Comment
Comments
Comments in green were written by me. Comments in blue were not written by me.
@Oleg: I pulled your code into Visual Studio 2026 and dealt with some warnings:
The COLLAPSES initializer was easy: Use (char) for 0x80
I changed unsafe localtime to:
struct tm buf;
auto err = localtime_s(&buf, &cur_time);
return std::put_time(&buf, "%F %T");
I did a quick change of __tzcnt_u32, __lzcnt32 to use _tzcnt_u32, _lzcnt_u32
Setting the compiler to use AVX and optimize for speed.
An '8' run worked, so I tried '9'
And it found only 1,729,930 solutions!
I suspected _tzcnt_u32, _lzcnt_u32 might be causing it, so I covered that:
#define __tzcnt_u32(v) ((v) ? (_tzcnt_u32(v)) : (32)) // Match BMI : should return 32 for value 0
#define __lzcnt32(v) ((v) ? (_lzcnt_u32(v)) : (32)) // Match BMI : should return 32 for value 0
but it didn't fix it.
Anyway, I decided to test multi-threading.
First I hunted for the initial depth sweet spot:
>Puzzle_Oleg.exe run 9 8 24
>Puzzle_Oleg.exe run 9 9 24
...
>Puzzle_Oleg.exe run 9 18 24
>Puzzle_Oleg.exe run 9 19 24
And found that 15 was fastest. [It didn't like 19]
'9' run [with the 1,729,930 problem] on Xeon E5-2697-v2
12t: 11m 52s
24t: 9m 1s
so HyperThreading is helping by about 48%
The COLLAPSES initializer was easy: Use (char) for 0x80
I changed unsafe localtime to:
struct tm buf;
auto err = localtime_s(&buf, &cur_time);
return std::put_time(&buf, "%F %T");
I did a quick change of __tzcnt_u32, __lzcnt32 to use _tzcnt_u32, _lzcnt_u32
Setting the compiler to use AVX and optimize for speed.
An '8' run worked, so I tried '9'
And it found only 1,729,930 solutions!
I suspected _tzcnt_u32, _lzcnt_u32 might be causing it, so I covered that:
#define __tzcnt_u32(v) ((v) ? (_tzcnt_u32(v)) : (32)) // Match BMI : should return 32 for value 0
#define __lzcnt32(v) ((v) ? (_lzcnt_u32(v)) : (32)) // Match BMI : should return 32 for value 0
but it didn't fix it.
Anyway, I decided to test multi-threading.
First I hunted for the initial depth sweet spot:
>Puzzle_Oleg.exe run 9 8 24
>Puzzle_Oleg.exe run 9 9 24
...
>Puzzle_Oleg.exe run 9 18 24
>Puzzle_Oleg.exe run 9 19 24
And found that 15 was fastest. [It didn't like 19]
'9' run [with the 1,729,930 problem] on Xeon E5-2697-v2
12t: 11m 52s
24t: 9m 1s
so HyperThreading is helping by about 48%
Lord Sméagol
on /blog/119
on /blog/119
@Lord Sméagol: Hello and happy New Year!
9 minutes is cool!
The answer is wrong because of _lzcnt instruction, as you suspected, as turns out it works differently on different cpus: https://nextmovesoftware.com/blog/2017...
With this error, solutions having 1x1 square directly in the center are not counted.
I guess, gcc/clang do it correctly because I specify -march=native (so it checks cpu and generates correct instruction), and run where I compile. But it's a potential problem I probably need to add some assertions to the code.
Maybe on your hardware you can either use WSL and clang compiler, or set constexpr bool USE_SSE_QUADRANT_FILL=false, to fall back to slower.
You could also try to use BitScanReverse instead of __lzcnt, but it has different input/output so I'm not sure how hard would that be to fix it.
9 minutes is cool!
The answer is wrong because of _lzcnt instruction, as you suspected, as turns out it works differently on different cpus: https://nextmovesoftware.com/blog/2017...
With this error, solutions having 1x1 square directly in the center are not counted.
