Help: About memcpy()

**naruto** · 06-21-2004

In my program, there is a iteration which does loop 1000000 times, each time it calls memcpy(). The iteration can't be improved or replaced for some reason, so I just want to use an optimized memcpy() to replace the standard version. I read some articles about memcpy optimization early years, but I can't remember the details. I have searched some sample code about optimization version, but almost all of them are useless in my program or just be optimized for AMD cpus. So who can give me some advice or sample whatever it is plain C code or embedded assemble code. My CPU is Intel P4. Thanks!

**linuxdude** · 06-21-2004

are you sure the iteration can't be improved or is this a task you have to do and that is why. If you want I guess instead of a character array you could have an allocated pointer then at the start set it to NULL

**learnc_kr** · 06-21-2004

hi,
instead of copying bye by byte, you can copy 4 bytes at a time until you need to copy more than 4 bytes, after you can copy byte by byte. it will be faster.

regards

**xErath** · 06-21-2004

What about this?

Code:

void *memcpy(void *dest, const void *src, int size){
	char *td = (char*)dest;
	char *ts = (char*)src;
	while(size--!=0)
		*td++=*ts++;
	return dest;
}

//EDIT
Or improved

Code:

void *memcpy(void *dest, const void *src, int size){
	int *td = (int*)dest;
	int *ts = (int*)src;
	char *cd, *cs;
	
	for(;size>4;size-=4)
		*td++=*ts++;

	cd=(char*)td;
	cs=(char*)ts;
	while(size-->0)
		*cd++=*cs++;
	return dest;
}

Or even better: with 8 byte integers

Code:

#ifdef WIN32
	typedef __int64 INT64;
#else
	typedef long long int INT64;
#endif

void *memcpy(void *dest, const void *src, int size){
	INT64 *td = (INT64*)dest;
	INT64 *ts = (INT64*)src;
	char *cd, *cs;
	
	for(;size>8;size-=8)
		*td++=*ts++;

	cd=(char*)td;
	cs=(char*)ts;
	while(size-->0)
		*cd++=*cs++;
	return dest;
}

Microsoft compilers don't suport "long long int"... bah

**naruto** · 06-21-2004

I just use MD5 algorithm, the code is from Internet. In wvMD5StoreDigest() and wvMD5Update(), there are totally three memcpy() calls. I have another function (has the large iteration) to call these MD5 function. Here is the function code:

Code:

int KeySearch(unsigned char salt[],  
                       unsigned char hashedsalt[],  // 16
	       unsigned char begin[],      // 5
	       unsigned long step,
	       unsigned char validkey[]    //5
			  )
{
	wvMD5_CTX mdContext,mdContext2;
	U8 pwarray[64];
	rc4_key key;
	unsigned long *pKey = (unsigned long *)pwarray;
	unsigned long i;

	memset (pwarray, 0, 64);
	memcpy (pwarray, begin, 5);
	pwarray[9] = 0x80;
	pwarray[56] = 0x48;
	salt[16] = 0x80;
	memset (salt + 17, 0, 47);
	salt[56] = 0x80;
	for (i=0; i < step; ++i) {

		wvMD5Init (&mdContext);
		wvMD5Update (&mdContext, pwarray, 64);
		wvMD5StoreDigest (&mdContext);
        	(*pKey)++;
     		prepare_key (mdContext.digest, 16, &key);
		rc4 (salt, 16, &key);
		rc4 (hashedsalt, 16, &key);


		wvMD5Init (&mdContext2);
		wvMD5Update (&mdContext2, salt, 64);
		wvMD5StoreDigest (&mdContext2);

	  	if ((memcmp (mdContext2.digest, hashedsalt, 16)) == 0) {
			memcpy(validkey, mdContext.digest, 5);
			return 1;
		}
	}
	return 0;
}

**xErath** · 06-21-2004

naruto please use code tags.
Do you know where all those functions come from???
Bah I find my code much simple.

**naruto** · 06-22-2004

Originally Posted by xErath

naruto please use code tags.
Do you know where all those functions come from???
Bah I find my code much simple.

Sorry, I'm new here and I have not read the readme of this forum, lazy, lazy

I' ve use your code (8bytes and 4bytes) in my program. Before using your code, the running time of my program is 43 days; by using your code to replace the standard memcpy(), the running time is 38 days. My complier is Borland C++ 5.5.

Would you mind give me your e-mail that I can send you the code.

