[Mesa-stable] [Mesa-dev] [PATCH] i965: Select pipeline and emit state base address in Gen8+ HiZ ops.

[Nanley Chery]

On Thu, Mar 16, 2017 at 05:34:13PM -0700, Kenneth Graunke wrote:
> On Wednesday, March 8, 2017 10:27:20 AM PDT Nanley Chery wrote:
> > On Wed, Mar 08, 2017 at 10:07:12AM -0800, Nanley Chery wrote:
> > > On Wed, Mar 08, 2017 at 02:17:59AM -0800, Kenneth Graunke wrote:
> > > > On Thursday, March 2, 2017 4:36:08 PM PST Nanley Chery wrote:
> > > > > On Mon, Feb 06, 2017 at 03:55:49PM -0800, Kenneth Graunke wrote:
> > > > > > If a HiZ op is the first thing in the batch, we should make sure
> > > > > > to select the render pipeline and emit state base address before
> > > > > > proceeding.
> > > > > > 
> > > > > > I believe 3DSTATE_WM_HZ_OP creates 3DPRIMITIVEs internally, and
> > > > > > dispatching those on the GPGPU pipeline seems a bit sketchy.  I'm
> > > > > 
> > > > > Yes, it does seem like we currently allow HZ_OPs within a GPGPU
> > > > > pipeline. This patch should fix that problem.
> > > > > 
> > > > > > not actually sure that STATE_BASE_ADDRESS is necessary, as the
> > > > > > depth related commands use graphics addresses, not ones relative
> > > > > > to the base address...but we're likely to set it as part of the
> > > > > > next operation anyway, so we should just do it right away.
> > > > > > 

Why should we do it right away if it will happen later on? I don't see
why this part of the patch is necessary.

> > > > > 
> > > > > I agree, re-emitting STATE_BASE_ADDRESS doesn't seem necessary. I think
> > > > > we should drop this part of the patch and add it back in later if we get
> > > > > some data that it's necessary. Leaving it there may be distracting to
> > > > > some readers and the BDW PRM warns that it's an expensive command:
> > > > > 
> > > > > 	Execution of this command causes a full pipeline flush, thus its
> > > > > 	use should be minimized for higher performance.
> > > > 
> > > > I think it should be basically free, actually.  We track a boolean,
> > > > brw->batch.state_base_address_emitted, to avoid emitting it multiple
> > > > times per batch.
> > > > 
> > > > Let's say the first thing in a fresh batch is a HiZ op, followed by
> > > > normal drawing.  Previously, we'd do:
> > > > 
> > > >     1. HiZ op commands
> > > >     2. STATE_BASE_ADDRESS (triggered by normal rendering upload)
> > > >     3. rest of normal drawing commands
> > > > 
> > > > Now we'd do:
> > > > 
> > > >     1. STATE_BASE_ADDRESS (triggered by HiZ op)
> > > >     2. HiZ op commands
> > > >     3. normal drawing commands (second SBA is skipped)
> > > > 
> > > > In other words...we're just moving it a bit earlier.  I suppose there
> > > > could be a batch containing only HiZ ops, at which point we'd pay for
> > > > a single STATE_BASE_ADDRESS...but that seems really unlikely.
> > > > 
> > > 
> > > Sorry for not stating it up front, but the special case you've mentioned
> > > is exactly what I'd like not to hurt unnecessarily.
> > > 
> 
> Why?  We really think there are going to be batches with only
> 3DSTATE_WM_HZ_OP and no normal rendering or BLORP?  It sounds
> really hypothetical to me.
> 

I've commented on the performance implications of that snippet because
it is the only functional change I can see from emitting SBA. That
unfortunately seems to have distracted us from the more important
question found above. Sorry about that.

> > Correct me if I'm wrong, but after thinking about it some more, it seems
> > that performance wouldn't suffer by emitting the SBA since the pipeline
> > was already flushed at the end of the preceding batch. It may also
> > improve since the pipelined HiZ op will likely be followed by other
> > pipelined commands. I'm not totally confident in my understanding on
> > pipeline flushes by the way. Is this why you'd like to emit the SBA here?
> > I think it's fine to leave it if we expound on the rationale.
> 
> Performance is not a motivation for this patch.  Having the GPU do
> work without a pipeline selected or state base addresses in place seems
> potentially dangerous.  I was hoping it would help with GPU hangs.  I'm
> not certain that it does, and it might be safe to skip this, but it
> seems like a lot of mental gymnastics to prove that it's safe for very
> little upside.
> 

I was only referring to the portion of the patch that emits SBA.

> I think you're right, though - doing the non-pipelined commands at the
> top may actually be better than kicking off work, stalling, and kicking
> off more work.  *shrug*
> 
> > -Nanley
> > 
> > > > > > Cc: "17.0" <mesa-stable at lists.freedesktop.org>
> > > > > > Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> > > > > > ---
> > > > > >  src/mesa/drivers/dri/i965/gen8_depth_state.c | 3 +++
> > > > > >  1 file changed, 3 insertions(+)
> > > > > > 
> > > > > > diff --git a/src/mesa/drivers/dri/i965/gen8_depth_state.c b/src/mesa/drivers/dri/i965/gen8_depth_state.c
> > > > > > index a7e61354fd5..620b32df8bb 100644
> > > > > > --- a/src/mesa/drivers/dri/i965/gen8_depth_state.c
> > > > > > +++ b/src/mesa/drivers/dri/i965/gen8_depth_state.c
> > > > > > @@ -404,6 +404,9 @@ gen8_hiz_exec(struct brw_context *brw, struct intel_mipmap_tree *mt,
> > > > > >     if (op == BLORP_HIZ_OP_NONE)
> > > > > >        return;
> > > > > >  
> > > > > 
> > > > > It would be helpful if you included the rationale here as a code
> > > > > comment. Something like the first two sentences of your commit message
> > > > > should work.
> > > > 
> > > > I can do that.
> > > > 
> > > > > > +   brw_select_pipeline(brw, BRW_RENDER_PIPELINE);
> > > > > 
> > > > > According to Vol07 of the BDW+ PRMs,
> > > > > 
> > > > > 	The previously active pipeline needs to be flushed via the
> > > > > 	MI_FLUSH command immediately before switching to a different
> > > > > 	pipeline via use of the PIPELINE_SELECT command.
> > > > > 
> > > > > However it doesn't look like MI_FLUSH is present after HSW. So there
> > > > > shouldn't be any additional work to do here.
> > > > 
> > > > Flushes are definitely required when switching the pipeline, but I
> > > > believe that brw_emit_select_pipeline() does that work.
> > > > 
> > > > FWIW, MI_FLUSH was replaced by PIPE_CONTROL many generations ago.
> > > > I believe the validation team stopped testing MI_FLUSH on Sandybridge.
> > > > 
> > > 
> > > Thanks for letting me know about the MI_FLUSH replacement. Do you know
> > > what bits must be set to perform an equivalent operation? From the looks of
> > > it, brw_emit_select_pipeline() actually avoids emitting PIPE_CONTROL BDW+.
> > > 
> > > -Nanley
> 
> Not exactly clear.  There are tables of what caches were invalidated by
> MI_FLUSH on previous platforms, and we can set equivalent bits.  I
> thought brw_emit_mi_flush() tried to capture that, but it might not.
> 
> Topi's proposed patch (which you mentioned in your other email) is
> definitely relevant here as well.

Thanks for letting me know!