Skip to contentSkip to conversation

Linus' sha1 is much faster!

PATCH Pádraig Brady 10 participants 21 messages Jun 15, 2016 – Apr 20, 2017 Library code

lore
Share & tools
lore
#0 lore raw

I've noticed before that coreutils hashing utils were a little behind in performance, but was prompted to look at it again when I noticed the recently updated sha1 implementation in git:
http://git.kernel.org/?p=git/git.git;a=history;f=block-sha1;h=d3121f7;hb=pu

Testing that with the attached program which I wrote in a couple of mins to try and match sha1sum's system calls shows that it's around 33% faster, as shown below:

$ gcc $(rpm -q --qf="%{OPTFLAGS}\n" coreutils) linus-sha1.c sha1.c -o linus-sha1
$ time ./linus-sha1 300MB_file
df1e19e245fee4f53087b50ef953ca2c8d1644d7  300MB_file
real    0m2.742s
user    0m2.516s
sys     0m0.206s
$ time ~/git/coreutils/src/sha1sum 300MB_file
df1e19e245fee4f53087b50ef953ca2c8d1644d7  300MB_file
real    0m4.166s
user    0m3.846s
sys     0m0.298s

So, could we use that code in coreutils?
Think of all the dead fish it would save.

I've also attached a trivial block-sha1 patch which doesn't affect performance, but does suppress a signed unsigned comparison warning which occurs with -Wextra for example.

cheers,
Pádraig.

/* gcc -O2 -Wall linus-sha1.c sha1.c -o linus-sha1 */ #include <stdio.h>
#include <stdlib.h>
#include "sha1.h"

int main(int argc, char** argv)

{
    if (argc != 2) return 1;
    const char* filename = argv[1];
    FILE *fp = fopen (filename, "r");
    if (!fp) return 1;

    #define BS 4096 /* match coreutils */

    blk_SHA_CTX ctx;
    blk_SHA1_Init(&ctx);
    size_t nr;
    char buf[BS];
    while ((nr=fread_unlocked(buf, 1, sizeof(buf), fp)))
        blk_SHA1_Update(&ctx, buf, nr);
    unsigned char hash[20];
    blk_SHA1_Final(hash, &ctx);
    int i;
    for (i=0; i<sizeof(hash); i++)
        printf("%02x",*(hash+i));
    printf("  %s\n", filename);

    return 0;
}
Quoted 1 line

From fa75e818836f763357ff9b7bbde3327e1aabbe47 Mon Sep 17 00:00:00 2001

From: =?utf-8?q?P=C3=A1draig=20Brady?= <[email protected]>
Date: Sat, 15 Aug 2009 00:17:30 +0100
Subject: [PATCH] block-sha1: suppress signed unsigned comparison warning

* block-sha1/sha1.c: Use unsigned ints as the values will never go negative.

---
 block-sha1/sha1.c |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
block-sha1/sha1.c +2 -2
index d3121f7..be763d8 100644
@@ -231,13 +231,13 @@ void blk_SHA1_Init(blk_SHA_CTX *ctx)
 
 void blk_SHA1_Update(blk_SHA_CTX *ctx, const void *data, unsigned long len)
 {
-	int lenW = ctx->size & 63;
+	unsigned int lenW = ctx->size & 63;
 
 	ctx->size += len;
 
 	/* Read the data into W and process blocks as they get full */
 	if (lenW) {
-		int left = 64 - lenW;
+		unsigned int left = 64 - lenW;
 		if (len < left)
 			left = len;
 		memcpy(lenW + (char *)ctx->W, data, left);
Signature · 1 line
1.6.2.5
#1 lore raw

2009/8/14 Pádraig Brady <[email protected]>:

Quoted 21 lines

I've noticed before that coreutils hashing utils were a little behind in performance, but was prompted to look at it again when I noticed the recently updated sha1 implementation in git:
http://git.kernel.org/?p=git/git.git;a=history;f=block-sha1;h=d3121f7;hb=pu

Testing that with the attached program which I wrote in a couple of mins to try and match sha1sum's system calls shows that it's around 33% faster, as shown below:

$ gcc $(rpm -q --qf="%{OPTFLAGS}\n" coreutils) linus-sha1.c sha1.c -o linus-sha1
$ time ./linus-sha1 300MB_file
df1e19e245fee4f53087b50ef953ca2c8d1644d7  300MB_file
real    0m2.742s
user    0m2.516s
sys     0m0.206s
$ time ~/git/coreutils/src/sha1sum 300MB_file
df1e19e245fee4f53087b50ef953ca2c8d1644d7  300MB_file
Quoted 6 lines

real    0m4.166s
user    0m3.846s
sys     0m0.298s

So, could we use that code in coreutils? Think of all the dead fish it would save.

coreutils is licensed under GPLv3, and git under GPLv2 (only), so you'd need permission from all contributors to the implementation in order to relicense under GPLv3. A quick grep of the history suggests these contributors to be:

Brandon Casey <[email protected]>
Junio C Hamano <[email protected]>
Linus Torvalds <[email protected]>
Nicolas Pitre <[email protected]>
(adding these people to the CC list)

Additionally, it was originally based on the code in mozilla-sha1/sha1.c, but that contains a license grant allowing it to be used under GPLv2 /or later/, so if GPLv3 relicensing is enough it shouldn't be necessary to get in contact with the original author. However if the FSF requires copyright assignment to accept the new implementation, it will be necessary to track down contributors to the original mozilla-sha1/sha1.c as well.

