[Mesa-dev] [PATCH] nv50/ir: saturate FRC result to avoid completely bogus values

[Ilia Mirkin]

On Tue, Nov 18, 2014 at 9:53 AM, Jose Fonseca <jfonseca at vmware.com> wrote:
> On 18/11/14 14:34, Roland Scheidegger wrote:
>>
>> Am 18.11.2014 um 15:05 schrieb Ilia Mirkin:
>>>
>>> On Tue, Nov 18, 2014 at 8:54 AM, Roland Scheidegger <sroland at vmware.com>
>>> wrote:
>>>>
>>>> Am 18.11.2014 um 05:03 schrieb Ilia Mirkin:
>>>>>
>>>>> For values above integer accuracy in floats, val - floor(val) might
>>>>> actually produce a value greater than 1. For such large floats, it's
>>>>> reasonable to be imprecise, but it's unreasonable for FRC to return a
>>>>> value that is not between 0 and 1.
>>>>>
>>>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>>>> ---
>>>>>   src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp | 3 ++-
>>>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
>>>>> b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
>>>>> index 41b91e8..e5b767f 100644
>>>>> --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
>>>>> +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_from_tgsi.cpp
>>>>> @@ -2512,7 +2512,8 @@ Converter::handleInstruction(const struct
>>>>> tgsi_full_instruction *insn)
>>>>>            src0 = fetchSrc(0, c);
>>>>>            val0 = getScratch();
>>>>>            mkOp1(OP_FLOOR, TYPE_F32, val0, src0);
>>>>> -         mkOp2(OP_SUB, TYPE_F32, dst0[c], src0, val0);
>>>>> +         mkOp2(OP_SUB, TYPE_F32, val0, src0, val0);
>>>>> +         mkOp1(OP_SAT, TYPE_F32, dst0[c], val0);
>>>>>         }
>>>>>         break;
>>>>>      case TGSI_OPCODE_ROUND:
>>>>>
>>>>
>>>> I don't understand the math behind this. For any such large number, as
>>>> far as I can tell floor(val) == val and hence the end result ought to be
>>>> zero. Or doesn't your floor work like that?
>>>
>>>
>>> I could be thinking about this backwards, but let's say that floats
>>> lose integer precision at 10.0. And I do floor(12.5)... normally this
>>> would be 12.0, but since that's not exactly representable, it might
>>> actually be 11.0. (Or would it be 11.9987? I didn't consider that
>>> possibility...) And then 12.5 - 11 = 1.5. Or am I thinking about this
>>> backwards? I guess ideally I'd do something along the lines of y = x -
>>> floor(x); return y - floor(y). That seems like it might be more
>>> accurate... not sure.
>>>
>> If your float is large enough that the next closest float is more than
>> 1.0 away, then that float would have been an exact integer, thus floor()
>> doing nothing.
>>
>> Roland
>
>
> Roland's right -- it takes less mantissa bits to represent an integer x,
> than a fractional number between x and x + 1
>
> The only case where `frac(x) = x - floor(x)` fails is when x is a negative
> denormal. It might give 1.0f instead of 0.0f, if the hardware is not setup
> to flush denormals to zero properly.

Very cool. This patch wasn't in response to an actual issue but rather
a theoretical one. Sounds like I got the theory a little wrong... time
to think more about floats :)

  -ilia