bsr64

[jj2007]

include \masm32\MasmBasic\MasmBasic.inc      ; download
.data
q1      QWORD 10000000000000000000000000000000000b      ; only bit 34 set
q2      QWORD 10010000000000000000000000000000000b      ; bits 34 and 31 set
q3      QWORD 00010000000000000000000000000000000b      ; only bit 31 set

bsr64 MACRO argQ
  mov eax, dword ptr argQ
  mov edx, dword ptr argQ[4]
  ; deb 4, "bsr", b:edx, b:eax
  bsr eax, edx            ; non-zero if edx is non-zero
  .if Zero?
      bsr eax, eax
  .else
      add eax, 32
  .endif
  exitm <eax>
ENDM

  Init
  Print Str$("bsr64 of q1 = %i\n", bsr64(q1))
  Print Str$("bsr64 of q2 = %i\n", bsr64(q2))
  Inkey Str$("bsr64 of q3 = %i\n", bsr64(q3))
EndOfCode

Output:
bsr64 of q1 = 34
bsr64 of q2 = 34
bsr64 of q3 = 31

I'd like to know whether anybody (e.g. on AMD cpu) sees a different output - exe is attached.

P.S.: See also Bit Manipulation Instruction Sets and the September 2013 bsr eax, eax with eax=0 thread.

Apparently it plays a role in chess programming, see Bsf/Bsr behavior with zero source:

Intel and AMD specify different behavior. In praxis there seems no difference so far. However, as long as Intel docs explicitly state content undefined, it is recommend to don't rely on a pre-initialized content of that target register, if the source is zero.

    Intel : If the content of the source operand is 0, the content of the destination operand is undefined. [35]
    AMD: If the second operand contains 0, the instruction sets ZF to 1 and does not change the contents of the destination register. [36]

Timings (b is the safe version):

Code Select Expand

Intel(R) Core(TM) i5-2450M CPU @ 2.50GHz (SSE4)
153     cycles for 100 * bsr64a, bit 31
153     cycles for 100 * bsr64b, bit 31
102     cycles for 100 * bsr64a, bit 34
165     cycles for 100 * bsr64b, bit 34

[sinsi]

Code Select Expand


bsr64 of q1 = 34
bsr64 of q2 = 34
bsr64 of q3 = 31


Yuck.

Code Select Expand


AMD A10-7850K Radeon R7, 12 Compute Cores 4C+8G (SSE4)

867     cycles for 100 * bsr64a, bit 31
893     cycles for 100 * bsr64b, bit 31
468     cycles for 100 * bsr64a, bit 34
460     cycles for 100 * bsr64b, bit 34

882     cycles for 100 * bsr64a, bit 31
850     cycles for 100 * bsr64b, bit 31
425     cycles for 100 * bsr64a, bit 34
445     cycles for 100 * bsr64b, bit 34

844     cycles for 100 * bsr64a, bit 31
856     cycles for 100 * bsr64b, bit 31
443     cycles for 100 * bsr64a, bit 34
440     cycles for 100 * bsr64b, bit 34

855     cycles for 100 * bsr64a, bit 31
847     cycles for 100 * bsr64b, bit 31
432     cycles for 100 * bsr64a, bit 34
437     cycles for 100 * bsr64b, bit 34

840     cycles for 100 * bsr64a, bit 31
842     cycles for 100 * bsr64b, bit 31
441     cycles for 100 * bsr64a, bit 34
445     cycles for 100 * bsr64b, bit 34

851     cycles for 100 * bsr64a, bit 31
852     cycles for 100 * bsr64b, bit 31
440     cycles for 100 * bsr64a, bit 34
439     cycles for 100 * bsr64b, bit 34

24      bytes for bsr64a, bit 31
26      bytes for bsr64b, bit 31
24      bytes for bsr64a, bit 34
26      bytes for bsr64b, bit 34

31      = eax bsr64a, bit 31
31      = eax bsr64b, bit 31
34      = eax bsr64a, bit 34
34      = eax bsr64b, bit 34

[jj2007]

Yuck.

Code Select Expand


AMD A10-7850K Radeon R7, 12 Compute Cores 4C+8G (SSE4)


It seems that AMD is not so fast for bsr. Thanks, it seems the additional push+pop doesn't slow the "safe" version down.

