[Mesa-stable] [Mesa-dev] [PATCH] st/glsl_to_tgsi: Translate float ir_unop_neg into float MOV with modifier.

Mon Mar 6 22:45:15 UTC 2017

Am 06.03.2017 um 22:58 schrieb Francisco Jerez:
> Roland Scheidegger <sroland at vmware.com> writes:
> 
>> Am 04.03.2017 um 20:45 schrieb Ilia Mirkin:
>>> Also, how is this happening in the first place? For example, we have:
>>>
>>>    case ir_unop_bitcast_f2i:
>>>    case ir_unop_bitcast_f2u:
>>>       /* Make sure we don't propagate the negate modifier to integer opcodes. */
>>>       if (op[0].negate || op[0].abs)
>>>          emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
>>>       else
>>>          result_src = op[0];
>>>
>>> Oh, but it's going directly into a ir_triop_csel, which is missing
>>> this logic. It should be added there instead, IMHO. OTOH, the same
>>> issue will hit in emit_block_mov() if you do. Would love to hear some
>>> other opinions... Marek, Brian, Roland?
>>
>> float modifiers used with UCMP are actually just fine in theory - the
>> UCMP sources are considered floats for the purposes of modifiers (this
>> is similar to mov which also has untyped sources, except that 1 of the
>> arguments of ucmp indeed is a uint, which cannot have modifiers).
>> (tgsi_opcode_infer_type() says it's untyped, but
>> tgsi_opcode_infer_src_type() says it's float though obviously this
>> doesn't apply to the condition argument.)
>>
> 
> Yikes...  That sounds terribly inconsistent...  Would it make more sense
> to fix UCMP to behave like other integer instructions? (i.e. as the TGSI
> docs claim it works)
Slightly inconsistent or not, I very much like it how it is, and would
rather improve the docs and fix tgsi exec. (Because a) it is probably
more useful that way, since int modifiers are rare, limited to neg and
I'm not sure that is used at all and b) I don't see a good reason to
deviate from dx10 here.)
Yes the docs say "Integer conditional mov" but that's really just
because the condition is an integer (it belongs to the native integer
operations).
(For TGSI_OPCODE_MOV, there's actually some special mention that the
operands count as floats.)

Roland


> 
>> Therefore, not handling float modifiers with ucmp src args is a bug of
>> the tgsi executor (I'm quite sure llvmpipe will handle it right). Looks
>> like all source arguments must be the same type there... Seems annoying
>> to fix actually...
>> Albeit the gallium docs don't really mention this (this is the same
>> behavior as dx10 movc). I don't know though if other drivers will handle
>> it correctly, but if they do it might be a better idea to just fix
>> tgsi_exec somehow.
>>
>> (I'm not sure if the opposite could happen, that is int modifiers
>> mistakenly going into a ucmp, or if this could be a problem with other
>> instructions.)
>>
>> Roland
>>
>>
>>>   -ilia
>>>
>>>
>>> On Sat, Mar 4, 2017 at 2:24 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
>>>> Hmmm... I wonder if this should only be done for the native_integers
>>>> case. I'm concerned that this will cause perf regressions on weaker hw
>>>> like nv30 and r300, as the neg will no longer be inserted as a
>>>> modifier into the next instruction. Any opinion on this?
>>>>
>>>> On Sat, Mar 4, 2017 at 2:16 PM, Francisco Jerez <currojerez at riseup.net> wrote:
>>>>> Otherwise result_src may be provided to an integer instruction whose
>>>>> negate modifier has different semantics.  Example is UCMP as in the
>>>>> bug linked below, where an unrelated change in the GLSL built-in
>>>>> lowering code for atan2 (e9ffd12827ac11a2d2002a42fa8eb1df847153ba)
>>>>> caused the generation of floating-point ir_unop_neg instructions
>>>>> followed by ir_triop_csel, which is lowered into UCMP on back-ends
>>>>> with native integer support.
>>>>>
>>>>> Fixes a regression of piglit glsl-fs-tan-1 on softpipe introduced by
>>>>> the above-mentioned glsl front-end commit.  Even though the commit
>>>>> that triggered the regression doesn't seem to have made it to any
>>>>> stable branches yet, this seems worth back-porting since I don't see
>>>>> any reason why the bug couldn't have been reproduced before that
>>>>> point.
>>>>>
>>>>> Bugzilla: https://urldefense.proofpoint.com/v2/url?u=https-3A__bugs.freedesktop.org_show-5Fbug.cgi-3Fid-3D99817&d=DwIGaQ&c=uilaK90D4TOVoH58JNXRgQ&r=_QIjpv-UJ77xEQY8fIYoQtr5qv8wKrPJc7v7_-CYAb0&m=ak9IRpzPmkscMdh984JOu0Zn2I-efj9Iy3K-Iw9vv04&s=VpboBTuXbZaDL1-iEDED9JFLZTZFsPVqjTSbTuGSuT8&e= 
>>>>> Tested-by: Vinson Lee <vlee at freedesktop.org>
>>>>> Cc: mesa-stable at lists.freedesktop.org
>>>>> ---
>>>>>  src/mesa/state_tracker/st_glsl_to_tgsi.cpp | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>>>> index af41bdb..6bf3c89 100644
>>>>> --- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>>>> +++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
>>>>> @@ -1633,7 +1633,7 @@ glsl_to_tgsi_visitor::visit_expression(ir_expression* ir, st_src_reg *op)
>>>>>           emit_asm(ir, TGSI_OPCODE_DNEG, result_dst, op[0]);
>>>>>        else {
>>>>>           op[0].negate = ~op[0].negate;
>>>>> -         result_src = op[0];
>>>>> +         emit_asm(ir, TGSI_OPCODE_MOV, result_dst, op[0]);
>>>>>        }
>>>>>        break;
>>>>>     case ir_unop_subroutine_to_int:
>>>>> --
>>>>> 2.10.2
>>>>>
>>>>> _______________________________________________
>>>>> mesa-dev mailing list
>>>>> mesa-dev at lists.freedesktop.org
>>>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.freedesktop.org_mailman_listinfo_mesa-2Ddev&d=DwIGaQ&c=uilaK90D4TOVoH58JNXRgQ&r=_QIjpv-UJ77xEQY8fIYoQtr5qv8wKrPJc7v7_-CYAb0&m=ak9IRpzPmkscMdh984JOu0Zn2I-efj9Iy3K-Iw9vv04&s=zkVQuaiGqon4ebXPnKm96V6UnqAVXfGc_ky9Lm_FMKY&e= 
>>> _______________________________________________
>>> mesa-dev mailing list
>>> mesa-dev at lists.freedesktop.org
>>> https://urldefense.proofpoint.com/v2/url?u=https-3A__lists.freedesktop.org_mailman_listinfo_mesa-2Ddev&d=DwIGaQ&c=uilaK90D4TOVoH58JNXRgQ&r=_QIjpv-UJ77xEQY8fIYoQtr5qv8wKrPJc7v7_-CYAb0&m=ak9IRpzPmkscMdh984JOu0Zn2I-efj9Iy3K-Iw9vv04&s=zkVQuaiGqon4ebXPnKm96V6UnqAVXfGc_ky9Lm_FMKY&e= 
>>>