Note that I'm not a lawyer, so there might be other roadblocks etc to this as well, etc :)

#2 lore raw

2009/8/15 Bryan Donlan <[email protected]>:

Quoted 4 lines

coreutils is licensed under GPLv3, and git under GPLv2 (only), so you'd need permission from all contributors to the implementation in order to relicense under GPLv3. A quick grep of the history suggests these contributors to be:

X11 also requires a fast SHA1 implementation. It uses this to check if two pixmaps are the same. So it would be really nice to relicense under a liberal enough license that xorg can use it.

John

#3 lore raw
Quoted 9 lines · John Tapsell

2009/8/15 Bryan Donlan <[email protected]>:

Quoted 4 lines

coreutils is licensed under GPLv3, and git under GPLv2 (only), so you'd need permission from all contributors to the implementation in order to relicense under GPLv3. A quick grep of the history suggests these contributors to be:

X11 also requires a fast SHA1 implementation. It uses this to check if two pixmaps are the same. So it would be really nice to relicense under a liberal enough license that xorg can use it.

If the checksum isn't being exposed in the protocol (i.e., it's just internal to the X server), one possibility for X11 is to consider to use the SHA-3 candidate Skein instead. After receiving a large amount of evaluation by cryptographic experts, it was one of the 18 algorithms (our of an original 64 entries) that have made it the 2nd round of the NIST competition. It's also *substantially* faster than SHA:

One exception to this is Skein, created by several well-known cryptographers and noted pundit Bruce Schneier. It was designed specifically to exploit all three of the Core 2 execution units and to run at a full 64-bits. This gives it roughly four to 10 times the logic density of competing submissions.

This is what I meant by the Matrix quote above. They didn't bend the spoon; they bent the crypto algorithm. They moved the logic operations around in a way that wouldn't weaken the crypto, but would strengthen its speed on the Intel Core 2.

In their paper (PDF), the authors of Skein express surprise that a custom silicon ASIC implementation is not any faster than the software implementation. They shouldn't be surprised. Every time you can redefine a problem to run optimally in software, you will reach the same speeds you get with optimized ASIC hardware. The reason software has a reputation of being slow is because people don't redefine the original problem.

http://www.darkreading.com/blog/archives/2008/11/bending_skein_c.html

For more information and some optimized implementation, see:

	http://www.skein-hash.info/

							- Ted
#4 lore raw
In reply to John Tapsell
Quoted 9 lines · John Tapsell

2009/8/15 Bryan Donlan <[email protected]>:

Quoted 4 lines

coreutils is licensed under GPLv3, and git under GPLv2 (only), so you'd need permission from all contributors to the implementation in order to relicense under GPLv3. A quick grep of the history suggests these contributors to be:

X11 also requires a fast SHA1 implementation. It uses this to check if two pixmaps are the same. So it would be really nice to relicense under a liberal enough license that xorg can use it.

I'm personally ok with retaining the mozilla-sha1 license.

There's not really anything _remaining_ of the mozilla code, but hey, I started from it. In retrospect I probably should have started from the PPC asm code that already did the blocking sanely - but that's a "20/20 hindsight" kind of thing.

Plus hey, the mozilla code being a horrid pile of crud was why I was so convinced that I could improve on things. So that's a kind of source for it, even if it's more about the motivational side than any actual remaining code ;)

That said, I don't know if the MPL is ok for X11. I've not looked at compatibility issues with MPL. For git, we could just ignore the MPL, since the GPLv2 was acceptable regardless of it.

Linus

#5 lore raw
Quoted 4 lines · Linus Torvalds

That said, I don't know if the MPL is ok for X11. I've not looked at compatibility issues with MPL. For git, we could just ignore the MPL, since the GPLv2 was acceptable regardless of it.

If MPL isn't ok for X11, then we'd need to make sure that even the silliest Mozilla crud has been rewritten. There really isn't much, but hey, the _history_ is based on the mozilla code, and who knows - the 'blk_SHA_CTX' struct has things like the fields in the same order as the Mozilla equivalent, for all those historical reasons.

(Heh. Looking at that, I probably should move the 'size' field first, since that would have different alignment rules, and the struct would be more tightly packed that way, and initialize better).

Afaik, none of the actual code remains (the mozilla SHA1 thing did the wrong thing for performance even for just the final bytes, and did those a byte at a time etc, so I rewrote even the trivial SHA1_Final parts).

Of course, maybe the Mozilla people would be interested in taking my faster version, and say that the new-BSD license is ok, and make everybody happy. The only listed author for the Mozilla SHA1 is Paul Kocher. I added him to the Cc.

Paul, for your information, we're talking about a faster rewritten "mostly portable" SHA1 routines that you can find at

	http://git.kernel.org/?p=git/git.git;a=tree;f=block-sha1;hb=pu

(follow the "blob" pointers to see sha1.c and sha1.h). I don't know if you're active with Mozilla/Firefox or whether you even care, but you seem to be the logical choice of person to ask.

