Intel has a nice document showing how one can optimize the inverse DCT when running on processors with the MMX instructions, such as the Pentium II. Listings include code you can drop into the Independent JPEG Group’s library.
DCL reader Neil H. had this to say: The drop-in replacement for the iDCT of the IJG works fine for 8-bit data (not 12-bit). It reduced total compression time to 2/3 or so of the original cross-platform code. The iDCT itself is supposed to be 4x faster.

http://cedar.intel.com/cgi-bin/ids.dll/content/content.jsp?cntKey=Legacy::irtp_886_11228&c
ntType=IDS_EDITORIAL