
IsInitialized = true;
}
public static extern nvjpegStatus nvjpegEncodeYUV(
nvjpegHandle handle,
nvjpegEncoderState encoder_state,
nvjpegEncoderParams encoder_params,
ref nvjpegImage source,
nvjpegChromaSubsampling chroma_subsampling,
int image_width,
int image_height,
CUstream stream);
public override void SetStream(int streamId)
{
if (streamId < 0)
{
throw new ArgumentOutOfRangeException("streamId");
}
CUstream cus = (CUstream)_gpu.GetStream(streamId);
SafeCall(_driver.SetStream(_gen, cus));
}
public void DestroyStream(CUstream stream)
{
if (_version >= 4000)
{
this.LastError = CUDADriver.cuStreamDestroy_v2(stream);
}
else
{
this.LastError = CUDADriver.cuStreamDestroy(stream);
}
}
/// <summary>
/// Creates a new Stream using <see cref="CUStreamFlags.None"/> and with the given priority<para/>
/// This API alters the scheduler priority of work in the stream. Work in a higher priority stream
/// may preempt work already executing in a low priority stream.<para/>
/// <c>priority</c> follows a convention where lower numbers represent higher priorities.<para/>
/// '0' represents default priority.
/// </summary>
/// <param name="priority">Stream priority. Lower numbers represent higher priorities.</param>
/// <param name="flags">Parameters for stream creation (must be <see cref="CUStreamFlags.None"/>)</param>
public CudaStream(int priority, CUStreamFlags flags)
{
_stream = new CUstream();
res = DriverAPINativeMethods.Streams.cuStreamCreateWithPriority(ref _stream, flags, priority);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuStreamCreateWithPriority", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
_isOwner = true;
}
/// <summary>
/// Asynchron copy device to host
/// </summary>
/// <param name="deviceVar"></param>
/// <param name="stream"></param>
public void AsyncCopyFromDevice(CudaDeviceVariable <T> deviceVar, CUstream stream)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
res = DriverAPINativeMethods.AsynchronousMemcpy_v2.cuMemcpyDtoHAsync_v2(_intPtr, deviceVar.DevicePointer, SizeInBytes, stream);
Debug.Write("");//Line(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuMemcpyDtoHAsync", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
}
/// <summary>
/// This function gets the stream to be used by the cudnn library to execute its routines.
/// </summary>
public CudaStream GetStream()
{
CUstream stream = new CUstream();
res = CudaDNNNativeMethods.cudnnGetStream(_handle, ref stream);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cudnnGetStream", res));
if (res != cudnnStatus.Success)
{
throw new CudaDNNException(res);
}
return(new CudaStream(stream));
}
/// <summary>
/// Asynchron copy 1D Array to host
/// </summary>
/// <param name="deviceArray"></param>
/// <param name="stream"></param>
/// <param name="offset">bytes</param>
public void AsyncCopyFromArray1D(CUarray deviceArray, CUstream stream, SizeT offset)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
res = DriverAPINativeMethods.AsynchronousMemcpy_v2.cuMemcpyAtoHAsync_v2(this._intPtr, deviceArray, offset, SizeInBytes, stream);
Debug.Write(""); //Line(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuMemcpyAtoHAsync", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
}
/// <summary>
/// Asynchron Copy host to device
/// </summary>
/// <param name="devicePtr"></param>
/// <param name="stream"></param>
public void AsyncCopyToDevice(CUdeviceptr devicePtr, CUstream stream)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
res = DriverAPINativeMethods.AsynchronousMemcpy_v2.cuMemcpyHtoDAsync_v2(devicePtr, _intPtr, SizeInBytes, stream);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuMemcpyHtoDAsync", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
}
/// <summary>
/// Unmaps the graphics resource.<para/>
/// Once unmapped, the resource may not be accessed by CUDA until they are mapped again.<para/>
/// This function provides the synchronization guarantee that any CUDA work issued in <c>stream</c> before <see cref="UnMap()"/>
/// will complete before any subsequently issued graphics work begins.<para/>
/// If the resource is not presently mapped for access by CUDA then <see cref="CUResult.ErrorNotMapped"/> exception is thrown.
