[FFmpeg-devel] [PATCH] Port mp=eq/eq2 to FFmpeg

[wm4]

On Thu, 22 Jan 2015 01:38:11 +0530
arwa arif <arwaarif1994 at gmail.com> wrote:

> From 703cc1887903c2868537e19e99b76927bec07884 Mon Sep 17 00:00:00 2001
> From: Arwa Arif <arwaarif1994 at gmail.com>
> Date: Mon, 19 Jan 2015 03:56:48 +0530
> Subject: [PATCH] Port mp=eq/eq2 to FFmpeg
> 
> Code adapted from James Darnley's previous commits

> [...]

> +#if HAVE_MMX && HAVE_6REGS
> +static void process_MMX(EQParameters *param, uint8_t *dst, int dst_stride,
> +                        uint8_t *src, int src_stride, int w, int h)
> +{
> +        int i;
> +        int pel;
> +        int dstep = dst_stride - w;
> +        int sstep = src_stride - w;
> +        short brvec[4];
> +        short contvec[4];
> +
> +        brvec[0] = brvec[1] = brvec[2] = brvec[3] = param->b;
> +        contvec[0] = contvec[1] = contvec[2] = contvec[3] = param->c;
> +
> +        while (h--) {
> +                __asm__ volatile (
> +                        "movq (%5), %%mm3      \n\t"
> +                        "movq (%6), %%mm4      \n\t"
> +                        "pxor %%mm0, %%mm0     \n\t"
> +                        "movl %4, %%eax        \n\t"
> +                        //ASMALIGN(4)
> +                        "1:                    \n\t"
> +                        "movq (%0), %%mm1      \n\t"
> +                        "movq (%0), %%mm2      \n\t"
> +                        "punpcklbw %%mm0, %%mm1\n\t"
> +                        "punpckhbw %%mm0, %%mm2\n\t"
> +                        "psllw $4, %%mm1       \n\t"
> +                        "psllw $4, %%mm2       \n\t"
> +                        "pmulhw %%mm4, %%mm1   \n\t"
> +                        "pmulhw %%mm4, %%mm2   \n\t"
> +                        "paddw %%mm3, %%mm1    \n\t"
> +                        "paddw %%mm3, %%mm2    \n\t"
> +                        "packuswb %%mm2, %%mm1 \n\t"
> +                        "add $8, %0            \n\t"
> +                        "movq %%mm1, (%1)      \n\t"
> +                        "add $8, %1            \n\t"
> +                        "decl %%eax            \n\t"
> +                        "jnz 1b                \n\t"
> +                        : "=r" (src), "=r" (dst)
> +                        : "0" (src), "1" (dst), "r" (w>>3), "r" (brvec), "r" (contvec)
> +                        : "%eax"
> +                );
> +
> +                for (i = w&7; i; i--) {
> +                        pel = ((*src++ * param->c) >> 12) + param->b;
> +                        if (pel & 768) 
> +                            pel = (-pel) >> 31;
> +                        *dst++ = pel;
> +                }
> +
> +                src += sstep;
> +                dst += dstep;
> +        }
> +        __asm__ volatile ( "emms \n\t" ::: "memory" );
> +}
> +#endif

IMO the asm shouldn't be added at all. It's not important enough for a
ported video filter, but has high maintenance overhead - there's a
reason the FFmpeg loathes inline asm, right?

How much does the inline asm help with speed? When I tested it on a
reasonably modern CPU a while ago, there was barely any advantage.

> +av_cold void ff_eq_init_x86(EQContext *eq)
> +{
> +#if HAVE_MMX_INLINE
> +    int cpu_flags = av_get_cpu_flags();
> +
> +    if (cpu_flags & AV_CPU_FLAG_MMX) {
> +        eq->process = process_MMX;
> +    }
> +#endif
> +}