[Mesa-dev] [PATCH 2/3] i965/fs: Make emit_uniformize a no-op for immediates

Thu Oct 22 14:45:40 PDT 2015

On Tue, Oct 20, 2015 at 11:56 AM, Francisco Jerez <currojerez at riseup.net> wrote:
> Kristian Høgsberg <krh at bitplanet.net> writes:
>
>> On Tue, Oct 20, 2015 at 3:16 AM, Francisco Jerez <currojerez at riseup.net> wrote:
>>> Kristian Høgsberg <krh at bitplanet.net> writes:
>>>
>>>> On Mon, Oct 19, 2015 at 4:19 AM, Francisco Jerez <currojerez at riseup.net> wrote:
>>>>> Neil Roberts <neil at linux.intel.com> writes:
>>>>>
>>>>>> Just a thought, would it be better to move this check into the
>>>>>> eliminate_find_live_channel optimisation? That way it could catch
>>>>>> sources that become immediates through later optimisations. One problem
>>>>>> with this that I've seen before is that eliminating the
>>>>>> FIND_LIVE_CHANNEL doesn't cause the subsequent BROADCAST to be
>>>>>> eliminated because the copy propagation doesn't work.
>>>>>
>>>>> I believe in this particular case the BROADCAST instruction should
>>>>> already be eliminated (by opt_algebraic), because its first source is an
>>>>> immediate, so this optimization would in fact be redundant with
>>>>> opt_algebraic+constant propagation if it weren't because the latter is
>>>>> unable to handle scalar copies.
>>>>
>>>> Yes, is happens as you say: first the FIND_LIVE_CHANNEL gets
>>>> eliminated because it's outside control flow
>>>> (fs_visitor::eliminate_find_live_channel). Then the broadcast gets
>>>> reduced to a MOV in opt_algebraic because src0 is uniform (immediate
>>>> constant). The problem then is the dst of the MOV has stride 0 which
>>>> can_propagate_from() then bails on.
>>>>
>>>>> Another possibility that would likely
>>>>> be more general than this (because it would handle cases in which, as
>>>>> Neil said, the argument becomes an immediate only after optimization,
>>>>> and because it would also avoid the issue with liveness analysis
>>>>> extending the live ranges of scalar values incorrectly [1]) would be to
>>>>> extend the destination register of the BROADCAST instructions to a
>>>>> vector [2] and then add a case to the switch statement of
>>>>> try_constant_propagate() so that the immediate MOV resulting from
>>>>> opt_algebraic is propagated into surface instructions (see attachment).
>>>>
>>>> I wrote exactly this code to make constant propagate work, plus adding
>>>> the extra opcodes to the switch statement. It works and I could
>>>> certainly send that out after this, but
>>>>
>>>> 1) This doesn't mean we shouldn't do the if (src.file == IMM)
>>>> shortcut. It saves the compiler a bit of work in the very common case
>>>> of
>>>> non-indirect buffer access.
>>>>
>>>> 2) I'm not even sure it makes sense to extend copy-propagation to do
>>>> this (which is why I went back to just the IMM test). Anything that
>>>> would be an immediate at this point should be an immediate, if not
>>>> we're missing something in nir.
>>>>
>>> Still this doesn't address the root of the problem, which is that
>>> emit_uniformize() emits scalar code that the rest of the compiler is not
>>> able to handle properly.
>>
>> This is not a case of papering over the root cause, it's about not
>> creating the root cause in the first place. The output of
>> emit_uniformize() always ends up as either a surface or a sampler
>> index, which we only look at later in the generator. There are no
>> other cases where the result of emit_uniformize() might be part of an
>> expression that we can copy propagate or otherwise optimize. If the
>> input to emit_uniformize() isn't an immediate where it could be, nir
>> optimization needs fixing. So if we add these two lines to
>> emit_uniformize() to pass immediates straight through, we avoid
>> generating code that we have to extend the copy prop pass to handle.
>>
>
> Kristian, there are legitimate uses of emit_uniformize() in which the
> argument is not an immediate but still can be optimized out later on --
> E.g. for images it will frequently be a uniform register, or a
> non-constant expression calculated within uniform control flow (in which
> case the FIND_LIVE_CHANNEL instruction emitted here will be reduced to a
> constant MOV by eliminate_find_live_chaso nnel()).  In such cases we still
> want copy-propagation to kick in, but it wont because of the problem I
> was talking about with scalar writes.  Even if the instructions emitted
> by emit_uniformize() cannot be optimized out, liveness analysis will
> overestimate the live ranges of the temporaries used by
> emit_uniformize() for the same reason, potentially causing the register
> allocator to spill or run out of registers.

