[x265] [PATCH] psyCost_pp: C code optimization, suitable for ASM conversion

[Steve Borho]

On 12/12, praveen at multicorewareinc.com wrote:
> # HG changeset patch
> # User Praveen Tiwari
> # Date 1418376812 -19800
> # Node ID ca805794c7519ef698776d0f717fd66b0d3bc626
> # Parent  b1c2ef980dfe59c486454a8838c2c1bb74bf4d32
> psyCost_pp: C code optimization, suitable for ASM conversion
> 
> psy_sum_std_4x4, is combined version of satd_4x4 and sad<4, 4>, eliminating unnecessary load operations, extra arguments(zeroBuf and 0).
> It also replaces SAD operation with low cost sum operation as the sencod pixel buffer is zero.

Do we actually measure psy energy for 4x4 pixel blocks? CU's are generally 8x8 to 64x64.

The short-short version which operates on residual blocks might be used
for 4x4 TUs, but I'm not certain about that.

> diff -r b1c2ef980dfe -r ca805794c751 source/common/pixel.cpp
> --- a/source/common/pixel.cpp	Thu Dec 11 16:52:06 2014 -0600
> +++ b/source/common/pixel.cpp	Fri Dec 12 15:03:32 2014 +0530
> @@ -801,6 +801,38 @@
>  #pragma warning(disable: 4127) // conditional expression is constant
>  #endif
>  
> +int psy_sum_std_4x4(const pixel* pix1, intptr_t stride_pix1)
> +{
> +    sum2_t tmp[4][2];
> +    sum2_t a0, a1, a2, a3, b0, b1;
> +    sum2_t sum = 0;
> +    int sum0 = 0;
> +    uint32_t sum1 = 0;
> +
> +    for (int i = 0; i < 4; i++, pix1 += stride_pix1)
> +    {
> +        a0 = pix1[0];
> +        a1 = pix1[1];
> +        a2 = pix1[2];
> +        a3 = pix1[3];
> +        sum1 = sum1 + (a0 + a1 + a2 + a3);
> +        b0 = (a0 + a1) + ((a0 - a1) << BITS_PER_SUM);
> +        b1 = (a2 + a3) + ((a2 - a3) << BITS_PER_SUM);
> +        tmp[i][0] = b0 + b1;
> +        tmp[i][1] = b0 - b1;
> +    }
> +
> +    for (int i = 0; i < 2; i++)
> +    {
> +        HADAMARD4(a0, a1, a2, a3, tmp[0][i], tmp[1][i], tmp[2][i], tmp[3][i]);
> +        a0 = abs2(a0) + abs2(a1) + abs2(a2) + abs2(a3);
> +        sum += ((sum_t)a0) + (a0 >> BITS_PER_SUM);
> +    }
> +
> +    sum0 = (int)(sum >> 1);
> +    return (int)(sum0 - (sum1 >> 2));
> +}
> +
>  template<int size>
>  int psyCost_pp(const pixel* source, intptr_t sstride, const pixel* recon, intptr_t rstride)
>  {
> @@ -828,8 +860,8 @@
>      else
>      {
>          /* 4x4 is too small for sa8d */
> -        int sourceEnergy = satd_4x4(source, sstride, zeroBuf, 0) - (sad<4, 4>(source, sstride, zeroBuf, 0) >> 2);
> -        int reconEnergy = satd_4x4(recon, rstride, zeroBuf, 0) - (sad<4, 4>(recon, rstride, zeroBuf, 0) >> 2);
> +        int sourceEnergy = psy_sum_std_4x4(source, sstride);
> +        int reconEnergy = psy_sum_std_4x4(recon, rstride);
>          return abs(sourceEnergy - reconEnergy);
>      }
>  }
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel

-- 
Steve Borho