[FFmpeg-devel] [PATCH v4 3/4] libavcodec/vulkan: Add modifications to common shader for VC2 vulkan encoder

[Andreas Rheinhardt]

IndecisiveTurtle:
> From: IndecisiveTurtle <geoster3d at gmail.com>
> 
> ---
>  libavcodec/vulkan/common.comp | 54 ++++++++++++++++++++++++++++-------
>  1 file changed, 44 insertions(+), 10 deletions(-)
> 
> diff --git a/libavcodec/vulkan/common.comp b/libavcodec/vulkan/common.comp
> index 10af9c0623..db216a2ac6 100644
> --- a/libavcodec/vulkan/common.comp
> +++ b/libavcodec/vulkan/common.comp
> @@ -18,6 +18,9 @@
>   * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
>   */
>  
> +#extension GL_EXT_buffer_reference : require
> +#extension GL_EXT_buffer_reference2 : require
> +
>  layout(buffer_reference, buffer_reference_align = 1) buffer u8buf {
>      uint8_t v;
>  };
> @@ -61,22 +64,20 @@ layout(buffer_reference, buffer_reference_align = 8) buffer u64buf {
>  #define mid_pred(a, b, c) \
>      max(min((a), (b)), min(max((a), (b)), (c)))
>  
> -/* TODO: optimize */
> +
>  uint align(uint src, uint a)
>  {
> -    uint res = src % a;
> -    if (res == 0)
> -        return src;
> -    return src + a - res;
> +    return (src + a - 1) & ~(a - 1);
> +}
> +
> +int align(int src, int a)
> +{
> +    return (src + a - 1) & ~(a - 1);
>  }
>  
> -/* TODO: optimize */
>  uint64_t align64(uint64_t src, uint64_t a)
>  {
> -    uint64_t res = src % a;
> -    if (res == 0)
> -        return src;
> -    return src + a - res;
> +    return (src + a - 1) & ~(a - 1);
>  }
>  
>  #define reverse4(src) \
> @@ -167,6 +168,39 @@ uint32_t flush_put_bits(inout PutBitContext pb)
>      return uint32_t(pb.buf - pb.buf_start);
>  }
>  
> +void skip_put_bytes(inout PutBitContext pb, int n)
> +{
> +    int bytes_left = pb.bit_left >> 3;
> +    if (n < bytes_left)
> +    {
> +        int n_bits = n << 3;
> +        int mask = (1 << n_bits) - 1;
> +        pb.bit_buf <<= n_bits;
> +        pb.bit_buf |= mask;
> +        pb.bit_left -= uint8_t(n_bits);
> +        return;
> +    }
> +    if (pb.bit_left < BUF_BITS)
> +    {
> +        int mask = (1 << pb.bit_left) - 1;
> +        pb.bit_buf <<= pb.bit_left;
> +        pb.bit_buf |= mask;
> +        u32vec2buf(pb.buf).v = BUF_REVERSE(pb.bit_buf);
> +        pb.buf += BUF_BYTES;
> +        n -= pb.bit_left >> 3;
> +    }
> +    int skip_dwords = n >> 2;
> +    while (skip_dwords > 0)
> +    {
> +        u8vec4buf(pb.buf).v = u8vec4(0xFF);
> +        pb.buf += 4;
> +        skip_dwords--;
> +    }
> +    int skip_bits = (n & 3) << 3;
> +    pb.bit_buf = (1 << skip_bits) - 1;
> +    pb.bit_left = uint8_t(BUF_BITS - skip_bits);
> +}

This differs quite a lot from the software implementation: It does not
presume that the PutBitContext is flushed and instead of simply skipping
over the buffer it actually fills the buffer with n 0xFF bytes,
effectively adding the memset used in the VC2 slice writing code to
skip_put_bytes(). But this file is (if I am not mistaken) supposed to be
generic, not vc2 specific, so this feels very wrong.

> +
>  void init_put_bits(out PutBitContext pb, u8buf data, uint64_t len)
>  {
>      pb.buf_start = uint64_t(data);