:: Re: [Frei0r] [PATCH] Optimize the b…
Góra strony
Delete this message
Reply to this message
Autor: Dan Dennedy
Data:  
Dla: Minimalistic plugin API for video effects
Temat: Re: [Frei0r] [PATCH] Optimize the balanc0r plugin using SSE2
On Wed, Sep 12, 2012 at 2:54 PM, Steinar H. Gunderson
<sgunderson@???> wrote:
> Hi,
>
> Before I abandoned the balanc0r plugin altogether, I made some efforts to
> optimize it (and to make the surrounding processes in Kdenlive/MLT less slow;
> see http://www.kdenlive.org/mantis/view.php?id=2727). Maybe the patch will be
> useful to someone, so I'm sending it on instead of letting it rot in my own
> directory.


Thank for you trying to help. To other committers, please do not apply
this because there is uncertainty about its correctness in another
thread. In that case, perhaps this should only be changed to somehow
indicate it is deprecated. Plus...

> /* Steinar */
>
>
> commit acdb08c2e3c63d256a73d6db119fe9a5a0b9567c
> Author: Steinar H. Gunderson <sgunderson@???>
> Date: Mon Sep 10 22:03:16 2012 +0200
>
>     Optimize the balanc0r plugin using SSE2.

>
>     There is no good reason why this plugin should work in floating point;
>     the conversions back and forth are very slow. Moving it to fixed point
>     also enabled SSE2-optimizing it.

>
>     Also make sure CLAMP0255() is inlined (it was not in my build).

>
>     On a rough estimate, this makes the plugin four times as fast on my
>     Core i5.

>
> diff --git a/include/frei0r_math.h b/include/frei0r_math.h
> index cf84eb4..ad625db 100644
> --- a/include/frei0r_math.h
> +++ b/include/frei0r_math.h
> @@ -18,7 +18,7 @@
>
> /* Clamps a int32-range int between 0 and 255 inclusive. */
> #ifndef CLAMP0255
> -unsigned char CLAMP0255(int32_t a)
> +static inline unsigned char CLAMP0255(int32_t a)


good idea

>  {
>    return (unsigned char)
>      ( (((-a) >> 31) & a)  // 0 if the number was negative
> diff --git a/src/filter/balanc0r/balanc0r.c b/src/filter/balanc0r/balanc0r.c
> index c52bc15..d08ff42 100644
> --- a/src/filter/balanc0r/balanc0r.c
> +++ b/src/filter/balanc0r/balanc0r.c
> @@ -27,6 +27,10 @@
>  #include <math.h>
>  #include <stdio.h>

>
> +#ifdef __SSE2__


where is that defined? Looks gcc-specific but also in clang, but is it
safe or dependent on some recent version of clang? I know some people
have submitted patches to frei0r for msvc. What about that compiler?

> +#include <emmintrin.h>
> +#endif
> +
> #include "frei0r.h"
> #include "frei0r_math.h"
>
> @@ -684,19 +688,51 @@ void f0r_get_param_value(f0r_instance_t instance,
>
> }
>
> +int clamp(int x, int low, int high)
> +{
> +       if (x < low) {
> +               return low;
> +       }
> +       if (x > high) {
> +               return high;
> +       }
> +       return x;
> +}
> +
>  void f0r_update(f0r_instance_t instance, double time,
>                 const uint32_t* inframe, uint32_t* outframe)
>  {
>         assert(instance);
>         balanc0r_instance_t* inst = (balanc0r_instance_t*)instance;
> -       unsigned int         len  = inst->width * inst->height + 1;
> +       unsigned int         len  = inst->width * inst->height;
>         unsigned char*       dst = (unsigned char*)outframe;
>         const unsigned char* src = (unsigned char*)inframe;
> +
> +       // Convert the multiplicative constants to 8.8 fixed point.
> +       int mr = clamp(lrint(inst->mr * 256.0), 0, 65535);
> +       int mg = clamp(lrint(inst->mg * 256.0), 0, 65535);
> +       int mb = clamp(lrint(inst->mb * 256.0), 0, 65535);
> +
> +#ifdef __SSE2__
> +       __m128i mulfac = _mm_setr_epi16(mr, mg, mb, 256, mr, mg, mb, 256);
> +       __m128i zero = _mm_setzero_si128();
> +       while (len >= 4) {
> +               __m128i lohi = _mm_loadu_si128((const __m128i *)src);
> +               __m128i lo = _mm_unpacklo_epi8(zero, lohi);
> +               __m128i hi = _mm_unpackhi_epi8(zero, lohi);
> +               lo = _mm_mulhi_epu16(lo, mulfac);
> +               hi = _mm_mulhi_epu16(hi, mulfac);
> +               _mm_storeu_si128((__m128i *)dst, _mm_packus_epi16(lo, hi));
> +               len -= 4;
> +               src += 16;
> +               dst += 16;


I guess it is not too big of deal if last few pixels are skipped when
(width * height) is not a multiple of 4.

> +       }
> +#endif

>
> -       while (--len) {
> -               *dst++ = CLAMP0255(*src++ * inst->mr);
> -               *dst++ = CLAMP0255(*src++ * inst->mg);
> -               *dst++ = CLAMP0255(*src++ * inst->mb);
> +       while (len--) {
> +               *dst++ = CLAMP0255((*src++ * mr) >> 8);
> +               *dst++ = CLAMP0255((*src++ * mg) >> 8);
> +               *dst++ = CLAMP0255((*src++ * mb) >> 8);
>                 *dst++ = *src++;  // copy alpha
>         }
>  }
> --
> Homepage: http://www.sesse.net/


--
+-DRD-+