/// <summary>
/// The GPUParams constructor.
/// </summary>
/// <param name="cuda">Specifies the CudaDnn connection to Cuda.</param>
/// <param name="log">Specifies the Log for output.</param>
/// <param name="root_solver">Specifies the root Solver.</param>
/// <param name="nDeviceID">Specifies the device ID to use for this instance.</param>
public GPUParams(CudaDnn <T> cuda, Log log, Solver <T> root_solver, int nDeviceID)
: base(root_solver)
{
m_cuda = cuda;
m_log = log;
m_nDeviceID = m_cuda.GetDeviceID();
if (nDeviceID != m_nDeviceID)
{
m_cuda.SetDeviceID(nDeviceID);
}
// Allocate device buffers
m_hData = m_cuda.AllocMemory(m_lCount);
// Copy blob values
BlobCollection <T> net = root_solver.net.learnable_parameters;
apply_buffers(net, m_hData, m_lCount, Op.copy);
m_hDiff = m_cuda.AllocMemory(m_lCount);
m_cuda.set((int)m_lCount, m_hDiff, 0);
m_hStream = m_cuda.CreateStream();
if (m_nDeviceID != nDeviceID)
{
m_cuda.SetDeviceID(m_nDeviceID);
}
}
/// <summary>
/// Allocate a number of items in GPU memory and save the handle.
/// </summary>
/// <param name="lCount">Specifies the number of items.</param>
public void Allocate(long lCount)
{
free();
m_nDeviceID = m_cuda.GetDeviceID();
m_hGpuData = m_cuda.AllocMemory(lCount);
m_lCapacity = lCount;
m_lCount = 0;
m_bOwnData = true;
return;
}
/// <summary>
/// The SyncedMemory constructor.
/// </summary>
/// <param name="cuda">Specifies the CudaDnn connection to Cuda.</param>
/// <param name="log">Specifies the Log for output.</param>
/// <param name="lCapacity">Optionally, specifies the capacity of the SyncedMemory (in items).</param>
/// <param name="tag">Optionally, specifies a tag used for debugging (the default = <i>null</i>).</param>
/// <param name="bUseHalfSize">Optionally, specifies to use half size (FP16) for both data and diff. This option is only available when using the <i>float</i> base type 'T'.</param>
public SyncedMemory(CudaDnn <T> cuda, Log log, long lCapacity = 0, object tag = null, bool bUseHalfSize = false)
{
if (bUseHalfSize && typeof(T) != typeof(float))
{
bUseHalfSize = false;
if (log != null)
{
log.WriteLine("Half size disabled for non 'float' basetypes!");
}
}
m_bHalfSize = bUseHalfSize;
m_cuda = cuda;
m_log = log;
m_tag = tag;
if (lCapacity > 0)
{
m_nDeviceID = m_cuda.GetDeviceID();
m_hGpuData = m_cuda.AllocMemory(lCapacity);
m_lCapacity = lCapacity;
m_lCount = lCapacity;
}
}
/// <summary>
/// Allocate a number of items in GPU memory and save the handle.
/// </summary>
/// <param name="lCount">Specifies the number of items.</param>
/// <param name="bUseHalfSize">Optionally, specifis to use half sized memory (default = false). This only applies to the 'float' base type.</param>
public void Allocate(long lCount, bool bUseHalfSize = false)
{
if (bUseHalfSize && typeof(T) != typeof(float))
{
bUseHalfSize = false;
if (m_log != null)
{
m_log.WriteLine("Half size disabled for non 'float' basetypes!");
}
}
free();
m_nDeviceID = m_cuda.GetDeviceID();
m_bHalfSize = bUseHalfSize;
m_hGpuData = m_cuda.AllocMemory(lCount, m_bHalfSize);
m_lCapacity = lCount;
m_lCount = 0;
m_bOwnData = true;
return;
}
/// <summary>
/// The SyncedMemory constructor.
/// </summary>
/// <param name="cuda">Specifies the CudaDnn connection to Cuda.</param>
/// <param name="log">Specifies the Log for output.</param>
/// <param name="lCapacity">Optionally, specifies the capacity of the SyncedMemory (in items).</param>
public SyncedMemory(CudaDnn <T> cuda, Log log, long lCapacity = 0)
{
m_cuda = cuda;
m_log = log;
if (lCapacity > 0)
{
m_nDeviceID = m_cuda.GetDeviceID();
m_hGpuData = m_cuda.AllocMemory(lCapacity);
m_lCapacity = lCapacity;
m_lCount = lCapacity;
}
}
static void Main(string[] args)
{
// This memory will reside on the GPU.
long hGpuMem = 0;
Console.WriteLine("Creating CudaCuDnn...");
CudaDnn <float> cuda = new CudaDnn <float>(0);
try
{
string strDeviceInfo = cuda.GetDeviceName(0);
Console.WriteLine(strDeviceInfo);
strDeviceInfo = cuda.GetDeviceP2PInfo(0);
Console.WriteLine(strDeviceInfo);
strDeviceInfo = cuda.GetDeviceInfo(0, true);
Console.WriteLine(strDeviceInfo);
List <long> rghGpuMem = new List <long>();
long lOffset = 0;
// You must first allocate the GPU memory to use.
// Below we will allocate an array of 1000 float values.
Console.WriteLine("Allocate 1000 items...");
hGpuMem = cuda.AllocMemory(1000);
cuda.set(1000, hGpuMem, 0.0);
Console.WriteLine("Create memory pointers...");
for (int i = 0; i < 10; i++)
{
long hMem1 = cuda.CreateMemoryPointer(hGpuMem, lOffset, 100);
cuda.set(100, hMem1, (double)(i + 1));
rghGpuMem.Add(hMem1);
lOffset += 100;
}
Console.WriteLine("Test memory...");
for (int i = 0; i < 10; i++)
{
long hMem1 = rghGpuMem[i];
float[] rgData = cuda.GetMemoryFloat(hMem1);
if (rgData.Length != 100)
{
throw new Exception("The data length should = 100!");
}
for (int j = 0; j < 100; j++)
{
if (rgData[j] != (float)(i + 1))
{
throw new Exception("The data at index " + j.ToString() + " is not correct!");
}
}
}
Console.WriteLine("Memory test passed successfully!");
}
catch (Exception excpt)
{
Console.WriteLine("ERROR: " + excpt.Message);
}
finally
{
// Clean-up and release all GPU memory used.
if (hGpuMem != 0)
{
cuda.FreeMemory(hGpuMem);
hGpuMem = 0;
}
cuda.Dispose();
}
Console.WriteLine("Press any key to exit.");
Console.Read();
}