Stockfish haswell optimized build

[j_romang]

Hello,
I just tried to make a haswell optimized build : https://www.dropbox.com/s/ghbs1vw18q6q4 ... 8_bmi2.zip
Thanks to Ronald de Man's code, I implemented BMI2 instructions in stockfish. This build also supports his syzygy's tablebases.
Please tell me if it works, you should have a ~4% speedup with the corresponding abrok.eu version. Of course you need a Haswell processor to run it !
Please note that this is NOT an official build, but just an experiment

[Gerd Isenberg]

Hello,
I just tried to make a haswell optimized build : https://www.dropbox.com/s/ghbs1vw18q6q4 ... 8_bmi2.zip
Thanks to Ronald de Man's code, I implemented BMI2 instructions in stockfish. This build also supports his syzygy's tablebases.
Please tell me if it works, you should have a ~4% speedup with the corresponding abrok.eu version. Of course you need a Haswell processor to run it !
Please note that this is NOT an official build, but just an experiment

This attack code instead of magics?
https://github.com/syzygy1/tb/blob/master/src/bmi2.h

Or is there something better in the meantime or more BMI2 changes elsewhere?

Thanks,
Gerd

[j_romang]

Here is the attack code, with no changes elsewhere :
https://github.com/jromang/Stockfish/bl ... ard.h#L271

[Gerd Isenberg]

Here is the attack code, with no changes elsewhere :
https://github.com/jromang/Stockfish/bl ... ard.h#L271

Thanks, yes Ronald's PDEP/PEXT 210.5k lookup approach. Wow, 4% is a huge speedup considering only attack-getters are changed, but of course affecting other memory and cache issues in other areas of the program. Assuming you also tried the PEXT only approach with 4 times greater tables, Ronald had the right sense, congratulations!

http://www.talkchess.com/forum/viewtopi ... 11&start=3

Cheers,
Gerd

[phenri]

I would probably have no utility of this patch before some months, or even some years.
But thank you for the future.

I have a question that has nothing to do with BMI2.
I'm asking you because I do not really expect a response from Marco.

I probably be wrong, but I want to understand.
Why in the makefile, POPCNT comes with the flag -msse3 instead of -msse4.2 while POPCNT is present only for architectures with a minimum SSE4.2.

And why the flag -mpopcnt is not included?

Code: Select all

### 3.9 popcnt
ifeq ($(popcnt),yes)
	CXXFLAGS += -msse3 -DUSE_POPCNT
endif

Same for preftech why the flag is so low?

Code: Select all

### 3.7 prefetch
ifeq ($(prefetch),yes)
	ifeq ($(sse),yes)
		CXXFLAGS += -msse
		DEPENDFLAGS += -msse
	endif
else
	CXXFLAGS += -DNO_PREFETCH
endif

Regards,

Paul

[syzygy]

It might be interesting to try this.
In bitboard.h change

Code: Select all

struct BMI2Info {
  unsigned short *data;
  uint64_t mask1;
  uint64_t mask2;
};

extern unsigned short attack_table[107648];
extern struct BMI2Info bishop_bmi2[64];
extern struct BMI2Info rook_bmi2[64];

template<PieceType Pt>
inline Bitboard attacks_bb&#40;Square s, Bitboard occ&#41; &#123;
  struct BMI2Info *info = &#40;Pt == ROOK ? &rook_bmi2&#91;s&#93; &#58; &bishop_bmi2&#91;s&#93;);
  return _pdep_u64&#40;info->data&#91;_pext_u64&#40;occ, info->mask1&#41;&#93;, info->mask2&#41;;
&#125;

into

Code: Select all

struct BMI2Info {
  uint64_t *data;
  uint64_t mask;
};

extern struct BMI2Info bishop_bmi2[64];
extern struct BMI2Info rook_bmi2[64];

template<PieceType Pt>
inline Bitboard attacks_bb&#40;Square s, Bitboard occ&#41; &#123;
  struct BMI2Info *info = &#40;Pt == ROOK ? &rook_bmi2&#91;s&#93; &#58; &bishop_bmi2&#91;s&#93;);
  return info->data&#91;_pext_u64&#40;occ, info->mask&#41;&#93;;
&#125;

In bitboard.cpp change the lines

Code: Select all

static unsigned short attacks_table[107648];
...
        info[sq].mask1 = bb
...
          if (i == 0)
        info[sq].mask2 = bb2;
          attacks_table[idx++] = _pext_u64(bb2, info[sq].mask2);

into

Code: Select all

static uint64_t attacks_table[107648];
...
        info[sq].mask = bb
...
          attacks_table[idx++] = bb2;

I did not test this, so maybe something is wrong or missing.

This might be faster and this might be slower, but it would be interesting to know.

[syzygy]

Assuming you also tried the PEXT only approach with 4 times greater tables, Ronald had the right sense, congratulations!

I don't think he tried, but we can now find out

[Gerd Isenberg]

Assuming you also tried the PEXT only approach with 4 times greater tables, Ronald had the right sense, congratulations!

I don't think he tried, but we can now find out

Would be nice if Jean-Francois or you could try and report. U64 versus pdep_u64(U16, mask2).

[j_romang]

I didn't try

[j_romang]

Wow, 4% is a huge speedup considering only attack-getters are changed, but of course affecting other memory and cache issues in other areas of the program.

According to my profiling experiments stockfish spends about 5-6% of time computing attack bitboards, that's we I wanted to give a try to the pext solution.