[FFmpeg-devel] [PATCH] lavc/aarch64: add pred16x16 10-bit functions

[Carl Eugen Hoyos]

Am Do., 8. Apr. 2021 um 21:10 Uhr schrieb Mikhail Nitenko <mnitenko at gmail.com>:
>
> here are the benchmarks https://0x1.st/kX.txt

Instead please add the relevant lines to the commit message.

> ---
>  libavcodec/aarch64/h264pred_init.c |  75 +++++++++++-------
>  libavcodec/aarch64/h264pred_neon.S | 123 +++++++++++++++++++++++++++++
>  2 files changed, 168 insertions(+), 30 deletions(-)
>
> diff --git a/libavcodec/aarch64/h264pred_init.c b/libavcodec/aarch64/h264pred_init.c
> index b144376f90..d74205c2de 100644
> --- a/libavcodec/aarch64/h264pred_init.c
> +++ b/libavcodec/aarch64/h264pred_init.c
> @@ -45,42 +45,57 @@ void ff_pred8x8_0lt_dc_neon(uint8_t *src, ptrdiff_t stride);
>  void ff_pred8x8_l00_dc_neon(uint8_t *src, ptrdiff_t stride);
>  void ff_pred8x8_0l0_dc_neon(uint8_t *src, ptrdiff_t stride);
>
> +void ff_pred16x16_128_dc_neon_10(uint8_t *src, ptrdiff_t stride);
> +void ff_pred16x16_top_dc_neon_10(uint8_t *src, ptrdiff_t stride);
> +void ff_pred16x16_left_dc_neon_10(uint8_t *src, ptrdiff_t stride);
> +void ff_pred16x16_dc_neon_10(uint8_t *src, ptrdiff_t stride);
> +void ff_pred16x16_hor_neon_10(uint8_t *src, ptrdiff_t stride);
> +void ff_pred16x16_vert_neon_10(uint8_t *src, ptrdiff_t stride);
> +
>  static av_cold void h264_pred_init_neon(H264PredContext *h, int codec_id,
>                                          const int bit_depth,
>                                          const int chroma_format_idc)
>  {
> -    const int high_depth = bit_depth > 8;
> -
> -    if (high_depth)
> -        return;
> -

> -    if (chroma_format_idc <= 1) {
> -        h->pred8x8[VERT_PRED8x8     ] = ff_pred8x8_vert_neon;
> -        h->pred8x8[HOR_PRED8x8      ] = ff_pred8x8_hor_neon;
> -        if (codec_id != AV_CODEC_ID_VP7 && codec_id != AV_CODEC_ID_VP8)
> -            h->pred8x8[PLANE_PRED8x8] = ff_pred8x8_plane_neon;
> -        h->pred8x8[DC_128_PRED8x8   ] = ff_pred8x8_128_dc_neon;
> -        if (codec_id != AV_CODEC_ID_RV40 && codec_id != AV_CODEC_ID_VP7 &&
> -            codec_id != AV_CODEC_ID_VP8) {
> -            h->pred8x8[DC_PRED8x8     ] = ff_pred8x8_dc_neon;
> -            h->pred8x8[LEFT_DC_PRED8x8] = ff_pred8x8_left_dc_neon;
> -            h->pred8x8[TOP_DC_PRED8x8 ] = ff_pred8x8_top_dc_neon;
> -            h->pred8x8[ALZHEIMER_DC_L0T_PRED8x8] = ff_pred8x8_l0t_dc_neon;
> -            h->pred8x8[ALZHEIMER_DC_0LT_PRED8x8] = ff_pred8x8_0lt_dc_neon;
> -            h->pred8x8[ALZHEIMER_DC_L00_PRED8x8] = ff_pred8x8_l00_dc_neon;
> -            h->pred8x8[ALZHEIMER_DC_0L0_PRED8x8] = ff_pred8x8_0l0_dc_neon;

> +    if (bit_depth == 8) {

> +        if (chroma_format_idc <= 1) {
> +            h->pred8x8[VERT_PRED8x8     ] = ff_pred8x8_vert_neon;
> +            h->pred8x8[HOR_PRED8x8      ] = ff_pred8x8_hor_neon;
> +            if (codec_id != AV_CODEC_ID_VP7 && codec_id != AV_CODEC_ID_VP8)
> +                h->pred8x8[PLANE_PRED8x8] = ff_pred8x8_plane_neon;
> +            h->pred8x8[DC_128_PRED8x8   ] = ff_pred8x8_128_dc_neon;
> +            if (codec_id != AV_CODEC_ID_RV40 && codec_id != AV_CODEC_ID_VP7 &&
> +                codec_id != AV_CODEC_ID_VP8) {
> +                h->pred8x8[DC_PRED8x8     ] = ff_pred8x8_dc_neon;
> +                h->pred8x8[LEFT_DC_PRED8x8] = ff_pred8x8_left_dc_neon;
> +                h->pred8x8[TOP_DC_PRED8x8 ] = ff_pred8x8_top_dc_neon;
> +                h->pred8x8[ALZHEIMER_DC_L0T_PRED8x8] = ff_pred8x8_l0t_dc_neon;
> +                h->pred8x8[ALZHEIMER_DC_0LT_PRED8x8] = ff_pred8x8_0lt_dc_neon;
> +                h->pred8x8[ALZHEIMER_DC_L00_PRED8x8] = ff_pred8x8_l00_dc_neon;
> +                h->pred8x8[ALZHEIMER_DC_0L0_PRED8x8] = ff_pred8x8_0l0_dc_neon;

Please do not re-indent these lines in the patch that adds the new functions,
feel free to send a separate patch for the re-indentation.

Carl Eugen