suggestion: separation of libraries

ixfd64 · 2010-10-19, 22:44

For those unaware, the size of the version 26.x binary is about 25 MB (~4 MB compressed) in size, compared to just below 5 MB (~1 MB compressed) for 25.x. According to George, this is due to Prime95 supporting more FFT sizes and new optimized code for different microarchitectures. In today's standards, 25 MB is hardly anything.

However, the smaller size of 25.x allowed users to easily install Prime95 using a floppy disk. Of course, everyone (well, almost everyone) uses USB flash drives these days, but it would still be nice to be able to keep the Prime95 file size below a few MB.

Therefore, users should be able to choose what libraries to download. The libraries would be stored in the /lib folder. Possible libraries would be:

all.lib - contains code for all microarchitectures
ge.lib - generic FFT code (from version 25)

amd-lg.lib - legacy AMD microarchitectures (K6 and earlier)
amd-k8.lib - AMD K8
amd-k9.lib - AMD K9
amd-k10.lib - AMD K10
amd-fu.lib - AMD Fusion (future)

int-lg.lib - legacy Intel microarchitectures (P6 and earlier)
int-mb.lib - Intel Atom and other mobile microarchitectures
int-nb.lib - NetBurst
int-nh.lib - Nehalem
int-sb.lib - Sandy Bridge (future)
int-hw.lib - Haswell (future)

Of course, libraries for GPU code would also be available:

mfaktc.lib - based on Oliver's code (CUDA TF)
cuda-ll.lib - based on msft's code (CUDA LL tests)
etc.

James Heinrich · 2011-01-19, 14:01

I like it, but I'd also suggest having the default client download as the be-all installer that contains everything, examines the system and installs the appropriate 32/64-bit client and appropriate libraries based on detected hardware. Optionally, a full install from the same installer could copy all libraries for the user to later pick-and-choose to copy to another machine.

ATH · 2011-01-19, 14:25

The compressed size is only 3.7 Mb which even if someone still has a dialup modem only would take 15-30min (I remember getting 10-15 Mb/hour at best back then).

And do you think anyone is using floppydisc to move Prime95 to different computers now? If someone did, they could split it up over just 3 floppydiscs using any compression software. I had games spanning 20+ floppydiscs just before I got CD drives.

I think it would be just be another source of confusion and errors for minimal or no gain.

Ralf Recker · 2011-01-19, 15:36

It's possible to recompress the archives* with 7z and transport them on a single old 1.44MB floppy disk (less than 900KB with standard compression - add a few K for a self extracting archive).

*With the exception of the recompressed source archive: It's size is 5.1 MB (standard compression, 3.5 MB with maximum compression -mx9). A little bit smaller than the 32.3 MB .zip file...

ixfd64 · 2011-01-19, 17:11

Another idea would be to have Prime95 detect what hardware the user is using and automatically download the libraries.

vsuite · 2011-01-20, 03:34

I agree with the idea of separate libraries to reduce the size of prime95. Version 26.x runs much slower on my Celeron than 25.x, which I suspect is largely due to the much larger exe size.

axn · 2011-01-20, 03:37

Quote:

Originally Posted by vsuite

I agree with the idea of separate libraries to reduce the size of prime95. Version 26.x runs much slower on my Celeron than 25.x, which I suspect is largely due to the much larger exe size.

No. George omitted a lot of P4-optimized FFTs (> 3072K?) from v26.

ixfd64 · 2011-01-20, 05:33

Since the Prime95 source code is publicly available, it's entirely possible to add the FFT optimization code back.

There are two problems, however:

1. Very few end users would know how to do something like this.
2. The public source code does not include the checksum-generating routines. As such, any user-compiled version of Prime95 will be mark as "untrusted."

Maybe someone can create a third-party patch that adds the P4 optimizations back to 26.x?

axn · 2011-01-20, 15:18

Quote:

Originally Posted by ixfd64

Maybe someone can create a third-party patch that adds the P4 optimizations back to 26.x?

Isn't that called v25?

ixfd64 · 2011-01-20, 18:04

Yes, but 25.11 still has quite a few bugs, mainly the infamous P-1 segfault crash.

George has recommended that P4 users switch to double-checking when exponents go beyond the 3072k FFT range. However, this may not be the best choice for computers that are only on for a few hours a day.

Prime95 · 2011-01-20, 23:02

Quote:

Originally Posted by ixfd64

George has recommended that P4 users switch to double-checking when exponents go beyond the 3072k FFT range.

That recommendation is for Northwood P4's only. There are P4 optimized FFTs up to at least 4096K if you have 1MB of L2 cache.

