[x265] [PATCH] dct: replaced partialButterfly8 vector class function with intrinsic

[Steve Borho]

On Thu, Oct 10, 2013 at 8:02 AM, <yuvaraj at multicorewareinc.com> wrote:

> # HG changeset patch
> # User Yuvaraj Venkatesh <yuvaraj at multicorewareinc.com>
> # Date 1381410080 -19800
> #      Thu Oct 10 18:31:20 2013 +0530
> # Node ID d43f21a8128f02c97aafde351162d00977f664b6
> # Parent  840229ed3794569f5e15d84289531c829b75dcd6
> dct: replaced partialButterfly8 vector class function with intrinsic
>

this new function uses pmulld, which is sse4.1, so this function needs to
be moved


>
> diff -r 840229ed3794 -r d43f21a8128f source/common/vec/dct-sse3.cpp
> --- a/source/common/vec/dct-sse3.cpp    Thu Oct 10 16:57:47 2013 +0530
> +++ b/source/common/vec/dct-sse3.cpp    Thu Oct 10 18:31:20 2013 +0530
> @@ -246,47 +246,48 @@
>      int j;
>      int add = 1 << (shift - 1);
>
> -    Vec4i zero_row(64, 64, 0, 0);
> -    Vec4i four_row(64, -64, 0, 0);
> -    Vec4i two_row(83, 36, 0, 0);
> -    Vec4i six_row(36, -83, 0, 0);
> +    __m128i zero_row = _mm_setr_epi32(64, 64, 0, 0);
> +    __m128i four_row = _mm_setr_epi32(64, -64, 0, 0);
> +    __m128i two_row = _mm_setr_epi32(83, 36, 0, 0);
> +    __m128i six_row = _mm_setr_epi32(36, -83, 0, 0);
>
> -    Vec4i one_row(89, 75, 50, 18);
> -    Vec4i three_row(75, -18, -89, -50);
> -    Vec4i five_row(50, -89, 18, 75);
> -    Vec4i seven_row(18, -50, 75, -89);
> +    __m128i one_row = _mm_setr_epi32(89, 75, 50, 18);
> +    __m128i three_row = _mm_setr_epi32(75, -18, -89, -50);
> +    __m128i five_row = _mm_setr_epi32(50, -89, 18, 75);
> +    __m128i seven_row = _mm_setr_epi32(18, -50, 75, -89);
>
>      for (j = 0; j < line; j++)
>      {
> -        Vec8s srcTmp;
> -        srcTmp.load(src);
> +        __m128i srcTmp;
> +        srcTmp = _mm_loadu_si128((__m128i*)(src));
>
> -        Vec4i E_first_half = extend_low(srcTmp);
> -        Vec4i E_second_half = extend_high(srcTmp);
> -        E_second_half = permute4i<3, 2, 1, 0>(E_second_half);
> +        __m128i sign = _mm_srai_epi16(srcTmp, 15);
> +        __m128i E_first_half = _mm_unpacklo_epi16(srcTmp, sign);
> +        __m128i E_second_half = _mm_unpackhi_epi16(srcTmp, sign);
> +        E_second_half = _mm_shuffle_epi32(E_second_half, 27);
>
> -        Vec4i E = E_first_half + E_second_half;
> -        Vec4i O = E_first_half - E_second_half;
> +        __m128i E = _mm_add_epi32(E_first_half, E_second_half);
> +        __m128i O = _mm_sub_epi32(E_first_half, E_second_half);
>
> -        Vec4i EE_first_half = permute4i<0, 1, -1, -1>(E);
> -        Vec4i EE_second_half = permute4i<3, 2, -1, -1>(E);
> -        Vec4i EE = EE_first_half + EE_second_half;
> -        Vec4i EO = EE_first_half - EE_second_half;
> +        __m128i EE_first_half = _mm_shuffle_epi32(E, 4);
> +        __m128i EE_second_half = _mm_shuffle_epi32(E, 11);
> +        __m128i EE = _mm_add_epi32(EE_first_half, EE_second_half);
> +        __m128i EO = _mm_sub_epi32(EE_first_half, EE_second_half);
>
> -        int dst0 = ((horizontal_add(zero_row * EE)) + add) >> shift;
> -        int dst4 = ((horizontal_add(four_row * EE)) + add) >> shift;
> -        int dst2 = ((horizontal_add(two_row * EO)) + add) >> shift;
> -        int dst6 = ((horizontal_add(six_row * EO)) + add) >> shift;
> +        int dst0 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(zero_row,
> EE), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst4 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(four_row,
> EE), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst2 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(two_row,
> EO), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst6 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(six_row,
> EO), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
>
>          dst[0] = dst0;
>          dst[4 * line] = dst4;
>          dst[2 * line] = dst2;
>          dst[6 * line] = dst6;
>
> -        int dst1 = ((horizontal_add(one_row * O)) + add) >> shift;
> -        int dst3 = ((horizontal_add(three_row * O)) + add) >> shift;
> -        int dst5 = ((horizontal_add(five_row * O)) + add) >> shift;
> -        int dst7 = ((horizontal_add(seven_row * O)) + add) >> shift;
> +        int dst1 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(one_row,
> O), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst3 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(three_row,
> O), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst5 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(five_row,
> O), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
> +        int dst7 =
> (_mm_cvtsi128_si32(_mm_hadd_epi32(_mm_hadd_epi32(_mm_mullo_epi32(seven_row,
> O), _mm_setzero_si128()), _mm_setzero_si128())) + add) >> shift;
>
>          dst[line] = dst1;
>          dst[3 * line] = dst3;
> _______________________________________________
> 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/20131010/eee8d6b5/attachment.html>