Update q_math.c Q_rsqrt 10x

[andreselcientifico]

Removing unnecessary variables increases speed 10x from 1000 ns to 100 ns.

[opkna]

Have you benchmarked this with optimizations turned on? Doubtful it will make a difference in the end.

[andreselcientifico]

Have you benchmarked this with optimizations turned on? Doubtful it will make a difference in the end.

Yes, I did a performance test, that's because the code in assembly is smaller, it also reduces the times in which it saves in memory and calls the memory again to do the other calculations, that makes it 10 times faster

[LegendaryGuard]

I did the tests in the online tools. I compared original vs new.
Original:
Compiler explorer (x86-64 gcc 14.2):

• With -std=c89: https://godbolt.org/z/zvYjcvrGj
• With -std=c89 -O3: https://godbolt.org/z/3e8MoYfW1

OnlineGDB:

• With -std=c89: https://onlinegdb.com/UGcMUdhgQW
• With -std=c89 -O3: https://onlinegdb.com/7CuqKpULJ5

New:
Compiler explorer (x86-64 gcc 14.2):

• With -std=c89: https://godbolt.org/z/bx6hK1f85
• With -std=c89 -O3: https://godbolt.org/z/dfx9bz8vj

OnlineGDB:

• With -std=c89: https://onlinegdb.com/_Yr7umoD-
• With -std=c89 -O3: https://onlinegdb.com/RyXe-gFSY

I don't see much difference using -std=c89 -O3 from both. Both have the same results.
But without -O3, the new one seems to have 0.002 seconds less than the original.
In compiler explorer, the new one has 34 assembly lines than the original which are 9 lines more (43 lines).
Compilers have optimization flags though, some projects don't have these flags set. I dunno what it could change that.

[andreselcientifico]

I did the tests in the online tools. I compared original vs new.
Original:
Compiler explorer (x86-64 gcc 14.2):

• With -std=c89: https://godbolt.org/z/zvYjcvrGj
• With std=c89 -O3: https://godbolt.org/z/3e8MoYfW1

OnlineGDB:

• With -std=c89: https://onlinegdb.com/UGcMUdhgQW
• With -std=c89 -O3: https://onlinegdb.com/7CuqKpULJ5

New:
Compiler explorer (x86-64 gcc 14.2):

• With -std=c89: https://godbolt.org/z/dfx9bz8vj
• With -std=c89 -O3: https://godbolt.org/z/dfx9bz8vj

OnlineGDB:

• With -std=c89: https://onlinegdb.com/_Yr7umoD-
• With -std=c89 -O3: https://onlinegdb.com/RyXe-gFSY

I don't see much difference using -std=c89 -O3 from both.
But without -O3, the new one seems to have 2 seconds less than the original.
Compilers have optimization flags though, some projects don't have these flags set. I dunno what it could change that.

In theory you can save the saving of 'y' in memory outside the game and place the return directly in the last function, that saves a step, and yes it is true some compilers already include optimization.

[LegendaryGuard]

I see, but it doesn't change anything with -std=c89 -O3 on compile explorer and OnlineGDB with that:

float Q_rsqrt( float number )
{
    long i;
    float x2;

    x2 = number * 0.5F;
    i  = 0x5f3759df - ( * ( long * ) &number >> 1 );
    
    return * ( float * ) &i * ( 1.5F - ( x2 * (* ( float * ) &i) * (* ( float * ) &i) ) );
}

Even with the very reduced version:

float Q_rsqrt( float number )
{
    float x2 = number * 0.5f;
    long i  = 0x5f3759df - ( *( long* )&number >> 1 );
    number  = *( float* )&i;
    return ( number * ( 1.5f - ( x2 * number * number ) ) );
}

Nothing changes. The result is the same.

[cultleader-f]

@andreselcientifico You're better off looking at _mm_rsqrt_ss (x86/x64) or vrsqrte_f32 (ARM), those provide a serious speed improvement and they are more precise

[hydrophobis]

Using an inline union seems to be much faster than normal, although it could just be the way that I'm timing them. these get about 9x faster than the source and the proposed changes in this PR when I run them

https://godbolt.org/z/TTPK8Yvc6
https://onlinegdb.com/OQt6Qwekn
And it runs about 3x faster on my local Win11 24H2 (0.018 vs 0.007)

[andreselcientifico]

El vie., 11 de julio de 2025 12:18, hydrophobis ***@***.***> escribió:

…

*hydrophobis* left a comment (id-Software/Quake-III-Arena#6) <#6 (comment)> Using an inline union seems to be much faster than normal, although it could just be the way that I'm timing them. these get about 9x faster than the source and the proposed changes in this PR when I run them https://godbolt.org/z/TTPK8Yvc6 https://onlinegdb.com/OQt6Qwekn And it runs about 3x faster on my local Win11 22H2 (0.018 vs 0.007) Screenshot.2025-07-11.121626.png (view on web) <https://github.com/user-attachments/assets/6f952c42-8a36-4a17-b5e3-48b418eba51b> Screenshot.2025-07-11.121635.png (view on web) <https://github.com/user-attachments/assets/6a228b3e-2b95-4c68-9490-fcbee385fac5> image.png (view on web) <https://github.com/user-attachments/assets/1ab3abf2-97a9-499f-8ec2-da131a7d41d4> — Reply to this email directly, view it on GitHub <#6 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AYNDVQLSE6PLO5VU27JT74T3H7WWBAVCNFSM6AAAAAB4ZXWXKWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTANRTGEZDAOJYG4> . You are receiving this because you were mentioned.Message ID: ***@***.***>