[FFmpeg-devel] [PATCH] avcodec/rv60dec: Avoid branch when decoding cbp16

[Andreas Rheinhardt]

Peter Ross:
> On Mon, May 19, 2025 at 12:06:02AM +0200, Andreas Rheinhardt wrote:
>> Patch attached.
>>
>> - Andreas
> 
>> From 02724d5792348bea618c049034dc0febf24a46ac Mon Sep 17 00:00:00 2001
>> From: Andreas Rheinhardt <andreas.rheinhardt at outlook.com>
>> Date: Sun, 18 May 2025 23:12:03 +0200
>> Subject: [PATCH] avcodec/rv60dec: Avoid branch when decoding cbp16
>>
>> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at outlook.com>
>> ---
>>  libavcodec/rv60dec.c | 11 ++++-------
>>  1 file changed, 4 insertions(+), 7 deletions(-)
>>
>> diff --git a/libavcodec/rv60dec.c b/libavcodec/rv60dec.c
>> index d704ae512c..2bbcb1d620 100644
>> --- a/libavcodec/rv60dec.c
>> +++ b/libavcodec/rv60dec.c
>> @@ -82,7 +82,7 @@ enum {
>>  };
>>  
>>  static const VLCElem * cbp8_vlc[7][4];
>> -static const VLCElem * cbp16_vlc[7][3][4];
>> +static const VLCElem * cbp16_vlc[7][4][4];
>>  
>>  typedef struct {
>>      const VLCElem * l0[2];
>> @@ -137,12 +137,12 @@ static av_cold void rv60_init_static_data(void)
>>  
>>      for (int i = 0; i < 7; i++)
>>          for (int j = 0; j < 4; j++)
>> -            cbp8_vlc[i][j] = gen_vlc(rv60_cbp8_lens[i][j], 64, &state);
>> +            cbp16_vlc[i][0][j] = cbp8_vlc[i][j] = gen_vlc(rv60_cbp8_lens[i][j], 64, &state);
>>  
>>      for (int i = 0; i < 7; i++)
>>          for (int j = 0; j < 3; j++)
>>              for (int k = 0; k < 4; k++)
>> -                cbp16_vlc[i][j][k] = gen_vlc(rv60_cbp16_lens[i][j][k], 64, &state);
>> +                cbp16_vlc[i][j + 1][k] = gen_vlc(rv60_cbp16_lens[i][j][k], 64, &state);
>>  
>>      build_coeff_vlc(rv60_intra_lens, intra_coeff_vlc, 5, &state);
>>      build_coeff_vlc(rv60_inter_lens, inter_coeff_vlc, 7, &state);
>> @@ -1650,10 +1650,7 @@ static int decode_super_cbp(GetBitContext * gb, const VLCElem * vlc[4])
>>  static int decode_cbp16(GetBitContext * gb, int subset, int qp)
>>  {
>>      int cb_set = rv60_qp_to_idx[qp];
>> -    if (!subset)
>> -        return decode_super_cbp(gb, cbp8_vlc[cb_set]);
>> -    else
>> -        return decode_super_cbp(gb, cbp16_vlc[cb_set][subset - 1]);
>> +    return decode_super_cbp(gb, cbp16_vlc[cb_set][subset]);
>>  }
>>  
>>  static int decode_cu_r(RV60Context * s, AVFrame * frame, ThreadContext * thread, GetBitContext * gb, int xpos, int ypos, int log_size, int qp, int sel_qp)
>> -- 
>> 2.45.2
> 
> Looks okay. What was the motivation for this change. Speed up; any numbers?
> 

I saw a branch that could be avoided. I don't think that this leads to
any measurable speedup.

- Andreas