Linus

#6 lore raw
Quoted 10 lines · Linus Torvalds
Quoted 1 line · Linus Torvalds

That said, I don't know if the MPL is ok for X11. I've not looked at compatibility issues with MPL. For git, we could just ignore the MPL, since the GPLv2 was acceptable regardless of it.

If MPL isn't ok for X11, then we'd need to make sure that even the silliest Mozilla crud has been rewritten. There really isn't much, but hey, the _history_ is based on the mozilla code, and who knows - the 'blk_SHA_CTX' struct has things like the fields in the same order as the Mozilla equivalent, for all those historical reasons.

(Heh. Looking at that, I probably should move the 'size' field first, since that would have different alignment rules, and the struct would be more tightly packed that way, and initialize better).

Afaik, none of the actual code remains (the mozilla SHA1 thing did the wrong thing for performance even for just the final bytes, and did those a byte at a time etc, so I rewrote even the trivial SHA1_Final parts).

Quoted 8 lines

Of course, maybe the Mozilla people would be interested in taking my faster version, and say that the new-BSD license is ok, and make everybody happy. The only listed author for the Mozilla SHA1 is Paul Kocher. I added him to the Cc.

Paul, for your information, we're talking about a faster rewritten "mostly portable" SHA1 routines that you can find at

	http://git.kernel.org/?p=git/git.git;a=tree;f=block-sha1;hb=pu

(follow the "blob" pointers to see sha1.c and sha1.h). I don't know if you're active with Mozilla/Firefox or whether you even care, but you seem to be the logical choice of person to ask.

I contacted Paul in february this year to get permission to use the mozilla sha1 code for libgit2. His reply then was:
"I'm not sure which version the diffs are relative to, so I haven't reviewed them. It's fine to distribute under BSD, GPL, or LGPL, however."

I also got explicit permission to relicense it under GPLv2 with the gcc exception.

I added the mail-address I used to contact him to CC as well. Sorry if you get this twice, Paul.

Naturally, I'd like to use the faster version for libgit2 as well. The people who Linus listed as contributors earlier (Brandon Casy, Linus, Junio and Nicolas Pitre) have already consented to relicense their git contributions for libgit2 use. If anyone would like to revoke that consent for this code, speak now please, or I'll patch it into libgit2 as well.

Signature · 7 lines
Andreas Ericsson                   [email protected]
OP5 AB                             www.op5.se
Tel: +46 8-230225                  Fax: +46 8-230231

Considering the successes of the wars on alcohol, poverty, drugs and
terror, I think we should give some serious thought to declaring war
on peace.
#7 lore raw
In reply to Linus Torvalds
Quoted 3 lines · Linus Torvalds

(Heh. Looking at that, I probably should move the 'size' field first, since that would have different alignment rules, and the struct would be more tightly packed that way, and initialize better).

I was about to suggest (i.e. post) a patch for that. This is indeed a good idea.

Quoted 3 lines

Afaik, none of the actual code remains (the mozilla SHA1 thing did the wrong thing for performance even for just the final bytes, and did those a byte at a time etc, so I rewrote even the trivial SHA1_Final parts).

Maybe a patch adding a proper header with the actual license would be a good idea too.

Nicolas

#11 lore raw
In reply to Pádraig Brady

Hi Pádraig,

I tried to reproduce your results but I wasn't able to do it. The biggest difference on a 300MB file I noticed was approximately 15% using on both implementations -O2, and 5% using -O3.
My GCC version is "gcc (Debian 4.3.3-14) 4.3.3" and the CPU is: Intel(R) Pentium(R) D CPU 3.20GHz.

I also spent some time trying to improve the gnulib SHA1 implementation and it seems a lookup table can improve things a bit.

Can you please try the patch that I have attached and tell me which performance difference (if any) you get?

Thanks,
Giuseppe

Quoted 1 line

From b975a5e0849eaa46e5cf410c5bf6e2308f044d61 Mon Sep 17 00:00:00 2001

From: Giuseppe Scrivano <[email protected]>
Date: Sun, 16 Aug 2009 20:53:54 +0200
Subject: [PATCH] SHA1: use a lookup table for faster hashing

* lib/sha1.c (struct sha1_pre): New member.
* lib/sha1.c (sha1_process_block): Use the lookup table to quickly find indices to use in the current round.

---
 lib/sha1.c |  160 ++++++++++++++++++++++++++++++++++-------------------------
 1 files changed, 92 insertions(+), 68 deletions(-)
lib/sha1.c +92 -68
index 9c6c7ae..ec18ba7 100644
@@ -283,6 +283,32 @@ sha1_process_bytes (const void *buffer, size_t len, struct sha1_ctx *ctx)
 #define F3(B,C,D) ( ( B & C ) | ( D & ( B | C ) ) )
 #define F4(B,C,D) (B ^ C ^ D)
 
