[Mesa-dev] [PATCH 3/9] st/mesa: make user index buffers optional

[Brian Paul]

On 04/26/2012 12:54 PM, Jose Fonseca wrote:
> ----- Original Message -----
>> ---
>>   src/mesa/state_tracker/st_context.c    |    4 +++-
>>   src/mesa/state_tracker/st_context.h    |    1 +
>>   src/mesa/state_tracker/st_draw.c       |    5 +++++
>>   src/mesa/state_tracker/st_extensions.c |    4 ++++
>>   4 files changed, 13 insertions(+), 1 deletions(-)
>>
>> diff --git a/src/mesa/state_tracker/st_context.c
>> b/src/mesa/state_tracker/st_context.c
>> index 84aae81..164cc45 100644
>> --- a/src/mesa/state_tracker/st_context.c
>> +++ b/src/mesa/state_tracker/st_context.c
>> @@ -155,7 +155,9 @@ st_create_context_priv( struct gl_context *ctx,
>> struct pipe_context *pipe )
>>      st->dirty.mesa = ~0;
>>      st->dirty.st = ~0;
>>
>> -   st->uploader = u_upload_create(st->pipe, 65536, 4,
>> PIPE_BIND_VERTEX_BUFFER);
>> +   st->uploader = u_upload_create(st->pipe, 128 * 1024, 4,
>> +                                  PIPE_BIND_VERTEX_BUFFER |
>> +                                  PIPE_BIND_INDEX_BUFFER);
>
> Marek,
>
> Instead of lumping this into the same hardware buffer, I think it would be better to use two separate uploaders so that the driver can effectively do optimization based on PIPE_BIND_VERTEX_BUFFER or PIPE_BIND_INDEX_BUFFER.  A quick look on current drivers showed that they do look at these bind flags.
>
> Otherwise I don't see anything wrong with this series. It seems a nice cleanup/speedup.
>
> Brian's OOTO till Monday, so allow more time for him to comment.
>
> Also, once you updated the series, please provide it in a clonable git branch for testing.

I'm testing the gallium-userbuf branch now.  With piglit's draw-batch 
test there's a failing assertion in svga_resource_buffer_upload.c at 
line 557 (the buffer in question is in a mapped state).  I disabled 
the assertion but then the test fails (it passes on master).  I'll try 
to dig a big deeper.

-Brian