/// <summary>
/// Copies data from host memory to device memory.
/// </summary>
/// <param name="src">The source array</param>
/// <param name="dest">The destination CUDA array</param>
public static void Copy <T>(T[,] src, CudaArray dest) where T : struct
{
if (src == null || dest.handle == IntPtr.Zero)
{
throw new NullPointerException();
}
int elemSize = Marshal.SizeOf(typeof(T));
int width = src.GetLength(1);
int height = src.GetLength(0);
CudaArray.Descriptor desc = dest.GetDescriptor();
if (desc.Width != width)
{
throw new ArgumentException("The source and destination width do not match.");
}
if (desc.Height != height)
{
throw new ArgumentException("The source and destination height do not match.");
}
IntPtr srcPtr = Marshal.AllocHGlobal(elemSize * width * height);
ArrayToPtr(src, srcPtr);
CudaUtil.Call(memcpy2DHelper(
MemoryType.Host, 0, 0, srcPtr, (uint)(elemSize * width),
MemoryType.Array, 0, 0, dest.handle, 0,
(uint)(elemSize * width), (uint)height)
);
Marshal.FreeHGlobal(srcPtr);
}
public DeviceBuffer GetConstantBuffer(string name)
{
DeviceBuffer result = new DeviceBuffer();
CudaUtil.Call(cuModuleGetGlobal(out result.ptr.ptr, out result.sizeInBytes, handle, name));
return(result);
}
private TexAddressMode GetTexAddressMode(int dim)
{
UInt32 result;
CudaUtil.Call(cuTexRefGetAddressMode(out result, handle, dim));
return((TexAddressMode)result);
}
public void GLUnmapBufferObject(UInt32 bufferId)
{
CudaUtil.Call(CudaMem.cuGLUnmapBufferObject(bufferId));
ptr = DevicePtr2D.Zero;
widthInBytes = 0;
height = 0;
}
/// <summary>
/// Maps a buffer to an OpenGL buffer object.
/// </summary>
/// <param name="bufferId">The unsigned integer name of the OpenGL buffer</param>
/// <returns>A pointer mapped to the OpenGL buffer</returns>
public static DeviceBuffer GLMapBufferObject(UInt32 bufferId)
{
DeviceBuffer result = new DeviceBuffer();
CudaUtil.Call(CudaMem.cuGLMapBufferObject(out result.ptr.ptr, out result.sizeInBytes, bufferId));
return(result);
}
/// <summary>
/// Creates a new buffer and allocates the specified ammount of memory.
/// </summary>
/// <param name="sizeInBytes">The number of bytes to allocate for the buffer</param>
/// <returns>A pointer to the allocated buffer</returns>
public static IntPtr Alloc(uint sizeInBytes)
{
IntPtr result = IntPtr.Zero;
CudaUtil.Call(CudaMem.cuMemAllocHost(out result, (UInt32)sizeInBytes));
return(result);
}
public static TexRef Create()
{
TexRef result = new TexRef();
CudaUtil.Call(cuTexRefCreate(out result.handle));
return(result);
}
public TexRef GetTexRef(string name)
{
TexRef result = new TexRef();
CudaUtil.Call(cuModuleGetTexRef(out result.handle, handle, name));
return(result);
}
/// <summary>
/// Allocates a block of memory and returns a pointer to it.
/// </summary>
/// <param name="sizeInBytes">The ammount of memory to allocate</param>
/// <returns>A pointer to the new memory</returns>
public static DevicePtr AllocRaw(uint sizeInBytes)
{
DevicePtr result = new DevicePtr();
CudaUtil.Call(CudaMem.cuMemAlloc(out result.ptr, sizeInBytes));
return(result);
}
private int GetAttr(DeviceAttribute attrib)
{
Int32 result;
CudaUtil.Call(cuDeviceGetAttribute(out result, (Int32)attrib, handle));
return(result);
}
/// <summary>
/// Fills bytes in this buffer with the given value.
