[Mesa-dev] [PATCH] i965/aa: fixing anti-aliasing bug for thinnest width lines - GEN6

[Ian Romanick]

On 03/17/2015 11:29 AM, Marius Predut wrote:
> On SNB and IVB hw, for 1 pixel line thickness or less, the general anti-aliasing algorithm give up - garbage line is generated.
> Setting a Line Width of 0.0 specifies the rasterization of the “thinnest” (one-pixel-wide), non-antialiased lines.
> Lines rendered with zero Line Width are rasterized using Grid Intersection Quantization rules as specified by bspec section 6.3.12.1 Zero-Width (Cosmetic) Line Rasterization.

You'll need to re-word wrap the commit message.  You'll get the commit
message right one of these days. :)

Also... when you send out new versions of a patch, please change the
patch subject to be something like "[PATCH v3] ...".  This makes it
easier for people to know which is the latest version to review.

You should also add notes to the commit message that explain what
changed from version to version.  For example, this commit message
should have something like:

v3: Fix = used instead of == in an if-statement.  Noticed by Daniel Stone.

This is helps people know that their review comments have been applied.
 It is also important to do this when the review changes are applied and
the patch committed without re-sending to the list.  Maintaining history
like this in commit messages helps us understand code in the future.

> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=28832

There are a number of bugs that have been closed as duplicates of this
bug.  Two of these, bug #27007 and bug #60797, have test cases.  Does
this fix also fix those?

I also have a more general question:  How are you testing this?  Daniel
noticed a bug in an earlier version of the patch that piglit should have
caught.  If you're doing a full piglit run and that didn't catch the
previous assignment-instead-of-comparison bug, it would be helpful if
you could craft a test that would detect that bug.  That may be
difficult, so don't spend a huge amount of time on it.

> Signed-off-by: Marius Predut <marius.predut at intel.com>
> ---
>  src/mesa/drivers/dri/i965/gen6_sf_state.c |   12 +++++++++---
>  1 file changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/src/mesa/drivers/dri/i965/gen6_sf_state.c b/src/mesa/drivers/dri/i965/gen6_sf_state.c
> index f9d8d27..1bed444 100644
> --- a/src/mesa/drivers/dri/i965/gen6_sf_state.c
> +++ b/src/mesa/drivers/dri/i965/gen6_sf_state.c
> @@ -367,9 +367,15 @@ upload_sf_state(struct brw_context *brw)
>        float line_width =
>           roundf(CLAMP(ctx->Line.Width, 0.0, ctx->Const.MaxLineWidth));
>        uint32_t line_width_u3_7 = U_FIXED(line_width, 7);
> -      /* TODO: line width of 0 is not allowed when MSAA enabled */
> -      if (line_width_u3_7 == 0)
> -         line_width_u3_7 = 1;
> +
> +      if (!(multisampled_fbo && ctx->Multisample.Enabled)) {

In Mesa there are often Enabled and _Enabled fields.  The Enabled field
is the setting made by the application via the OpenGL API.  The _Enabled
field means that feature in question is actually enabled, and this may
depend on other state.

In this case, the application may enable multisampling, but
multisampling may not occur of there is not a multisample buffer.  This
means gl_multisample_attrib::Enabled would be set but
gl_multisample_attrib::_Enabled would not.  Instead of open-coding that
check, just check gl_multisample_attrib::_Enabled directly:

      if (!ctx->Multisample._Enabled) {

I actually think it's more clear if you leave the original comment and
implement this as:

      /* Line width of 0 is not allowed when MSAA enabled */
      if (ctx->Multisample._Enabled) {
         if (line_width_u3_7 == 0)
             line_width_u3_7 = 1;
      } else if (ctx->Line.SmoothFlag && ctx->Line.Width <= 1.0) {
         /* For lines less than 1 pixel thick, the general
          * anti-aliasing algorithm gives up, and a garbage line is
          * generated.  Setting a Line Width of 0.0 specifies the
          * rasterization of the "thinnest" (one-pixel-wide),
          * non-antialiased lines.
          *
          * Lines rendered with zero Line Width are rasterized using
          * Grid Intersection Quantization rules as specified by
          * bspec section 6.3.12.1 Zero-Width (Cosmetic) Line
          * Rasterization.
          */
         line_width_u3_7 = 0;
      }

I reworded the comment a bit (from the commit message), and I changed
the Line.Width comparison to compare with 1.0.

One final question.  Does it produce better results to use 0 or 1.0?  It
sounds like using 1.0 would still enable line antialiasing, and the
resulting line shouldn't be appreciably thicker.

> +        if (ctx->Line.SmoothFlag && ctx->Line.Width <=1)
> +              line_width_u3_7 = 0;
> +      } else {
> +            if (line_width_u3_7 == 0)
> +                line_width_u3_7 = 1;
> +      }
> +
>        dw3 |= line_width_u3_7 << GEN6_SF_LINE_WIDTH_SHIFT;
>     }
>     if (ctx->Line.SmoothFlag) {