[FFmpeg-devel] [PATCH 2/5] fate: avoid framemd5, use framecrc its faster
Michael Niedermayer
michaelni at gmx.at
Fri May 17 20:40:12 CEST 2013
On Fri, May 17, 2013 at 07:58:54PM +0200, Giorgio Vazzana wrote:
> 2013/5/17 Michael Niedermayer <michaelni at gmx.at>:
> > On Fri, May 17, 2013 at 03:15:59PM +0200, Giorgio Vazzana wrote:
> >> Hi,
> >>
> >> about a year ago I studied ffmpeg's implementation of md5, and I
> >> noticed it was a slow compared to GNU md5sum. So I managed to make it
> >> almost as fast as the GNU implementation, but I never sent a patch
> >> because I wanted to work on it some more... of course I never did that
> >> because of lack of free time. Anyway, I found my old patch, since it's
> >> relatively simple I thought I'd send it in, hoping you will find it
> >> useful, maybe as a temporary solution until someone else comes up with
> >> something faster.
> >>
> >> Quick benchmark:
> >> # create test file and compile test program
> >> $ dd if=/dev/zero of=data bs=1M count=500
> >> $ gcc -Wall avutil_md5_test.c -O2 -o avutil_md5_test -L ./libavutil/ -lavutil
> >>
> >>
> >> # without the patch:
> >> $ for i in $(seq 1 3); do time ./avutil_md5_test < data; done
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m2.906s
> >> user 0m2.680s
> >> sys 0m0.180s
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m2.936s
> >> user 0m2.560s
> >> sys 0m0.330s
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m2.903s
> >> user 0m2.690s
> >> sys 0m0.170s
> >>
> >>
> >> # with the patch:
> >> $ for i in $(seq 1 3); do time ./avutil_md5_test < data; done
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m1.896s
> >> user 0m1.580s
> >> sys 0m0.280s
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m1.895s
> >> user 0m1.650s
> >> sys 0m0.210s
> >> d8b61b2c0025919d5321461045c8226f
> >>
> >> real 0m1.891s
> >> user 0m1.650s
> >> sys 0m0.200s
> >>
> >>
> >> # for comparison:
> >> $ for i in $(seq 1 3); do time md5sum < data; done
> >> d8b61b2c0025919d5321461045c8226f -
> >>
> >> real 0m1.842s
> >> user 0m1.510s
> >> sys 0m0.280s
> >> d8b61b2c0025919d5321461045c8226f -
> >>
> >> real 0m1.803s
> >> user 0m1.530s
> >> sys 0m0.250s
> >> d8b61b2c0025919d5321461045c8226f -
> >>
> >> real 0m1.821s
> >> user 0m1.620s
> >> sys 0m0.160s
> >>
> >> Giorgio Vazzana
> >
> >> /* compile: gcc -Wall avutil_md5_test.c -O2 -o avutil_md5_test -L ./libavutil/ -lavutil */
> >> #include <stdio.h>
> >> #include <stdlib.h>
> >> #include <string.h>
> >> #include "libavutil/md5.h"
> >>
> >> #define BUFSIZE (64*1024)
> >>
> >> static void print_md5(uint8_t *md5)
> >> {
> >> int i;
> >> for (i = 0; i < 16; i++)
> >> printf("%02x", md5[i]);
> >> printf("\n");
> >> }
> >>
> >> int main()
> >> {
> >> size_t nread;
> >> struct AVMD5 *ctx;
> >> uint8_t buf[BUFSIZE];
> >> uint8_t md5val[16];
> >>
> >> ctx = av_md5_alloc();
> >>
> >> av_md5_init(ctx);
> >> while ((nread = fread(buf, 1, BUFSIZE, stdin)) > 0)
> >> av_md5_update(ctx, buf, nread);
> >> av_md5_final(ctx, md5val);
> >> free(ctx);
> >>
> >> print_md5(md5val);
> >>
> >> return 0;
> >> }
> >
> >> md5.c | 41 ++++++++++++++++++++++++++++++++++++++---
> >> 1 file changed, 38 insertions(+), 3 deletions(-)
> >> 69e18ab824e95e1279af1fa6004d68d24358c2f9 0001-md5-speed-optimizations.patch
> >> From f7fbb6cfd6d4e70f66c21ec732902d5431e45bbc Mon Sep 17 00:00:00 2001
> >> From: Giorgio Vazzana <mywing81 at gmail.com>
> >> Date: Fri, 17 May 2013 14:51:45 +0200
> >> Subject: [PATCH] md5: speed optimizations
