[FFmpeg-devel] [PATCH 2/5] fate: avoid framemd5, use framecrc its faster

Michael Niedermayer michaelni at gmx.at
Fri May 17 19:43:04 CEST 2013


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

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Why not whip the teacher when the pupil misbehaves? -- Diogenes of Sinope
-------------- 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/bc499bb1/attachment.asc>


More information about the ffmpeg-devel mailing list