[x265] [PATCH] pixel16: converted sad_4 from vector class to intrinsic

[Steve Borho]

On Thu, Oct 17, 2013 at 7:17 AM, <yuvaraj at multicorewareinc.com> wrote:

> # HG changeset patch
> # User Yuvaraj Venkatesh <yuvaraj at multicorewareinc.com>
> # Date 1382012201 -19800
> #      Thu Oct 17 17:46:41 2013 +0530
> # Node ID 77f60b1e4441ab947f75291eadf199d2f3ad1057
> # Parent  fc9dbd798ac37ec1acc0596aa179f0deb586c092
> pixel16: converted sad_4 from vector class to intrinsic
>

Queued, but this is not where you should be spending your time.

We have an urgent need to convert the remaining vector class 8bpp
primitives (HIGH_BIT_DEPTH=0) to intrinsics, and we are almost done.  The
only ones left are in intra-sse3.cpp, blockcopy-avx2.cpp, and
pixel-avx2.cpp.

Of those, the intra DC and planar primitives have the highest priority,
followed by AVX2, followed by the intra-angular functions.

I am very tempted to just delete the 8bpp vector class intra-angular
functions and use the C references until we generate assembly for those
because I think their general concept needs to be redesigned.

The HIGH_BIT_DEPTH=1 primitives will all go directly to assembly code
because we are in no rush for those.


> diff -r fc9dbd798ac3 -r 77f60b1e4441 source/common/vec/pixel16-sse41.cpp
> --- a/source/common/vec/pixel16-sse41.cpp       Thu Oct 17 14:14:40 2013
> +0530
> +++ b/source/common/vec/pixel16-sse41.cpp       Thu Oct 17 17:46:41 2013
> +0530
> @@ -41,42 +41,38 @@
>  template<int ly>
>  int sad_4(pixel * fenc, intptr_t fencstride, pixel * fref, intptr_t
> frefstride)
>  {
> -    Vec8s m1, n1;
> +    __m128i sum1 = _mm_setzero_si128();
> +    __m128i T00, T01, T02, T03;
> +    __m128i T10, T11, T12, T13;
> +    __m128i T20, T21;
>
> -    Vec4i sum(0);
> -    Vec8us sad(0);
> -    int max_iterators = (ly >> 4) << 4;
> -    int row;
> +    for (int i = 0; i < ly; i += 4)
> +    {
> +        T00 = _mm_loadl_epi64((__m128i*)(fenc + (i + 0) * fencstride));
> +        T01 = _mm_loadl_epi64((__m128i*)(fenc + (i + 1) * fencstride));
> +        T01 = _mm_unpacklo_epi64(T00, T01);
> +        T02 = _mm_loadl_epi64((__m128i*)(fenc + (i + 2) * fencstride));
> +        T03 = _mm_loadl_epi64((__m128i*)(fenc + (i + 3) * fencstride));
> +        T03 = _mm_unpacklo_epi64(T02, T03);
>
> -    for (row = 0; row < max_iterators; row += 16)
> -    {
> -        for (int i = 0; i < 16; i++)
> -        {
> -            m1.load_a(fenc);
> -            n1.load(fref);
> -            sad += abs(m1 - n1);
> +        T10 = _mm_loadl_epi64((__m128i*)(fref + (i + 0) * frefstride));
> +        T11 = _mm_loadl_epi64((__m128i*)(fref + (i + 1) * frefstride));
> +        T11 = _mm_unpacklo_epi64(T10, T11);
> +        T12 = _mm_loadl_epi64((__m128i*)(fref + (i + 2) * frefstride));
> +        T13 = _mm_loadl_epi64((__m128i*)(fref + (i + 3) * frefstride));
> +        T13 = _mm_unpacklo_epi64(T12, T13);
> +        T20 = _mm_sub_epi16(T01, T11);
> +        T20 = _mm_abs_epi16(T20);
> +        T21 = _mm_sub_epi16(T03, T13);
> +        T21 = _mm_abs_epi16(T21);
> +        T21 = _mm_add_epi16(T20, T21);
> +        sum1 = _mm_add_epi16(sum1, T21);
> +    }
> +    sum1 = _mm_hadd_epi16(sum1, sum1);
> +    sum1 = _mm_unpacklo_epi16(sum1, _mm_setzero_si128());
> +    sum1 = _mm_hadd_epi32(_mm_hadd_epi32(sum1, sum1), sum1);
>
> -            fenc += fencstride;
> -            fref += frefstride;
> -        }
> -
> -        sum += extend_low(sad);
> -        sad = 0;
> -    }
> -
> -    while (row++ < ly)
> -    {
> -        m1.load_a(fenc);
> -        n1.load(fref);
> -        sad += abs(m1 - n1);
> -
> -        fenc += fencstride;
> -        fref += frefstride;
> -    }
> -
> -    sum += extend_low(sad);
> -
> -    return horizontal_add(sum);
> +    return _mm_cvtsi128_si32(sum1);
>  }
>
>  template<int ly>
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>



-- 
Steve Borho
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20131017/2450049f/attachment.html>