[FFmpeg-devel] [PATCH] avfilter/vf_transpose: rewrite for x86 SIMD

[James Almer]

On 8/15/2017 5:25 PM, Paul B Mahol wrote:
> From f25f0022fbc675affd65b95f097fa62e55788a37 Mon Sep 17 00:00:00 2001
> From: Paul B Mahol <onemda at gmail.com>
> Date: Tue, 15 Aug 2017 20:12:32 +0200
> Subject: [PATCH] avfilter/vf_transpose: rewrite for x86 SIMD
> 
> Transpose first in chunks of 8x8 blocks.
> 15% faster overall.
> ---
>  libavfilter/vf_transpose.c | 184 +++++++++++++++++++++++++++++++++++----------
>  1 file changed, 143 insertions(+), 41 deletions(-)
> 
> diff --git a/libavfilter/vf_transpose.c b/libavfilter/vf_transpose.c
> index 75b4dda41f..2fa751c925 100644
> --- a/libavfilter/vf_transpose.c
> +++ b/libavfilter/vf_transpose.c
> @@ -58,6 +58,12 @@ typedef struct TransContext {
>  
>      int passthrough;    ///< PassthroughType, landscape passthrough mode enabled
>      int dir;            ///< TransposeDir
> +
> +    void (*transpose_8x8)(uint8_t *src, ptrdiff_t src_linesize,
> +                          uint8_t *dst, ptrdiff_t dst_linesize);
> +    void (*transpose_block)(uint8_t *src, ptrdiff_t src_linesize,
> +                            uint8_t *dst, ptrdiff_t dst_linesize,
> +                            int w, int h);
>  } TransContext;
>  
>  static int query_formats(AVFilterContext *ctx)
> @@ -79,6 +85,109 @@ static int query_formats(AVFilterContext *ctx)
>      return ff_set_common_formats(ctx, pix_fmts);
>  }
>  
> +static av_always_inline void transpose_block_8_c(uint8_t *src, ptrdiff_t src_linesize,

Is always_inline needed? Shouldn't inline be enough for the 8x8 functions?

> +                                                 uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                 int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize, src++)
> +        for (x = 0; x < w; x++)
> +            dst[x] = src[x*src_linesize];
> +}
> +
> +static void transpose_8x8_8_c(uint8_t *src, ptrdiff_t src_linesize,
> +                              uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_8_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +
> +static av_always_inline void transpose_block_16_c(uint8_t *src, ptrdiff_t src_linesize,
> +                                                  uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                  int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize, src += 2)
> +        for (x = 0; x < w; x++)
> +            *((uint16_t *)(dst + 2*x)) = *((uint16_t *)(src + x*src_linesize));

Use local uint16_t* pointers instead of casting inside the loop. It will
probably make no difference for compilers but it helps readability.

Same for the cases below.

> +}
> +
> +static void transpose_8x8_16_c(uint8_t *src, ptrdiff_t src_linesize,
> +                               uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_16_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +
> +static av_always_inline void transpose_block_24_c(uint8_t *src, ptrdiff_t src_linesize,
> +                                                  uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                  int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize) {
> +        for (x = 0; x < w; x++) {
> +            int32_t v = AV_RB24(src + x*src_linesize + y*3);
> +            AV_WB24(dst + 3*x, v);
> +        }
> +    }
> +}
> +
> +static void transpose_8x8_24_c(uint8_t *src, ptrdiff_t src_linesize,
> +                               uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_24_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +
> +static av_always_inline void transpose_block_32_c(uint8_t *src, ptrdiff_t src_linesize,
> +                                                  uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                  int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize, src += 4) {
> +        for (x = 0; x < w; x++)
> +            *((uint32_t *)(dst + 4*x)) = *((uint32_t *)(src + x*src_linesize));
> +    }
> +}
> +
> +static void transpose_8x8_32_c(uint8_t *src, ptrdiff_t src_linesize,
> +                               uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_32_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +
> +static av_always_inline void transpose_block_48_c(uint8_t *src, ptrdiff_t src_linesize,
> +                                                  uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                  int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize, src += 6) {
> +        for (x = 0; x < w; x++) {
> +            int64_t v = AV_RB48(src + x*src_linesize);
> +            AV_WB48(dst + 6*x, v);
> +        }
> +    }
> +}
> +
> +static void transpose_8x8_48_c(uint8_t *src, ptrdiff_t src_linesize,
> +                               uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_48_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +
> +static av_always_inline void transpose_block_64_c(uint8_t *src, ptrdiff_t src_linesize,
> +                                                  uint8_t *dst, ptrdiff_t dst_linesize,
> +                                                  int w, int h)
> +{
> +    int x, y;
> +    for (y = 0; y < h; y++, dst += dst_linesize, src += 8)
> +        for (x = 0; x < w; x++)
> +            *((uint64_t *)(dst + 8*x)) = *((uint64_t *)(src + x*src_linesize));
> +}
> +
> +static void transpose_8x8_64_c(uint8_t *src, ptrdiff_t src_linesize,
> +                               uint8_t *dst, ptrdiff_t dst_linesize)
> +{
> +    transpose_block_64_c(src, src_linesize, dst, dst_linesize, 8, 8);
> +}
> +