September 30, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to monarch_dodra | > I don't really agree with that. floating point operations are just inexact, regardless of optimizations. That's how they work, period.
It is true that they are inexact, but inexact and non-deterministic are not the same thing. Floating point operations are deterministic. Doing the same operation twice on the same machine with the same rounding mode and parameters will give you the same result. This should also be true when you do those operations using D, and using == on the two results should return true.
|
September 30, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to jerro | On Sunday, 30 September 2012 at 20:47:41 UTC, jerro wrote:
>> I don't really agree with that. floating point operations are just inexact, regardless of optimizations. That's how they work, period.
>
> It is true that they are inexact, but inexact and non-deterministic are not the same thing. Floating point operations are deterministic.
Technically (AFAIK), IEEE754 does need require reproducibility, ergo determinism.
You can open an ER requesting the ability to specify the FP behavior want (like rounding behavior), the way vstudio does it.
|
September 30, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to monarch_dodra | > Technically (AFAIK), IEEE754 does need require reproducibility, ergo determinism.
It actually requires more than that:
"Algebraic operations covered by IEEE 754, namely + , - , * , / , square root and Binary <-> Decimal Conversion with rare exceptions, must be Correctly Rounded to the precision of the operationâs destination unless the programmer has specified a rounding other than the default. If it does not Overflow, a correctly rounded operationâs error cannot exceed half the gap between adjacent floating-point numbers astride the operationâs ideal ( unrounded ) result. Half-way cases are rounded to Nearest Even, which means that the neighbor with last digit 0 is chosen."
I don't know if implementation conform exactly to IEEE 754, but I doubt there is any commonly used implementation that isn't deterministic.
|
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to jerro | On Sunday, 30 September 2012 at 06:20:56 UTC, jerro wrote:
>> The second one uses fstp twice, then fld twice. I don't know, maybe
>> this could be a bug.
>
> You're right the lack of one fst/fld in the first case is a bug. x87 floating point registers are 80 bit. This:
>
> fstp dword [ebp-0CH]
>
> Converts the value in ST0 to single precision float and stores it to memory (and pops ST0). When it is later loaded with fld, it is not the same as before storing since some precision is lost (because the D code compares floats and not reals, this is the correct behavior). In the first example, this storing and loading only happens for the first function call. For the second call the value is returned in ST0 and stays in x87 registers until it is compared with fucompp so it is not truncated as the result of the first function call was. That's why the compared values are not equal.
Can you or anyone report this bug? I don't know how to do this.
|
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to Ivan Agafonov | On Monday, 1 October 2012 at 04:10:25 UTC, Ivan Agafonov wrote: > On Sunday, 30 September 2012 at 06:20:56 UTC, jerro wrote: >>> The second one uses fstp twice, then fld twice. I don't know, maybe >>> this could be a bug. >> >> You're right the lack of one fst/fld in the first case is a bug. x87 floating point registers are 80 bit. This: >> >> fstp dword [ebp-0CH] >> >> Converts the value in ST0 to single precision float and stores it to memory (and pops ST0). When it is later loaded with fld, it is not the same as before storing since some precision is lost (because the D code compares floats and not reals, this is the correct behavior). In the first example, this storing and loading only happens for the first function call. For the second call the value is returned in ST0 and stays in x87 registers until it is compared with fucompp so it is not truncated as the result of the first function call was. That's why the compared values are not equal. > > Can you or anyone report this bug? I don't know how to do this. I'm not really sure that this is a bug anymore. Apparently c++ does this like this too (http://www.parashift.com/c++-faq/floating-point-arith2.html). I do think it would be more useful if the result of floating point comparison would be defined in a cases like this, but for floating point operation in general it makes sense to keep intermediate results in registers. So I don't know whether introducing a special case for this is worth it. If you care about this, you should probably make a thread on digitalmars.D, where more people will see it and comment on it. |
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to monarch_dodra | On Sunday, 30 September 2012 at 17:07:19 UTC, monarch_dodra wrote:
> As a rule of thumb, NEVER use opEqual with floating point types aniways. You need to use some sort of comparison with leway for error, such as std.math.approxEqual.
