Since Version 6.0, Intel IPP introduced a new Deferred Mode Image Processing (DMIP) library, which organizes Intel IPP functions into image processing pipeline and provides additional threading and memory access optimization. For large images, DMIP library can significantly improve the image processing performance. DMIP library is provided in Intel IPP Sample code.
Three main features of DMIP image processing improve the performance : 1) processing by the image fragments so data can fit into cache 2) highly optimized IPP to process image fragments. 3) parallel processing to different fragments or execution of different independent branches of a graph.
To start such processing, you create a task description as a DAG, (directed acyclic graph), and translate the graph a sequence of IPP calls. To complete (deferred), processing you run the generated sequence of IPP, calls. For example, you need to filter harmonization filter, you can build the DAG as bellow:
The graph can be compiled and executed. For large images, compared with traditional IPP processing, DMIP can significantly improve the image processing performance.
Here are some useful resource about Deferred Library:
A introduction paper, DEFERRED IMAGE PROCESSING IN INTEL IPP LIBRARY:
DIMP Library summary in Intel Software Insight Magazine(Page 19-20): A Landmark in Image Processing:DMIP
Deferred Mode Optimization article:
New features on DMIP in IPP 6.1:
DIMP Libirary Reference Manual document at Intel IPP sample code:
To download DIMP library, users can login into Intel registration center:
Select "Intel Integrated Performance Primitives for Windows* ,and then click "DMIP Sample" to get the package.