What is the most efficient way of implement a rotation of ubyte[4] array?

By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]

TIA,
Tom;

Tom:

> What is the most efficient way of implement a rotation of ubyte[4] array?
> 
> By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]

Two versions, I have done no benchmarks so far:

import std.c.stdio: printf;

union Four {
    ubyte[4] a;
    uint u;
}

void showFour(Four f) {
    printf("f.u: %u\n", f.u);
    printf("f.a: [%d, %d, %d, %d]\n",
           cast(int)f.a[0], cast(int)f.a[1],
           cast(int)f.a[2], cast(int)f.a[3]);
}

void main() {
    Four f;
    f.a[] = [1, 2, 3, 4];
    showFour(f);
    f.u = (f.u << 8) | (f.u >> 24);
    showFour(f);
    printf("\n");

    // alternative
    f.a[] = [1, 2, 3, 4];
    uint u2 = f.u;
    showFour(f);
    printf("u2: %u\n", u2);
    asm {
        rol  u2, 8;
    }
    f.u = u2;
    showFour(f);
}

/*

dmd -O -release test.d

__Dmain comdat
        push    EBP
        mov EBP,ESP
        sub ESP,8
        push    4
        mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
        push    4
        push    3
        push    2
        push    1
        push    4
        mov dword ptr -8[EBP],0
        push    EAX
        call    near ptr __d_arrayliteralT
        add ESP,018h
        push    EAX
        lea EAX,-8[EBP]
        push    EAX
        call    near ptr _memcpy
        mov EAX,-8[EBP]
        call    near ptr _D4test8showFourFS4test4FourZv
        mov EAX,-8[EBP]
        mov ECX,-8[EBP]
        shl EAX,8        ; <=========
        shr ECX,018h
        or  EAX,ECX
        mov -8[EBP],EAX
        mov EAX,-8[EBP]
        call    near ptr _D4test8showFourFS4test4FourZv
        mov EAX,offset FLAT:_DATA[024h]
        push    EAX
        call    near ptr _printf
        mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
        push    4
        push    4
        push    3
        push    2
        push    1
        push    4
        push    EAX
        call    near ptr __d_arrayliteralT
        add ESP,018h
        push    EAX
        lea EAX,-8[EBP]
        push    EAX
        call    near ptr _memcpy
        mov EAX,-8[EBP]
        mov -4[EBP],EAX
        mov EAX,-8[EBP]
        call    near ptr _D4test8showFourFS4test4FourZv
        mov EAX,offset FLAT:_DATA[028h]
        push    dword ptr -4[EBP]
        push    EAX
        call    near ptr _printf
        add ESP,024h
        rol -4[EBP],8   ; <=========
        mov EAX,-4[EBP]
        mov -8[EBP],EAX
        mov EAX,-4[EBP]
        call    near ptr _D4test8showFourFS4test4FourZv
        mov ESP,EBP
        pop EBP
        ret
*/

In theory a C/C++/D compiler has to compile an expression like (x<< 8)|(x>>24) with a ROL instruction, in practice DMD doesn't do it. Months ago I have asked the two (four in X86) roll instructions to be added to the Phobos core intrinsics module, but I am not sure what Walter answered me.

Bye,
bearophile

On 03/09/2011 03:41 PM, Tom wrote:
> What is the most efficient way of implement a rotation of ubyte[4] array?
>
> By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]
>
> TIA,
> Tom;

I don't know of anything more efficient than:
    ubyte[4] bytes = [1,2,3,4];
    bytes = bytes[$-1] ~ bytes[0..$-1]; // Rotate left
    bytes = bytes[1..$] ~ bytes[0]; // Rotate right

Both static arrays and dynamic arrays (ubyte[] bytes = [1,2,3,4];) perform about the same between 1 and 10 milling rotations in either direction. I think a temporary array might be created for the rhs, and then have the values of the rhs array copied to the lhs array, but I don't know. With static arrays, I'm not sure there would be a way to get around it with out at least a temporary value for the one that's moving between the first and last positions.

