Have you tried using the profile directed compilation option? Reports are
that it can make quite a difference to a programs performance (30% increase is not so uncommon).

Tristan wrote:

Have you tried using the profile directed compilation option? Reports are
that it can make quite a difference to a programs performance (30% increase is not so uncommon).

Haven't tried it with DOSBox, yet, but I was meddling with the profile-driven and parallelism with POV-Ray, but I didn't really get much gains. (Benchmark may have been 5-10min better, but that's about it)

Just tried the profile guided optimization....

Compling for P4 only, using Parallelism, turning off /Op adding /O3 on my P4 2.6 HT, I can get up to 10,000 instructions per loop, with no skipping on the various Sierra games that have MT-32 support! 😁

Niiiice...

It is quite amazing how sensitive virtual machines (like Dosbox) are to compiler optimisation settings. A wrote a virtual machine a while ago and discovered that by changing one small option on the compiler(GCC) I could get an average improvement of 300% in performance. It seems that there is not much instruction level parallelism in an average virtual machine hence the sensitivity to compiler settings. I was thinking of releasing binaries using the Intel compiler I am just not sure if that is legal with the free version.

Compling for P4 only, using Parallelism, turning off /Op adding /O3 on my P4 2.6 HT, I can get up to 10,000 instructions per loop, with no skipping on the various Sierra games that have MT-32 support! Very Happy

With optimization like that I'm going to have to try the Intel Compiler for use with the driver code! I'm going to give a shot here this weekend. I'll let you know how it goes.

canadacow wrote:

With optimization like that I'm going to have to try the Intel Compiler for use with the driver code! I'm going to give a shot here this weekend. I'll let you know how it goes.

Just a small thought, don't do straight P4 (/QxN), try /QaxN, otherwise it will not run on PIII and below systems. Also -- The parallelism is only good for Multiprocessor and hyperthreaded machines, on a normal uniprocessor P4 (pre-net burst) and earlier generation processors, and those who are not hyperthreaded, there will be some performance degradation, but it may be livable.. experiment. When I ran the profile, I ran DOSBox through a few clips of music in firehawk, then let a clip run for a bit with a high CPU count and 0 frameskip. It WILL run frighteningly slow. With firehawk, there seems to be some trouble panning the drums, but other games seem to work just fine (Maybe they have the panning set up a certain way for the game, but, I dunno..

Overall, I plan to purchase the ICL compiler after experimenting with it some, but at $400, I'll have to wait a while.

For true spooniness, you can use the maximum dynamic options, heheheh. /QaxiMKWPB - W/N are interchangeable.

kode54 wrote:

W/N are interchangeable.

According to the ICL doc, N enables some additional optimizations for the P4 Northwood. (W only enables SSE2 instructions...)