Date   : Tue, 31 Jan 2017 17:47:56 +0100
From   : kortink@... (John Kortink)
Subject: Pi-based second processor - fast, flexible,

On Tue, 31 Jan 2017 16:29:23 +0000, Ed Spittles <percy.p.person@...>
wrote:

>Indeed, although it's very fast, more testing might say more about exactly
>how fast it is.
>
>It's not clear to me that the results do show that we're measuring overhead
>- the instruction mixes in each test will differ, so the speedup will
>differ. And each additional optimisation will generally affect different
>instructions differently. But anyhow, ClockSp is just one possible workload.

Well, it /should/ produce roughly the same number for each test.
That in itself is not a sign of measuring overhead.

But it is odd that at the start of development the numbers mentioned
were actually quite different. Measuring error or some magic speedup
of a very frequently used instruction (that disappeared later on) ?
Benchmarking is tricky stuff, and the accuracy of using 'TIME' is not
that great (probably a 10 ms error on average).

I'm sure the overall figure is reasonable though.

>[...]
>
>As a single interesting data point, Dominic said (elsewhere)
>
>The fastest speedup instruction I think is :
>Opcode 91 - STA ($00),Y
>This I think in 6502 world is 6 cycles and 12 ARM cycle in our world.

Okay, so that is equivalent to 500 MHz. Not bad ...

>[...]


John Kortink