Help needed to test Athlon64 code

[geoff]

I am trying to implement 64 bit modular multiplication for the Athlon 64 but I don't have access to one for testing. If someone with gcc, make, etc. installed and running a 64 bit OS could compile and run this test program for me it would be a great help: k8test.zip

Just unzip and type `make', it will run two test programs, one in 32 bit mode and one in 64 bit mode, and print the results to a file `k8.txt'. Post that file or any errors here, or send me a PM.

Thanks,
Geoff.

[jasonp]

Originally Posted by geoff

Just unzip and type `make', it will run two test programs, one in 32 bit mode and one in 64 bit mode, and print the results to a file `k8.txt'. Post that file or any errors here, or send me a PM.

You may want to use a higher-resolution timer...

jasonp

uint64
read_clock( void ) {

#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
	uint32 lo, hi;
	asm("rdtsc":"=d"(hi),"=a"(lo));
	return (uint64)hi << 32 | lo;

#elif defined(_MSC_VER)
	LARGE_INTEGER ret;
	QueryPerformanceCounter(&ret);
	return ret.QuadPart;

#else
	struct timeval thistime;   
	gettimeofday(&thistime, NULL);
	return thistime.tv_sec * 1000000 + thistime.tv_usec;
#endif
}

[dsouza123]

Bochs might be a solution.

Bochs emulates an x86 CPU and some other PC chips
and uses hard drive images.

Among the CPUs emulated are a Pentium Pro and an Athlon 64.

I have used both emulated CPUs with good success.

Used it to run Win98 SE with an emulated Athlon 64 (for SSE2 support)
on top of Win XP Pro with a real Athlon (early 32 bit model no SSE2 ).

There are various Linux images available.

http://bochs.sourceforge.net/

[geoff]

Thanks for the help everyone.

From the looks of it in 64 bit mode the generic C mulmod using the FPU is much faster than the assembler for both 32 and 64 bit arguments, but in 32 bit mode the assembler integer version is a little faster for 32 bit arguments and the assembler FPU version for 64 bit arguments.

[c00ler]

Something strange with latest prime95 x64 under windows 2003 x64 (athlon 64 3200+, 2gb ram).

1) Start torture test or prime testing, then (maybe even after some time) start playing video (wmplayer x32, divx player x32)
Exactly when video starts i get illegal sumout, or bad rounding (0.4993896484 or even 0.5)

2) VirtualDub x64 playing/compressing video doesn't crash prime95 x64

Maybe the reason is http://www.intel.com/support/perform.../cs-020438.htm ?

But why x32 and not x64 codecs cause errors is a mystery to me =)

[dsouza123]

If the SSE2 registers aren't saved/restored correctly
so they become corrupted, does it also affect the FPU ?

In local.ini add
CpuSupportsSSE=0
CpuSupportsSSE2=0
then it would simulate an earlier Athlon using only the FPU.

Without disabling SSE/SSE2 support,
does it also affect Prime95 x86-32 run under Windows 2003 x64 ?

[c00ler]

Tried
CpuSupportsSSE=0
CpuSupportsSSE2=0

These options don't affect corruption, also there is not even a 1% difference in benchmark in both 32 bit and x64.

32 bit prime95 is not affected by corruption, with and without SSE/SSE2.

P.S. maybe i can recompile prime95 with intel c++ 9.1, but i need instructions =)

[Prime95]

Originally Posted by dsouza123

In local.ini add
CpuSupportsSSE=0
CpuSupportsSSE2=0
then it would simulate an earlier Athlon using only the FPU.

Nice try, but this won't work on the 64-bit prime95 version. It assumes (and requires) SSE2 support.