2010-10-19, 22:44	#1
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2·5·239 Posts	suggestion: separation of libraries For those unaware, the size of the version 26.x binary is about 25 MB (~4 MB compressed) in size, compared to just below 5 MB (~1 MB compressed) for 25.x. According to George, this is due to Prime95 supporting more FFT sizes and new optimized code for different microarchitectures. In today's standards, 25 MB is hardly anything. However, the smaller size of 25.x allowed users to easily install Prime95 using a floppy disk. Of course, everyone (well, almost everyone) uses USB flash drives these days, but it would still be nice to be able to keep the Prime95 file size below a few MB. Therefore, users should be able to choose what libraries to download. The libraries would be stored in the /lib folder. Possible libraries would be: all.lib - contains code for all microarchitectures ge.lib - generic FFT code (from version 25) amd-lg.lib - legacy AMD microarchitectures (K6 and earlier) amd-k8.lib - AMD K8 amd-k9.lib - AMD K9 amd-k10.lib - AMD K10 amd-fu.lib - AMD Fusion (future) int-lg.lib - legacy Intel microarchitectures (P6 and earlier) int-mb.lib - Intel Atom and other mobile microarchitectures int-nb.lib - NetBurst int-nh.lib - Nehalem int-sb.lib - Sandy Bridge (future) int-hw.lib - Haswell (future) Of course, libraries for GPU code would also be available: mfaktc.lib - based on Oliver's code (CUDA TF) cuda-ll.lib - based on msft's code (CUDA LL tests) etc. Last fiddled with by ixfd64 on 2010-10-19 at 22:45

2011-01-19, 14:25	#3
ATH Einyen Dec 2003 Denmark 3159₁₀ Posts	The compressed size is only 3.7 Mb which even if someone still has a dialup modem only would take 15-30min (I remember getting 10-15 Mb/hour at best back then). And do you think anyone is using floppydisc to move Prime95 to different computers now? If someone did, they could split it up over just 3 floppydiscs using any compression software. I had games spanning 20+ floppydiscs just before I got CD drives. I think it would be just be another source of confusion and errors for minimal or no gain. Last fiddled with by ATH on 2011-01-19 at 14:29

2011-01-19, 15:36	#4
Ralf Recker Oct 2010 BF₁₆ Posts	It's possible to recompress the archives* with 7z and transport them on a single old 1.44MB floppy disk (less than 900KB with standard compression - add a few K for a self extracting archive). With the exception of the recompressed source archive: It's size is 5.1 MB (standard compression, 3.5 MB with maximum compression -mx9). A little bit smaller than the 32.3 MB .zip file... Last fiddled with by Ralf Recker on 2011-01-19 at 16:03 Reason: replaced binaries with archives and added data for source264.zip*

2011-01-19, 17:11	#5
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 956₁₆ Posts	Another idea would be to have Prime95 detect what hardware the user is using and automatically download the libraries. Last fiddled with by ixfd64 on 2011-01-19 at 17:11

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Separation of Church and State	rogue	Soap Box	39	2011-04-16 21:20
Separation of GPU contributions	cheesehead	Data	6	2010-08-24 23:21
Interested in C/C++ libraries for factoring	PythonPower	Factoring	27	2009-05-28 17:08
Just a suggestion...	bearnol	Miscellaneous Math	0	2006-04-23 07:50
Large Integer Libraries	andi314	Programming	18	2004-06-21 22:37

2011-01-19, 14:01	#2
James Heinrich "James Heinrich" May 2004 ex-Northern Ontario 110101010000₂ Posts	I like it, but I'd also suggest having the default client download as the be-all installer that contains everything, examines the system and installs the appropriate 32/64-bit client and appropriate libraries based on detected hardware. Optionally, a full install from the same installer could copy all libraries for the user to later pick-and-choose to copy to another machine.

2011-01-20, 03:34	#6
vsuite Jan 2010 162₈ Posts	I agree with the idea of separate libraries to reduce the size of prime95. Version 26.x runs much slower on my Celeron than 25.x, which I suspect is largely due to the much larger exe size.

2011-01-20, 05:33	#8
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 4526₈ Posts	Since the Prime95 source code is publicly available, it's entirely possible to add the FFT optimization code back. There are two problems, however: 1. Very few end users would know how to do something like this. 2. The public source code does not include the checksum-generating routines. As such, any user-compiled version of Prime95 will be mark as "untrusted." Maybe someone can create a third-party patch that adds the P4 optimizations back to 26.x?

2011-01-20, 18:04	#10
ixfd64 Bemusing Prompter "Danny" Dec 2002 California 2·5·239 Posts	Yes, but 25.11 still has quite a few bugs, mainly the infamous P-1 segfault crash. George has recommended that P4 users switch to double-checking when exponents go beyond the 3072k FFT range. However, this may not be the best choice for computers that are only on for a few hours a day.