[Mesa-dev] [PATCH] i965/gen6/gs: Handle case where a GS doesn't allocate VUE
Iago Toral
itoral at igalia.com
Wed Jun 20 12:19:26 UTC 2018
On Tue, 2018-06-19 at 17:06 +0300, Andrii Simiklit wrote:
> We can not use the VUE Dereference flags combination for EOT
> message under ILK and SNB because the threads are not initialized
> there with initial VUE handle unlike Pre-IL.
> So to avoid GPU hangs on SNB and ILK we need
> to avoid usage of the VUE Dereference flags combination.
> (Was tested only on SNB but according to the specification
> SNB Volume 2 Part 1: 1.6.5.3, 1.6.5.6
> the ILK must behave itself in the similar way)
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105399
>
> Signed-off-by: Andrii Simiklit <andrii.simiklit at globallogic.com>
> ---
> src/intel/compiler/gen6_gs_visitor.cpp | 56
> +++++++++++++++++++++++++++++-----
> 1 file changed, 49 insertions(+), 7 deletions(-)
>
> diff --git a/src/intel/compiler/gen6_gs_visitor.cpp
> b/src/intel/compiler/gen6_gs_visitor.cpp
> index 66c69fb..ac3ba55 100644
> --- a/src/intel/compiler/gen6_gs_visitor.cpp
> +++ b/src/intel/compiler/gen6_gs_visitor.cpp
> @@ -300,10 +300,11 @@ gen6_gs_visitor::emit_urb_write_opcode(bool
> complete, int base_mrf,
> /* Otherwise we always request to allocate a new VUE handle.
> If this is
> * the last write before the EOT message and the new handle
> never gets
> * used it will be dereferenced when we send the EOT message.
> This is
> - * necessary to avoid different setups for the EOT message
> (one for the
> + * necessary to avoid different setups (under Pre-IL only) for
> the EOT message (one for the
> * case when there is no output and another for the case when
> there is)
> * which would require to end the program with an
> IF/ELSE/ENDIF block,
> - * something we do not want.
> + * something we do not want.
> + * But for ILK and SNB we can not avoid the end the program
> with an IF/ELSE/ENDIF block.
> */
> inst = emit(GS_OPCODE_URB_WRITE_ALLOCATE);
> inst->urb_write_flags = BRW_URB_WRITE_COMPLETE;
> @@ -449,8 +450,11 @@ gen6_gs_visitor::emit_thread_end()
> if (prog->info.has_transform_feedback_varyings)
> xfb_write();
> }
> - emit(BRW_OPCODE_ENDIF);
> -
> + const bool common_eot_approach_can_be_used = (devinfo->gen < 5);
We don't implement GS before gen6, and I don't think there are plans
for it at this point, so I think we can just simplify the patch by
assuming that devinfo->gen is always going to be 6 here (later gens use
a different implementation of GS).
> + if(common_eot_approach_can_be_used)
> + {
> + emit(BRW_OPCODE_ENDIF);
> + }
> /* Finally, emit EOT message.
> *
> * In gen6 we need to end the thread differently depending on
> whether we have
> @@ -463,8 +467,32 @@ gen6_gs_visitor::emit_thread_end()
> * VUE handle every time we do a URB WRITE, even for the last
> vertex we emit.
> * With this we make sure that whether we have emitted at least
> one vertex
> * or none at all, we have to finish the thread without writing
> to the URB,
> - * which works for both cases by setting the COMPLETE and UNUSED
> flags in
> - * the EOT message.
> + * which works for both cases (but only under Pre-IL) by setting
> + * the COMPLETE and UNUSED flags in the EOT message.
> + *
> + * But under ILK or SNB we must not use combination COMPLETE and
> UNUSED
> + * because this combination could be used only for already
> allocated VUE.
> + * But unlike Pre-IL in the ILK and SNB
> + * the initial VUE is not passed to threads.
> + * This behaver mentioned in specification:
> + * SNB Volume 2 Part 1:
> + * "1.6.5.3 VUE Allocation (GS, CLIP) [DevIL]"
> + * "1.6.5.4 VUE Allocation (GS) [DevSNB+]"
> + * "The threads are not passed an initial handle.