I guess, gcc/clang do it correctly because I specify -march=native (so it checks cpu and generates correct instruction), and run where I compile. But it's a potential problem I probably need to add some assertions to the code.
Maybe on your hardware you can either use WSL and clang compiler, or set constexpr bool USE_SSE_QUADRANT_FILL=false, to fall back to slower.
You could also try to use BitScanReverse instead of __lzcnt, but it has different input/output so I'm not sure how hard would that be to fix it.
Oleg
on /blog/119
on /blog/119
@Oleg:
I removed my macros:
#define __tzcnt_u32(v) ((v) ? (_tzcnt_u32(v)) : (32))
#define __lzcnt32(v) ((v) ? (_lzcnt_u32(v)) : (32))
replacing them with simple inline code
Util.h
//#include // not including all
#include // just include what's needed
#include // just include what's needed
#include // just include what's needed
#if 1 // use safe localtime
struct tm buf; // use safe localtime
auto err = localtime_s(&buf, &cur_time); // use safe localtime
return std::put_time(&buf, "%F %T"); // use safe localtime
#else // use safe localtime
return std::put_time(std::localtime(&cur_time), "%F %T");
#endif // use safe localtime
State.h
changed _mm_set_epi8(0x80 to -0x80 to stop warnings
inline replacement:
//int i = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
int i = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
inline replacement:
//int last_idx_before_mid = 31 - __lzcnt32(off_mask); // for no BMI; without zero test, as not needed here
int last_idx_before_mid = _lzcnt_u32(off_mask); // for no BMI; without zero test, as not needed here
Solver.h
inline replacement:
//return ini.size(); // to stop warning
return (int)ini.size(); // to stop warning
inline replacement:
//const int dim = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
const int dim = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
I tried '9' runs: with asserts: 10:31, without: 10:18 (saved 2%)
A minute slower than the faulty version, but still not too bad for a 2013 (Q3) CPU :)
I removed my macros:
#define __tzcnt_u32(v) ((v) ? (_tzcnt_u32(v)) : (32))
#define __lzcnt32(v) ((v) ? (_lzcnt_u32(v)) : (32))
replacing them with simple inline code
Util.h
//#include // not including all
#include // just include what's needed
#include // just include what's needed
#include // just include what's needed
#if 1 // use safe localtime
struct tm buf; // use safe localtime
auto err = localtime_s(&buf, &cur_time); // use safe localtime
return std::put_time(&buf, "%F %T"); // use safe localtime
#else // use safe localtime
return std::put_time(std::localtime(&cur_time), "%F %T");
#endif // use safe localtime
State.h
changed _mm_set_epi8(0x80 to -0x80 to stop warnings
inline replacement:
//int i = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
int i = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
inline replacement:
//int last_idx_before_mid = 31 - __lzcnt32(off_mask); // for no BMI; without zero test, as not needed here
int last_idx_before_mid = _lzcnt_u32(off_mask); // for no BMI; without zero test, as not needed here
Solver.h
inline replacement:
//return ini.size(); // to stop warning
return (int)ini.size(); // to stop warning
inline replacement:
//const int dim = __tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
const int dim = _tzcnt_u32(mask); // for no BMI; without zero test, as not needed here
I tried '9' runs: with asserts: 10:31, without: 10:18 (saved 2%)
A minute slower than the faulty version, but still not too bad for a 2013 (Q3) CPU :)
Lord Sméagol
on /blog/119
on /blog/119
@Oleg: Happy new year!
