Bitcoin x64 for Windows

[Bitquux]

So --- Bitquux, that was Olipro's binary that found the coin?

Negative, that was before the switch. And at this rate, it will be a long time before I can tell if it will. Even if it's a dirty bin it's clever enough to earn a donation.

[Olipro]

So --- Bitquux, that was Olipro's binary that found the coin?

Negative, that was before the switch. And at this rate, it will be a long time before I can tell if it will. Even if it's a dirty bin it's clever enough to earn a donation.

It shouldn't be dirty, the SHA256 is untouched and I don't think the compiler has introduced any errors.

[Bitquux]

It shouldn't be dirty, the SHA256 is untouched and I don't think the compiler has introduced any errors.

Yes, let me clarify. This is not a dirty bin. There is no reason so far to believe it's anything other than what Olipro says it is. The performance of my previously mentioned machine has even gone up since I first posted those numbers.

[knightmb]

Here's a weird question.. has anyone actually generated a block with this faster version?

I have a few machines that used to regularly generate, and since switching to this version -- zip.  I know these things are subject to random variation, and it could be a dry spell, and the difficulty is going up, but... could it be a bug?

Yes, about 10 so far. The other "regular" clients acknowledge test Coin transfer between the two, so far it appears good. I also have a packet sniffer running on my test machine and it's basically doing what the other clients do, nothing out of the ordinary so far.

[Olipro]

This worked perfectly! On linux and windows if you compile yourself. I went from 380khash/s on one core to over 1200khash/s!

yeah, because dividing a number by 300 instead of 1000 makes it look a lot bigger, doesn't *actually* improve performance.

[Olipro]

OK, I've made a new build of Bitcoin, this one is compiled using the Intel compiler which is considerably more advanced than the standard MS compiler.

Please note this build is 64-bit only since I see no reason to compile a build for 32-bit when most 32-bit processors lack the newer SSE instructions anyway.

Performance? my MSVC build averages about 2400K for me, this averages 2900K so you're looking at an improvement of about 125k per core or thereabouts although this does come at the cost of a larger EXE, improved performance is worth it in my opinion

My next goal is to see if this PolarSSL SHA256 algo really is faster. for now however, download Bitcoin x64 ICC optimised build here

[Ground Loop]

Runs about the same speed as the VC++ version here (Xeon, Win7 64-bit).

[knightmb]

OK, I've made a new build of Bitcoin, this one is compiled using the Intel compiler which is considerably more advanced than the standard MS compiler.

Please note this build is 64-bit only since I see no reason to compile a build for 32-bit when most 32-bit processors lack the newer SSE instructions anyway.

Performance? my MSVC build averages about 2400K for me, this averages 2900K so you're looking at an improvement of about 125k per core or thereabouts although this does come at the cost of a larger EXE, improved performance is worth it in my opinion

My next goal is to see if this PolarSSL SHA256 algo really is faster. for now however, download Bitcoin x64 ICC optimised build here

Awesome, gives a 41% speed boost over the stock binary.

My experience so far.

Stock = Stock
Last x64 Build = 28% speed increase over stock
This x64 Build = 41% speed increase over stock

Nice stuff!

[Olipro]

Runs about the same speed as the VC++ version here (Xeon, Win7 64-bit).

which Xeon, there's quite a few.

I don't think I need to point this out but here goes anyway: the significance of the performance benefit from using any of my builds will depend entirely on how many cores your computer has.

Also, in case anyone missed it: the PolarSSL algo is *not* faster.

next step... CUDA.

[ichi]

What would I expect for two Xeon 5570?

[Olipro]

What would I expect for two Xeon 5570?

a 4 core processor with hyperthreading... and two of them? a pretty significant performance increase.

[Olipro]

OK, I've made a new EXE, this one seems to get me an extra 100-200k (or about 25-50k per core).

Difference? I modified the ByteSwap function to operate on 64-bit integers; it does this by using the bswap intrinsic on a 64 bit register followed by rotate right through 32 bits to put the result in the correct order. it also initializes the SHA256 vectors using unsigned 64 bit values (however, the actual hashing still uses 32-bit so I doubt this is making much of a difference) and yes, I did convert the 32 bit numbers to 64 bit correctly (i.e. 0x12345678UL 0xabcdef0UL -> 0xabcdef012345678ULL) if that appears wrong to you, think about how little endian machines store 32 bit integers in memory.

anyway, grab it here

[Quantumplation]

Wow, that tweaked one... I'm getting stunning rates.  I have a quad core Intel laptop.

I have two different number sets, 1 while operating as usual, with other programs running, and one while everything but bitcoin (including explorer.exe) is shut down.

	Stock:	x64 v1	x64 v2	x64 v2 Tweaked
Standard Usage:	500-1200	1000-1800	750-1500	2200-2700
Optimized Usage:	1500-1800	2000-2500	1500-2000	2700-3400

O_O

[m0mchil]

May I kindly ask you, mighty Olipro, to patch bitcoin 0.3.1 or may be 0.3.2? There were some issues in 0.3

[knightmb]

OK, I've made a new EXE, this one seems to get me an extra 100-200k (or about 25-50k per core).

Difference? I modified the ByteSwap function to operate on 64-bit integers; it does this by using the bswap intrinsic on a 64 bit register followed by rotate right through 32 bits to put the result in the correct order. it also initializes the SHA256 vectors using unsigned 64 bit values (however, the actual hashing still uses 32-bit so I doubt this is making much of a difference) and yes, I did convert the 32 bit numbers to 64 bit correctly (i.e. 0x12345678UL 0xabcdef0UL -> 0xabcdef012345678ULL) if that appears wrong to you, think about how little endian machines store 32 bit integers in memory.

anyway, grab it here

I'll give this one a run, the last build would crash randomly after a few hours 

[Olipro]

May I kindly ask you, mighty Olipro, to patch bitcoin 0.3.1 or may be 0.3.2? There were some issues in 0.3

my builds are from the latest SVN source code, of course, had you checked the about page, you would notice it clearly states itself as 0.3.2

[m0mchil]

hmmmmm, that's strange... because I am experiencing the problem already fixed in http://bitcoin.svn.sourceforge.net/viewvc/bitcoin?view=revision&revision=102

[Olipro]

hmmmmm, that's strange... because I am experiencing the problem already fixed in http://bitcoin.svn.sourceforge.net/viewvc/bitcoin?view=revision&revision=102

then either it got re-broken or he also modified something in wxWidgets itself which I don't have.

on an unrelated note: I have CUDA working for SHA256 hashing... currently the host has to iterate the nonce so performance is terrible, working on moving the whole kit and kaboodle into the card and then performance should be pretty damn sexy.

[sgtstein]

DEFINITELY in to test CUDA!

[Quantumplation]

Oli: One alternative is to start the CUDA part hashing with a nonce of MAXINT and subtract one, and the host client start at one and go upwards.  Then, either one could find a hash and you're not repeating work.