[Spice-devel] [PATCH 12/15] worker: don't use weird RedCompressedBuf nbytes shifting

[Marc-André Lureau]

----- Original Message -----
> > 
> > From: Marc-André Lureau <marcandre.lureau at gmail.com>
> > 
> > ---
> >  server/display-channel.h |  3 +--
> >  server/red_worker.c      | 18 +++++++++---------
> >  2 files changed, 10 insertions(+), 11 deletions(-)
> > 
> > diff --git a/server/display-channel.h b/server/display-channel.h
> > index 7c62a62..f1f4e3a 100644
> > --- a/server/display-channel.h
> > +++ b/server/display-channel.h
> > @@ -70,10 +70,9 @@
> >  #define NUM_STREAMS 50
> >  #define NUM_SURFACES 10000
> >  
> > -#define RED_COMPRESS_BUF_SIZE (1024 * 64)
> >  typedef struct RedCompressBuf RedCompressBuf;
> >  struct RedCompressBuf {
> > -    uint32_t buf[RED_COMPRESS_BUF_SIZE / 4];
> > +    uint8_t buf[64 * 1024];
> >      RedCompressBuf *next;
> >      RedCompressBuf *send_next;
> >  };
> 
> Note that adding fields before buf could make buf not aligned to 4 bytes
> while we want it aligned to 4 bytes. Quic code for portability require it.
> On GCC we could force the alignment with an attribute.
> Another portable way to make it aligned would be

Note g_slice_alloc() is aligned by default. 
> 
>   union {
>     uint8_t bytes[RED_COMPRESS_BUF_SIZE];
>     uint32_t words[RED_COMPRESS_BUF_SIZE / 4];
>   } buf;
> 

looks good to me too.

> > diff --git a/server/red_worker.c b/server/red_worker.c
> > index cf72100..83e0fa0 100644
> > --- a/server/red_worker.c
> > +++ b/server/red_worker.c
> > @@ -3888,7 +3888,7 @@ static void glz_usr_free(GlzEncoderUsrContext *usr,
> > void *ptr)
> >      free(ptr);
> >  }
> >  
> > -static inline int encoder_usr_more_space(EncoderData *enc_data, uint32_t
> > **io_ptr)
> > +static int encoder_usr_more_space(EncoderData *enc_data, uint8_t **io_ptr)
> >  {
> 
> The declaration suggests io_ptr can be one byte aligned but this is false.
> A comment should be enough.
> 
> >      RedCompressBuf *buf;
> >  
> > @@ -3899,31 +3899,31 @@ static inline int
> > encoder_usr_more_space(EncoderData
> > *enc_data, uint32_t **io_pt
> >      enc_data->bufs_tail = buf;
> >      buf->send_next = NULL;
> >      *io_ptr = buf->buf;
> > -    return sizeof(buf->buf) >> 2;
> > +    return sizeof(buf->buf);
> >  }
> >  
> 
> to remove just the shift here we could use G_N_ELEMENTS macro.

Why not

> 
> >  static int quic_usr_more_space(QuicUsrContext *usr, uint32_t **io_ptr, int
> >  rows_completed)
> >  {
> >      EncoderData *usr_data = &(((QuicData *)usr)->data);
> > -    return encoder_usr_more_space(usr_data, io_ptr);
> > +    return encoder_usr_more_space(usr_data, (uint8_t **)io_ptr) /
> > sizeof(uint32_t);
> >  }
> >  
> >  static int lz_usr_more_space(LzUsrContext *usr, uint8_t **io_ptr)
> >  {
> >      EncoderData *usr_data = &(((LzData *)usr)->data);
> > -    return (encoder_usr_more_space(usr_data, (uint32_t **)io_ptr) << 2);
> > +    return encoder_usr_more_space(usr_data, io_ptr);
> >  }
> >  
> >  static int glz_usr_more_space(GlzEncoderUsrContext *usr, uint8_t **io_ptr)
> >  {
> >      EncoderData *usr_data = &(((GlzData *)usr)->data);
> > -    return (encoder_usr_more_space(usr_data, (uint32_t **)io_ptr) << 2);
> > +    return encoder_usr_more_space(usr_data, io_ptr);
> >  }
> >  
> >  static int jpeg_usr_more_space(JpegEncoderUsrContext *usr, uint8_t
> >  **io_ptr)
> >  {
> >      EncoderData *usr_data = &(((JpegData *)usr)->data);
> > -    return (encoder_usr_more_space(usr_data, (uint32_t **)io_ptr) << 2);
> > +    return encoder_usr_more_space(usr_data, io_ptr);
> >  }
> >  
> >  #ifdef USE_LZ4
> > @@ -3937,7 +3937,7 @@ static int lz4_usr_more_space(Lz4EncoderUsrContext
> > *usr, uint8_t **io_ptr)
> >  static int zlib_usr_more_space(ZlibEncoderUsrContext *usr, uint8_t
> >  **io_ptr)
> >  {
> >      EncoderData *usr_data = &(((ZlibData *)usr)->data);
> > -    return (encoder_usr_more_space(usr_data, (uint32_t **)io_ptr) << 2);
> > +    return encoder_usr_more_space(usr_data, io_ptr);
> >  }
> >  
> >  static inline int encoder_usr_more_lines(EncoderData *enc_data, uint8_t
> >  **lines)
> > @@ -4591,8 +4591,8 @@ static inline int
> > red_quic_compress_image(DisplayChannelClient *dcc, SpiceImage
> >          stride = -src->stride;
> >      }
> >      size = quic_encode(quic, type, src->x, src->y, NULL, 0, stride,
> > -                       quic_data->data.bufs_head->buf,
> > -                       sizeof(quic_data->data.bufs_head->buf) >> 2);
> > +                       (uint32_t*)quic_data->data.bufs_head->buf,
> > +                       sizeof(quic_data->data.bufs_head->buf) /
> > sizeof(uint32_t));
> >  
> 
> Here G_N_ELEMENTS could be used. With the union it would became
> 
>    size = quic_encode(quic, type, src->x, src->y, NULL, 0, stride,
>                       quic_data->data.bufs_head->buf.words,
>                       G_N_ELEMENTS(quic_data->data.bufs_head->buf.words);
> 
> 
> >      // the compressed buffer is bigger than the original data
> >      if ((size << 2) > (src->y * src->stride)) {
> > --
> > 2.4.3
> > 
> 
> Beside the alignment I think other stuff are just style so very
> personal opinions.
> 
> I won't suggest some definitive code, let's see other comments.
> 
> Frediano
> _______________________________________________
> Spice-devel mailing list
> Spice-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/spice-devel
>