I just added this:
#if 0
int last_idx_before_mid = 31 - __lzcnt32(off_mask); // 31 - LZCNT ==> index of MSb
#else
// if off_mask can never be zero, no need for check to override BSR result
assert(off_mask);
// a '9' run didn't reveal any 0 [you would know for sure for other sizes]
// need unsigned long result
unsigned long last_idx_before_mid;
// get index of MSb [no need for adjustment if off_mask can never be zero]
_BitScanReverse(&last_idx_before_mid, off_mask);
#endif
a run of '9' now produces the correct result: 1,730,280 :)
I just added this:
#if 0
int last_idx_before_mid = 31 - __lzcnt32(off_mask); // 31 - LZCNT ==> index of MSb
#else
// if off_mask can never be zero, no need for check to override BSR result
assert(off_mask);
// a '9' run didn't reveal any 0 [you would know for sure for other sizes]
// need unsigned long result
unsigned long last_idx_before_mid;
// get index of MSb [no need for adjustment if off_mask can never be zero]
_BitScanReverse(&last_idx_before_mid, off_mask);
#endif
a run of '9' now produces the correct result: 1,730,280 :)
Lord Sméagol
on /blog/119
on /blog/119
@Oleg: Great!
Going by your time differences between single and multi-thread, I assume you are not using SMT.
My performance is limited by my old tech (Ivy Bridge E5-2697-v2) !
I have got multi-threading working in C and it looks like there are no bugs :)
But I am still not getting any benefit from HyperThreading, maybe 16 general purpose registers aren't enough for the compiler!
I have just started converting my inner search function to asm, which will give me full control of ALL registers and let me use some coding tricks that are not availble using C.
Once I get it going, I will try a '10' run, which should verify your results.
Going by your time differences between single and multi-thread, I assume you are not using SMT.
My performance is limited by my old tech (Ivy Bridge E5-2697-v2) !
I have got multi-threading working in C and it looks like there are no bugs :)
But I am still not getting any benefit from HyperThreading, maybe 16 general purpose registers aren't enough for the compiler!
I have just started converting my inner search function to asm, which will give me full control of ALL registers and let me use some coding tricks that are not availble using C.
Once I get it going, I will try a '10' run, which should verify your results.
Lord Sméagol
on /blog/119
on /blog/119
@Lord Sméagol: Right, I wasn't using SMT (I meant it when I said without multithreading, sorry for the bad wording). I tried to use intel-based VM before with 4 cpu/8 threads, but the speedup was about 5.5 times only, and the price was only 20% less (I used Azure spot instances which are not so expensive, but some automation is needed to restart them every time they are stopped by Azure).
To give you all details, my program spent 21 days on an AMD EPYC 9004 (8 cores without SMT, Azure spot instance Standard F8als v6) using 8 threads (that is, about 160 CPU-days!)
I've published the source code, still planning to write about the optimizations: https://github.com/lightln2/partridge-...
To give you all details, my program spent 21 days on an AMD EPYC 9004 (8 cores without SMT, Azure spot instance Standard F8als v6) using 8 threads (that is, about 160 CPU-days!)
I've published the source code, still planning to write about the optimizations: https://github.com/lightln2/partridge-...
(anonymous)
on /blog/119
on /blog/119
@(anonymous): Thanks for the clarification.
I'm still (slowly) building my asm funcion. I think I have settled on register allocation, leaving only rcx as a 'scratch' register because cl will be needed for some variable shifts.
I also use the xmm registers (14 so far) to minimize memory operations to hopefully let HT/SMT get some decent gains.
How long would Matt Parker's 'terrible Python code' take to solve this problem ? :)
Ok, his maths knowledge might produce some decent algorithms, but it would help him a lot to use something that compiles to native code.
I'm still (slowly) building my asm funcion. I think I have settled on register allocation, leaving only rcx as a 'scratch' register because cl will be needed for some variable shifts.
I also use the xmm registers (14 so far) to minimize memory operations to hopefully let HT/SMT get some decent gains.
How long would Matt Parker's 'terrible Python code' take to solve this problem ? :)
Ok, his maths knowledge might produce some decent algorithms, but it would help him a lot to use something that compiles to native code.
Lord Sméagol
on /blog/119
on /blog/119
For comparison: my program runs 2 seconds for n=8 (16 seconds single-threaded), and 15 minutes for n=9 (2 hours single-threaded). It has a number of quite tricky optimizations, I should write about those sometime.
Of course the result needs to be verified as the program is quite complex and might have bugs.
on /blog/119