NimbusML
NimbusML copied to clipboard
Add support for batch transfers
Currently, each return value is transferred from managed code to native code one at a time. See NativeDataInterop.cs.
for (int i = 0; i < fillers.Length; i++)
{
fillers[i].Set();
}
...
_poker(value, _colIndex, _input.Position, 0);
Could performance be increased by batching the values and sending over a batch at a time? This would reduce the amount of managed/unmanaged transitions.