private static void ExecuteSingleVectorOperation <T>(BinaryExpression operation,
LinearAlgebraProvider provider,
ArrayPoolStack <T> arrayPoolStack, List <NArray <T> > localsToFree,
OutputGetter <T> getter, Aggregator aggregator,
int vectorLength,
int chunkIndex, int startIndex, ExecutionTimer timer)
{
if (operation == null || operation.NodeType != ExpressionType.Assign)
{
return;
}
if (operation == null)
{
return;
}
NArray <T> result;
var left = operation.Left as ReferencingVectorParameterExpression <T>;
NArray <T> aggregationTarget = null;
if (left.ParameterType == ParameterType.Local)
{
result = left.Array;
var chunkyStorage = result.Storage as ChunkyStorage <T>;
if (chunkyStorage != null)
{
var newStorage = arrayPoolStack.Pop();
chunkyStorage.SetChunk(chunkIndex, newStorage);
}
aggregationTarget = getter.TryGetNext(left.Index);
}
else
{
result = (operation.Left as ReferencingVectorParameterExpression <T>).Array;
}
if (operation.Right is UnaryMathsExpression)
{
var unaryOperation = operation.Right as UnaryMathsExpression;
if (unaryOperation.UnaryType == UnaryElementWiseOperation.ScaleOffset)
{
var scaleOffset = unaryOperation as ScaleOffsetExpression <T>;
(provider as IElementWise <T>).ScaleOffset(Slice <T>(unaryOperation.Operand, chunkIndex, startIndex, vectorLength),
scaleOffset.Scale, scaleOffset.Offset,
Slice(result, chunkIndex, startIndex, vectorLength));
}
else if (unaryOperation.UnaryType == UnaryElementWiseOperation.ScaleInverse)
{
var scaleInverse = unaryOperation as ScaleInverseExpression <T>;
(provider as IElementWise <T>).ScaleInverse(Slice <T>(unaryOperation.Operand, chunkIndex, startIndex, vectorLength),
scaleInverse.Scale,
Slice(result, chunkIndex, startIndex, vectorLength));
}
else
{
(provider as IElementWise <T>).UnaryElementWiseOperation(
Slice <T>(unaryOperation.Operand, chunkIndex, startIndex, vectorLength),
Slice(result, chunkIndex, startIndex, vectorLength),
unaryOperation.UnaryType);
}
}
if (operation.Right is BinaryExpression)
{
var binaryOperation = operation.Right as BinaryExpression;
(provider as IElementWise <T>).BinaryElementWiseOperation(Slice <T>(binaryOperation.Left, chunkIndex, startIndex, vectorLength),
Slice <T>(binaryOperation.Right, chunkIndex, startIndex, vectorLength),
Slice(result, chunkIndex, startIndex, vectorLength),
binaryOperation.NodeType);
}
if (aggregationTarget != null)
{
if (aggregator != Aggregator.ElementwiseAdd)
{
throw new NotImplementedException();
}
if (aggregationTarget.IsScalar)
{
throw new Exception();
}
var slice = Slice <T>(aggregationTarget, chunkIndex, startIndex, vectorLength);
(provider as IElementWise <T>).BinaryElementWiseOperation(
slice,
Slice(result, chunkIndex, startIndex, vectorLength),
slice, ExpressionType.Add);
}
foreach (var item in localsToFree)
{
arrayPoolStack.Push((item.Storage as ChunkyStorage <T>).GetChunk(chunkIndex));
}
}
/// <summary>
/// Here we execute the expression, but without attempting to compile it.
/// The performance gain comes only from reducing the amount of memory allocation required, not
/// from compiling an optimised kernel.
/// </summary>
/// <param name="block"></param>
/// <param name="provider"></param>
/// <param name="vectorOptions"></param>
public static void RunNonCompiling(BlockExpressionBuilder _builder,
LinearAlgebraProvider provider, VectorExecutionOptions vectorOptions,
NArray[] outputs, int[] outputsIndices, Aggregator aggregator, ExecutionTimer timer)
{
var block = _builder.ToBlock();
if (block.Operations.Count == 0)
{
return;
}
// we will cycle through arrays in order of increasing index
var getter = new OutputGetter <double>(outputs, outputsIndices);
int chunksLength = _builder.VectorLength; //5000;
var arrayPoolStack = ExecutionContext.ArrayPool.GetStack(chunksLength);
int length = (block.ArgumentParameters.First() as ReferencingVectorParameterExpression <double>)
.Array.Length;
int chunkCount = length / chunksLength;
if (length % chunksLength != 0)
{
chunkCount++;
}
List <NArray <double> >[] localsToFree; // the storage that can be freed up after each operation is complete
AssignNArrayStorage <double>(block, chunkCount, chunksLength, length, out localsToFree);
// for integer support, add here
var options = new ParallelOptions();
if (!vectorOptions.MultipleThreads)
{
options.MaxDegreeOfParallelism = 1;
}
timer.MarkExecutionTemporaryStorageAllocationComplete();
// can multi-thread here, but better to do so at higher level
//Parallel.For(0, chunkCount, options, (i) =>
for (int i = 0; i < chunkCount; ++i)
{
int startIndex = i * chunksLength;
int vectorLength = Math.Min(chunksLength, length - startIndex);
for (int j = 0; j < block.Operations.Count; ++j)
{
var operation = block.Operations[j];
if (operation.Type == typeof(NArray))
{
var newOperation = _builder.SimplifyOperation(operation); // deal with any expressions containing scalars that can be simplified
ExecuteSingleVectorOperation <double>(newOperation, provider,
arrayPoolStack, localsToFree[j],
getter, aggregator,
vectorLength,
i, startIndex, timer);
}
}
}
;
if (!arrayPoolStack.StackCountEqualsCreated)
{
throw new Exception("not all storage arrays created returned to stack");
}
}