> + * Instead, they request a first handle (if any)
> + * via the URB shared function’s FF_SYNC message (see Shared
> Functions).
> + * If additional handles are required,
> + * the URB_WRITE allocate mechanism (mentioned above) is
> used."
> + *
> + * So for ILK and for SNB we must use only UNUSED flag.
> + * This is accepteble combination according to:
> + * SNB Volume 4 Part 2:
> + * "2.4.2 Message Descriptor"
> + * "Table lists the valid and invalid combinations of
> + * the Complete, Used, Allocate and EOT bits"
> + * "Thread terminate non-write of URB"
> + * SNB Volume 2 Part 1:
> + * "1.6.5.6 Thread Termination"
> */
I am not sure why you conclude all this from the PRM. This is what I
see:
Section 1.6.5.5 VUE Dereference (GS) (vol2, part1) says:
"It is possible and legal for a thread to produce no output
or subsequently allocate a destination VUE that
was not required (e.g., the thread allocated ahead).
Therefore, there is a mechanism by which a thread can “give back”
(dereference) an a llocated VUE. This mechanism must be used if
the VUE is not written before the thread terminates. A kernel can
explicitly dereference a VUE by issuing a URB_WRITE message
(specifying the to-be-dereference handle) with the Complete
bit set and the Used bit clear."
This is explicitly saying that COMPLETE + UNUSED is a valid
combination, and one that is in fact created for this very purpose.
Nothing in that text states that this is Pre-ILK or that this is only
for thread pre-allocated VUEs alone.
Then in 2.4.2 Message Descriptor (vol4, part2), it says:
" Used:
If set, this signals that the URB entry(s) referenced by
the handle(s) are valid outputs of the thread. In
all likelihood this means that that entry(s) contains
complete & valid data to be subject to further
processing by the pipeline.
If clear, this signals that the URB entry(s) referenced by
the handle(s) are not valid outputs of the thread.
Use of this setting will result in the handle(s)
being immediately dereferenced by the owning FF unit.
This setting is to be used by GS or CLIP threads to
dereference handles it obtained (either in the initial
thread payload or subsequent allocation writebacks)
but subsequently determined were not required (e.g.,
the object was completely clipped out)."
Again, there is no mention of this being Pre-ILK only and on top of
that, the text explicitly states that this combination is used to
deference handles obtained either in the initial thread payload or
subsequent allocation writebacks.
And finally, it also says the following:
"Complete: (...)
Programming Notes:
The following message descriptor fields are only valid when
Complete is set: Used"
Which I understand means that 'Used' is only applicable when Complete
is set, or in other words, that the only possible combinations where
Used is accounted for are those in which we we also have Complete set.
So I am not sure why you understand that COMPLETE + UNUSED is pre-ILK
or for thread allocated handles only only. Could you provide a specific
pointer to the exact place in the documentation where such thing is
clearly stated?
Iago
> this->current_annotation = "gen6 thread end: EOT";
>
> @@ -480,8 +508,22 @@ gen6_gs_visitor::emit_thread_end()
> inst->urb_write_flags = BRW_URB_WRITE_COMPLETE |
> BRW_URB_WRITE_UNUSED;
> inst->base_mrf = base_mrf;
> inst->mlen = 1;
> -}
> +
> + if(!common_eot_approach_can_be_used)
> + {
> + emit(BRW_OPCODE_ELSE);
> +
> + this->current_annotation = "gen6 thread end: EOT";
> +
> + vec4_instruction *unused_urb_inst =
> emit(GS_OPCODE_THREAD_END);
> + unused_urb_inst->urb_write_flags = BRW_URB_WRITE_UNUSED;
> + unused_urb_inst->base_mrf = base_mrf;
> + unused_urb_inst->mlen = 1;
>
> + emit(BRW_OPCODE_ENDIF);
> + }
> +}
> +
> void
> gen6_gs_visitor::setup_payload()
> {
More information about the mesa-dev
mailing list