PerfCount

As teased a while ago, I’ve been working on a module to provide easy access to the performance counters that are present on most/all ARMv5+ CPUs. This should help programmers to measure and tune the performance of their code (e.g. I made heavy use of an early version of the module when working on NBlood’s vertical line routines)

The first version of PerfCount is now available at http://www.phlamethrower.co.uk/riscos/perfcount.php

Features:

• Reporting of which events are available, using a common ID scheme to provide some consistency between different CPUs
• Support for multiple contexts, each measuring different events. Note: Currently, only one context can be active at a time.
• Automatic & manual multipass/multiplexing system to allow you to gather data for as many events as you want, bypassing the limits on how many events the CPU can track at once
• 64bit counter width to avoid overflow
• Optional per-context Wimp task association to pause/resume monitoring depending on whether a specific task is active (e.g. for system-wide profiling)
• Filtering by privileged & unprivileged CPU mode (if supported by the CPU)

Feedback welcome!

It runs on my Titanium, and WimpPerf produces results. I am not (yet) sure what it all means!

One question about WimpPerf – it includes lines
tasks%(A)=buf%!(16*A)
SYS "OS_IntOn",buf%!(16*A+4) TO tasks$(A)
which make no sense to me, especially as IntOn (wherever I look) takes no input and produces no output!
Care to explain?!

SYS "OS_IntOn",buf%!(16*A+4) TO tasks$(A)

That was just a hack to convert a buffer containing a null terminated string into a BASIC string. I suppose I should have taken a closer look at that test code before deciding to release it as an example!

One question about WimpPerf – it includes lines

tasks%(A)=buf%!(16*A)
SYS "OS_IntOn",buf%!(16*A+4) TO tasks$(A)

I think it’s doing the old

SYS "XOS_GenerateError",zero_terminated_string% TO basic_string$

trick, to take task names from the block returned by TaskManager_EnumerateTasks and place them into the tasks$() array.

I guess the assumption is that in BASIC, where you’d expect interrupts to be on, OS_IntOn is assumed to be a SWI with no effect – which allows the SYS command’s string wrangling to be used.

Looking at WimpPerf, how is it that lots of task have BUS_ACCESS but no CPU_CYCLES? Is this the Wimp paging in a task to get some data from it, but not giving it control?

But the denominator for the other events is CPU_CYCLES, so how can BUS_ACCESS have been monitored for &23B4 cycles say, without that showing up in CPU_CYCLES?