+struct lookup_t
+{
+  unsigned char l1 : 4;
+  unsigned char l2 : 4;
+  unsigned char l3 : 4;
+  unsigned char l4 : 4;
+};
+
+const static struct lookup_t
+sha1_pre[16] = {{(0 - 3) & 0x0f, (0 - 8) & 0x0f, (0 - 14) & 0x0f},
+                {(1 - 3) & 0x0f, (1 - 8) & 0x0f, (1 - 14) & 0x0f},
+                {(2 - 3) & 0x0f, (2 - 8) & 0x0f, (2 - 14) & 0x0f},
+                {(3 - 3) & 0x0f, (3 - 8) & 0x0f, (3 - 14) & 0x0f},
+                {(4 - 3) & 0x0f, (4 - 8) & 0x0f, (4 - 14) & 0x0f},
+                {(5 - 3) & 0x0f, (5 - 8) & 0x0f, (5 - 14) & 0x0f},
+                {(6 - 3) & 0x0f, (6 - 8) & 0x0f, (6 - 14) & 0x0f},
+                {(7 - 3) & 0x0f, (7 - 8) & 0x0f, (7 - 14) & 0x0f},
+                {(8 - 3) & 0x0f, (8 - 8) & 0x0f, (8 - 14) & 0x0f},
+                {(9 - 3) & 0x0f, (9 - 8) & 0x0f, (9 - 14) & 0x0f},
+                {(10 - 3) & 0x0f, (10 - 8) & 0x0f, (10 - 14) & 0x0f},
+                {(11 - 3) & 0x0f, (11 - 8) & 0x0f, (11 - 14) & 0x0f},
+                {(12 - 3) & 0x0f, (12 - 8) & 0x0f, (12 - 14) & 0x0f},
+                {(13 - 3) & 0x0f, (13 - 8) & 0x0f, (13 - 14) & 0x0f},
+                {(14 - 3) & 0x0f, (14 - 8) & 0x0f, (14 - 14) & 0x0f},
+                {(15 - 3) & 0x0f, (15 - 8) & 0x0f, (15 - 14) & 0x0f}};
+
 /* Process LEN bytes of BUFFER, accumulating context into CTX.
    It is assumed that LEN % 64 == 0.
    Most of this code comes from GnuPG's cipher/sha1.c.  */
@@ -309,9 +335,8 @@ sha1_process_block (const void *buffer, size_t len, struct sha1_ctx *ctx)
 
 #define rol(x, n) (((x) << (n)) | ((uint32_t) (x) >> (32 - (n))))
 
-#define M(I) ( tm =   x[I&0x0f] ^ x[(I-14)&0x0f] \
-		    ^ x[(I-8)&0x0f] ^ x[(I-3)&0x0f] \
-	       , (x[I&0x0f] = rol(tm, 1)) )
+#define M(I) (x[I] = rol (x[sha1_pre[I].l1] ^ x[sha1_pre[I].l2] \
+                          ^ x[sha1_pre[I].l3] ^ x[I], 1))
 
 #define R(A,B,C,D,E,F,K,M)  do { E += rol( A, 5 )     \
 				      + F( B, C, D )  \
@@ -322,7 +347,6 @@ sha1_process_block (const void *buffer, size_t len, struct sha1_ctx *ctx)
 
