unsafe static void Main(string[] args)
{
int nStreams = 8;
cudaStream_t[] streams = new cudaStream_t[nStreams];
//create streams
int N = 1024 * 1024 * 32;
float[] a = new float[N];
float[] b = new float[N];
for (int k = 0; k < N; ++k)
{
a[k] = (float)k;
b[k] = 1.0F;
}
IntPtr d_a, d_b; // device pointers
cuda.Malloc(out d_a, N * sizeof(float));
cuda.Malloc(out d_b, N * sizeof(float));
GCHandle handle_a = GCHandle.Alloc(a, GCHandleType.Pinned);
GCHandle handle_b = GCHandle.Alloc(b, GCHandleType.Pinned);
IntPtr h_a = handle_a.AddrOfPinnedObject();
IntPtr h_b = handle_b.AddrOfPinnedObject();
cuda.DeviceSynchronize();
cuda.Memcpy(d_a, h_a, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
cuda.Memcpy(d_b, h_b, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
int slice = N / nStreams; // size of the array compute by each stream
dynamic wrapped = HybRunner.Cuda().Wrap(new Program());
int start;
int stop;
// call kernel with each stream
// copy data device to host
// synchronize and destroy streams
for (int i = 0; i < N; ++i)
{
if (a[i] != (float)i + (1.0F * 100.0F))
{
Console.Error.WriteLine("ERROR at {0} -- {1} != {2}", i, a[i], i + 1);
Environment.Exit(6); // abort
}
}
handle_a.Free();
handle_b.Free();
}
public static NPPImage Load(string path, cudaStream_t stream)
{
NPPImage result = new NPPImage();
byte[] rawData;
if (Path.GetExtension(path).Contains("pgm"))
{
using (FileStream fs = new FileStream(path, FileMode.Open))
{
using (TextReader tReader = new StreamReader(fs))
using (BinaryReader bReader = new BinaryReader(fs))
{
string formatLine = tReader.ReadLine(); // skip
string sizeLine = tReader.ReadLine();
string[] splitted = sizeLine.Split(' ');
result.width = int.Parse(splitted[0]);
result.height = int.Parse(splitted[1]);
string maxValueLine = tReader.ReadLine(); // skip
int pos = formatLine.Length + sizeLine.Length + maxValueLine.Length + 3;
fs.Seek(pos, SeekOrigin.Begin);
// TODO: optimize that part
rawData = bReader.ReadBytes((int)(fs.Length - pos));
}
}
}
else if (Path.GetExtension(path).Contains("png"))
{
Bitmap image = Bitmap.FromFile(path) as Bitmap;
result.width = image.Width;
result.height = image.Height;
rawData = new byte[result.width * result.height];
int index = 0;
for (int j = 0; j < result.height; ++j)
{
for (int i = 0; i < result.width; ++i, ++index)
{
rawData[index] = image.GetPixel(i, j).R;
}
}
}
else
{
throw new NotSupportedException("unsupported file format");
}
IntPtr deviceData;
size_t p;
cuda.ERROR_CHECK(cuda.MallocPitch(out deviceData, out p, result.width * sizeof(ushort), result.height));
result.pitch = (int)p;
result.hostData = new ushort[result.height * result.width];
for (int j = 0; j < result.height; ++j)
{
for (int i = 0; i < result.width; ++i)
{
result.hostData[j * result.width + i] = rawData[j * result.width + i];
}
}
var handle = GCHandle.Alloc(result.hostData, GCHandleType.Pinned);
cuda.ERROR_CHECK(cuda.Memcpy2DAsync(deviceData, p, handle.AddrOfPinnedObject(), result.width * sizeof(ushort), result.width * sizeof(ushort), result.height, cudaMemcpyKind.cudaMemcpyHostToDevice, stream));
handle.Free();
result.deviceData = deviceData;
return(result);
}
public static extern int NPP_ImageSegmentationx46Programx46ColorizeLabels_ExternCWrapperStream_CUDA(
int gridDimX, int gridDimY, int blockDimX, int blockDimY, int blockDimZ, int shared, cudaStream_t stream,
IntPtr segments, IntPtr colors, IntPtr colormap, int maxLabel, int count, int width, int pitch);
