[x265] [PATCH] avx2: 'integral4h' asm code -> 6.01x faster than 'C' version

[Pradeep Ramachandran]

On Thu, Jun 8, 2017 at 6:51 PM, <jayashri at multicorewareinc.com> wrote:

> # HG changeset patch
> # User Jayashri Murugan <jayashri at multicorewareinc.com>
> # Date 1496062876 -19800
> #      Mon May 29 18:31:16 2017 +0530
> # Node ID dd8b50b2d936a5f21972563e117101f82a1ae9c4
> # Parent  e75d5f5eeae3413057437af9f7d3ba9bc10fa3fa
> avx2: 'integral4h' asm code -> 6.01x faster than 'C' version
>
>     integral_init4h  6.01x    353.40          2122.37
>

Pushed to default branch of x265.


>
> diff -r e75d5f5eeae3 -r dd8b50b2d936 source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp      Mon Jun 05 15:20:44 2017
> +0530
> +++ b/source/common/x86/asm-primitives.cpp      Mon May 29 18:31:16 2017
> +0530
> @@ -3709,6 +3709,7 @@
>          p.integral_initv[INTEGRAL_16] = PFX(integral16v_avx2);
>          p.integral_initv[INTEGRAL_24] = PFX(integral24v_avx2);
>          p.integral_initv[INTEGRAL_32] = PFX(integral32v_avx2);
> +        p.integral_inith[INTEGRAL_4] = PFX(integral4h_avx2);
>
>      }
>  #endif
> diff -r e75d5f5eeae3 -r dd8b50b2d936 source/common/x86/seaintegral.asm
> --- a/source/common/x86/seaintegral.asm Mon Jun 05 15:20:44 2017 +0530
> +++ b/source/common/x86/seaintegral.asm Mon May 29 18:31:16 2017 +0530
> @@ -152,8 +152,67 @@
>  ;static void integral_init4h_c(uint32_t *sum, pixel *pix, intptr_t stride)
>  ;-----------------------------------------------------------
> ------------------
>  INIT_YMM avx2
> -cglobal integral4h, 3, 3, 0
> -
> +
> +%macro INTEGRAL_FOUR_HORIZONTAL_16 0
> +    pmovzxbw       m0, [r1]
> +    pmovzxbw       m1, [r1 + 1]
> +    paddw          m0, m1
> +    pmovzxbw       m1, [r1 + 2]
> +    paddw          m0, m1
> +    pmovzxbw       m1, [r1 + 3]
> +    paddw          m0, m1
> +%endmacro
> +
> +cglobal integral4h, 3, 5, 3
> +    lea            r3, [4 * r2]
> +    sub            r0, r3
> +    sub            r2, 4                      ;stride - 4
> +    mov            r4, r2
> +    shr            r4, 4
> +
> +.loop_16:
> +    INTEGRAL_FOUR_HORIZONTAL_16
> +    vperm2i128     m2, m0, m0, 1
> +    pmovzxwd       m2, xm2
> +    pmovzxwd       m0, xm0
> +    movu           m1, [r0]
> +    paddd          m0, m1
> +    movu           [r0 + r3], m0
> +    movu           m1, [r0 + 32]
> +    paddd          m2, m1
> +    movu           [r0 + r3 + 32], m2
> +    add            r1, 16
> +    add            r0, 64
> +    sub            r2, 16
> +    sub            r4, 1
> +    jnz            .loop_16
> +    cmp            r2, 12
> +    je             .loop_12
> +    cmp            r2, 4
> +    je             .loop_4
> +
> +.loop_12:
> +    INTEGRAL_FOUR_HORIZONTAL_16
> +    vperm2i128     m2, m0, m0, 1
> +    pmovzxwd       xm2, xm2
> +    pmovzxwd       m0, xm0
> +    movu           m1, [r0]
> +    paddd          m0, m1
> +    movu           [r0 + r3], m0
> +    movu           xm1, [r0 + 32]
> +    paddd          xm2, xm1
> +    movu           [r0 + r3 + 32], xm2
> +    jmp             .end
> +
> +.loop_4:
> +    INTEGRAL_FOUR_HORIZONTAL_16
> +    pmovzxwd       xm0, xm0
> +    movu           xm1, [r0]
> +    paddd          xm0, xm1
> +    movu           [r0 + r3], xm0
> +    jmp            .end
> +
> +.end
>      RET
>
>  ;-----------------------------------------------------------
> ------------------
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20170612/5dd6ff86/attachment.html>