/// <summary>
/// Allocates a 2D block of memory, and returns a pointer pointing to it.
/// </summary>
/// <param name="elemSizeInBytes">The byte size of each element.</param>
/// <param name="width">The number of columns</param>
/// <param name="height">The number of rows</param>
/// <returns>A pointer to the block of new memory</returns>
public static DeviceBuffer2D Alloc(int elemSizeInBytes, int width, int height)
{
DeviceBuffer2D result = new DeviceBuffer2D();
result.widthInBytes = (uint)elemSizeInBytes * (uint)width;
result.height = (uint)height;
CudaUtil.Call(CudaMem.cuMemAllocPitch(out result.ptr.ptr, out result.ptr.pitch,
result.widthInBytes, result.height, (uint)elemSizeInBytes));
return(result);
}
public static DeviceBuffer2D GLMapBufferObject(UInt32 bufferId, uint widthInBytes)
{
DeviceBuffer2D result = new DeviceBuffer2D();
UInt32 size;
CudaUtil.Call(CudaMem.cuGLMapBufferObject(out result.ptr.ptr, out size, bufferId));
result.ptr.pitch = widthInBytes;
result.widthInBytes = widthInBytes;
result.height = size / widthInBytes;
return(result);
}
/// <summary>
/// Copies data from device memory to device memory.
/// </summary>
/// <param name="src">The source array</param>
/// <param name="dest">The destination array</param>
public static void Copy(DeviceBuffer2D src, DeviceBuffer2D dest)
{
if (src.ptr == DevicePtr2D.Zero || dest.ptr == DevicePtr2D.Zero)
{
throw new NullPointerException();
}
CudaUtil.Call(memcpy2DUnalignedHelper(
MemoryType.Device, 0, 0, (IntPtr)src.ptr.ptr, src.ptr.pitch,
MemoryType.Device, 0, 0, (IntPtr)dest.ptr.ptr, dest.ptr.pitch,
src.widthInBytes, src.height)
);
}
/// <summary>
/// Copies data from device memory to host memory.
/// </summary>
/// <param name="src">The source array</param>
/// <param name="dest">The destination array</param>
public static void Copy <T>(DeviceBuffer2D src, T[,] dest) where T : struct
{
if (src.ptr == DevicePtr2D.Zero || dest == null)
{
throw new NullPointerException();
}
IntPtr destPtr = AllocArrayFor(dest, src.widthInBytes, src.height);
CudaUtil.Call(memcpy2DHelper(
MemoryType.Device, 0, 0, (IntPtr)src.ptr.ptr, src.ptr.pitch,
MemoryType.Host, 0, 0, destPtr, src.widthInBytes,
src.widthInBytes, src.height)
);
PtrToArray(destPtr, dest);
Marshal.FreeHGlobal(destPtr);
}
/// <summary>
/// Copies data from device memory to host memory asynchronously.
/// </summary>
/// <param name="src">The source pointer</param>
/// <param name="dest">The destination buffer</param>
/// <param name="stream">The stream in which to queue the copy operation</param>
public static void CopyAsync <T>(DeviceBuffer2D src, HostBuffer2D <T> dest, Stream stream) where T : struct
{
if (src.ptr == DevicePtr2D.Zero || dest.IsNull())
{
throw new NullPointerException();
}
if (src.widthInBytes != (uint)dest.RowSizeBytes || src.height != (uint)dest.height)
{
throw new ArgumentException("The source and destination sizes do not match.");
}
CudaUtil.Call(memcpy2DAsyncHelper(
MemoryType.Device, 0, 0, (IntPtr)src.ptr.ptr, src.ptr.pitch,
MemoryType.Host, 0, 0, dest.Ptr, (uint)dest.RowSizeBytes,
(uint)dest.RowSizeBytes, (uint)dest.Height, stream.Handle)
);
}
/// <summary>
/// Copies data from host memory to device memory.
/// </summary>
/// <param name="src">The source buffer</param>
/// <param name="dest">The destination buffer</param>
public static void Copy <T>(HostBuffer2D <T> src, DeviceBuffer2D dest) where T : struct
{
if (src.IsNull() || dest.ptr == DevicePtr2D.Zero)
{
throw new NullPointerException();
}
if ((uint)src.RowSizeBytes != dest.WidthInBytes || (uint)src.Height != dest.Height)
{
throw new ArgumentException("The source and destination sizes do not match.");
}
CudaUtil.Call(CudaMem.memcpy2DHelper(
MemoryType.Host, 0, 0, src.ptr, (uint)src.RowSizeBytes,
MemoryType.Device, 0, 0, (IntPtr)dest.ptr.ptr, dest.ptr.pitch,
(uint)src.RowSizeBytes, (uint)src.height)
);
}
/// <summary>
/// Copies data from host memory to device memory.
/// </summary>
/// <param name="src">The source array</param>
/// <param name="dest">The destination pointer</param>
public static void Copy <T>(T[,] src, DeviceBuffer2D dest) where T : struct
{
if (src == null || dest.ptr == DevicePtr2D.Zero)
{
throw new NullPointerException();
}
IntPtr srcPtr = AllocArrayFor(src, dest.widthInBytes, dest.height);
ArrayToPtr(src, srcPtr);
CudaUtil.Call(memcpy2DHelper(
MemoryType.Host, 0, 0, srcPtr, dest.widthInBytes,
MemoryType.Device, 0, 0, (IntPtr)dest.ptr.ptr, dest.ptr.pitch,
dest.widthInBytes, dest.height)
);
Marshal.FreeHGlobal(srcPtr);
}
