private string GetSpecific(KernelMethodInfo kmi)
{
StringBuilder sb = new StringBuilder();
if (kmi != null)
{
sb.AppendLine("Parameters : " + kmi.GetParametersString());
}
return sb.ToString();
}
private void lbFunctions_SelectedIndexChanged(object sender, EventArgs e)
{
KernelMethodInfo item = lbFunctions.SelectedItem as KernelMethodInfo;
if (item != null)
{
string text = GetCommon(item as KernelMemberInfo);
tbFunctions.Text = text;
text = GetSpecific(item);
tbFunctions.AppendText(text);
}
}
/// <summary>
/// Does the launch.
/// </summary>
/// <param name="gridSize">Size of the grid.</param>
/// <param name="blockSize">Size of the block.</param>
/// <param name="streamId">Stream id, or -1 for non-async.</param>
/// <param name="gpuMethodInfo">The gpu method info.</param>
/// <param name="arguments">The arguments.</param>
protected override void DoLaunch(dim3 gridSize, dim3 blockSize, int streamId, KernelMethodInfo gpuMethodInfo, params object[] arguments)
{
if (streamId > -1 && !_streams.ContainsKey(streamId))
{
_streams.Add(streamId, streamId);
}
MethodInfo mi = gpuMethodInfo.Method;
if (mi == null)
{
throw new CudafyHostException(CudafyHostException.csX_NOT_SET, gpuMethodInfo.Name);
}
object instance = mi.IsStatic ? null : Activator.CreateInstance(mi.DeclaringType);
if (gpuMethodInfo.IsDummy)
{
object[] argsCopy = new object[arguments.Length];
for (int i = 0; i < arguments.Length; i++)
{
if (arguments[i].GetType().IsArray)
{
var v = TryGetDeviceMemory(arguments[i]) as EmuDevicePtrEx;
if (v != null)
{
if (v.Offset == 0)
{
argsCopy[i] = v.DevPtr;
}
else
{
throw new CudafyHostException(CudafyHostException.csX_NOT_CURRENTLY_SUPPORTED, "Offsets in arrays passed to dummy functions");
}
}
else
{
argsCopy[i] = arguments[i];
}
}
else
{
argsCopy[i] = arguments[i];
}
}
mi.Invoke(instance, argsCopy);
return;
}
GGrid grid = new GGrid(gridSize);
Dictionary <Array, EmuDevicePtrEx> dic;
object[] pList = BuildParameterList2(mi, arguments, out dic);
//object[] pListCopy = new object[0];
if (gridSize.z > 1)
{
throw new CudafyHostException(CudafyHostException.csX_NOT_SUPPORTED, "3D grid sizes");
}
if (blockSize.z > 1)
{
throw new CudafyHostException(CudafyHostException.csX_NOT_SUPPORTED, "3D block sizes");
}
for (int x = 0; x < gridSize.x; x++)
{
for (int y = 0; y < gridSize.y; y++)
{
int totalSize = blockSize.x * blockSize.y * blockSize.z;
Thread[] threads = new Thread[totalSize];
GBlock blk2lnch = new GBlock(grid, blockSize, x, y);
int tCtr = 0;
int pListLen = pList.Length;
for (int tx = 0; tx < blockSize.x; tx++)
{
for (int ty = 0; ty < blockSize.y; ty++)
{
GThread ht = new GThread(tx, ty, blk2lnch);
object[] pListCopy = new object[pListLen];
for (int pc = 0; pc < pListLen; pc++)
{
if (pList[pc] is GThread)
{
pListCopy[pc] = ht;
}
else
{
pListCopy[pc] = pList[pc];
}
}
#warning OPTIMIZATION if there is no synchronize then start and join threads in multiple of processor count - check this in disassembly and put flag in gpuMethodInfo
threads[tCtr] = new Thread(() =>
{
mi.Invoke(instance, pListCopy);
});
threads[tCtr].Name = string.Format("Grid_{0}_{1}_Thread_{2}_{3}", x, y, tx, ty);
threads[tCtr].Start();
//if (ctr % 16 == 0)
// Console.WriteLine("Ctr=" + ctr.ToString());
//ctr++;
tCtr++;
}
}
for (int i = 0; i < totalSize; i++)
{
threads[i].Join();
//Console.WriteLine("Thread {0} exited.", threads[i].Name);
}
}
}
int iArgs = 0;
ParameterInfo[] piArray = mi.GetParameters();
for (int iParams = 0; iParams < piArray.Length; iParams++)
{
ParameterInfo pi = piArray[iParams];
if (pi.ParameterType == typeof(GThread))
{
continue;
}
else if (iArgs < pList.Length)
{
object o = pList[iArgs++];
if (!(o is GThread))
{
if (o.GetType().IsArray)
{
if (dic.ContainsKey(o as Array))
{
EmuDevicePtrEx ptrEx = dic[o as Array];
DoCopy(o as Array, 0, ptrEx.DevPtr, ptrEx.Offset, ptrEx.TotalSize, pi.ParameterType.GetElementType());
}
}
}
else
{
iParams--;
}
}
}
}