   while (words < endp)
     {
-      uint32_t tm;
       int t;
       for (t = 0; t < 16; t++)
 	{
@@ -346,70 +370,70 @@ sha1_process_block (const void *buffer, size_t len, struct sha1_ctx *ctx)
       R( c, d, e, a, b, F1, K1, x[13] );
       R( b, c, d, e, a, F1, K1, x[14] );
       R( a, b, c, d, e, F1, K1, x[15] );
-      R( e, a, b, c, d, F1, K1, M(16) );
-      R( d, e, a, b, c, F1, K1, M(17) );
-      R( c, d, e, a, b, F1, K1, M(18) );
-      R( b, c, d, e, a, F1, K1, M(19) );
-      R( a, b, c, d, e, F2, K2, M(20) );
-      R( e, a, b, c, d, F2, K2, M(21) );
-      R( d, e, a, b, c, F2, K2, M(22) );
-      R( c, d, e, a, b, F2, K2, M(23) );
-      R( b, c, d, e, a, F2, K2, M(24) );
-      R( a, b, c, d, e, F2, K2, M(25) );
-      R( e, a, b, c, d, F2, K2, M(26) );
-      R( d, e, a, b, c, F2, K2, M(27) );
-      R( c, d, e, a, b, F2, K2, M(28) );
-      R( b, c, d, e, a, F2, K2, M(29) );
-      R( a, b, c, d, e, F2, K2, M(30) );
-      R( e, a, b, c, d, F2, K2, M(31) );
-      R( d, e, a, b, c, F2, K2, M(32) );
-      R( c, d, e, a, b, F2, K2, M(33) );
-      R( b, c, d, e, a, F2, K2, M(34) );
-      R( a, b, c, d, e, F2, K2, M(35) );
-      R( e, a, b, c, d, F2, K2, M(36) );
-      R( d, e, a, b, c, F2, K2, M(37) );
-      R( c, d, e, a, b, F2, K2, M(38) );
-      R( b, c, d, e, a, F2, K2, M(39) );
-      R( a, b, c, d, e, F3, K3, M(40) );
-      R( e, a, b, c, d, F3, K3, M(41) );
-      R( d, e, a, b, c, F3, K3, M(42) );
-      R( c, d, e, a, b, F3, K3, M(43) );
-      R( b, c, d, e, a, F3, K3, M(44) );
-      R( a, b, c, d, e, F3, K3, M(45) );
-      R( e, a, b, c, d, F3, K3, M(46) );
-      R( d, e, a, b, c, F3, K3, M(47) );
-      R( c, d, e, a, b, F3, K3, M(48) );
-      R( b, c, d, e, a, F3, K3, M(49) );
-      R( a, b, c, d, e, F3, K3, M(50) );
-      R( e, a, b, c, d, F3, K3, M(51) );
-      R( d, e, a, b, c, F3, K3, M(52) );
-      R( c, d, e, a, b, F3, K3, M(53) );
-      R( b, c, d, e, a, F3, K3, M(54) );
-      R( a, b, c, d, e, F3, K3, M(55) );
-      R( e, a, b, c, d, F3, K3, M(56) );
-      R( d, e, a, b, c, F3, K3, M(57) );
-      R( c, d, e, a, b, F3, K3, M(58) );
-      R( b, c, d, e, a, F3, K3, M(59) );
-      R( a, b, c, d, e, F4, K4, M(60) );
-      R( e, a, b, c, d, F4, K4, M(61) );
-      R( d, e, a, b, c, F4, K4, M(62) );
-      R( c, d, e, a, b, F4, K4, M(63) );
-      R( b, c, d, e, a, F4, K4, M(64) );
-      R( a, b, c, d, e, F4, K4, M(65) );
-      R( e, a, b, c, d, F4, K4, M(66) );
-      R( d, e, a, b, c, F4, K4, M(67) );
-      R( c, d, e, a, b, F4, K4, M(68) );
-      R( b, c, d, e, a, F4, K4, M(69) );
-      R( a, b, c, d, e, F4, K4, M(70) );
-      R( e, a, b, c, d, F4, K4, M(71) );
-      R( d, e, a, b, c, F4, K4, M(72) );
-      R( c, d, e, a, b, F4, K4, M(73) );
-      R( b, c, d, e, a, F4, K4, M(74) );
-      R( a, b, c, d, e, F4, K4, M(75) );
-      R( e, a, b, c, d, F4, K4, M(76) );
-      R( d, e, a, b, c, F4, K4, M(77) );
-      R( c, d, e, a, b, F4, K4, M(78) );
-      R( b, c, d, e, a, F4, K4, M(79) );
+      R( e, a, b, c, d, F1, K1, M( 0) );
+      R( d, e, a, b, c, F1, K1, M( 1) );
+      R( c, d, e, a, b, F1, K1, M( 2) );
+      R( b, c, d, e, a, F1, K1, M( 3) );
+      R( a, b, c, d, e, F2, K2, M( 4) );
+      R( e, a, b, c, d, F2, K2, M( 5) );
+      R( d, e, a, b, c, F2, K2, M( 6) );
+      R( c, d, e, a, b, F2, K2, M( 7) );
+      R( b, c, d, e, a, F2, K2, M( 8) );
+      R( a, b, c, d, e, F2, K2, M( 9) );
+      R( e, a, b, c, d, F2, K2, M(10) );
+      R( d, e, a, b, c, F2, K2, M(11) );
+      R( c, d, e, a, b, F2, K2, M(12) );
+      R( b, c, d, e, a, F2, K2, M(13) );
+      R( a, b, c, d, e, F2, K2, M(14) );
+      R( e, a, b, c, d, F2, K2, M(15) );
+      R( d, e, a, b, c, F2, K2, M( 0) );
+      R( c, d, e, a, b, F2, K2, M( 1) );
+      R( b, c, d, e, a, F2, K2, M( 2) );
+      R( a, b, c, d, e, F2, K2, M( 3) );
+      R( e, a, b, c, d, F2, K2, M( 4) );
+      R( d, e, a, b, c, F2, K2, M( 5) );
+      R( c, d, e, a, b, F2, K2, M( 6) );
+      R( b, c, d, e, a, F2, K2, M( 7) );
+      R( a, b, c, d, e, F3, K3, M( 8) );
+      R( e, a, b, c, d, F3, K3, M( 9) );
+      R( d, e, a, b, c, F3, K3, M(10) );
+      R( c, d, e, a, b, F3, K3, M(11) );
+      R( b, c, d, e, a, F3, K3, M(12) );
+      R( a, b, c, d, e, F3, K3, M(13) );
+      R( e, a, b, c, d, F3, K3, M(14) );
+      R( d, e, a, b, c, F3, K3, M(15) );
+      R( c, d, e, a, b, F3, K3, M( 0) );
+      R( b, c, d, e, a, F3, K3, M( 1) );
+      R( a, b, c, d, e, F3, K3, M( 2) );
+      R( e, a, b, c, d, F3, K3, M( 3) );
+      R( d, e, a, b, c, F3, K3, M( 4) );
+      R( c, d, e, a, b, F3, K3, M( 5) );
+      R( b, c, d, e, a, F3, K3, M( 6) );
+      R( a, b, c, d, e, F3, K3, M( 7) );
+      R( e, a, b, c, d, F3, K3, M( 8) );
+      R( d, e, a, b, c, F3, K3, M( 9) );
+      R( c, d, e, a, b, F3, K3, M(10) );
+      R( b, c, d, e, a, F3, K3, M(11) );
+      R( a, b, c, d, e, F4, K4, M(12) );
+      R( e, a, b, c, d, F4, K4, M(13) );
+      R( d, e, a, b, c, F4, K4, M(14) );
+      R( c, d, e, a, b, F4, K4, M(15) );
+      R( b, c, d, e, a, F4, K4, M( 0) );
+      R( a, b, c, d, e, F4, K4, M( 1) );
+      R( e, a, b, c, d, F4, K4, M( 2) );
+      R( d, e, a, b, c, F4, K4, M( 3) );
+      R( c, d, e, a, b, F4, K4, M( 4) );
+      R( b, c, d, e, a, F4, K4, M( 5) );
+      R( a, b, c, d, e, F4, K4, M( 6) );
+      R( e, a, b, c, d, F4, K4, M( 7) );
+      R( d, e, a, b, c, F4, K4, M( 8) );
+      R( c, d, e, a, b, F4, K4, M( 9) );
+      R( b, c, d, e, a, F4, K4, M(10) );
+      R( a, b, c, d, e, F4, K4, M(11) );
+      R( e, a, b, c, d, F4, K4, M(12) );
+      R( d, e, a, b, c, F4, K4, M(13) );
+      R( c, d, e, a, b, F4, K4, M(14) );
+      R( b, c, d, e, a, F4, K4, M(15) );
 
