On 6/20/2025 10:21 AM, Niklas Haas wrote:
> From: Niklas Haas <git@haasn.dev>
> 
> Processes two channels in parallel, using 128-bit XMM registers.
> 
> In theory, we could go up to YMM registers to process 4 channels, but this is
> not a gain except for relatively high channel counts (e.g. 7.1), and also
> complicates the sample load/store operations considerably.
> 
> I decided to only add an AVX variant, since the C code is not substantially
> slower enough to justify a separate function just for ancient CPUs.
> ---
>   libavfilter/f_ebur128.c          |  15 ++--
>   libavfilter/f_ebur128.h          |  16 ++++
>   libavfilter/x86/Makefile         |   4 +
>   libavfilter/x86/f_ebur128.asm    | 141 +++++++++++++++++++++++++++++++
>   libavfilter/x86/f_ebur128_init.c |  35 ++++++++
>   5 files changed, 206 insertions(+), 5 deletions(-)
>   create mode 100644 libavfilter/x86/f_ebur128.asm
>   create mode 100644 libavfilter/x86/f_ebur128_init.c
> 
> diff --git a/libavfilter/f_ebur128.c b/libavfilter/f_ebur128.c
> index b9e210c05a..2d94cefce7 100644
> --- a/libavfilter/f_ebur128.c
> +++ b/libavfilter/f_ebur128.c
> @@ -579,6 +579,11 @@ static av_cold int init(AVFilterContext *ctx)
>       /* summary */
>       av_log(ctx, AV_LOG_VERBOSE, "EBU +%d scale\n", ebur128->meter);
>   
> +    ebur128->dsp.filter_channels = ff_ebur128_filter_channels_c;
> +#if ARCH_X86
> +    ff_ebur128_init_x86(&ebur128->dsp);
> +#endif
> +
>       return 0;
>   }
>   
> @@ -692,11 +697,11 @@ static int filter_frame(AVFilterLink *inlink, AVFrame *insamples)
>           MOVE_TO_NEXT_CACHED_ENTRY(400);
>           MOVE_TO_NEXT_CACHED_ENTRY(3000);
>   
> -        ff_ebur128_filter_channels_c(dsp, &samples[idx_insample * nb_channels],
> -                                     &ebur128->i400.cache[bin_id_400 * nb_channels],
> -                                     &ebur128->i3000.cache[bin_id_3000 * nb_channels],
> -                                     ebur128->i400.sum, ebur128->i3000.sum,
> -                                     nb_channels);
> +        dsp->filter_channels(dsp, &samples[idx_insample * nb_channels],
> +                             &ebur128->i400.cache[bin_id_400 * nb_channels],
> +                             &ebur128->i3000.cache[bin_id_3000 * nb_channels],
> +                             ebur128->i400.sum, ebur128->i3000.sum,
> +                             nb_channels);
>   
>   #define FIND_PEAK(global, sp, ptype) do {                        \
>       int ch;                                                      \
> diff --git a/libavfilter/f_ebur128.h b/libavfilter/f_ebur128.h
> index 7b8e876576..1889e28bdd 100644
> --- a/libavfilter/f_ebur128.h
> +++ b/libavfilter/f_ebur128.h
> @@ -22,6 +22,9 @@
>   #ifndef AVFILTER_F_EBUR128_H
>   #define AVFILTER_F_EBUR128_H
>   
> +#include <assert.h>
> +#include <stddef.h>
> +
>   typedef struct EBUR128Biquad {
>       double b0, b1, b2;
>       double a1, a2;
> @@ -35,8 +38,21 @@ typedef struct EBUR128DSPContext {
>       /* Cache of 3 samples for each channel */
>       double *y; /* after pre-filter */
>       double *z; /* after RLB-filter */
> +
> +    /* DSP functions */
> +    void (*filter_channels)(const struct EBUR128DSPContext *dsp,
> +                            const double *samples,
> +                            double *cache_400, double *cache_3000,
> +                            double *sum_400, double *sum_3000,
> +                            int nb_channels);
>   } EBUR128DSPContext;
>   
> +static_assert(offsetof(EBUR128DSPContext, pre) == 0,                   "struct layout mismatch");
> +static_assert(offsetof(EBUR128DSPContext, rlb) == 5  * sizeof(double), "struct layout mismatch");
> +static_assert(offsetof(EBUR128DSPContext, y)   == 10 * sizeof(double), "struct layout mismatch");
> +
> +void ff_ebur128_init_x86(EBUR128DSPContext *dsp);
> +
>   void ff_ebur128_filter_channels_c(const EBUR128DSPContext *, const double *,
>                                     double *, double *, double *, double *, int);
>   
> diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
> index 0d9a28a935..e5f0c55a5e 100644
> --- a/libavfilter/x86/Makefile
> +++ b/libavfilter/x86/Makefile
> @@ -7,6 +7,7 @@ OBJS-$(CONFIG_BLEND_FILTER)                  += x86/vf_blend_init.o
>   OBJS-$(CONFIG_BWDIF_FILTER)                  += x86/vf_bwdif_init.o
>   OBJS-$(CONFIG_COLORSPACE_FILTER)             += x86/colorspacedsp_init.o
>   OBJS-$(CONFIG_CONVOLUTION_FILTER)            += x86/vf_convolution_init.o
> +OBJS-$(CONFIG_EBUR128_FILTER)                += x86/f_ebur128_init.o
>   OBJS-$(CONFIG_EQ_FILTER)                     += x86/vf_eq_init.o
>   OBJS-$(CONFIG_FSPP_FILTER)                   += x86/vf_fspp_init.o
>   OBJS-$(CONFIG_GBLUR_FILTER)                  += x86/vf_gblur_init.o
> @@ -52,6 +53,9 @@ X86ASM-OBJS-$(CONFIG_BLEND_FILTER)           += x86/vf_blend.o
>   X86ASM-OBJS-$(CONFIG_BWDIF_FILTER)           += x86/vf_bwdif.o
>   X86ASM-OBJS-$(CONFIG_COLORSPACE_FILTER)      += x86/colorspacedsp.o
>   X86ASM-OBJS-$(CONFIG_CONVOLUTION_FILTER)     += x86/vf_convolution.o
> +ifdef ARCH_X86_64

nit: The way we do this usually is by adding this check to the asm file, 
to cover whatever is needed after the x86util.asm include.

Also, a checkasm test would be nice.