====== Which build has the best decoding performance when using filters? ======

Optimized builds of //ffdshow.ax// can have an impact on performance when certain filters are used in ffdshow.
Below is a list of - so far tested - filters that benefit from optimizations:
  * xsharpen
  * denoise3d

The following chart illustrates the possible increase in decoding performance when using optimized builds **and** filters:

{{faq:benchmark-en.png}}

Test setup: Pentium 4 2.8GHz (Northwood), Windows XP SP2, using null renderer\\
Test sample: [[http://trailers.divx.com/Dreamworks/Madagascar_HD.zip|Madagascar.avi]] (DX50)

To sum it up:

  * The results when decoding without any filters applied are **indistinguishable**. All builds perform within 2% of each other.
  * The ICL9 (SSE) build proved to be the fastest when using the //xsharpen// filter. The decoding speed increased by 88% over generic MSVC builds.
  * SSE2-optimized builds showed no improvement over SSE builds.
  * The improvements over the generic builds differ from system to system though. It is recommended to test the provided builds on your own particular system.
      For benchmarking purposes, you will need Haali's [[http://haali.cs.msu.ru/mkv/timeCodec.exe|timeCodec.exe]] [96KB]. For timeCodec.exe to work, [[http://haali.cs.msu.ru/mkv/|Haali Media Splitter]] is needed.