Thumb cycle counting

[thrust26]

Maybe we could extend the Thumbulator code to (roughly) count cycles? While not perfect, this might still help developer to identify slow code or code running out of time.

https://developer.arm.com/documentation/ddi0234/b/instruction-cycle-timings/instruction-cycle-count-summary

[sa666666]

We've discussed this in the past, and found that there is really no optimal solution. But I am happy to revisit it to see if we can come up with an approximation.

[thrust26]

At the moment I cannot even match our code with the description in the link. Besides that, it doesn't look too complicated. 😄

Probably someone needs at least basic ARM assembler knowledge to implement this.

[thrust26]

@sa666666 @DirtyHairy
After having some coarse calculation available already, I added code (0c51180) which makes use of it during emulation in developer mode. Now the calculated cycles are multiplied with a user defined factor (default 1.25) and increase the system cycles. So that a timer overrun can be detected.

I have no clue how good the approximation is, but by some testing with existing ROMs I found the upper limit of the factor to be 1.79. Probably we should ask some ARM game developers (e.g. @SpiceWare) if they can provide a test program which helps getting an idea about the correctness of our calculations.

Probably I will add some UI elements for easier changing the new options. Not to the settings dialog, but to the debugger widgets.

[andrew-davie]

I can send you a confidential copy of my project for testing. It’s pushing the ARM quite hard and timing is quite critical.

…

On 13 Jun 2021, at 10:03 pm, Thomas Jentzsch ***@***.***> wrote: After having some coarse calculation available already, I added code (0c51180 <0c51180>) which makes use of it during emulation in developer mode. Now the calculated cycles are multiplied with a user defined factor (default 1.25) and increase the system cycles. So that a timer overrun can be detected. I have no clue how good the approximation is, but by some testing with existing ROMs I found the upper limit of the factor to be 1.79. Probably we should ask some ARM game developers (e.g. @SpiceWare <https://github.com/SpiceWare>) if they can provide a test program which helps getting an idea about the correctness of our calculations. Probably I will add some UI elements for easier changing the new options. Not to the settings dialog, but to the debugger widgets. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#757 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD5JZCYUQMHDC2RN576MYETTSSNCJANCNFSM4WSFVCSA>.

[thrust26]

I can send you a confidential copy of my project for testing. It’s pushing the ARM quite hard and timing is quite critical.

Sure. But probably I need some help for setting up the development environment. Also, if you can compile Stella, you can check the changes yourself too.

[andrew-davie]

Wouldn’t a binary be sufficient?

…

On 13 Jun 2021, at 10:57 pm, Thomas Jentzsch ***@***.***> wrote: I can send you a confidential copy of my project for testing. It’s pushing the ARM quite hard and timing is quite critical. Sure. But probably I need some help for setting up the development environment. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#757 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD5JZC2H6KI2OSM2UL2O4GDTSSTNDANCNFSM4WSFVCSA>.

[thrust26]

If you let me know which situations are close to the edge, yes.

[andrew-davie]

Have sent to your yahoo email.

…

On 13 Jun 2021, at 11:04 pm, Thomas Jentzsch ***@***.***> wrote: If you let me know which situations are close to the edge, yes. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#757 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD5JZC67Z72VCIC6BYFVAEDTSSUG3ANCNFSM4WSFVCSA>.

[thrust26]

Got it. Thanks!

[thrust26]

Meanwhile I committed a number of updates. The cycle count has become a lot better, but there are still issues to solve:

• no ARM pipeline emulation (this should reduce the cycle counts in case of slow Flash memory access)
• missing driver cycles (these cycles can only be estimated, since they are not emulated)
  • IRQ driven audio regularly interrupts the Thumb code with some ARM code and 5 cycles of 6507 (currently 10% are added to the count if IRQ driven audio is enabled)
  • the Thumb code regularly switches to driver ARM code for updating audio values
• adding known good cycle counts still making the game roll in Stella (mysterious)

Overall I estimate that the cycle count is ~10% too high.