[FFmpeg-devel] [PATCHv2 1/5] lavu/float_dsp: add double-precision scalar product

[James Almer]

On 5/30/2024 4:06 PM, Rémi Denis-Courmont wrote:
> The function pointer is appended to the structure for backward binary
> compatibility. Fortunately, this is allocated by libavutil, not by the
> user, so increasing the structure size is safe.
> ---
>   libavutil/float_dsp.c | 12 ++++++++++++
>   libavutil/float_dsp.h | 31 ++++++++++++++++++++++++++++++-
>   2 files changed, 42 insertions(+), 1 deletion(-)
> 
> diff --git a/libavutil/float_dsp.c b/libavutil/float_dsp.c
> index e9fb023466..08bbc85e3e 100644
> --- a/libavutil/float_dsp.c
> +++ b/libavutil/float_dsp.c
> @@ -132,6 +132,17 @@ float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len)
>       return p;
>   }
>   
> +double ff_scalarproduct_double_c(const double *v1, const double *v2,
> +                                 size_t len)
> +{
> +    double p = 0.0;
> +
> +    for (size_t i = 0; i < len; i++)
> +        p += v1[i] * v2[i];
> +
> +    return p;
> +}
> +
>   av_cold AVFloatDSPContext *avpriv_float_dsp_alloc(int bit_exact)
>   {
>       AVFloatDSPContext *fdsp = av_mallocz(sizeof(AVFloatDSPContext));
> @@ -149,6 +160,7 @@ av_cold AVFloatDSPContext *avpriv_float_dsp_alloc(int bit_exact)
>       fdsp->vector_fmul_reverse = vector_fmul_reverse_c;
>       fdsp->butterflies_float = butterflies_float_c;
>       fdsp->scalarproduct_float = avpriv_scalarproduct_float_c;
> +    fdsp->scalarproduct_double = ff_scalarproduct_double_c;
>   
>   #if ARCH_AARCH64
>       ff_float_dsp_init_aarch64(fdsp);
> diff --git a/libavutil/float_dsp.h b/libavutil/float_dsp.h
> index 342a8715c5..5053aa240d 100644
> --- a/libavutil/float_dsp.h
> +++ b/libavutil/float_dsp.h
> @@ -19,6 +19,8 @@
>   #ifndef AVUTIL_FLOAT_DSP_H
>   #define AVUTIL_FLOAT_DSP_H
>   
> +#include <stddef.h>
> +
>   typedef struct AVFloatDSPContext {
>       /**
>        * Calculate the entry wise product of two vectors of floats and store the result in
> @@ -187,19 +189,46 @@ typedef struct AVFloatDSPContext {
>        */
>       void (*vector_dmul)(double *dst, const double *src0, const double *src1,
>                           int len);
> +
> +    /**
> +     * Calculate the scalar product of two vectors of doubles.
> +     *
> +     * @param v1  first vector
> +     * @param v2  second vector
> +     * @param len length of vectors
> +     *
> +     * @return inner product of the vectors
> +     */
> +    double (*scalarproduct_double)(const double *v1, const double *v2,
> +                                   size_t len);
>   } AVFloatDSPContext;
>   
>   /**
> - * Return the scalar product of two vectors.
> + * Return the scalar product of two vectors of floats.
>    *
>    * @param v1  first input vector
> + *            constraints: 32-byte aligned
>    * @param v2  first input vector
> + *            constraints: 32-byte aligned
>    * @param len number of elements
> + *            constraints: multiple of 16

Why are you adding this to the doxy for scalarproduct_float()? Those 
constrains are not correct for it. They are for scalarproduct_double() 
which you're adding now.

>    *
>    * @return sum of elementwise products
>    */
>   float avpriv_scalarproduct_float_c(const float *v1, const float *v2, int len);
>   
> +/**
> + * Return the scalar product of two vectors of doubles.
> + *
> + * @param v1  first input vector
> + * @param v2  first input vector
> + * @param len number of elements
> + *
> + * @return inner product of the vectors
> + */
> +double ff_scalarproduct_double_c(const double *v1, const double *v2,
> +                                 size_t len);
> +
>   void ff_float_dsp_init_aarch64(AVFloatDSPContext *fdsp);
>   void ff_float_dsp_init_arm(AVFloatDSPContext *fdsp);
>   void ff_float_dsp_init_ppc(AVFloatDSPContext *fdsp, int strict);