/// </summary>
/// <param name="value">The value to fill the bytes with</param>
/// <param name="widthInBytes">The number of bytes in each row</param>
/// <param name="height">The number of rows</param>
public void MemSet(byte value)
{
if (ptr.ptr == 0)
{
throw new NullPointerException();
}
CudaUtil.Call(CudaMem.cuMemsetD2D8(ptr.ptr, ptr.pitch, value, widthInBytes, height));
}
/// <summary>
/// Releases any memory used by this buffer, and sets it to null. This method may be
/// called on a null pointer, but will have no effect.
/// </summary>
/// <param name="ptr">A pointer to the buffer to deallocate</param>
public static void Free(ref IntPtr ptr)
{
if (ptr != IntPtr.Zero)
{
CudaUtil.Call(CudaMem.cuMemFreeHost(ptr));
ptr = IntPtr.Zero;
}
}
/// <summary>
/// Allocates a block of memory sized to hold a 1D array and returns a pointer to it.
/// </summary>
/// <param name="elemSizeBytes">The size of each array element in bytes</param>
/// <param name="width">The number of elements to allocate memory for</param>
/// <returns>A pointer to the array</returns>
public static DeviceBuffer Alloc(int elemSizeBytes, int width)
{
DeviceBuffer result = new DeviceBuffer();
result.sizeInBytes = (uint)elemSizeBytes * (uint)width;
CudaUtil.Call(CudaMem.cuMemAlloc(out result.ptr.ptr, result.sizeInBytes));
return(result);
}
/// <summary>
/// Releases memory pointed to by this pointer. This method may be called on a null
/// pointer, but will have no effect.
/// </summary>
public void Free()
{
if (ptr != 0)
{
CudaUtil.Call(CudaMem.cuMemFree(ptr));
ptr = 0;
}
}
protected virtual void Dispose(bool disposing)
{
if (disposing)
{
CudaUtil.Call(cuModuleUnload(handle));
handle = IntPtr.Zero;
}
}
public void Free()
{
if (handle != IntPtr.Zero)
{
CudaUtil.Call(cuArrayDestroy(handle));
handle = IntPtr.Zero;
}
}
/// <summary>
/// Copies data from device memory to host memory asynchronously. No size checks are performed, it is
/// up to the programmer to ensure the destination is large enough to hold the data.
/// </summary>
/// <param name="src">The source pointer</param>
/// <param name="dest">The destination pointer, must be to page locked memory</param>
/// <param name="sizeInBytes">The number of bytes to copy</param>
/// <param name="stream">The stream in which to queue the copy operation</param>
public static void CopyRawAsync(DevicePtr src, IntPtr destPtr, IntPtr sizeInBytes, Stream stream)
{
if (src == DevicePtr.Zero || destPtr == IntPtr.Zero)
{
throw new NullPointerException();
}
CudaUtil.Call(cuMemcpyDtoHAsync(destPtr, src.ptr, (uint)sizeInBytes, stream.Handle));
}
/// <summary>
/// Copies data from device memory to host memory. No size checks are performed, it is
/// up to the programmer to ensure the destination is large enough to hold the data.
/// </summary>
/// <param name="src">The source pointer</param>
/// <param name="dest">The destination buffer</param>
/// <param name="sizeInBytes">The number of bytes to copy</param>
public static void CopyRaw(DevicePtr src, IntPtr destPtr, IntPtr sizeInBytes)
{
if (src == DevicePtr.Zero || destPtr == IntPtr.Zero)
{
throw new NullPointerException();
}
CudaUtil.Call(cuMemcpyDtoH(destPtr, src.ptr, (uint)sizeInBytes));
}
protected virtual void Dispose(bool disposing)
{
if (disposing)
{
CudaUtil.Call(cuCtxDestroy(handle));
handle = IntPtr.Zero;
}
}
/// <summary>
/// Allocates a 2D block of memory, and returns a pointer pointing to it. Rows are NOT padded,
/// and may not meet the hardware's alignment requirements.
