amibench answers "how much faster is machine A than machine B" with numbers
that survive scrutiny. Seven phases: integer ALU (dependency-chained, so it
cannot be pipelined away), FPU, memcpy, memset and a linear read in Fast RAM,
and copy and read in Chip RAM. It takes well under a minute and needs nothing
else installed.
NEW IN 2.0
* FPU phase. Run only when exec reports a 68881/68882 or an 040/060 FPU;
on a machine without one it prints "fpu skipped". Only that phase is
compiled for the FPU, so the binary still starts on a plain 020 or 030.
(A first build linked the FPU C library, whose startup ran FPU
instructions before main() and crashed FPU-less 030s.)
* Chip RAM phases, copy and read. Everything else allocates Fast RAM on
purpose, so this is the only part that measures the chipset bus - where
an accelerated machine can be surprisingly slow.
* Works with any amount of Chip RAM. The Chip buffers want 512 KB each,
but a 1 MB Chip machine has less than that free once the OS is in.
They now halve until both fit, down to 16 KB, and the size used is
printed as chipsz. Before, such a machine skipped the Chip phases.
* Ctrl-C stops it between attempts, and every way out frees its memory -
AllocMem'd memory is otherwise gone until the next reboot.
* Version string for the Version command.
TO RUN
From a Shell:
amibench
WHAT IT PRINTS
amibench 3 attnflags 8037 bufsz 4194304 chipsz 262144 minticks 100 fpu 1 chip 1
int reps 6400000 ticks 367
fpu reps 400000 ticks 224
copy reps 10 ticks 151
set reps 40 ticks 398
read reps 10 ticks 111
chipcopy reps 10 ticks 183
chipread reps 40 ticks 386
Deliberately raw: reps and ticks, no rates. Divide reps by (ticks/50) for
operations per second. For the Fast memory phases multiply reps by bufsz
first to get bytes, for the Chip phases by chipsz. There are 50 ticks to
the second.
Doing that arithmetic on the Amiga would drag in soft-float or an FPU
dependency and change what is being measured, and the byte totals overflow
32 bits on a fast machine. So the numbers come out raw.
THREE THINGS IT DOES DELIBERATELY
Self-calibrating. Timing comes from DateStamp, which resolves to 1/50 s.
A fixed workload is useless across a wide speed range: on a PiStorm the
first version finished an entire 50 MB memcpy phase inside a SINGLE tick,
while sizing that phase for a PiStorm would leave a 68060 grinding for
minutes. Each phase quadruples its rep count until it runs at least two
seconds, then reports the count it settled on.
Buffers bigger than any cache. 4 MB per Fast buffer, 8 MB total. At 256 KB
the whole working set sat inside a Raspberry Pi's L2 cache, so an emulated
Amiga was clocking cache bandwidth against a 68060 clocking main memory.
Fast RAM explicitly. MEMF_ANY could land in Chip RAM on a machine short of
Fast, which would measure the chipset bus rather than the CPU. The Chip
phases measure that bus on purpose, and separately.
COMPARING TWO MACHINES
Run the SAME BINARY on both. This build targets 68020, so one executable
runs on an 020, an 030, an 040, an 060 and on Emu68's JIT alike. If you
rebuild it for 68000 to run on an A500, its numbers are not comparable
with these - you would be measuring the compiler as much as the machine.
Record the clock with the result. A doubled score between two runs is
more often a changed clock than anything else.
SOME RESULTS
A2000 + 68030/68882 at 50 MHz against an A1200 + ACA1234 (68030 at
25 MHz, no FPU), both idle, November 2026:
ALU 0.86 M iter/s vs 0.41 M iter/s 2.1x
memcpy 12.9 MB/s vs 8.8 MB/s 1.5x
memset 19.8 MB/s vs 9.4 MB/s 2.1x
read 18.0 MB/s vs 10.7 MB/s 1.7x
Chip copy 1.45 MB/s vs 2.72 MB/s 0.53x
Chip read 2.78 MB/s vs 3.37 MB/s 0.82x
Twice the clock shows as twice the score where the CPU is the limit (ALU,
memset), less where Fast RAM is (memcpy, read), and not at all in Chip
RAM - the chipset bus does not care about the CPU clock, and there the
A1200's AGA bus wins. The same runs with a player, a CPU meter and a dock
open on both machines came out within 2% of these.
A1200 + PiStorm32-lite (Emu68 on a Pi 4 at 1.8 GHz) against an A4000/060
at its stock 50 MHz (amibench 1.0, Fast RAM only):
ALU 151.7 M iter/s vs 2.8 M iter/s 54x
memcpy 1.19 GB/s vs 17.6 MB/s 67x
read 3.98 GB/s vs 35.5 MB/s 112x
That cross-checks against SysInfo 4.4, which puts the same two machines
50.1x apart. But the PiStorm reaches the real chipset across its bus, so
anything Chip-bound will not see that CPU win and may run slower.
Synthetic ratios are not application ratios.
Source (tools/amibench), and the amimcp project it comes from:
https://github.com/thomas-luebker/amimcp
Distributed under the licence in the archive.
|