Yeah, fair point. I went and took a look at non-constant surface index
and sent out a new series that make that work better as well as ssbo
write optimizations. I took out the imm shortcut and put in the
generic optimizations. It makes non-const surface index a little
better, but as I wrote in the cover letter, it still doesn't work that
well as we don't propagate the surface index into the send
instructions.

> Anyway don't take me wrong, I'm not NAK-ing your patch or anything, I
> just have the feeling that by fixing this more generally you could've
> saved us [or most likely me ;)] work in the near future.

I may be saving you some time, but I've  been working on this for
longer than I expected ;-)

Kristian

>> Kristian
>>
>>> Re-implementing a special case of the
>>> optimization already done by opt_algebraic() might have helped in this
>>> specific case, but it won't help when the argument is of some other kind
>>> of uniform value which are also frequently encountered in practice and
>>> could also be copy-propagated, and it won't help avoid the liveness
>>> analysis bug [1] in cases where the argument is not an immediate (the
>>> bug is reported to break some thousands SSBO dEQP tests, although I
>>> don't know what fraction of them actually use non-constant indexing).
>>> The alternative solution I provided patches for (and you seem to have
>>> implemented yourself independently) would address all these issues at
>>> once.
>>>
>>>> Kristian
>>>>
>>>>>> I made this patch a while ago but I never posted it anywhere because
>>>>>> it's a of a kludge and it would probably be better to fix the copy
>>>>>> propagation:
>>>>>>
>>>>>> https://github.com/bpeel/mesa/commit/e4c3286075f891f466fb8558106d2aaa
>>>>>>
>>>>> Heh, yeah, I'd rather fix copy propagation instead, which I believe will
>>>>> become much easier with the use-def-chain analysis pass I'm working on.
>>>>>
>>>>>> Either way though I don't think it would do any harm to have Kristian's
>>>>>> patch as well even if we did improve elimintate_find_live_channel so it
>>>>>> is:
>>>>>>
>>>>>> Reviewed-by: Neil Roberts <neil at linux.intel.com>
>>>>>>
>>>>>> - Neil
>>>>>>
>>>>>> Kristian Høgsberg Kristensen <krh at bitplanet.net> writes:
>>>>>>
>>>>>>> An immdiate is already uniform so just return it up front. Without this,
>>>>>>> brw_fs_surface_builder ends up passing immediate surface indices through
>>>>>>> SHADER_OPCODE_BROADCAST. This writes to a stride 0 dst, which we can't
>>>>>>> constant propagate out of, and further, we don't constant propagate into
>>>>>>> the typed/untype read/write opcodes at all.  The end result is that all
>>>>>>> typed/untyped read/write/atomics end up as indirect sends.
>>>>>>>
>>>>>>> Code generation should always produce either an immediate or an actual
>>>>>>> indirect surface index, so we can fix this by just special casing
>>>>>>> immediates in emit_uniformize.
>>>>>>>
>>>>>>> Signed-off-by: Kristian Høgsberg Kristensen <krh at bitplanet.net>
>>>>>>> ---
>>>>>>>  src/mesa/drivers/dri/i965/brw_fs_builder.h | 3 +++
>>>>>>>  1 file changed, 3 insertions(+)
>>>>>>>
>>>>>>> diff --git a/src/mesa/drivers/dri/i965/brw_fs_builder.h b/src/mesa/drivers/dri/i965/brw_fs_builder.h
>>>>>>> index df10a9d..98ce71e 100644
>>>>>>> --- a/src/mesa/drivers/dri/i965/brw_fs_builder.h
>>>>>>> +++ b/src/mesa/drivers/dri/i965/brw_fs_builder.h
>>>>>>> @@ -394,6 +394,9 @@ namespace brw {
>>>>>>>           const dst_reg chan_index = component(vgrf(BRW_REGISTER_TYPE_UD), 0);
>>>>>>>           const dst_reg dst = component(vgrf(src.type), 0);
>>>>>>>
>>>>>>> +         if (src.file == IMM)
>>>>>>> +            return src;
>>>>>>> +
>>>>>>>           ubld.emit(SHADER_OPCODE_FIND_LIVE_CHANNEL, chan_index);
>>>>>>>           ubld.emit(SHADER_OPCODE_BROADCAST, dst, src, chan_index);
>>>>>>>
>>>>>>> --
>>>>>>> 2.6.2
>>>>>>>
>>>>>>> _______________________________________________
>>>>>>> mesa-dev mailing list
>>>>>>> mesa-dev at lists.freedesktop.org
>>>>>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>>>>> _______________________________________________
>>>>>> mesa-dev mailing list
>>>>>> mesa-dev at lists.freedesktop.org
>>>>>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>>>>
>>>>>
>>>>> [1] http://lists.freedesktop.org/archives/mesa-dev/2015-October/097085.html
>>>>> [2] http://lists.freedesktop.org/archives/mesa-dev/attachments/20151014/25dd38dc/attachment.patch
>>>>>