youmuliu/stgsharp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Mat mul mat is now on full function

Browse Source
This commit is contained in:
Youmu Liu
2025-12-16 16:15:30 +08:00
parent 2c7277f3f9
commit 88ed5a95c5
14 changed files with 357 additions and 488 deletions
Download Patch File Download Diff File Expand all files Collapse all files
StgSharp.Mathematics/Numeric/MatrixParallel.Publish.cs
+1 -76
View File
@@ -51,7 +51,7 @@ namespace StgSharp.Mathematics.Numeric
MatrixParallelTaskPackage* nonGeneric = package;
switch (nonGeneric->ComputeMode.OperationStyle)
{
case MatrixOperationStyle.BUFFER_INDEX:
case MatrixIndexStyle.BUFFER_INDEX:
return ScheduleBufferTask(nonGeneric);
default:
throw new InvalidOperationException($"Unknown mode {nonGeneric->ComputeMode.OperationStyle}.");
@@ -110,80 +110,5 @@ namespace StgSharp.Mathematics.Numeric
return handle;
}
private static unsafe MatrixParallelHandle ScheduleKernelTask(
MatrixParallelTaskPackage* package)
{
int primCount = package->PrimCount;
int secCount = package->SecCount;
int tileWidth = (1024 / secCount) + 1;
int tileCount = (primCount + tileWidth - 1) / tileWidth;
MatrixParallelThread[] pool = LeadParallel(tileCount);
/*
if (pool.Length == 0)
{
// no need
// LeadParallel has contained capacity check
}
/**/
MatrixParallelHandle handle = new MatrixParallelHandle(pool);
if (handle.Wraps.Length == 1)
{
handle.Wraps[0].WrapTask.Enqueue((nint)package);
return handle;
}
int threadCount = pool.Length;
int tilePerThreadLess = tileCount / threadCount; // tiles per thread without extra slice
int columnAfterBaseTile = tileCount % threadCount * tileWidth; // tile sliced into columns
int extraColumnPerThread = columnAfterBaseTile / threadCount;
int extraColumnCount = columnAfterBaseTile % threadCount; // columns to be sliced
int total = primCount;
MatrixParallelTaskPackage* current = package;
foreach (MatrixParallelWrap wrap in handle.Wraps)
{
bool hasSlice = columnAfterBaseTile > 0;
bool isLargeSlice = extraColumnCount > 0;
// _count tiles for wrap
int wrapWidth = wrap!.WrapTaskCapacity/**/;
int columnFromSlicingCount = (hasSlice ? extraColumnPerThread : 0) * wrapWidth;
int baseColumn = wrapWidth * tilePerThreadLess * tileWidth;
int columnInWrap = baseColumn + columnFromSlicingCount + (isLargeSlice ? 1 : 0);
// set tasks
if (columnInWrap == 0)
{
#if DEBUG
throw new InvalidOperationException("Zero tasks published.");
#else
break;
#endif
}
// update current task
current->PrimColumnCountInTile = columnInWrap;
current->SecColumnCountInTile = current->SecCount;
current->SecTileOffset = 0;
// remove tiles _count
extraColumnCount--;
columnAfterBaseTile -= columnFromSlicingCount;
total -= columnInWrap;
// set to wrap and push
wrap.WrapTask.Enqueue((nint)current);
if (total <= 0)
{
break;
}
current = MatrixParallelFactory.FromExistPackage(current);
current->PrimTileOffset += columnInWrap;
}
return handle;
}
}
}