       a = ctx->A += a;
       b = ctx->B += b;
Signature · 1 line
1.6.3.3
#12 lore raw
Quoted 8 lines · Giuseppe Scrivano

Hi Pádraig,

I tried to reproduce your results but I wasn't able to do it. The biggest difference on a 300MB file I noticed was approximately 15% using on both implementations -O2, and 5% using -O3.
My GCC version is "gcc (Debian 4.3.3-14) 4.3.3" and the CPU is: Intel(R) Pentium(R) D CPU 3.20GHz.

Quoted 5 lines

I also spent some time trying to improve the gnulib SHA1 implementation and it seems a lookup table can improve things a bit.

Can you please try the patch that I have attached and tell me which performance difference (if any) you get?

Thanks for looking at this Giuseppe
and sorry for not mentioning my GCC and CPU.

Note the binaries below is compiled with
$(rpm -q --qf="%{OPTFLAGS}\n" coreutils)
for consistency, which on my F11 machines is:

  -O2 -g -pipe -Wall -Wp,-D_FORTIFY_SOURCE=2 -fexceptions
  -fstack-protector --param=ssp-buffer-size=4 -m32 -march=i586
  -mtune=generic -fasynchronous-unwind-tables -D_GNU_SOURCE=1

Testing on 2 machines I have here:

$ rpm -q gcc
gcc-4.4.1-2.fc11.i586
$ grep "model name" /proc/cpuinfo | head -n1 | tr -s '[:blank:]' ' '
model name : Intel(R) Pentium(R) M processor 1.70GHz
$ truncate -s300MB sha1.test
$ time sha1sum sha1.test
real    0m3.540s
$ time linus-sha1 sha1.test
real    0m2.319s (-34%)
$ time  giuseppe-sha1sum sha1.test
real    0m3.513s (-.8%)
$ rpm -q gcc
gcc-4.4.1-2.fc11.i586
$ grep "model name" /proc/cpuinfo | head -n1 | tr -s '[:blank:]' ' '
model name : Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
$ truncate -s300MB sha1.test
$ time sha1sum sha1.test
real    0m1.857s
$ time linus-sha1 sha1.test
real    0m1.102s (-40%)
$ time giuseppe-sha1sum sha1.test
real    0m1.932s (+ 4%)

cheers,
Pádraig.

#13 lore raw
Quoted 3 lines · Pádraig Brady
  -O2 -g -pipe -Wall -Wp,-D_FORTIFY_SOURCE=2 -fexceptions
  -fstack-protector --param=ssp-buffer-size=4 -m32 -march=i586
  -mtune=generic -fasynchronous-unwind-tables -D_GNU_SOURCE=1

thanks. I did again all tests on my machine using these same options. I repeated each test 6 times and I took the median without consider the first result. Except the first run that it is not considered, I didn't report a big variance on results of the same test.

gcc 4.3.3

gnulib sha1:            real	0m2.543s
gnulib sha1 lookup:     real	0m1.906s (-25%)

linus's sha1: real 0m2.468s (-3%)
linus's sha1 no asm: real 0m2.289s (-9%)

gcc 4.4.1

gnulib sha1:            real	0m3.386s
gnulib sha1 lookup:     real	0m3.110s (-8%)

linus's sha1: real 0m1.701s (-49%)
linus's sha1 no asm: real 0m1.284s (-62%)

I don't see such big differences in asm generated by gcc 4.4.1 and gcc 4.3.3 to explain this performance difference, what I noticed immediately is that in the gcc-4.4 generated asm there are more "lea" instructions (+30%), but I doubt this is the reason of these poor results. Anyway, I haven't yet looked much in details.