El 09/03/2011 20:25, bearophile escribió:
> Tom:
>
>> What is the most efficient way of implement a rotation of ubyte[4] array?
>>
>> By rotation I mean: rotateRight([1, 2, 3, 4]) ->  [4, 1, 2, 3]
>
> Two versions, I have done no benchmarks so far:
>
> import std.c.stdio: printf;
>
> union Four {
>      ubyte[4] a;
>      uint u;
> }
>
> void showFour(Four f) {
>      printf("f.u: %u\n", f.u);
>      printf("f.a: [%d, %d, %d, %d]\n",
>             cast(int)f.a[0], cast(int)f.a[1],
>             cast(int)f.a[2], cast(int)f.a[3]);
> }
>
> void main() {
>      Four f;
>      f.a[] = [1, 2, 3, 4];
>      showFour(f);
>      f.u = (f.u<<  8) | (f.u>>  24);
>      showFour(f);
>      printf("\n");
>
>      // alternative
>      f.a[] = [1, 2, 3, 4];
>      uint u2 = f.u;
>      showFour(f);
>      printf("u2: %u\n", u2);
>      asm {
>          rol  u2, 8;
>      }
>      f.u = u2;
>      showFour(f);
> }
>
> /*
>
> dmd -O -release test.d
>
> __Dmain comdat
>          push    EBP
>          mov EBP,ESP
>          sub ESP,8
>          push    4
>          mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
>          push    4
>          push    3
>          push    2
>          push    1
>          push    4
>          mov dword ptr -8[EBP],0
>          push    EAX
>          call    near ptr __d_arrayliteralT
>          add ESP,018h
>          push    EAX
>          lea EAX,-8[EBP]
>          push    EAX
>          call    near ptr _memcpy
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,-8[EBP]
>          mov ECX,-8[EBP]
>          shl EAX,8        ;<=========
>          shr ECX,018h
>          or  EAX,ECX
>          mov -8[EBP],EAX
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,offset FLAT:_DATA[024h]
>          push    EAX
>          call    near ptr _printf
>          mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
>          push    4
>          push    4
>          push    3
>          push    2
>          push    1
>          push    4
>          push    EAX
>          call    near ptr __d_arrayliteralT
>          add ESP,018h
>          push    EAX
>          lea EAX,-8[EBP]
>          push    EAX
>          call    near ptr _memcpy
>          mov EAX,-8[EBP]
>          mov -4[EBP],EAX
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,offset FLAT:_DATA[028h]
>          push    dword ptr -4[EBP]
>          push    EAX
>          call    near ptr _printf
>          add ESP,024h
>          rol -4[EBP],8   ;<=========
>          mov EAX,-4[EBP]
>          mov -8[EBP],EAX
>          mov EAX,-4[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov ESP,EBP
>          pop EBP
>          ret
> */
>
> In theory a C/C++/D compiler has to compile an expression like (x<<  8)|(x>>24) with a ROL instruction, in practice DMD doesn't do it. Months ago I have asked the two (four in X86) roll instructions to be added to the Phobos core intrinsics module, but I am not sure what Walter answered me.
>
> Bye,
> bearophile

Wow, thanks b, didn't know of the rol instruction...

Tom;

On 03/09/2011 04:25 PM, bearophile wrote:
> Tom:
>
>> What is the most efficient way of implement a rotation of ubyte[4] array?
>>
>> By rotation I mean: rotateRight([1, 2, 3, 4]) ->  [4, 1, 2, 3]
>
> Two versions, I have done no benchmarks so far:
>
> import std.c.stdio: printf;
>
> union Four {
>      ubyte[4] a;
>      uint u;
> }
>
> void showFour(Four f) {
>      printf("f.u: %u\n", f.u);
>      printf("f.a: [%d, %d, %d, %d]\n",
>             cast(int)f.a[0], cast(int)f.a[1],
>             cast(int)f.a[2], cast(int)f.a[3]);
> }
>
> void main() {
>      Four f;
>      f.a[] = [1, 2, 3, 4];
>      showFour(f);
>      f.u = (f.u<<  8) | (f.u>>  24);
>      showFour(f);
>      printf("\n");
>
>      // alternative
>      f.a[] = [1, 2, 3, 4];
>      uint u2 = f.u;
>      showFour(f);
>      printf("u2: %u\n", u2);
>      asm {
>          rol  u2, 8;
>      }
>      f.u = u2;
>      showFour(f);
> }
>
> /*
>
> dmd -O -release test.d
>
> __Dmain comdat
>          push    EBP
>          mov EBP,ESP
>          sub ESP,8
>          push    4
>          mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
>          push    4
>          push    3
>          push    2
>          push    1
>          push    4
>          mov dword ptr -8[EBP],0
>          push    EAX
>          call    near ptr __d_arrayliteralT
>          add ESP,018h
>          push    EAX
>          lea EAX,-8[EBP]
>          push    EAX
>          call    near ptr _memcpy
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,-8[EBP]
>          mov ECX,-8[EBP]
>          shl EAX,8        ;<=========
>          shr ECX,018h
>          or  EAX,ECX
>          mov -8[EBP],EAX
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,offset FLAT:_DATA[024h]
>          push    EAX
>          call    near ptr _printf
>          mov EAX,offset FLAT:_D12TypeInfo_xAh6__initZ
>          push    4
>          push    4
>          push    3
>          push    2
>          push    1
>          push    4
>          push    EAX
>          call    near ptr __d_arrayliteralT
>          add ESP,018h
>          push    EAX
>          lea EAX,-8[EBP]
>          push    EAX
>          call    near ptr _memcpy
>          mov EAX,-8[EBP]
>          mov -4[EBP],EAX
>          mov EAX,-8[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov EAX,offset FLAT:_DATA[028h]
>          push    dword ptr -4[EBP]
>          push    EAX
>          call    near ptr _printf
>          add ESP,024h
>          rol -4[EBP],8   ;<=========
>          mov EAX,-4[EBP]
>          mov -8[EBP],EAX
>          mov EAX,-4[EBP]
>          call    near ptr _D4test8showFourFS4test4FourZv
>          mov ESP,EBP
>          pop EBP
>          ret
> */
>
> In theory a C/C++/D compiler has to compile an expression like (x<<  8)|(x>>24) with a ROL instruction, in practice DMD doesn't do it. Months ago I have asked the two (four in X86) roll instructions to be added to the Phobos core intrinsics module, but I am not sure what Walter answered me.
>
> Bye,
> bearophile

