[OpenRISC] FPU benchmarking

[BAndViG]

I have tried to find an open sourced FPU benchmark and faced with a number
of problems.
The FPBench of www.eembc.org (CoreMark holder) is accessible for registered
membership. (By the way the CoreMark itself is available for free
registration, but the registration doesn't grant you membership status).
With Google II 've found:
a) Floating Point Benchmarks by John Walker
(https://www.fourmilab.ch/fbench). But is written with doubles and have to
be ported
b) FPU tests: Whetstone and Sons, Ltd.
(http://www.tux.org/~balsa/linux/benchmarking/articles/html/Article1d-3.html).
But the mentined ftp (ftp://ftp.nosc.mil/pub/aburto) is not accessible.

Does anybody knows another open sourced FPU benchmarks?

WBR
Andrey

[Sebastian Macke]

Hi,

I took a look at the other CPU benchmarks which I know and which don't
appear on the first page on Google,
but couldn't find anything suitable.

nbench:
http://www.tux.org/~mayer/linux/bmark.html
Unfortunately only one test runs with floats. And they only publish
total floating point scores

povray:
http://www.povray.org/download/benchmark.php
Probably also mainly doubles.

Sebastian

Post by BAndViG
I have tried to find an open sourced FPU benchmark and faced with a
number of problems.
The FPBench of www.eembc.org (CoreMark holder) is accessible for
registered membership. (By the way the CoreMark itself is available
for free registration, but the registration doesn't grant you
membership status).
a) Floating Point Benchmarks by John Walker
(https://www.fourmilab.ch/fbench). But is written with doubles and
have to be ported
b) FPU tests: Whetstone and Sons, Ltd.
(http://www.tux.org/~balsa/linux/benchmarking/articles/html/Article1d-3.html).
But the mentined ftp (ftp://ftp.nosc.mil/pub/aburto) is not accessible.
Does anybody knows another open sourced FPU benchmarks?
WBR
Andrey
_______________________________________________
OpenRISC mailing list
http://lists.openrisc.net/listinfo/openrisc

[BAndViG]

Additionally, I've found classic benchmarks on
http://www.roylongbottom.org.uk including floating point focused
'whetstone'. It could be configured either for single or double precision by
just one macro definition. However its time measurement is based on
clock_gettime() method that (as far as I understand) isn't implemented for
OR1K newlib tool chain. So, the whetstone's time measurement engine have to
be ported on OR1K timer basis.
Am I right?
Or is there shorter way?

Andrey


-----Исходное сообщение-----
From: Sebastian Macke
Sent: Saturday, September 20, 2014 8:58 PM
To: ***@lists.openrisc.net
Subject: Re: [OpenRISC] FPU benchmarking

Hi,

I took a look at the other CPU benchmarks which I know and which don't
appear on the first page on Google,
but couldn't find anything suitable.

nbench:
http://www.tux.org/~mayer/linux/bmark.html
Unfortunately only one test runs with floats. And they only publish
total floating point scores

povray:
http://www.povray.org/download/benchmark.php
Probably also mainly doubles.

Sebastian

Post by BAndViG
I have tried to find an open sourced FPU benchmark and faced with a number
of problems.
The FPBench of www.eembc.org (CoreMark holder) is accessible for
registered membership. (By the way the CoreMark itself is available for
free registration, but the registration doesn't grant you membership
status).
a) Floating Point Benchmarks by John Walker
(https://www.fourmilab.ch/fbench). But is written with doubles and have to
be ported
b) FPU tests: Whetstone and Sons, Ltd.
(http://www.tux.org/~balsa/linux/benchmarking/articles/html/Article1d-3.html).
But the mentined ftp (ftp://ftp.nosc.mil/pub/aburto) is not accessible.
Does anybody knows another open sourced FPU benchmarks?
WBR
Andrey
_______________________________________________
OpenRISC mailing list
http://lists.openrisc.net/listinfo/openrisc

[Jakob Sandgren]

Hi,

Post by BAndViG
Does anybody knows another open sourced FPU benchmarks?

I/we work a lot with FPU-heavy users/customers in the High Performance
Computing sector and one benchmark is by far the most common and it's
Open Source; HPL /Linpack.

http://www.netlib.org/benchmark/hpl/

There is a few variants of this, both the "core" BLAS implementation
and there is also more "complete" code from Intel for their
cpu:s. It's such an important benchmark that both Intel and AMD has
released optimized code/libraries for this.
For common cores (Intel/AMD) the implementations are now so optimized
that I would say it's close to what you possible can get out of these
cores.
I would just _love_ to see an optimized fpu core for this, however,
thats beyond both my time and knowledge! ;)

BR
Jakob

[BAndViG]

Thanks, Jackob
Even so the HPC focused benchmark looks "too cool" for core like or1k :) ,
I'll add it into my collection in some time.

Andrey

-----Исходное сообщение-----
From: Jakob Sandgren
Sent: Monday, September 22, 2014 12:01 PM
To: BAndViG
Cc: ***@lists.openrisc.net
Subject: Re: [OpenRISC] FPU benchmarking

Hi,

Post by BAndViG
Does anybody knows another open sourced FPU benchmarks?

I/we work a lot with FPU-heavy users/customers in the High Performance
Computing sector and one benchmark is by far the most common and it's
Open Source; HPL /Linpack.

http://www.netlib.org/benchmark/hpl/

There is a few variants of this, both the "core" BLAS implementation
and there is also more "complete" code from Intel for their
cpu:s. It's such an important benchmark that both Intel and AMD has
released optimized code/libraries for this.
For common cores (Intel/AMD) the implementations are now so optimized
that I would say it's close to what you possible can get out of these
cores.
I would just _love_ to see an optimized fpu core for this, however,
thats beyond both my time and knowledge! ;)

BR
Jakob

[BAndViG]

I've ported Whetstone benchmark and run it both with soft & hard float
options for or1k compiler.
The original description of the benchmark could be found on
http://www.roylongbottom.org.uk/whetstone.pdf .
The results are (I've removed the column containing calculated values as
they are the same for both cases):

case #1: -msoft-float
Single Precision C/C++ Whetstone Benchmark

Loop content MFLOPS MOPS Seconds

N1 floating point 0.392 0.049
N2 floating point 0.320 0.420
N3 if then else 3.450 0.030
N4 fixed point 0.984 0.320
N5 sin,cos etc. 0.019 4.330
N6 floating point 0.242 2.230
N7 assignments 1.680 0.110
N8 exp,sqrt etc. 0.009 4.300

MWIPS 0.848 11.789


case #2: -mhard-float, fpu32_v1.0:
Single Precision C/C++ Whetstone Benchmark

Loop content MFLOPS MOPS Seconds

N1 floating point 2.400 0.008
N2 floating point 2.240 0.060
N3 if then else 3.450 0.030
N4 fixed point 3.938 0.080
N5 sin,cos etc. 0.019 4.300
N6 floating point 1.199 0.450
N7 assignments 1.680 0.110
N8 exp,sqrt etc. 0.009 4.300

MWIPS 1.071 9.338


It is interesting to see that N5 (sin, cos, etc) and N8 (exp,sqrt, etc) have
the same performance as for hard as for soft float operation.

Does it mean that these functions are based on soft-float routines
independently on -mfloat option?

PS. I haven't found any license agreement in source package. That is why I
doubt about publishing the modified code. So, if anybody is interested in
sources of or1k port of Whetstone, e-mail me and I'll send you 6.5 KB zip
archive.

BR
Andrey