/// </summary>
/// <param name="stream"></param>
public void UnMap(CUstream stream)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
res = DriverAPINativeMethods.GraphicsInterop.cuGraphicsUnmapResources(1, ref _cudaResource, stream);
Debug.Write("");//Line(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuGraphicsUnmapResources", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
_IsMapped = false;
}
/// <summary>
/// Sets the stream.
/// </summary>
/// <param name="plan">The plan to set the stream for.</param>
/// <param name="streamId">The stream id.</param>
public override void SetStream(FFTPlan plan, int streamId)
{
if (streamId < 0)
{
throw new ArgumentOutOfRangeException("streamId");
}
CUstream cus = (CUstream)_gpu.GetStream(streamId);
FFTPlanEx planEx = Plans[plan];
cudaStream cs = new cudaStream();
//cs.Value = cus.Pointer.ToInt32();
CUFFTResult res = _driver.cufftSetStream(planEx.CudaFFTHandle, cs);
if (res != CUFFTResult.Success)
{
throw new CudafyMathException(CudafyMathException.csCUDA_EXCEPTION_X, res);
}
}
/// <summary>
/// Unmaps all graphics resources.<para/>
/// Once unmapped, the resources may not be accessed by CUDA until they are mapped again.<para/>
/// This function provides the synchronization guarantee that any CUDA work issued in <c>stream</c> before <see cref="UnmapAllResources()"/>
/// will complete before any subsequently issued graphics work begins.<para/>
/// If any of the resources are not presently mapped for access by CUDA then <see cref="CUResult.ErrorNotMapped"/> exception is thrown.
/// </summary>
/// <param name="stream"></param>
public void UnmapAllResources(CUstream stream)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
CUResult res;
res = DriverAPINativeMethods.GraphicsInterop.cuGraphicsUnmapResources((uint)_CUResources.Count, _CUResources.ToArray(), stream);
Debug.Write("");//Line(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuGraphicsUnmapResources", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
foreach (var elem in _resources)
{
elem.SetIsUnmapped();
}
}
public void RunAsync(CudaStream stream, params object[] args)
{
for (int i = 0; i < args.Length; i++)
{
if (args[i] is MyAbstractMemoryBlock)
{
args[i] = (args[i] as MyAbstractMemoryBlock).GetDevicePtr(m_GPU);
if (((CUdeviceptr)args[i]).Pointer == 0)
{
// TODO(Premek): this is now handled in observers, should be also handled in the simulation
throw new InvalidOperationException("Memory block resolved to null device ptr (not allocated on device?).");
}
}
}
CUstream cuStream = CUstream.NullStream;
if (stream != null)
{
cuStream = stream.Stream;
}
m_kernel.RunAsync(cuStream, args);
}
/// <summary>
/// Asynchron copy 2D Array to host
/// </summary>
/// <param name="deviceArray"></param>
/// <param name="stream"></param>
public void AsyncCopyFromArray2D(CUarray deviceArray, CUstream stream)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
CUDAMemCpy2D cpyProps = new CUDAMemCpy2D();
cpyProps.srcArray = deviceArray;
cpyProps.srcMemoryType = CUMemoryType.Array;
cpyProps.dstHost = _intPtr;
cpyProps.dstMemoryType = CUMemoryType.Host;
cpyProps.dstPitch = _pitchInBytes;
cpyProps.WidthInBytes = _width * _typeSize;
cpyProps.Height = _height;
res = DriverAPINativeMethods.AsynchronousMemcpy_v2.cuMemcpy2DAsync_v2(ref cpyProps, stream);
Debug.Write("");//Line(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuMemcpy2DAsync", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
}
/// <summary>
/// Creates a new wrapper for an existing stream
/// </summary>
public CudaStream(CUstream _stream)
{
_isOwner = false;
}
public curandStatus SetStream(RandGenerator generator, CUstream stream)
{
return(curandSetStream(generator, stream));
}
private static extern curandStatus curandSetStream(RandGenerator generator, CUstream stream);
public static extern cufftResult cufftSetStream([In] cufftHandle plan, [In] CUstream stream);
