Path: utzoo!utgpu!news-server.csri.toronto.edu!rutgers!psuvax1!julius.cs.uiuc.edu!usc!samsung!uunet!rocket!adcock From: adcock@rocket.uucp (John Adcock) Newsgroups: comp.dsp Subject: Re: C compilers for DSPs Message-ID: Date: 9 Nov 90 15:25:25 GMT References: <1990Nov5.230918.17597@sono.uucp> <1990Nov6.165814.5669@esseye.UUCP> <32450@netnews.upenn.edu> Sender: news@rocket.UUCP Distribution: usa Organization: Lockheed Sanders Inc. Lines: 36 In-reply-to: hvs@ee.upenn.edu's message of 7 Nov 90 16:26:30 GMT Version 4.00 of the tms320c30 C compiler does actually do optimization. That is to say that is has an optimization switch with several levels of optimization (it didn't before). We just got it here and at least one person I know who's used it got very dramatic speedups over what he was getting with the 3.00 compiler: (these times are for the particular algorithm he is running) version 3.00 : 48 ms version 4.00 : 21.3 ms (without optimization) version 4.00 : 8.7 ms (with all available optimization) On the other hand, the programs I'm running on the C30 call heavily on a library of assembler coded vector routines and I saw no speedup. If you want speed, code it in assembler. No compiler is going to figure out that the loop you wrote to do a dot product can be set up to use a parallel multiply/add instruction. John Adcock Lockheed Sanders Inc. Member of Technical Staff 144 Daniel Webster Hwy. Signal Processing Center of Technology Room MER24-158C Merrimack, NH. 03054 (usenet) ...!uunet!rocket!adcock phone (603) 885-2803 (internet) fax (603) 885-0631 -- John Adcock Lockheed Sanders Inc. Member of Technical Staff 144 Daniel Webster Hwy. Signal Processing Center of Technology Room MER24-158C Merrimack, NH. 03054 (usenet) ...!uunet!rocket!adcock phone (603) 885-2803 (internet) fax (603) 885-0631