From lucasvr at gobolinux.org  Wed Oct 25 14:00:51 2006
From: lucasvr at gobolinux.org (Lucas C. Villa Real)
Date: Wed, 25 Oct 2006 12:00:51 -0200
Subject: [gobolinux-arm] ARM soft-float
Message-ID: <2c03f9590610250700o6fc27ecap112b8cf0855f3e2@mail.gmail.com>

Hi there,

On the last week I decided to measure how much penalty was added on
math emulation in the ARM land. I took the lmbench 3 suite and
measured the lat_op test, which basically performs a lot of int and
floating point operations trying to figure out how much time was spent
on each one.

I must admit that I was pretty impressed by the times I saw there.
I've measured the standard NWFPE emulation done in kernel (through
traps) and the emulation done by GCC soft float (by modification of FP
instructions by their correspondent integer implementation).

These are the times I got on my XScale (PXA255), in nano-seconds:

                   Sum        Multiplication  Division    BogoMflops
GCC:           182.58    124.69            355.65     865.80
NWFPE:       1075.23  1330.92          1815.00   6739.44
Difference:  5.9x        10.2x               5.11x       7.8x

I think this is a good reason to start working on a Soft-Float
distribution. I've already compiled a toolchain based on GCC 4.1.1 +
Glibc 2.4 and started to adapt our recipe for Glibc 2.4 to the ARM.

-- 
Lucas
powered by /dev/dsp

From lucasvr at gobolinux.org  Wed Oct 25 14:06:50 2006
From: lucasvr at gobolinux.org (Lucas C. Villa Real)
Date: Wed, 25 Oct 2006 12:06:50 -0200
Subject: [gobolinux-arm] Development documentation
Message-ID: <2c03f9590610250706i35c531a0jab3fa8b80c0b7291@mail.gmail.com>

There's an updated documentation on ARM development in the wiki now.

For those of you interested, here's the link to it:
http://gobo.kundor.org/wiki/GoboLinux_Embedded

-- 
Lucas
powered by /dev/dsp

From lucasvr at gobolinux.org  Wed Oct 25 14:14:04 2006
From: lucasvr at gobolinux.org (Lucas C. Villa Real)
Date: Wed, 25 Oct 2006 12:14:04 -0200
Subject: [gobolinux-arm] ARM soft-float
In-Reply-To: <2c03f9590610250700o6fc27ecap112b8cf0855f3e2@mail.gmail.com>
References: <2c03f9590610250700o6fc27ecap112b8cf0855f3e2@mail.gmail.com>
Message-ID: <2c03f9590610250714j59b24918t8cc183b23e66a657@mail.gmail.com>

On 10/25/06, Lucas C. Villa Real <lucasvr at gobolinux.org> wrote:
> These are the times I got on my XScale (PXA255), in nano-seconds:
>
>                    Sum        Multiplication  Division    BogoMflops
> GCC:           182.58    124.69            355.65     865.80
> NWFPE:       1075.23  1330.92          1815.00   6739.44
> Difference:  5.9x        10.2x               5.11x       7.8x

Just clarifying: less is better. BogoMflops also isn't a good name, as
it's actually a combo of fpu operations. Less is better again (the
same FPU operations were executed in both tests - the additional time
shows the overhead).

-- 
Lucas
powered by /dev/dsp