[FFmpeg-devel] [PATCH] h264: integrate clear_blocks calls with IDCT.

[Ronald S. Bultje]

Hi,

On Sun, Feb 17, 2013 at 6:04 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> On Sun, Feb 17, 2013 at 02:52:54PM -0800, Ronald S. Bultje wrote:
>> From: "Ronald S. Bultje" <rsbultje at gmail.com>
>>
>> The non-intra-pcm branch in hl_decode_mb (simple, 8bpp) goes from 700
>> to 672 cycles, and the complete loop of decode_mb_cabac and hl_decode_mb
>> (in the decode_slice loop) goes from 1759 to 1733 cycles on the clip
>> tested (cathedral), i.e. almost 30 cycles per mb faster.
>>
>
>> Arm assembly changes untested.
>
> fate-h264 (h264-conformance-ba_mw_d in this case but its not the only
> one)
>
> Program received signal SIGSEGV, Segmentation fault.
> ff_h264_idct_add8_neon () at ffmpeg/libavcodec/arm/h264idct_neon.S:166
> 166         ldrsh           r8,  [r1]

diff --git a/libavcodec/arm/h264idct_neon.S b/libavcodec/arm/h264idct_neon.S
index a2521b7..8d85227 100644
--- a/libavcodec/arm/h264idct_neon.S
+++ b/libavcodec/arm/h264idct_neon.S
@@ -105,6 +105,7 @@ function ff_h264_idct_add16_neon, export=1
         ldr             r0,  [r5], #4
         ldrb            r8,  [r6, r8]
         subs            r8,  r8,  #1
+        mov             r3,  r1
         blt             2f
         ldrsh           lr,  [r1]
         add             r0,  r0,  r4
@@ -116,7 +117,7 @@ function ff_h264_idct_add16_neon, export=1
         adreq           lr,  ff_h264_idct_add_neon    + CONFIG_THUMB
         blx             lr
 2:      subs            ip,  ip,  #1
-        add             r1,  r1,  #32
+        add             r1,  r3,  #32
         bne             1b
         pop             {r4-r8,pc}
 endfunc
@@ -136,13 +137,14 @@ function ff_h264_idct_add16intra_neon, export=1
         add             r0,  r0,  r4
         cmp             r8,  #0
         ldrsh           r8,  [r1]
+        mov             r3,  r1
         iteet           ne
         adrne           lr,  ff_h264_idct_add_neon    + CONFIG_THUMB
         adreq           lr,  ff_h264_idct_dc_add_neon + CONFIG_THUMB
         cmpeq           r8,  #0
         blxne           lr
         subs            ip,  ip,  #1
-        add             r1,  r1,  #32
+        add             r1,  r3,  #32
         bne             1b
         pop             {r4-r8,pc}
 endfunc

?

Ronald