[x265] [PATCH Review only] asm: 10bpp code for transpose 4x4 and 8x8

[Murugan Vairavel]

Ignore this patch. Need modifications for 16x16.



On Tue, Dec 3, 2013 at 7:08 PM, <murugan at multicorewareinc.com> wrote:

> # HG changeset patch
> # User Murugan Vairavel <murugan at multicorewareinc.com>
> # Date 1386077908 -19800
> #      Tue Dec 03 19:08:28 2013 +0530
> # Node ID 1ae4e8ae04d0792db6590a62272990d83f49a265
> # Parent  126f3aefc79dad37e7985953c404ccff370d2729
> asm: 10bpp code for transpose 4x4 and 8x8
>
> diff -r 126f3aefc79d -r 1ae4e8ae04d0 source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp      Tue Dec 03 18:33:13 2013
> +0530
> +++ b/source/common/x86/asm-primitives.cpp      Tue Dec 03 19:08:28 2013
> +0530
> @@ -520,6 +520,9 @@
>          p.sse_ss[LUMA_32x32] = x265_pixel_ssd_ss_32x32_sse2;
>          p.sse_ss[LUMA_32x64] = x265_pixel_ssd_ss_32x64_sse2;
>
> +        p.transpose[BLOCK_4x4] = x265_transpose4_sse2;
> +        p.transpose[BLOCK_8x8] = x265_transpose8_sse2;
> +
>          p.ssim_4x4x2_core = x265_pixel_ssim_4x4x2_core_sse2;
>          PIXEL_AVG(sse2);
>          PIXEL_AVG_W4(mmx2);
> diff -r 126f3aefc79d -r 1ae4e8ae04d0 source/common/x86/pixel-util8.asm
> --- a/source/common/x86/pixel-util8.asm Tue Dec 03 18:33:13 2013 +0530
> +++ b/source/common/x86/pixel-util8.asm Tue Dec 03 19:08:28 2013 +0530
> @@ -830,7 +830,20 @@
>  ;-----------------------------------------------------------------
>  INIT_XMM sse2
>  cglobal transpose4, 3, 3, 4, dest, src, stride
> -
> +%if HIGH_BIT_DEPTH
> +    add          r2,    r2
> +    movh         m0,    [r1]
> +    movh         m1,    [r1 + r2]
> +    movh         m2,    [r1 + 2 * r2]
> +    lea          r1,    [r1 + 2 * r2]
> +    movh         m3,    [r1 + r2]
> +    punpcklwd    m0,    m1
> +    punpcklwd    m2,    m3
> +    punpckhdq    m1,    m0,    m2
> +    punpckldq    m0,    m2
> +    movu         [r0],       m0
> +    movu         [r0 + 16],  m1
> +%else
>      movd         m0,    [r1]
>      movd         m1,    [r1 + r2]
>      movd         m2,    [r1 + 2 * r2]
> @@ -841,26 +854,61 @@
>      punpcklbw    m2,    m3
>      punpcklwd    m0,    m2
>      movu         [r0],    m0
> -
> +%endif
>      RET
>
>  ;-----------------------------------------------------------------
>  ; void transpose_8x8(pixel *dst, pixel *src, intptr_t stride)
>  ;-----------------------------------------------------------------
>  INIT_XMM sse2
> -cglobal transpose8, 3, 3, 8, dest, src, stride
> -
> +%if HIGH_BIT_DEPTH
> +%macro TRANSPOSE_4x4 1
>      movh         m0,    [r1]
>      movh         m1,    [r1 + r2]
>      movh         m2,    [r1 + 2 * r2]
>      lea          r1,    [r1 + 2 * r2]
>      movh         m3,    [r1 + r2]
> -    movh         m4,    [r1 + 2 * r2]
> -    lea          r1,    [r1 + 2 * r2]
> +    punpcklwd    m0,    m1
> +    punpcklwd    m2,    m3
> +    punpckhdq    m1,    m0,    m2
> +    punpckldq    m0,    m2
> +    movlps         [r0],             m0
> +    movhps         [r0 + %1],        m0
> +    movlps         [r0 + 2 * %1],    m1
> +    lea            r0,               [r0 + 2 * %1]
> +    movhps         [r0 + %1],        m1
> +%endmacro
> +cglobal transpose8_internal
> +    TRANSPOSE_4x4 r5
> +    lea    r1,    [r1 + 2 * r2]
> +    lea    r0,    [r3 + 8]
> +    TRANSPOSE_4x4 r5
> +    lea    r1,    [r4 + 8]
> +    lea    r0,    [r3 + 4 * r5]
> +    TRANSPOSE_4x4 r5
> +    lea    r1,    [r1 + 2 * r2]
> +    lea    r0,    [r3 + 8 + 4 * r5]
> +    TRANSPOSE_4x4 r5
> +    ret
> +cglobal transpose8, 3, 6, 4, dest, src, stride
> +    add    r2,    r2
> +    mov    r3,    r0
> +    mov    r4,    r1
> +    mov    r5,    16
> +    call   transpose8_internal
> +%else
> +cglobal transpose8, 3, 5, 8, dest, src, stride
> +    lea          r3,    [2 * r2]
> +    lea          r4,    [3 * r2]
> +    movh         m0,    [r1]
> +    movh         m1,    [r1 + r2]
> +    movh         m2,    [r1 + r3]
> +    movh         m3,    [r1 + r4]
> +    movh         m4,    [r1 + 4 * r2]
> +    lea          r1,    [r1 + 4 * r2]
>      movh         m5,    [r1 + r2]
> -    movh         m6,    [r1 + 2 * r2]
> -    lea          r1,    [r1 + 2 * r2]
> -    movh         m7,    [r1 + r2]
> +    movh         m6,    [r1 + r3]
> +    movh         m7,    [r1 + r4]
>
>      punpcklbw    m0,    m1
>      punpcklbw    m2,    m3
> @@ -880,7 +928,7 @@
>      movu         [r0 + 16],    m2
>      movu         [r0 + 32],    m1
>      movu         [r0 + 48],    m3
> -
> +%endif
>      RET
>
>  %macro TRANSPOSE_8x8 1
>



-- 
With Regards,

Murugan. V
+919659287478
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20131203/9cade489/attachment.html>