unsafe static void Main(string[] args)
{
int nStreams = 8;
cudaStream_t[] streams = new cudaStream_t[nStreams];
for (int k = 0; k < nStreams; ++k)
{
cuda.StreamCreate(out streams[k]);
}
int N = 1024 * 1024 * 32;
float[] a = new float[N];
float[] b = new float[N];
for (int k = 0; k < N; ++k)
{
a[k] = (float)k;
b[k] = 1.0F;
}
IntPtr d_a, d_b; // device pointers
cuda.Malloc(out d_a, N * sizeof(float));
cuda.Malloc(out d_b, N * sizeof(float));
GCHandle handle_a = GCHandle.Alloc(a, GCHandleType.Pinned);
GCHandle handle_b = GCHandle.Alloc(b, GCHandleType.Pinned);
IntPtr h_a = handle_a.AddrOfPinnedObject();
IntPtr h_b = handle_b.AddrOfPinnedObject();
cuda.DeviceSynchronize();
cuda.Memcpy(d_a, h_a, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
cuda.Memcpy(d_b, h_b, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
int slice = N / nStreams;
dynamic wrapped = HybRunner.Cuda().Wrap(new Program());
for (int k = 0; k < nStreams; ++k)
{
int start = k * slice;
int stop = start + slice;
wrapped.SetStream(streams[k]).Add(d_a, d_b, start, stop);
}
for (int k = 0; k < nStreams; ++k)
{
int start = k * slice;
cuda.MemcpyAsync(h_a + start * sizeof(float), d_a + start * sizeof(float), slice * sizeof(float), cudaMemcpyKind.cudaMemcpyDeviceToHost, streams[k]);
}
for (int k = 0; k < nStreams; ++k)
{
cuda.StreamSynchronize(streams[k]);
cuda.StreamDestroy(streams[k]);
}
for (int i = 0; i < N; ++i)
{
if (a[i] != (float)i + 1.0F)
{
Console.Error.WriteLine("ERROR at {0} -- {1} != {2}", i, a[i], i + 1);
Environment.Exit(6); // abort
}
}
handle_a.Free();
handle_b.Free();
}
unsafe static void Main(string[] args)
{
int nStreams = 8;
cudaStream_t[] streams = new cudaStream_t[nStreams];
HybRunner[] runners = new HybRunner[nStreams];
dynamic[] wrapped = new dynamic[nStreams];
cudaDeviceProp prop;
cuda.GetDeviceProperties(out prop, 0);
for (int k = 0; k < nStreams; ++k)
{
cuda.StreamCreate(out streams[k]);
runners[k] = HybRunner.Cuda("Streams_CUDA.dll", streams[k], CudaMarshaler.Instance).SetDistrib(16 * prop.multiProcessorCount, 128);
wrapped[k] = runners[k].Wrap(new Program());
}
int N = 1024 * 1024 * 32;
IntPtr d_a, d_b; // device pointers
float[] a = new float[N];
float[] b = new float[N];
cuda.Malloc(out d_a, N * sizeof(float));
cuda.Malloc(out d_b, N * sizeof(float));
for (int k = 0; k < N; ++k)
{
a[k] = (float)k;
b[k] = 1.0F;
}
GCHandle handle_a = GCHandle.Alloc(a, GCHandleType.Pinned);
GCHandle handle_b = GCHandle.Alloc(b, GCHandleType.Pinned);
IntPtr h_a = handle_a.AddrOfPinnedObject();
IntPtr h_b = handle_b.AddrOfPinnedObject();
int slice = N / nStreams;
cuda.DeviceSynchronize();
cuda.Memcpy(d_a, h_a, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
cuda.Memcpy(d_b, h_b, N * sizeof(float), cudaMemcpyKind.cudaMemcpyHostToDevice);
for (int k = 0; k < nStreams; ++k)
{
int start = k * slice;
int stop = start + slice;
wrapped[k].Add(d_a, d_b, start, stop, 100);
}
for (int k = 0; k < nStreams; ++k)
{
int start = k * slice;
cuda.MemcpyAsync(h_a + start * sizeof(float), d_a + start * sizeof(float), slice * sizeof(float), cudaMemcpyKind.cudaMemcpyDeviceToHost, streams[k]);
}
for (int k = 0; k < nStreams; ++k)
{
cuda.StreamSynchronize(streams[k]);
cuda.StreamDestroy(streams[k]);
}
for (int k = 0; k < 10; ++k)
{
Console.WriteLine(a[k]);
}
handle_a.Free();
handle_b.Free();
}
public void SetStream(cudaStream_t stream)
{
cuRAND.SetStream(generator, stream);
}