Cheers,
Giuseppe

#14 lore raw
Quoted 28 lines · Giuseppe Scrivano
Quoted 3 lines · Pádraig Brady

  -O2 -g -pipe -Wall -Wp,-D_FORTIFY_SOURCE=2 -fexceptions -fstack-protector --param=ssp-buffer-size=4 -m32 -march=i586 -mtune=generic -fasynchronous-unwind-tables -D_GNU_SOURCE=1

thanks.  I did again all tests on my machine using these same options. I repeated each test 6 times and I took the median without consider the first result.  Except the first run that it is not considered, I didn't report a big variance on results of the same test.

gcc 4.3.3

gnulib sha1:            real    0m2.543s
gnulib sha1 lookup:     real    0m1.906s (-25%)
linus's sha1:           real    0m2.468s (-3%)
linus's sha1 no asm:    real    0m2.289s (-9%)

gcc 4.4.1

gnulib sha1:            real    0m3.386s
gnulib sha1 lookup:     real    0m3.110s (-8%)
linus's sha1:           real    0m1.701s (-49%)
linus's sha1 no asm:    real    0m1.284s (-62%)

Quoted 8 lines

I don't see such big differences in asm generated by gcc 4.4.1 and gcc 4.3.3 to explain this performance difference, what I noticed immediately is that in the gcc-4.4 generated asm there are more "lea" instructions (+30%), but I doubt this is the reason of these poor results.  Anyway, I haven't yet looked much in details.

Cheers,
Giuseppe

Interesting. I compared Linus' implementation to the public domain one by Steve Reid[1], which is used in OpenLDAP and a few other projects. Anyone with some experience testing these kinds of things in a statistically sound manner want to try it out? In my tests, I got this:

(average of 5 runs)
Linus' sha1: 283MB/s
Steve Reid's sha1: 305MB/s

- Steven

[1] http://gpl.nas-central.org/SYNOLOGY/x07-series/514_UNTARED/source/openldap-2.3.11/libraries/liblutil/sha1.c

#15 lore raw

Hi,

These are the results I reported (median of 5 plus an additional not considered first run) on the Steve Reid's SHA1 implementation using the same flags to the compiler that I used for previous tests.

GCC 4.3.3:  real	0m2.627s
GCC 4.4.1:  real	0m3.742s

In both cases it showed to be slower than other implementations I have already tried.

Additional note: as for gnulib SHA1, GCC 4.4.1 produced slower code than GCC 4.3.3.

Cheers,
Giuseppe

Quoted 14 lines · Steven Noonan

Interesting. I compared Linus' implementation to the public domain one by Steve Reid[1], which is used in OpenLDAP and a few other projects. Anyone with some experience testing these kinds of things in a statistically sound manner want to try it out? In my tests, I got this:

(average of 5 runs)
Linus' sha1: 283MB/s
Steve Reid's sha1: 305MB/s

- Steven

[1] http://gpl.nas-central.org/SYNOLOGY/x07-series/514_UNTARED/source/openldap-2.3.11/libraries/liblutil/sha1.c

#16 lore raw
In reply to Steven Noonan
Quoted 3 lines · Steven Noonan

Interesting. I compared Linus' implementation to the public domain one by Steve Reid[1]

You _really_ need to talk about what kind of environment you have.

There are three major issues:
- Netburst vs non-netburst
- 32-bit vs 64-bit
- compiler version

Steve Reid's code looks great, but the way it is coded, gcc makes a mess of it, which is exactly what my SHA1 tries to avoid.

[ In contrast, gcc does very well on just about _any_ straightforward
unrolled SHA1 C code if the target architecture is something like PPC or ia64 that has enough registers to keep it all in registers.

I haven't really tested other compilers - a less aggressive compiler would actually do _better_ on SHA1, because the problem with gcc is that it turns the whole temporary 16-entry word array into register accesses, and tries to do register allocation on that _array_.

That is wonderful for the above-mentioned PPC and IA64, but it makes gcc create totally crazy code when there aren't enough registers, and then gcc starts spilling randomly (ie it starts spilling a-e etc). This is why the compiler and version matters so much. ]

Quoted 3 lines

(average of 5 runs)
Linus' sha1: 283MB/s
Steve Reid's sha1: 305MB/s

So I get very different results:

# TIME[s] SPEED[MB/s]
Reid 2.742 222.6
linus 1.464 417

this is Intel Nehalem, but compiled for 32-bit mode (which is the more challenging one because x86-32 only has 7 general-purpose registers), and with gcc-4.4.0.

Linus

#17 lore raw
Quoted 7 lines · Linus Torvalds
Quoted 3 lines · Steven Noonan

Interesting. I compared Linus' implementation to the public domain one by Steve Reid[1]

Quoted 6 lines

You _really_ need to talk about what kind of environment you have.

There are three major issues:
 - Netburst vs non-netburst
 - 32-bit vs 64-bit
 - compiler version

Right. I'm running a Core 2 "Merom" 2.33GHz. The code was compiled for x86_64 with GCC 4.2.1. I didn't _expect_ it to compile for x86_64, but apparently the version of GCC that ships with Xcode 3.2 defaults to compiling 64-bit code on machines that are capable of running it.

Quoted 22 lines

Steve Reid's code looks great, but the way it is coded, gcc makes a mess of it, which is exactly what my SHA1 tries to avoid.

[ In contrast, gcc does very well on just about _any_ straightforward unrolled SHA1 C code if the target architecture is something like PPC or ia64 that has enough registers to keep it all in registers.