It is possible to compare exactly floating point types by binary comparison, if it provides some benefits.
import std.stdio;
import std.math;
@property float getFloat()
{
return sqrt(1.1);
}
void main()
{
writeln(getFloat is getFloat); // doesn't fail
}
|
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to Maxim Fomin | On Monday, 1 October 2012 at 11:36:43 UTC, Maxim Fomin wrote:
> On Sunday, 30 September 2012 at 17:07:19 UTC, monarch_dodra wrote:
>> As a rule of thumb, NEVER use opEqual with floating point types aniways. You need to use some sort of comparison with leway for error, such as std.math.approxEqual.
>
> It is possible to compare exactly floating point types by binary comparison, if it provides some benefits.
>
> import std.stdio;
> import std.math;
>
> @property float getFloat()
> {
> return sqrt(1.1);
> }
>
> void main()
> {
> writeln(getFloat is getFloat); // doesn't fail
> }
I think that what you are comparing here is the functions (the address), and not the results of the call. Try
writeln(getFloat() is getFloat()); //*May* fail
Also, "is" works like opEqual on built in types, AFAIK, it doesn't use any "binary" magic or anything like that.
|
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to monarch_dodra | 2012/10/1 monarch_dodra <monarchdodra@gmail.com>: > On Monday, 1 October 2012 at 11:36:43 UTC, Maxim Fomin wrote: >> >> On Sunday, 30 September 2012 at 17:07:19 UTC, monarch_dodra wrote: >>> >>> As a rule of thumb, NEVER use opEqual with floating point types aniways. You need to use some sort of comparison with leway for error, such as std.math.approxEqual. >> >> >> It is possible to compare exactly floating point types by binary comparison, if it provides some benefits. >> >> import std.stdio; >> import std.math; >> >> @property float getFloat() >> { >> return sqrt(1.1); >> } >> >> void main() >> { >> writeln(getFloat is getFloat); // doesn't fail >> } > > > I think that what you are comparing here is the functions (the address), and > not the results of the call. Try > writeln(getFloat() is getFloat()); //*May* fail > http://dpaste.dzfl.pl/1f94c0b1 It works with -m32 too. _Dmain: 0x0806d0e4 <+0>: push %ebp 0x0806d0e5 <+1>: mov %esp,%ebp 0x0806d0e7 <+3>: sub $0x10,%esp 0x0806d0ea <+6>: push %esi 0x0806d0eb <+7>: push %edi 0x0806d0ec <+8>: call 0x806d0d4 <_D4test8getFloatFNdZf> 0x0806d0f1 <+13>: fstps -0x10(%ebp) 0x0806d0f4 <+16>: lea -0x10(%ebp),%esi 0x0806d0f7 <+19>: call 0x806d0d4 <_D4test8getFloatFNdZf> 0x0806d0fc <+24>: fstps -0xc(%ebp) 0x0806d0ff <+27>: lea -0xc(%ebp),%edi 0x0806d102 <+30>: mov $0x4,%ecx 0x0806d107 <+35>: xor %eax,%eax 0x0806d109 <+37>: repz cmpsb %es:(%edi),%ds:(%esi) 0x0806d10b <+39>: je 0x806d112 <_Dmain+46> 0x0806d10d <+41>: sbb %eax,%eax 0x0806d10f <+43>: sbb $0xffffffff,%eax 0x0806d112 <+46>: neg %eax 0x0806d114 <+48>: sbb %eax,%eax 0x0806d116 <+50>: inc %eax 0x0806d117 <+51>: call 0x806d164 <_D3std5stdio14__T7writelnTbZ7writelnFbZv> 0x0806d11c <+56>: call 0x806d0d4 <_D4test8getFloatFNdZf> 0x0806d121 <+61>: fstps -0x8(%ebp) 0x0806d124 <+64>: lea -0x8(%ebp),%esi 0x0806d127 <+67>: call 0x806d0d4 <_D4test8getFloatFNdZf> 0x0806d12c <+72>: fstps -0x4(%ebp) 0x0806d12f <+75>: lea -0x4(%ebp),%edi 0x0806d132 <+78>: mov $0x4,%ecx 0x0806d137 <+83>: xor %eax,%eax 0x0806d139 <+85>: repz cmpsb %es:(%edi),%ds:(%esi) 0x0806d13b <+87>: je 0x806d142 <_Dmain+94> 0x0806d13d <+89>: sbb %eax,%eax 0x0806d13f <+91>: sbb $0xffffffff,%eax 0x0806d142 <+94>: neg %eax 0x0806d144 <+96>: sbb %eax,%eax 0x0806d146 <+98>: inc %eax 0x0806d147 <+99>: call 0x806d164 <_D3std5stdio14__T7writelnTbZ7writelnFbZv> 0x0806d14c <+104>: call 0x806d0d4 <_D4test8getFloatFNdZf> 0x0806d151 <+109>: sub $0x4,%esp 0x0806d154 <+112>: fstps (%esp) 0x0806d157 <+115>: call 0x806d588 <_D3std5stdio14__T7writelnTfZ7writelnFfZv> 0x0806d15c <+120>: xor %eax,%eax 0x0806d15e <+122>: pop %edi 0x0806d15f <+123>: pop %esi 0x0806d160 <+124>: leave 