**xErath** · 06-22-2004

Hum, it depends on what I can do... :P
5days?? that in fact is an improval

Still.. 38 days???
You'll have to merge function to prevent stack push/pops and function calls, plus you should optimize the function for each diferent call context. What about assembly??
I don't know if Borland suports this, but microsft compilers do, plus this is Microsofts assembly sintax

Code:

void *memcpy(void *_dest, const void *_src, int _size){
__asm{
	push eax
	push ecx
	push edi
	push esi
	mov edi, _dest	;edi destiny, esi source
	mov esi, _src
	mov ecx, _size

	cmp ecx, 4	;if size<4
	jl _lbl1
_cicle1:		;copy 4bytes by 4 bytes
		mov eax, dword ptr [esi]
		mov dword ptr [edi], eax
		add edi, 4
		add esi, 4
		sub ecx, 4
		cmp ecx, 4
		jge _cicle1
_lbl1:
	cmp ecx, 0	;if size<=0
	jle _lbl2_end
_cicle2:		;copy each byte at a time
		mov al, byte ptr [esi]
		mov byte ptr [edi], al
		inc edi
		inc esi
		loop _cicle2
_lbl2_end:
	pop esi
	pop edi
	pop ecx
	pop eax
	}
	return _dest;
}

This is extremetly optimized compared to C... Now you'll have to embed this with you source for Borland compilers.

**naruto** · 06-22-2004

As I know the fastest running speed can be 20 days( The code must be different but the aim is same). All code of this program is not written by me. I have no experience on code optimzation although I'm doing this job. The motivation is that I've read Michael Abrash's book(game programming black book) and I remember that he mentioned in his book about the optimzation of memcpy() but I can't find the article of this topic in that book.

You say you found your code much simple, what code?

**Salem** · 06-22-2004

Mmm, brute force attack on a crypto system.
Doesn't matter what you do, you'll never turn those days into hours unless you have a room full of PC's all working on the same problem.

**xErath** · 06-22-2004

Using this main, it took only 9 seconds to cpy 50*10^8 bytes of data

Code:

main(){
	char s[51],d[51];
	int i, t;

	t=time(0);
	for(i=0;i<100000000;i++)
		memcpy(d, s, 50);
	
	printf("TIME: %d_\n", time(0)-t);
}

Do you think that it is something worth?? Things can't get ant better than assembly.

**xErath** · 06-22-2004

Originally Posted by naruto

You say you found your code much simple, what code?

The copy stuff with 8 byte integers.

**naruto** · 06-22-2004

Originally Posted by Salem

Mmm, brute force attack on a crypto system.
Doesn't matter what you do, you'll never turn those days into hours unless you have a room full of PC's all working on the same problem.

Bingo! But it is a legal "attack"

It's a distributed program.

**naruto** · 06-22-2004

Originally Posted by xErath

Using this main, it took only 9 seconds to cpy 50*10^8 bytes of data

Code:

main(){
	char s[51],d[51];
	int i, t;

	t=time(0);
	for(i=0;i<100000000;i++)
		memcpy(d, s, 50);
	
	printf("TIME: %d_\n", time(0)-t);
}

Do you think that it is something worth?? Things can't get ant better than assembly.

Some guys has modified it by using assembly to replace some parts of the code (nobody re-write memcpy), but the assembly edition is not effective than what I have done. I'm not understand the zen of code optimization, but the only thing I've learn from michael abrash's book is "do not to optimize a program which is not good design. If you use assembly in a bad design program, it may be not faster than C. The right situation to use assembly to optimize is that you have really understood the effect of each instruction which you have written.".

I only can remember the thinking he has said, but I can't do according to his thinking

**xErath** · 06-22-2004

Basicly, what he's saying is "If you use assembly in a bad design program", if you write lame assemlby you get crapy code... Duh... That's for every stuff, even worse if you have a lame compiler. Believe me ... the assembly code I posted is quite small.. probably you could cut 2 or 3 instructions... It does the same stuf as the memcpy version with 4 bit integers copy. If you have any doubts, try both.
Then again if something is pourly designed, re-design it well.

//edit
I did that main stuff with both the C and assembly memcpy, this time to copy 50*10^9 bytes of data (50GB): the C function did it in 134 second, the assembly in 81 seconds. Nice!

Thread: Help: About memcpy()

Thread Tools

Search Thread

Display

Help: About memcpy()

memcpy

Similar Threads

Disagreement about memcpy

Okay, giant issue, I think I'm treading into territory outside of cboards maybe...

Memcpy(); Errors...

memcpy with 128 bit registers

memcpy