diff options
| author | Michael Niedermayer <michaelni@gmx.at> | 2007-08-21 16:29:40 +0000 |
|---|---|---|
| committer | Michael Niedermayer <michaelni@gmx.at> | 2007-08-21 16:29:40 +0000 |
| commit | ce611a27be7c9201b5920d8232e68209529065c4 (patch) | |
| tree | 7d02fb42432300f1ca74277ab0aa77abab7b9fd8 /libavcodec/i386 | |
| parent | 7506d47aa320beb00369325b2e868f4a84c36af0 (diff) | |
Change rounding of the horizontal DWT to match the vertical one.
This allows some simplifications and optimizations and should
not have any effect on quality.
Originally committed as revision 10172 to svn://svn.ffmpeg.org/ffmpeg/trunk
Diffstat (limited to 'libavcodec/i386')
| -rw-r--r-- | libavcodec/i386/snowdsp_mmx.c | 44 |
1 files changed, 17 insertions, 27 deletions
diff --git a/libavcodec/i386/snowdsp_mmx.c b/libavcodec/i386/snowdsp_mmx.c index f2eb14b2b3..03f622b756 100644 --- a/libavcodec/i386/snowdsp_mmx.c +++ b/libavcodec/i386/snowdsp_mmx.c @@ -111,8 +111,7 @@ void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width){ i = 0; asm volatile( - "pcmpeqd %%xmm7, %%xmm7 \n\t" - "psrad $29, %%xmm7 \n\t" + "pslld $1, %%xmm7 \n\t" ::); for(; i<w_l-7; i+=8){ asm volatile( @@ -157,25 +156,21 @@ void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width){ "movdqu 20(%1), %%xmm6 \n\t" "paddd (%1), %%xmm2 \n\t" "paddd 16(%1), %%xmm6 \n\t" - "movdqa %%xmm2, %%xmm0 \n\t" - "movdqa %%xmm6, %%xmm4 \n\t" - "pslld $2, %%xmm2 \n\t" - "pslld $2, %%xmm6 \n\t" - "psubd %%xmm2, %%xmm0 \n\t" - "psubd %%xmm6, %%xmm4 \n\t" - "psrad $1, %%xmm0 \n\t" - "psrad $1, %%xmm4 \n\t" - "movdqu (%0), %%xmm2 \n\t" - "movdqu 16(%0), %%xmm6 \n\t" - "psubd %%xmm0, %%xmm2 \n\t" - "psubd %%xmm4, %%xmm6 \n\t" + "movdqu (%0), %%xmm0 \n\t" + "movdqu 16(%0), %%xmm4 \n\t" + "paddd %%xmm2, %%xmm0 \n\t" + "paddd %%xmm6, %%xmm4 \n\t" + "psrad $1, %%xmm2 \n\t" + "psrad $1, %%xmm6 \n\t" + "paddd %%xmm0, %%xmm2 \n\t" + "paddd %%xmm4, %%xmm6 \n\t" "movdqa %%xmm2, (%2) \n\t" "movdqa %%xmm6, 16(%2) \n\t" :: "r"(&src[i]), "r"(&b[i]), "r"(&temp[i]) : "memory" ); } - snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); + snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS); } { @@ -291,10 +286,9 @@ void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){ DWTELEM * const ref = b+w2 - 1; i = 1; - b[0] = b[0] + (((2 * ref[1] + W_BO-1) + 4 * b[0]) >> W_BS); + b[0] = b[0] + (((2 * ref[1] + W_BO) + 4 * b[0]) >> W_BS); asm volatile( - "pcmpeqd %%mm7, %%mm7 \n\t" - "psrld $29, %%mm7 \n\t" + "pslld $1, %%mm7 \n\t" ::); for(; i<w_l-3; i+=4){ asm volatile( @@ -333,16 +327,12 @@ void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){ "movq 12(%1), %%mm6 \n\t" "paddd (%1), %%mm2 \n\t" "paddd 8(%1), %%mm6 \n\t" - "pxor %%mm0, %%mm0 \n\t" //note: the 2 xor could be avoided if we would flip the rounding direction - "pxor %%mm4, %%mm4 \n\t" - "psubd %%mm2, %%mm0 \n\t" - "psubd %%mm6, %%mm4 \n\t" - "psrad $1, %%mm0 \n\t" - "psrad $1, %%mm4 \n\t" - "psubd %%mm0, %%mm2 \n\t" - "psubd %%mm4, %%mm6 \n\t" "movq (%0), %%mm0 \n\t" "movq 8(%0), %%mm4 \n\t" + "paddd %%mm2, %%mm0 \n\t" + "paddd %%mm6, %%mm4 \n\t" + "psrad $1, %%mm2 \n\t" + "psrad $1, %%mm6 \n\t" "paddd %%mm0, %%mm2 \n\t" "paddd %%mm4, %%mm6 \n\t" "movq %%mm2, (%2) \n\t" @@ -351,7 +341,7 @@ void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width){ : "memory" ); } - snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO, W_AS); + snow_horizontal_compose_lift_lead_out(i, temp, src, b, width, w_r, 1, -W_AM, W_AO+1, W_AS); } { |