Btw there is a hilarious debate at SOF on What is the fastest/most efficient way to find the highest set bit (msb) in an integer in C?. At about half the page, Josh posts "some (simple) benchmarks". Compiles fine with GCC, using e.g. using c++ -O1 -o Tmp.exe  "Tmp.cpp"

Now it turns out that the inline assembly part is not particularly fast. Given that the other approaches are, ehm, not as simple and straightforward as bsr eax, eax, I inserted an int 3, and Olly told me a secret 8)

unsigned OUT;
#define POS_OF_HIGHESTBITasm(a) (({asm("bsrl %1,%0" : "=r"(OUT) : "r"(a));}), OUT)
...
  start = clock();
  for (ui = 0; ui < LOOPS; ++ui)
    asm("int $3");
    n = NUM_OF_HIGHESTBITasm(ui);
  end = clock();
  printf("asm:\t%e\n", (double)(end-start)/CLOCKS_PER_SEC);

So far so good. But why is that simple bsr eax, eax as slow as the acrobatic C algos???

Code Select Expand

004016F4   ³.  E8 7F0A0000        call <jmp.&msvcrt.clock>                 ; Jump to msvcrt.clock
004016F9   ³.  89C3               mov ebx, eax
004016FB   ³.  BA 00E1F505        mov edx, 5F5E100                         ; edx is decimal 100000000
00401700   ³>  CC                 Úint3
00401701   ³.  83EA 01            ³sub edx, 1
00401704   ³. 75 FA              Àjnz short 00401700
00401706   ³.  B8 00E1F505        mov eax, 5F5E100
0040170B   ³.  0FBDC0             bsr eax, eax
0040170E   ³.  A3 08604000        mov [406008], eax
00401713   ³.  E8 600A0000        call <jmp.&msvcrt.clock>                 ; Jump to msvcrt.clock

Because GCC inserts a little loop with 100000000 iterations there

[qWord]

  for (ui = 0; ui < LOOPS; ++ui)
    asm("int $3");
    n = NUM_OF_HIGHESTBITasm(ui);


So far so good. But why is that simple bsr eax, eax as slow as the acrobatic C algos???
[...]
Because GCC inserts a little loop with 100000000 iterations there

No, because you insert a loop with 100000000 iterations  ;)

[mineiro]

Code Select Expand

bsr64 of q1 = 34
bsr64 of q2 = 34
bsr64 of q3 = 31

Code Select Expand

Intel(R) Core(TM) i3-3110M CPU @ 2.40GHz (SSE4)

171     cycles for 100 * bsr64a, bit 31
176     cycles for 100 * bsr64b, bit 31
113     cycles for 100 * bsr64a, bit 34
181     cycles for 100 * bsr64b, bit 34

172     cycles for 100 * bsr64a, bit 31
173     cycles for 100 * bsr64b, bit 31
113     cycles for 100 * bsr64a, bit 34
180     cycles for 100 * bsr64b, bit 34

171     cycles for 100 * bsr64a, bit 31
174     cycles for 100 * bsr64b, bit 31
114     cycles for 100 * bsr64a, bit 34
181     cycles for 100 * bsr64b, bit 34

196     cycles for 100 * bsr64a, bit 31
201     cycles for 100 * bsr64b, bit 31
115     cycles for 100 * bsr64a, bit 34
182     cycles for 100 * bsr64b, bit 34

173     cycles for 100 * bsr64a, bit 31
173     cycles for 100 * bsr64b, bit 31
113     cycles for 100 * bsr64a, bit 34
182     cycles for 100 * bsr64b, bit 34

171     cycles for 100 * bsr64a, bit 31
172     cycles for 100 * bsr64b, bit 31
113     cycles for 100 * bsr64a, bit 34
182     cycles for 100 * bsr64b, bit 34

24      bytes for bsr64a, bit 31
26      bytes for bsr64b, bit 31
24      bytes for bsr64a, bit 34
26      bytes for bsr64b, bit 34

31      = eax bsr64a, bit 31
31      = eax bsr64b, bit 31
34      = eax bsr64a, bit 34
34      = eax bsr64b, bit 34

--- ok ---


[jj2007]

No, because you insert a loop with 100000000 iterations  ;)

Oops, you are right - at war again <with> (the) [goddamn] {C brackets} :icon_redface: