[Mesa-dev] [PATCH] vbo: Set FLUSH_UPDATE_CURRENT when setting vertex attibutes

[Kristian Høgsberg]

2010/9/14 Chia-I Wu <olvaffe at gmail.com>:
> 2010/9/14 Kristian Høgsberg <krh at bitplanet.net>:
>> 2010/9/14 Chia-I Wu <olvaffe at gmail.com>:
>>> 2010/9/14 Kristian Høgsberg <krh at bitplanet.net>:
>>>> 2010/9/13 keith whitwell <keith.whitwell at gmail.com>:
>>>>> Hey Kristian,
>>>>>
>>>>> The first question is whether this is necessary - from vague memory I
>>>>> have an idea that current attributes need not be updated by vertex
>>>>> buffer rendering - ie. it's optional/implementation-dependent.
>>>>>
>>>>> I assume you're concerned with the case where you have something like
>>>>>
>>>>>   // ctx->Current.Color is xyz
>>>>>
>>>>>   glDrawArrays();
>>>>>
>>>>>   // has ctx->Current.Color been updated??
>>>>>
>>>>> But assuming I'm wrong about that & we really do want to make
>>>>> DrawArrays set the current values, the patch looks good...
>>>>
>>>> No, what I'm seeing is that the code in question sets three generic
>>>> vertex attributes and then calls glDrawArrays().  The value of the
>>>> last attribute is not propagates into the shader.
>>>>
>>>> The problem is that the vertex array code keeps the values in
>>>> exec->vtx.vertex, but the implementation of glDrawArrays looks in
>>>> ctx->Current (that's what I assume, I didn't track that down).  When
>>>> the code hits a case where the size of an attribute is smaller that
>>>> what we're trying to set, it recomputes the layout of the
>>>> exec->vtx.vertex values and as a side effect copies the
>>>> exec->vtx.vertex values to ctx->Current.  Since we start out with
>>>> attrsz == 0 for all attributes, each new attribute will trigger this
>>>> recomputation and thus effectively flushes all previous values to
>>>> ctx->Current.  Which is why all but the last attribute make it to the
>>>> shader.
>>>>
>>>> Note that the ATTR macro is defined differently, depending on
>>>> FEATURE_beginend - the !FEATURE_beginend case sets the
>>>> FLUSH_UPDATE_CURRENT flag too.  I don't know why we wouldn't also set
>>>> it in the FEATURE_beginend case, not using begin/end in that case is
>>>> still an option.
>>> The way glColor4f is dispatched depends on whether it is GL or ES:
>>>
>>>  GL (with FEATURE_beginend): glColor4f -> neutral_Color4f -> vbo_Color4f
>>>  ES (w/o  FEATURE_beginend): glColor4f -> _es_Color4f -> vbo_Color4f
>>>
>>> In the former case, FLUSH_UPDATE_CURRENT should have been set by
>>> vbo_exec_BeginVertices which is called by neutral_Color4f.  In the latter case,
>>> the flag must be set in vbo_Color4f.  Could it be a bug some where in vtxfmt?
>>>
>>> One issue I noticed just last week is that the current scheme does not take
>>> into account "ES with FEATURE_beginend".  This happens with --enable-gles2
>>> build.  Since FEATURE_beginend is enabled in such build, vbo_Color4f does not
>>> set FLUSH_UPDATE_CURRENT.  Yet, an ES context does not use neutral_Color4f.
>>
>> And that's exactly the problem I have.  Setting FLUSH_UPDATE_CURRENT
>> in the ATTR macro makes sure that the FLUSH_CURRENT in
>> vbo_exec_DrawArrays (and other array draw funcs) ends up calling
>> vbo_exec_copy_to_current(), which then pushes the values into
>> ctx->Current before the draw call.  I don't see a problem with this
>> approach; in a begin/end, this flag is already set, so there's no
>> overhead, outside begin/end (ES2) it's required for correct behaviour.
> If this patch is to be applied, it makes sense to also stop setting
> FLUSH_UPDATE_CURRENT in vbo_exec_BeginVertices, and remove the
> !FEATURE_beginend case.
>
> But then, does it make sense to remove exec vtxfmt in main/vtxfmt.c and the
> related code (a big chunk!), and define ATTR to
>
>  #define ATTR(...) do {
>     struct vbo_exec_context *exec = &vbo_context(ctx)->exec;
>
>     if (!exec->ctx->Driver.NeedFlush)
>        vbo_exec_BeginVertices(ctx);
>
>     /* ... */
>  }
>
> From what I can tell, and I am be wrong, exec vtxfmt and PRE_LOOPBACK are to
> avoid any unnecessary operation in vtxfmt functions, even as cheap as setting
> FLUSH_UPDATE_CURRENT repeatedly.  If that cannot be avoided, there seems to be
> no need to have exec vtxfmt.

Yeah, it seems like we can just do that.  We can use the unlikely()
macro to reduce the overhead of the test.

Kristian