[FFmpeg-devel] [PATCH v3] SH4 mpegaudio decoder optimizations
Michael Niedermayer
michaelni
Sun Dec 26 15:01:13 CET 2010
On Fri, Dec 17, 2010 at 05:54:56PM +0100, Guennadi Liakhovetski wrote:
> Hi all
>
> This is version 3 of the following patch:
>
> This patch implements several mpegaudio optimizations for SH4 FPU-enabled
> SoCs. Verified to provide more than 40% acceleration, when decoding a
> 128kbps stereo mp3 audio file to /dev/null.
>
> Changes listed in the patch header, thanks to all who commented!
>
> Thanks
> Guennadi
> ---
> Guennadi Liakhovetski, Ph.D.
> Freelance Open-Source Software Developer
> http://www.open-technology.de/
> mathops.h | 2 +
> mpegaudiodec.c | 26 ++++++++++++---
> sh4/mathops.h | 97 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> 3 files changed, 120 insertions(+), 5 deletions(-)
> ad802c254a9f25cf487eba7d3cb584112c6c59b9 ff_mpegaudio_sh4-v3.patch
> From: Guennadi Liakhovetski <g.liakhovetski at gmx.de>
> Subject: [PATCH v3] SH4 mpegaudio decoder optimizations
>
> This patch implements several mpegaudio optimizations for SH4 FPU-enabled SoCs.
> Verified to provide more than 40% acceleration, when decoding a 128kbps stereo
> mp3 audio file to /dev/null.
> ---
>
> v3:
> 1. added a "&" constraint to input-output assembly parameters
>
> v2:
> 1. fixed copyright year to 2010
> 2. consistent lower-case register names
> 3. joined inline assembly macros into one
>
> diff --git a/libavcodec/mathops.h b/libavcodec/mathops.h
> index 4d88ed1..b234ae4 100644
> --- a/libavcodec/mathops.h
> +++ b/libavcodec/mathops.h
> @@ -34,6 +34,8 @@
> # include "mips/mathops.h"
> #elif ARCH_PPC
> # include "ppc/mathops.h"
> +#elif ARCH_SH4
> +# include "sh4/mathops.h"
> #elif ARCH_X86
> # include "x86/mathops.h"
> #endif
> diff --git a/libavcodec/mpegaudiodec.c b/libavcodec/mpegaudiodec.c
> index 769be89..513fa1b 100644
> --- a/libavcodec/mpegaudiodec.c
> +++ b/libavcodec/mpegaudiodec.c
> @@ -584,6 +584,22 @@ static inline int round_sample(int64_t *sum)
> op(sum, (w)[7 * 64], (p)[7 * 64]); \
> }
>
> +#ifndef SUM8_MACS
> +#define SUM8_MACS(sum, w, p) SUM8(MACS, sum, w, p)
> +#endif
> +
> +#ifndef SUM8_MLSS
> +#define SUM8_MLSS(sum, w, p) SUM8(MLSS, sum, w, p)
> +#endif
> +
> +#ifndef SUM8P2_MACS_MLSS
> +#define SUM8P2_MACS_MLSS(sum, sum2, w, w2, p) SUM8P2(sum, MACS, sum2, MLSS, w, w2, p)
> +#endif
> +
> +#ifndef SUM8P2_MLSS_MLSS
> +#define SUM8P2_MLSS_MLSS(sum, sum2, w, w2, p) SUM8P2(sum, MLSS, sum2, MLSS, w, w2, p)
> +#endif
> +
> #define SUM8P2(sum1, op1, sum2, op2, w1, w2, p) \
> { \
> INTFLOAT tmp;\
> @@ -666,9 +682,9 @@ static void apply_window_mp3_c(MPA_INT *synth_buf, MPA_INT *window,
>
> sum = *dither_state;
> p = synth_buf + 16;
> - SUM8(MACS, sum, w, p);
> + SUM8_MACS(sum, w, p);
> p = synth_buf + 48;
> - SUM8(MLSS, sum, w + 32, p);
> + SUM8_MLSS(sum, w + 32, p);
> *samples = round_sample(&sum);
> samples += incr;
> w++;
This is ugly
You should implement the API not change it in one file and implement the
modified API
[...]
> +
> +#define SH_SUM8_MACS(sum, w, p) \
> +do { \
> + register const int32_t *wp = (w), *pp = (p); \
> + register int32_t tmp = 63 * 4; \
> + union {int64_t x; int32_t u32[2];} u = \
> + {.x = (sum),}; \
> + __asm__ volatile( \
> + " lds %[hi], mach\n" \
> + " lds %[lo], macl\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 0 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 1 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 2 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 3 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 4 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 5 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 6 */ \
> + " add %[tmp], %[wp]\n" \
> + " add %[tmp], %[pp]\n" \
> + " mac.l @%[wp]+, @%[pp]+\n" /* 7 */ \
> + " sts mach, %[hi]\n" \
> + " sts macl, %[lo]" \
> + : [hi] "+&r" (u.u32[1]), \
> + [lo] "+&r" (u.u32[0]), \
> + [wp] "+&r" (wp), \
> + [pp] "+&r" (pp) \
> + : [tmp] "r" (tmp) \
> + : "mach", "macl"); \
> + sum = u.x; \
> +} while (0)
Write the whole loop in asm, leaving it to gcc is not a good idea it can just
mess up.
also i assume there are no multiply and subtract instructions in SH4 ?
btw do you have a instruction set reference you can share or a url to one ?
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
When you are offended at any man's fault, turn to yourself and study your
own failings. Then you will forget your anger. -- Epictetus
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20101226/dcf47c55/attachment.pgp>
More information about the ffmpeg-devel
mailing list