Building sr1sieve for RPi

wombatman · 2017-03-02, 04:21

I'm trying to build sr1sieve on the RPi Model 3B running Raspbian Jessie 4.4 (32-bit), and encountering the following error:

Code:

gcc -O2 -ffast-math -march=armv8-a -mtune=cortex-a53  -Wall -DNDEBUG  -c -o sr1sieve.o sr1sieve.c
In file included from sr1sieve.c:23:0:
arithmetic.h: In function ‘PRE2_MULMOD64_INIT’:
arithmetic.h:97:3: error: ‘b_over_p’ undeclared (first use in this function)
   b_over_p = (double)((int64_t)b) * one_over_p;

The relevant line from arithmetic.h is:

Code:

extern double _b_over_p;

GCC is version 4.9.2.

This obviously isn't an emergency, but I'd like to learn more about the Pi and what it can do. Thanks for any help.

ET_ · 2017-03-02, 09:43

Quote:

Originally Posted by wombatman

I'm trying to build sr1sieve on the RPi Model 3B running Raspbian Jessie 4.4 (32-bit), and encountering the following error:

Code:

gcc -O2 -ffast-math -march=armv8-a -mtune=cortex-a53  -Wall -DNDEBUG  -c -o sr1sieve.o sr1sieve.c
In file included from sr1sieve.c:23:0:
arithmetic.h: In function ‘PRE2_MULMOD64_INIT’:
arithmetic.h:97:3: error: ‘b_over_p’ undeclared (first use in this function)
   b_over_p = (double)((int64_t)b) * one_over_p;

The relevant line from arithmetic.h is:

Code:

extern double _b_over_p;

GCC is version 4.9.2.

This obviously isn't an emergency, but I'd like to learn more about the Pi and what it can do. Thanks for any help.

I *suppose* it's an issue related to the 32-bit OS not having 64-bit instructions. I had a similar error on my RPI 3 when I tried to compile my own sieve (but here I also had an asm instruction I had to recode in C, so I could be proved wrong).

wombatman · 2017-03-02, 14:45

It could be, but I can't even get the 32-bit versions to build. I had no problems getting the GGNFS sievers to build a 32-bit version, but this one has stumped me so far.

Thomas11 · 2017-03-02, 20:36

You can easily change this line:

Code:

extern double _b_over_p;

into

Code:

extern double b_over_p;

Then it should compile without any problems.

I tried this on a Pi2 using the source code version 1.4.5 of sr1sieve and got a stable binary.
The single core performance is about 1/20th of a single i3770 core.

Thomas11 · 2017-03-02, 20:41

I forgot to mention that in the Makefile you need to disable the x86 assembly code:

Code:

ASSEMBLER=no

And also unselect any predefined architecture, e.g.:

Code:

#ARCH=x86_64

wombatman · 2017-03-02, 20:48

Fantastic! I'll try this at home tonight.

wombatman · 2017-03-03, 00:07

Just writing to confirm that your steps worked, and that the performance of the binary is atrocious. 4-threads on an i7-4930K gives about 125-130 million p/sec. 4-threads on the Model 3B gives about 5 million p/sec. Still, now have something to mess with, and I can try some other compilation with it. Thanks again!

ET_ · 2017-03-03, 16:30

Quote:

Originally Posted by wombatman

Just writing to confirm that your steps worked, and that the performance of the binary is atrocious. 4-threads on an i7-4930K gives about 125-130 million p/sec. 4-threads on the Model 3B gives about 5 million p/sec. Still, now have something to mess with, and I can try some other compilation with it. Thanks again!

Just to be sure... did you compare the results of the sieve between Inter and Arm versions? Do they sieve the same results and find the same factors?

Thomas11 · 2017-03-03, 16:46

Quote:

Originally Posted by ET_

Just to be sure... did you compare the results of the sieve between Inter and Arm versions? Do they sieve the same results and find the same factors?

I did it at my end on the Pi2B, and the results and factors are matching.

wombatman · 2017-03-03, 17:05

Quote:

Originally Posted by ET_

Just to be sure... did you compare the results of the sieve between Inter and Arm versions? Do they sieve the same results and find the same factors?

I did not, but I won't be able to for now. I'm currently working on getting the Gentoo 64-bit image loaded and working properly. I'm having some overheating issues at the moment.

ET_ · 2017-03-03, 17:52

Quote:

Originally Posted by wombatman

I did not, but I won't be able to for now. I'm currently working on getting the Gentoo 64-bit image loaded and working properly. I'm having some overheating issues at the moment.

Keep me informed about it :-) I suppose you could get a 2.5x speedup from using the Arm8 code, the FPUs and the 64bits.

2017-03-02, 04:21	#1
wombatman I moo ablest echo power! May 2013 1855₁₀ Posts	Building sr1sieve for RPi I'm trying to build sr1sieve on the RPi Model 3B running Raspbian Jessie 4.4 (32-bit), and encountering the following error: Code: gcc -O2 -ffast-math -march=armv8-a -mtune=cortex-a53 -Wall -DNDEBUG -c -o sr1sieve.o sr1sieve.c In file included from sr1sieve.c:23:0: arithmetic.h: In function ‘PRE2_MULMOD64_INIT’: arithmetic.h:97:3: error: ‘b_over_p’ undeclared (first use in this function) b_over_p = (double)((int64_t)b) * one_over_p; The relevant line from arithmetic.h is: Code: extern double _b_over_p; GCC is version 4.9.2. This obviously isn't an emergency, but I'd like to learn more about the Pi and what it can do. Thanks for any help.

2017-03-02, 20:36	#4
Thomas11 Feb 2003 1920₁₀ Posts	You can easily change this line: Code: extern double _b_over_p; into Code: extern double b_over_p; Then it should compile without any problems. I tried this on a Pi2 using the source code version 1.4.5 of sr1sieve and got a stable binary. The single core performance is about 1/20th of a single i3770 core.

2017-03-02, 20:41	#5
Thomas11 Feb 2003 2⁷·3·5 Posts	I forgot to mention that in the Makefile you need to disable the x86 assembly code: Code: ASSEMBLER=no And also unselect any predefined architecture, e.g.: Code: #ARCH=x86_64

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
sr1sieve with work files?	masser	Conjectures 'R Us	1	2017-12-28 17:43
32-bit of sr1sieve and sr2sieve for Win	pepi37	Software	5	2013-08-09 22:31
Building gcc 4.4.0	CRGreathouse	Software	1	2009-07-07 22:25
sr1sieve and big P	barcode	Software	5	2007-12-27 04:05
sr1sieve 1.2.0	arminius	Riesel Prime Search	14	2007-12-02 22:53

2017-03-02, 14:45	#3
wombatman I moo ablest echo power! May 2013 5×7×53 Posts	It could be, but I can't even get the 32-bit versions to build. I had no problems getting the GGNFS sievers to build a 32-bit version, but this one has stumped me so far.

2017-03-02, 20:48	#6
wombatman I moo ablest echo power! May 2013 5·7·53 Posts	Fantastic! I'll try this at home tonight.

2017-03-03, 00:07	#7
wombatman I moo ablest echo power! May 2013 1855₁₀ Posts	Just writing to confirm that your steps worked, and that the performance of the binary is atrocious. 4-threads on an i7-4930K gives about 125-130 million p/sec. 4-threads on the Model 3B gives about 5 million p/sec. Still, now have something to mess with, and I can try some other compilation with it. Thanks again!