[FFmpeg-devel] [PATCH 1/2] avcodec/dnxhdenc: Move PutBitContext from ctx to stack

[Andreas Rheinhardt]

Vittorio Giovara:
> On Sat, Jun 8, 2024 at 9:45 PM Andreas Rheinhardt <
> andreas.rheinhardt at outlook.com> wrote:
> 
>> Andreas Rheinhardt:
>>> Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt at outlook.com>
>>> ---
>>>  libavcodec/dnxhdenc.c | 33 +++++++++++++++++----------------
>>>  1 file changed, 17 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c
>>> index 0cb25d7714..4760a2932c 100644
>>> --- a/libavcodec/dnxhdenc.c
>>> +++ b/libavcodec/dnxhdenc.c
>>> @@ -559,7 +559,7 @@ static int dnxhd_write_header(AVCodecContext *avctx,
>> uint8_t *buf)
>>>      return 0;
>>>  }
>>>
>>> -static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int
>> diff)
>>> +static av_always_inline void dnxhd_encode_dc(PutBitContext *pb,
>> DNXHDEncContext *ctx, int diff)
>>>  {
>>>      int nbits;
>>>      if (diff < 0) {
>>> @@ -568,19 +568,19 @@ static av_always_inline void
>> dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
>>>      } else {
>>>          nbits = av_log2_16bit(2 * diff);
>>>      }
>>> -    put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
>>> +    put_bits(pb, ctx->cid_table->dc_bits[nbits] + nbits,
>>>               (ctx->cid_table->dc_codes[nbits] << nbits) +
>>>               av_mod_uintp2(diff, nbits));
>>>  }
>>>
>>>  static av_always_inline
>>> -void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block,
>>> -                        int last_index, int n)
>>> +void dnxhd_encode_block(PutBitContext *pb, DNXHDEncContext *ctx,
>>> +                        int16_t *block, int last_index, int n)
>>>  {
>>>      int last_non_zero = 0;
>>>      int slevel, i, j;
>>>
>>> -    dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
>>> +    dnxhd_encode_dc(pb, ctx, block[0] - ctx->m.last_dc[n]);
>>>      ctx->m.last_dc[n] = block[0];
>>>
>>>      for (i = 1; i <= last_index; i++) {
>>> @@ -589,14 +589,14 @@ void dnxhd_encode_block(DNXHDEncContext *ctx,
>> int16_t *block,
>>>          if (slevel) {
>>>              int run_level = i - last_non_zero - 1;
>>>              int rlevel = slevel * (1 << 1) | !!run_level;
>>> -            put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel],
>> ctx->vlc_codes[rlevel]);
>>> +            put_bits(pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
>>>              if (run_level)
>>> -                put_bits(&ctx->m.pb, ctx->run_bits[run_level],
>>> +                put_bits(pb, ctx->run_bits[run_level],
>>>                           ctx->run_codes[run_level]);
>>>              last_non_zero = i;
>>>          }
>>>      }
>>> -    put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
>>> +    put_bits(pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB
>>>  }
>>>
>>>  static av_always_inline
>>> @@ -879,9 +879,10 @@ static int dnxhd_encode_thread(AVCodecContext
>> *avctx, void *arg,
>>>                                 int jobnr, int threadnr)
>>>  {
>>>      DNXHDEncContext *ctx = avctx->priv_data;
>>> +    PutBitContext pb0, *const pb = &pb0;
>>>      int mb_y = jobnr, mb_x;
>>>      ctx = ctx->thread[threadnr];
>>> -    init_put_bits(&ctx->m.pb, (uint8_t *)arg + ctx->data_offset +
>> ctx->slice_offs[jobnr],
>>> +    init_put_bits(pb, (uint8_t *)arg + ctx->data_offset +
>> ctx->slice_offs[jobnr],
>>>                    ctx->slice_size[jobnr]);
>>>
>>>      ctx->m.last_dc[0] =
>>> @@ -892,8 +893,8 @@ static int dnxhd_encode_thread(AVCodecContext
>> *avctx, void *arg,
>>>          int qscale = ctx->mb_qscale[mb];
>>>          int i;
>>>
>>> -        put_bits(&ctx->m.pb, 11, qscale);
>>> -        put_bits(&ctx->m.pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
>>> +        put_bits(pb, 11, qscale);
>>> +        put_bits(pb, 1, avctx->pix_fmt == AV_PIX_FMT_YUV444P10);
>>>
>>>          dnxhd_get_blocks(ctx, mb_x, mb_y);
>>>
>>> @@ -904,13 +905,13 @@ static int dnxhd_encode_thread(AVCodecContext
>> *avctx, void *arg,
>>>                                                   ctx->is_444 ? (((i >>
>> 1) % 3) < 1 ? 0 : 4): 4 & (2*i),
>>>                                                   qscale, &overflow);
>>>
>>> -            dnxhd_encode_block(ctx, block, last_index, n);
>>> +            dnxhd_encode_block(pb, ctx, block, last_index, n);
>>>          }
>>>      }
>>> -    if (put_bits_count(&ctx->m.pb) & 31)
>>> -        put_bits(&ctx->m.pb, 32 - (put_bits_count(&ctx->m.pb) & 31), 0);
>>> -    flush_put_bits(&ctx->m.pb);
>>> -    memset(put_bits_ptr(&ctx->m.pb), 0, put_bytes_left(&ctx->m.pb, 0));
>>> +    if (put_bits_count(pb) & 31)
>>> +        put_bits(pb, 32 - (put_bits_count(pb) & 31), 0);
>>> +    flush_put_bits(pb);
>>> +    memset(put_bits_ptr(pb), 0, put_bytes_left(pb, 0));
>>>      return 0;
>>>  }
>>>
>>
>> Will apply this patchset tomorrow unless there are objections.
>>
> 
> No objections, but what is the rationale here? Just reducing variables
> scope?

Avoiding usage of MpegEncContext (which is only supposed to be used for
quantization dsp functions). And putting transient variables on the
stack is always good practice (unless they are too big). It also has the
advantage that the compiler can better reason about them, because their
address does not escape (currently, the compiler has to presume that all
the dsp calls in dnxhd_get_blocks() can modify the PutBitContext).

- Andreas