I love it.
I've done a little benchmark that just repeats the rotate left a certain number of times, and then a rotate right a certain number of times. It looks like shifting ( << | >> ) is faster than the process of copying the uint value, shifting it, and copying it back. If I move the assignment for the rol outside of the for loop, the rol is about twice as fast.
http://dl.dropbox.com/u/12135920/rotate.d

Both are anywhere from 30 to 80 times faster than the slicing method I proposed (also included in the rotate.d file).


------
Rotating static array left to right 5000000 times
Finished in    1971.46 milliseconds
Rotating static array right to left 5000000 times
Finished in    1987.60 milliseconds
Rotating dynamic array left to right 5000000 times
Finished in    1932.40 milliseconds
Rotating dynamic array right to left 5000000 times
Finished in    1981.71 milliseconds
Shifting Union left to right 5000000 times
Finished in      33.46 milliseconds
Shifting Union right to left 5000000 times
Finished in      34.26 milliseconds
Rolling Union left to right 5000000 times
Finished in      67.51 milliseconds
Rolling Union right to left 5000000 times
Finished in      67.47 milliseconds
Rolling Union left to right 5000000 times with assignment with temporary variable outside of the loop
Finished in      28.81 milliseconds
Rolling Union right to left 5000000 times with assignment with temporary variable outside of the loop
Finished in      25.57 milliseconds

On Wednesday, March 09, 2011 15:35:29 Kai Meyer wrote:
> On 03/09/2011 03:41 PM, Tom wrote:
> > What is the most efficient way of implement a rotation of ubyte[4] array?
> > 
> > By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]
> > 
> > TIA,
> > Tom;
> 
> I don't know of anything more efficient than:
>      ubyte[4] bytes = [1,2,3,4];
>      bytes = bytes[$-1] ~ bytes[0..$-1]; // Rotate left

I'm stunned that this works. I'd even consider reporting it as a bug. You're concatenating a ubyte[] ont a ubyte...

>      bytes = bytes[1..$] ~ bytes[0]; // Rotate right

You're concatenating a ubyte onto a slice of the array (so it's ubyte[] instead of ubyte[4]). That will result in a temporary whose value will then be assigned to the original ubyte[4].

> Both static arrays and dynamic arrays (ubyte[] bytes = [1,2,3,4];) perform about the same between 1 and 10 milling rotations in either direction. I think a temporary array might be created for the rhs, and then have the values of the rhs array copied to the lhs array, but I don't know. With static arrays, I'm not sure there would be a way to get around it with out at least a temporary value for the one that's moving between the first and last positions.

Honestly, given that this is 4 ubytes, I would fully expect that the fastest way to do this would involve casting it to a unit and shifting it - something along the lines of what Bearophile suggested. I'd be _very_ suprised if this implementation were faster, since it involves creating a temporary array.