0x0806d161 <+125>: ret > Also, "is" works like opEqual on built in types, AFAIK, it doesn't use any "binary" magic or anything like that. I don't understand what you are trying to say. Is operator at runtime compares two objects without calling opEquals functions (if applied on user-defined types). For built-in and derived types it is similar to == operator. Although, I am suprised that TDPL and spec doesn't mention it (focused only on CT usage), there is a paragraph (http://ddili.org/ders/d.en/null_is.html) from Turkish D book which clearly shows such usage - so, I think this a valid D feature. Object comparison at low-level (repz cmpsb) means binary comparison. |
October 01, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to Maxim Fomin | On Monday, 1 October 2012 at 13:08:07 UTC, Maxim Fomin wrote: > 2012/10/1 monarch_dodra <monarchdodra@gmail.com>: >> On Monday, 1 October 2012 at 11:36:43 UTC, Maxim Fomin wrote: >>> >>> On Sunday, 30 September 2012 at 17:07:19 UTC, monarch_dodra wrote: >>>> >>>> As a rule of thumb, NEVER use opEqual with floating point types aniways. >>>> You need to use some sort of comparison with leway for error, such as >>>> std.math.approxEqual. >>> >>> >>> It is possible to compare exactly floating point types by binary >>> comparison, if it provides some benefits. >>> >>> import std.stdio; >>> import std.math; >>> >>> @property float getFloat() >>> { >>> return sqrt(1.1); >>> } >>> >>> void main() >>> { >>> writeln(getFloat is getFloat); // doesn't fail >>> } >> >> >> I think that what you are comparing here is the functions (the address), and >> not the results of the call. Try >> writeln(getFloat() is getFloat()); //*May* fail >> > > http://dpaste.dzfl.pl/1f94c0b1 > [SNIP] Hum, yes, I guess I was wrong about the comparison of functions. Sorry! >> Also, "is" works like opEqual on built in types, AFAIK, it doesn't use any >> "binary" magic or anything like that. > > I don't understand what you are trying to say. Is operator at runtime > compares two objects without calling opEquals functions (if applied on > user-defined types). For built-in and derived types it is similar to > == operator. Although, I am suprised that TDPL and spec doesn't > mention it (focused only on CT usage), there is a paragraph > (http://ddili.org/ders/d.en/null_is.html) from Turkish D book which > clearly shows such usage - so, I think this a valid D feature. Object > comparison at low-level (repz cmpsb) means binary comparison. What I was saying is that for built in types such a floats, "is" is (should be) no different from "==". But you catch something interesting: the fact that it provides different results is (IMO), a bug. Looking at it, I'd say the bug is probably that "==" is overly sensitive to extended precision. I've filed a BR: http://d.puremagic.com/issues/show_bug.cgi?id=8745 Please feel free to add anything to it. We'll see if Walter will react to it for a more definite answer. |
October 02, 2012 Re: Very strange problem with comparing floating point numbers | ||||
---|---|---|---|---|
| ||||
Posted in reply to monarch_dodra | On Monday, 1 October 2012 at 21:23:31 UTC, monarch_dodra wrote:
> What I was saying is that for built in types such a floats, "is" is (should be) no different from "==".
>
> But you catch something interesting: the fact that it provides different results is (IMO), a bug. Looking at it, I'd say the bug is probably that "==" is overly sensitive to extended precision.
>
> I've filed a BR:
> http://d.puremagic.com/issues/show_bug.cgi?id=8745
>
> Please feel free to add anything to it. We'll see if Walter will react to it for a more definite answer.
It looks like dmd uses x87 comparison instructions which are inexact comparing to is. So, similarity of is and == operators on built-in and user-defined types may be subject to float/double/real exception.
|
Copyright © 1999-2021 by the D Language Foundation