> >>
> >> ---
> >> libavutil/md5.c | 41 ++++++++++++++++++++++++++++++++++++++---
> >> 1 files changed, 38 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/libavutil/md5.c b/libavutil/md5.c
> >> index f8f08f1..84c01f9 100644
> >> --- a/libavutil/md5.c
> >> +++ b/libavutil/md5.c
> >> @@ -49,12 +49,28 @@ struct AVMD5 *av_md5_alloc(void)
> >> return av_mallocz(sizeof(struct AVMD5));
> >> }
> >>
> >> +#if CONFIG_SMALL
> >> static const uint8_t S[4][4] = {
> >> { 7, 12, 17, 22 }, /* round 1 */
> >> { 5, 9, 14, 20 }, /* round 2 */
> >> { 4, 11, 16, 23 }, /* round 3 */
> >> { 6, 10, 15, 21 } /* round 4 */
> >> };
> >> +#else
> >> +static const uint32_t G[64] = { /* i = 0..63 */
> >> + 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, /* i */
> >> + 1, 6, 11, 0, 5, 10, 15, 4, 9, 14, 3, 8, 13, 2, 7, 12, /* (5*i + 1) % 16 */
> >> + 5, 8, 11, 14, 1, 4, 7, 10, 13, 0, 3, 6, 9, 12, 15, 2, /* (3*i + 5) % 16 */
> >> + 0, 7, 14, 5, 12, 3, 10, 1, 8, 15, 6, 13, 4, 11, 2, 9 /* (7*i) % 16 */
> >> +};
> >> +
> >> +static const uint32_t S[64] = {
> >> + 7, 12, 17, 22, 7, 12, 17, 22, 7, 12, 17, 22, 7, 12, 17, 22, /* round 1 */
> >> + 5, 9, 14, 20, 5, 9, 14, 20, 5, 9, 14, 20, 5, 9, 14, 20, /* round 2 */
> >> + 4, 11, 16, 23, 4, 11, 16, 23, 4, 11, 16, 23, 4, 11, 16, 23, /* round 3 */
> >> + 6, 10, 15, 21, 6, 10, 15, 21, 6, 10, 15, 21, 6, 10, 15, 21 /* round 4 */
> >> +};
> >> +#endif
> >>
> >> static const uint32_t T[64] = { // T[i]= fabs(sin(i+1)<<32)
> >> 0xd76aa478, 0xe8c7b756, 0x242070db, 0xc1bdceee, /* round 1 */
> >> @@ -78,6 +94,7 @@ static const uint32_t T[64] = { // T[i]= fabs(sin(i+1)<<32)
> >> 0xf7537e82, 0xbd3af235, 0x2ad7d2bb, 0xeb86d391,
> >> };
> >>
> >> +#if CONFIG_SMALL
> >> #define CORE(i, a, b, c, d) do { \
> >> t = S[i >> 4][i & 3]; \
> >> a += T[i]; \
> >> @@ -91,6 +108,21 @@ static const uint32_t T[64] = { // T[i]= fabs(sin(i+1)<<32)
> >> } \
> >> a = b + (a << t | a >> (32 - t)); \
> >> } while (0)
> >> +#else
> >> +#define CORE(i, a, b, c, d) do { \
> >> + t = S[i]; \
> >> + a += T[i]; \
> >> + \
> >> + if (i < 32) { \
> >> + if (i < 16) a += (d ^ (b & (c ^ d))) + X[G[i]]; \
> >> + else a += (c ^ (d & (c ^ b))) + X[G[i]]; \
> >> + } else { \
> >> + if (i < 48) a += (b ^ c ^ d) + X[G[i]]; \
> >> + else a += (c ^ (b | ~d)) + X[G[i]]; \
> >> + } \
> >> + a = b + (a << t | a >> (32 - t)); \
> >> + } while (0)
> >> +#endif
> >
> > i wonder why these changes make any difference at all
> > gcc should be optimizing both to the same resulting code
>
> Actually, the memcpy() trick seems to be the only responsible for the
> speed boost. If I apply this:
>
> diff --git a/libavutil/md5.c b/libavutil/md5.c
> index f8f08f1..a212150 100644
> --- a/libavutil/md5.c
> +++ b/libavutil/md5.c
> @@ -146,8 +146,12 @@ void av_md5_update(AVMD5 *ctx, const uint8_t
> *src, const int len)
> j = ctx->len & 63;
> ctx->len += len;
>
> - for (i = 0; i < len; i++) {
> - ctx->block[j++] = src[i];
> + i = 0;
> + while (i < len) {
> + int l = (len-i <= 64-j) ? len-i : 64-j;
> + memcpy(ctx->block + j, src + i, l);
> + i += l;
> + j += l;
> if (j == 64) {
> body(ctx->ABCD, (uint32_t *) ctx->block);
> j = 0;
>
> I get 1.9s instead of 2.9s with the same test file as above:
very nice
great work!
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Avoid a single point of failure, be that a person or equipment.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20130517/181cc189/attachment.asc>
More information about the ffmpeg-devel
mailing list