[Mesa-dev] [PATCH 10/18] i965: Speculatively flush the batch after transform feedback

[Chris Wilson]

On Tue, Jul 07, 2015 at 10:31:07AM -0700, Kenneth Graunke wrote:
> On Tuesday, July 07, 2015 04:46:22 PM Chris Wilson wrote:
> > On Tue, Jul 07, 2015 at 10:12:20AM +0100, Chris Wilson wrote:
> > > On Mon, Jul 06, 2015 at 09:05:18PM -0700, Kristian Høgsberg wrote:
> > > > On Mon, Jul 6, 2015 at 12:36 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> > > > > On Monday, July 06, 2015 11:33:15 AM Chris Wilson wrote:
> > > > >> Since the purpose of transform feedback tends to be for the client to
> > > > >> act upon the results to change the geometry in the scene, it is likely
> > > > >> that the client will soon be waiting upon the results. Flush the batch
> > > > >> early so that we don't build up a long queue of commands afterwards that
> > > > >> could delay the readback.
> > > > >> ---
> > > > >>  src/mesa/drivers/dri/i965/gen7_sol_state.c | 6 ++++++
> > > > >>  1 file changed, 6 insertions(+)
> > > > >>
> > > > >> diff --git a/src/mesa/drivers/dri/i965/gen7_sol_state.c b/src/mesa/drivers/dri/i965/gen7_sol_state.c
> > > > >> index 857ebe5..13dbe5b 100644
> > > > >> --- a/src/mesa/drivers/dri/i965/gen7_sol_state.c
> > > > >> +++ b/src/mesa/drivers/dri/i965/gen7_sol_state.c
> > > > >> @@ -494,6 +494,12 @@ gen7_end_transform_feedback(struct gl_context *ctx,
> > > > >>
> > > > >>     brw_batch_end(&brw->batch);
> > > > >>
> > > > >> +   /* We will likely want to read the results in the very near future, so
> > > > >> +    * push this primitive to hardware if it is currently idle.
> > > > >> +    */
> > > > >> +   if (!brw_batch_busy(&brw->batch))
> > > > >> +      brw_batch_flush(&brw->batch);
> > > > >> +
> > > > >>     /* EndTransformFeedback() means that we need to update the number of
> > > > >>      * vertices written.  Since it's only necessary if DrawTransformFeedback()
> > > > >>      * is called and it means mapping a buffer object, we delay computing it
> > > > >>
> > > > >
> > > > > We need some data to justify this change.
> > > > 
> > > > I think even the theory is not correct - transform feedback is
> > > > typically fed back into the GPU (as new geometry, eg) rather than
> > > > consumed by the CPU, and in that case the flush is not helpful. But at
> > > > the end of the day, data will tell.
> > > 
> > > How are they fed back? Can the xfb buffer be bound to the vertex buffer?
> > > (Genuine question! The only examples I've seen were for testing by the
> > > CPU.)
> 
> Yes, it can.  Just glBindBuffer() some buffers around.  Or, I suspect
> one could bind it as a texture buffer object or SSBO and then use a
> compute shader on the results.
> 
> With GL 4.x, the "avoid synchronizing with the CPU" mentality is a lot
> more prevalent, due to the advent of compute shaders.
> 
> > 
> > I've reviewed the code again, and gen7_end_transform_feedback() is always
> > followed by brw_compute_xfb_vertices_written (and a read of the sol
> > buffer) afaict, maybe not immediately but always before the next
> > transform feedback.
> 
> Sadly, yes.  We have a primitive count and we need a vertex count - so,
> a tiny bit of math.  Ideally, we would use the Gen7.5 MI_MATH+ feature
> to do this, eliminating the CPU-GPU synchronization point.
> 
> > Also afaict it is not possible to map the sol buffer directly into the
> > application.
> > -Chris
> 
> It definitely is - the application creates GL buffer objects and binds
> them for use with transform feedback.  They can certainly
> glMapBufferRange() those buffers.

The trouble I see is that the values stored currently are implementation
dependent and often reset. How is the application meant to use them
directly?

(Just trying to understand a bit better. If it is that the current
implementation is stalling when not required, then trying to speed
those stalls up really is just lipstick on a pig and irrelevant. The
patch was just trying to make a suggestion that feeding the gpu around
expected stall points works best with the current batch-level granularity
of our fences. Using intrabatch semaphores for the query objects seems a
more promising avenue than doing batch flushes anyway.)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre