[x265] [PATCH] avx2: 'integral12h' asm code -> 2.42x faster than 'C' version
Ashok Kumar Mishra
ashok at multicorewareinc.com
Thu Jun 15 15:08:40 CEST 2017
pushed
On Wed, Jun 14, 2017 at 4:07 PM, <jayashri at multicorewareinc.com> wrote:
> # HG changeset patch
> # User Jayashri Murugan <jayashri at multicorewareinc.com>
> # Date 1496834769 -19800
> # Wed Jun 07 16:56:09 2017 +0530
> # Node ID 6d1fea9c34c5a1b7cb062ce3126df7de50170419
> # Parent 324ee113f48943791ec295013012d30fcbe16338
> avx2: 'integral12h' asm code -> 2.42x faster than 'C' version
>
> BIT_DEPTH = 8 : integral_init12h 2.42x 807.20 1950.56
> BIT_DEPTH = 10|12 : integral_init12h 1.55x 1260.02 1958.27
>
> diff -r 324ee113f489 -r 6d1fea9c34c5 source/common/x86/asm-primitives.cpp
> --- a/source/common/x86/asm-primitives.cpp Wed Jun 07 15:37:49 2017
> +0530
> +++ b/source/common/x86/asm-primitives.cpp Wed Jun 07 16:56:09 2017
> +0530
> @@ -2166,6 +2166,7 @@
> p.integral_initv[INTEGRAL_32] = PFX(integral32v_avx2);
> p.integral_inith[INTEGRAL_4] = PFX(integral4h_avx2);
> p.integral_inith[INTEGRAL_8] = PFX(integral8h_avx2);
> + p.integral_inith[INTEGRAL_12] = PFX(integral12h_avx2);
>
> /* TODO: This kernel needs to be modified to work with
> HIGH_BIT_DEPTH only
> p.planeClipAndMax = PFX(planeClipAndMax_avx2); */
> @@ -3713,6 +3714,7 @@
> p.integral_initv[INTEGRAL_32] = PFX(integral32v_avx2);
> p.integral_inith[INTEGRAL_4] = PFX(integral4h_avx2);
> p.integral_inith[INTEGRAL_8] = PFX(integral8h_avx2);
> + p.integral_inith[INTEGRAL_12] = PFX(integral12h_avx2);
>
> }
> #endif
> diff -r 324ee113f489 -r 6d1fea9c34c5 source/common/x86/seaintegral.asm
> --- a/source/common/x86/seaintegral.asm Wed Jun 07 15:37:49 2017 +0530
> +++ b/source/common/x86/seaintegral.asm Wed Jun 07 16:56:09 2017 +0530
> @@ -389,14 +389,204 @@
> RET
> %endif
>
> +%macro INTEGRAL_TWELVE_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
> + pmovzxbw m1, [r1 + 4]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 5]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 6]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 7]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 8]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 9]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 10]
> + paddw m0, m1
> + pmovzxbw m1, [r1 + 11]
> + paddw m0, m1
> +%endmacro
> +
> +%macro INTEGRAL_TWELVE_HORIZONTAL_4 0
> + movd xm0, [r1]
> + movd xm1, [r1 + 1]
> + pmovzxbw xm0, xm0
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 2]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 3]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 4]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 5]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 6]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 7]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 8]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 9]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 10]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> + movd xm1, [r1 + 11]
> + pmovzxbw xm1, xm1
> + paddw xm0, xm1
> +%endmacro
> +
> +%macro INTEGRAL_TWELVE_HORIZONTAL_8_HBD 0
> + pmovzxwd m0, [r1]
> + pmovzxwd m1, [r1 + 2]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 4]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 6]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 8]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 10]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 12]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 14]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 16]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 18]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 20]
> + paddd m0, m1
> + pmovzxwd m1, [r1 + 22]
> + paddd m0, m1
> +%endmacro
> +
> +%macro INTEGRAL_TWELVE_HORIZONTAL_4_HBD 0
> + pmovzxwd xm0, [r1]
> + pmovzxwd xm1, [r1 + 2]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 4]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 6]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 8]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 10]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 12]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 14]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 16]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 18]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 20]
> + paddd xm0, xm1
> + pmovzxwd xm1, [r1 + 22]
> + paddd xm0, xm1
> +%endmacro
> +
> ;-----------------------------------------------------------
> ------------------
> ;static void integral_init12h_c(uint32_t *sum, pixel *pix, intptr_t
> stride)
> ;-----------------------------------------------------------
> ------------------
> INIT_YMM avx2
> -cglobal integral12h, 3, 3, 0
> -
> +%if HIGH_BIT_DEPTH
> +cglobal integral12h, 3, 5, 3
> + lea r3, [4 * r2]
> + sub r0, r3
> + sub r2, 12 ;stride - 12
> + mov r4, r2
> + shr r4, 3
> +
> +.loop:
> + INTEGRAL_TWELVE_HORIZONTAL_8_HBD
> + movu m1, [r0]
> + paddd m0, m1
> + movu [r0 + r3], m0
> + add r1, 16
> + add r0, 32
> + sub r2, 8
> + sub r4, 1
> + jnz .loop
> + INTEGRAL_TWELVE_HORIZONTAL_4_HBD
> + movu xm1, [r0]
> + paddd xm0, xm1
> + movu [r0 + r3], xm0
> RET
>
> +%else
> +cglobal integral12h, 3, 5, 3
> + lea r3, [4 * r2]
> + sub r0, r3
> + sub r2, 12 ;stride - 12
> + mov r4, r2
> + shr r4, 4
> +
> +.loop_16:
> + INTEGRAL_TWELVE_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_TWELVE_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_TWELVE_HORIZONTAL_4
> + pmovzxwd xm0, xm0
> + movu xm1, [r0]
> + paddd xm0, xm1
> + movu [r0 + r3], xm0
> + jmp .end
> +
> +.end
> + RET
> +%endif
> +
> ;-----------------------------------------------------------
> ------------------
> ;static void integral_init16h_c(uint32_t *sum, pixel *pix, intptr_t
> stride)
> ;-----------------------------------------------------------
> ------------------
> _______________________________________________
> 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/20170615/eb9e3671/attachment.html>
More information about the x265-devel
mailing list