 I haven't really tested other compilers - a less aggressive compiler would actually do _better_ on SHA1, because the problem with gcc is that it turns the whole temporary 16-entry word array into register accesses, and tries to do register allocation on that _array_.

 That is wonderful for the above-mentioned PPC and IA64, but it makes gcc create totally crazy code when there aren't enough registers, and then gcc starts spilling randomly (ie it starts spilling a-e etc). This is why the compiler and version matters so much. ]

Quoted 3 lines

(average of 5 runs)
Linus' sha1: 283MB/s
Steve Reid's sha1: 305MB/s

Quoted 5 lines

So I get very different results:

       #             TIME[s] SPEED[MB/s]
       Reid            2.742       222.6
       linus           1.464         417

Added -m32:

Steve Reid: 156MB/s
Linus: 209MB/s

So on x86, your code really kicks butt.

Quoted 6 lines

this is Intel Nehalem, but compiled for 32-bit mode (which is the more challenging one because x86-32 only has 7 general-purpose registers), and with gcc-4.4.0.

                       Linus

#18 lore raw
In reply to Giuseppe Scrivano
Quoted 3 lines · Giuseppe Scrivano

My GCC version is "gcc (Debian 4.3.3-14) 4.3.3" and the CPU is: Intel(R) Pentium(R) D CPU 3.20GHz.

Netburst is very sensitive to random spill effects, and you can basically tune things by just code shuffling that just has random effects on the generated asm code.

Quoted 2 lines

I also spent some time trying to improve the gnulib SHA1 implementation and it seems a lookup table can improve things a bit.

I pretty much can guarantee you that it improves things only because it makes gcc generate crap code, which then hides some of the P4 issues.

I'd also suggest you try gcc-4.4, since that apparently fixes some of the oddest spill issues.

Linus

#19 lore raw
Quoted 5 lines · Linus Torvalds

I pretty much can guarantee you that it improves things only because it makes gcc generate crap code, which then hides some of the P4 issues.

I'd also suggest you try gcc-4.4, since that apparently fixes some of the oddest spill issues.

Thanks for the hint. I tried gcc-4.4 and it produces slower code than 4.3 on the gnulib SHA1 implementation and my patch makes it even more!

I noticed that on my machine your implementation is ~30-40% faster using SHA_ROT for rol/ror instructions than inline assembly, at least with the test-case Pádraig wrote. Am I the only one reporting it?

Cheers,
Giuseppe

#20 lore raw
Quoted 3 lines · Giuseppe Scrivano

Thanks for the hint. I tried gcc-4.4 and it produces slower code than 4.3 on the gnulib SHA1 implementation and my patch makes it even more!

Check out the asm, see if you can see why. One of the most common problems with P4's is literally that you end up loading from the same stack slot that you just stored to (gcc can do some really crazy spills), and that causes a store buffer hazard replay.

My personal opinion is that Netburst is useless for trying to optimize C code for. It's just too random.

Quoted 3 lines

I noticed that on my machine your implementation is ~30-40% faster using SHA_ROT for rol/ror instructions than inline assembly, at least with the test-case Pádraig wrote. Am I the only one reporting it?

I bet it's the same thing. Small perturbations of the source causing small changes to register allocation and thus spilling, and then Netburst goes crazy one way or another. It's interestign trying to fix it, and very frustrating.

My workstation is a Nehalem (but Core 2 will have pretty much the same behavior), and it doesn't have the crazy netburst behavior. Shorter and simpler code generally performs better (which is _not_ true on Netburst).

On my machine, for example, forcing gcc to do those rotates on registers is the difference between ~381MB/s and 415MB/s. And that's mainly because it makes gcc keep A-E in registers, rather than trying to cache the array[] references.

Linus

#8 lore raw
In reply to Nicolas Pitre
Quoted 9 lines · Nicolas Pitre
Quoted 3 lines · Linus Torvalds

(Heh. Looking at that, I probably should move the 'size' field first, since that would have different alignment rules, and the struct would be more tightly packed that way, and initialize better).

I was about to suggest (i.e. post) a patch for that. This is indeed a good idea.

Quoted 6 lines
Quoted 3 lines

Afaik, none of the actual code remains (the mozilla SHA1 thing did the wrong thing for performance even for just the final bytes, and did those a byte at a time etc, so I rewrote even the trivial SHA1_Final parts).

Maybe a patch adding a proper header with the actual license would be a good idea too.

So have you decided on a final licence,
and if so update the headers accordingly?

cheers!
Pádraig.

#9 lore raw

I also wanted to include Linus' sha1 in our software at work. But the GPLv2 license was incompatible.
Too bad it is just just in the public domain.
I grabbed Steve Reid's public domain code from 1999
and ran it. It produced the same output.
I ran it on a 3GB input file, and Linus' code from 2009 takes 37 to 40 seconds.
(Just reading the file in the same 4k buffers only takes 3 seconds so disk reading does not dominate the time.)
When I ran Steve's old version on the same input it was taking just 36 or 37 seconds.
So it is slightly faster.
Have compilers improved?
I am using gcc 4.4.7-17.

--
View this message in context: http://git.661346.n2.nabble.com/Linus-sha1-is-much-faster-tp3448007p7657473.html
Sent from the git mailing list archive at Nabble.com.