/// </summary>
/// <param name="widthInBytes">The number of bytes to allocate for each row.</param>
/// <param name="height">The number of rows to allocate</param>
/// <returns>A pointer to the new block</returns>
public static DevicePtr2D AllocRaw(uint widthInBytes, uint height)
{
DevicePtr2D result = new DevicePtr2D();
result.pitch = widthInBytes;
CudaUtil.Call(CudaMem.cuMemAlloc(out result.ptr, widthInBytes * height));
return(result);
}
public static HostBuffer1D <T> Alloc(int length)
{
HostBuffer1D <T> result = new HostBuffer1D <T>();
uint elemSize = (uint)Marshal.SizeOf(typeof(T));
CudaUtil.Call(CudaMem.cuMemAllocHost(out result.ptr, elemSize * (uint)length));
result.length = length;
return(result);
}
/// <summary>
/// Releases any memory used by this buffer. This method may be called on a null
/// buffer, but will have no effect.
/// </summary>
public void Free()
{
if (ptr != IntPtr.Zero)
{
CudaUtil.Call(CudaMem.cuMemFreeHost(ptr));
ptr = IntPtr.Zero;
length = 0;
}
}
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>
/// 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 Descriptor GetDescriptor()
{
if (handle == IntPtr.Zero)
{
return(new Descriptor());
}
else
{
Descriptor result;
CudaUtil.Call(cuArrayGetDescriptor(out result, handle));
return(result);
}
}
public static CudaArray Allocate(int width, int height, CudaArrayFormat format, int numChannels)
{
CudaArray result = new CudaArray();
Descriptor desc = new Descriptor();
desc.Width = (UInt32)width;
desc.Height = (UInt32)height;
desc.Format = format;
desc.NumChannels = (UInt32)numChannels;
CudaUtil.Call(cuArrayCreate(out result.handle, ref desc));
return(result);
}
/// <summary>
/// Copies data from host memory to device memory. No size checks are performed, it is
/// up to the programmer to ensure the destination is large enough to hold the data.
/// </summary>
/// <param name="src">The source buffer</param>
/// <param name="dest">The destination pointer</param>
/// <param name="widthInBytes">The size of each row in bytes</param>
/// <param name="height">The number of rows</param>
public static void CopyRaw2D(IntPtr srcPtr, DevicePtr2D dest, IntPtr widthInBytes, IntPtr height)
{
if (srcPtr == IntPtr.Zero || dest == DevicePtr2D.Zero)
{
throw new NullPointerException();
}
CudaUtil.Call(CudaMem.memcpy2DHelper(
MemoryType.Host, 0, 0, srcPtr, (uint)widthInBytes,
MemoryType.Device, 0, 0, (IntPtr)dest.ptr, (uint)dest.Pitch,
(uint)widthInBytes, (uint)height)
);
}
/// <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>(DeviceBuffer src, T[] dest) where T : struct
{
if (src.ptr == DevicePtr.Zero || dest == null)
{
throw new NullPointerException();
}
IntPtr destPtr = AllocArrayFor(dest, src.sizeInBytes);
CudaUtil.Call(cuMemcpyDtoH(destPtr, src.ptr.ptr, src.sizeInBytes));
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>(DeviceBuffer src, HostBuffer1D <T> dest, Stream stream) where T : struct
{
if (src.ptr == DevicePtr.Zero || dest.IsNull())
{
throw new NullPointerException();
}
if (src.sizeInBytes != (uint)dest.SizeInBytes)
{
throw new ArgumentException("The source and destination sizes do not match.");
}
CudaUtil.Call(cuMemcpyDtoHAsync(dest.ptr, src.ptr.ptr, (uint)dest.SizeInBytes, stream.Handle));
}