- Jonathan M Davis

On 03/10/2011 12:55 AM, Jonathan M Davis wrote:
>> I don't know of anything more efficient than:
>> >        ubyte[4] bytes = [1,2,3,4];
>> >        bytes = bytes[$-1] ~ bytes[0..$-1]; // Rotate left
> I'm stunned that this works. I'd even consider reporting it as a bug. You're
> concatenating a ubyte[] ont a ubyte...

This works for other arrays as well. dmd understands.

Denis
-- 
_________________
vita es estrany
spir.wikidot.com

== Quote from bearophile (bearophileHUGS@lycos.com)'s article
> Tom:
> > What is the most efficient way of implement a rotation of ubyte[4] array?
> >
> > By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]
> Two versions, I have done no benchmarks so far:
> import std.c.stdio: printf;
> union Four {
>     ubyte[4] a;
>     uint u;
> }
> void showFour(Four f) {
>     printf("f.u: %u\n", f.u);
>     printf("f.a: [%d, %d, %d, %d]\n",
>            cast(int)f.a[0], cast(int)f.a[1],
>            cast(int)f.a[2], cast(int)f.a[3]);
> }
> void main() {
>     Four f;
>     f.a[] = [1, 2, 3, 4];
>     showFour(f);
>     f.u = (f.u << 8) | (f.u >> 24);
>     showFour(f);
>     printf("\n");
>     // alternative
>     f.a[] = [1, 2, 3, 4];
>     uint u2 = f.u;
>     showFour(f);
>     printf("u2: %u\n", u2);
>     asm {
>         rol  u2, 8;
>     }
>     f.u = u2;
>     showFour(f);
> }
> Bye,
> bearophile

I am offend!

On Wed, Mar 9, 2011 at 7:25 PM, U2 fan <iam@u2fan.com> wrote:
> == Quote from bearophile (bearophileHUGS@lycos.com)'s article
>> Tom:
>> > What is the most efficient way of implement a rotation of ubyte[4] array?
>> >
>> > By rotation I mean: rotateRight([1, 2, 3, 4]) -> [4, 1, 2, 3]
>> Two versions, I have done no benchmarks so far:
>> import std.c.stdio: printf;
>> union Four {
>>     ubyte[4] a;
>>     uint u;
>> }
>> void showFour(Four f) {
>>     printf("f.u: %u\n", f.u);
>>     printf("f.a: [%d, %d, %d, %d]\n",
>>            cast(int)f.a[0], cast(int)f.a[1],
>>            cast(int)f.a[2], cast(int)f.a[3]);
>> }
>> void main() {
>>     Four f;
>>     f.a[] = [1, 2, 3, 4];
>>     showFour(f);
>>     f.u = (f.u << 8) | (f.u >> 24);
>>     showFour(f);
>>     printf("\n");
>>     // alternative
>>     f.a[] = [1, 2, 3, 4];
>>     uint u2 = f.u;
>>     showFour(f);
>>     printf("u2: %u\n", u2);
>>     asm {
>>         rol  u2, 8;
>>     }
>>     f.u = u2;
>>     showFour(f);
>> }
>> Bye,
>> bearophile
>
> I am offend!
>

Once I figured it out, I lol'd quite a bit.

While creating the rotation code I have found two things I don't understand. Maybe some of you is able to help me understand.

This version of the code:


union Four {
    uint u;
    ubyte[4] a;
}
void main() {
    Four f;
    asm {
        rol f.u, 8;
    }
}


DMD 2.052 gives this error, do you know why?

test.d(8): bad type/size of operands 'f.u'

----------------------------------

So to avoid wasting load asm instructions I have tried to write it like this:


union Four {
    ubyte[4] arr;
    uint ui;
}
void main() {
    Four fo;
    fo.arr[0] = 1;
    fo.arr[1] = 2;
    fo.arr[2] = 3;
    fo.arr[3] = 4;
    uint* uptr = &(fo.ui);
    asm {
        rol [uptr], 8;
    }
    asm {
        rol uptr, 8;
    }
}


but looking at the asm it produces, do you know why the rol with [uptr] and uptr are translated to the same instruction (so it rotates the pointer instead of the pointed uint)?


__Dmain comdat
    push EBP
    mov EBP,ESP
    sub ESP,8
    mov dword ptr -8[EBP],0
    lea EAX,-8[EBP]
    mov byte ptr -8[EBP],1
    mov byte ptr -7[EBP],2
    mov byte ptr -6[EBP],3
    mov byte ptr -5[EBP],4
    mov -4[EBP],EAX
    rol -4[EBP],8  ; <======
    rol -4[EBP],8  ; <======
    mov ESP,EBP
    pop EBP
    ret

Bye and thank you,
bearophile