Factor is found by GMP-ECM but not by GMP-ECM[GWNUM]

[YuL]

I wanted to switch from
GMP-ECM 7.0.4 [configured with GMP 6.1.2, --enable-asm-redc]
to
GMP-ECM 7.0.4 [configured with GMP 6.1.2, GWNUM 29.8, --enable-asm-redc]
for my factoring work.
So I was doing some testing and it appears that a factor found by the first one
is not found by the second one:

GMP-ECM 7.0.4 [configured with GMP 6.1.2, --enable-asm-redc] [ECM]
Tuned for x86_64/k8/params.h
Input number is (103*10^286-1)/(3*43*47*9283691*60137568222544788641951) (255 digits)
[Thu Dec  3 11:31:56 2020]
Using MODMULN [mulredc:0, sqrredc:2]
Using B1=6500000, B2=14271500890, polynomial Dickson(12), sigma=0:8646756833320245055
dF=16384, k=5, d=158340, d2=11, i0=31
Expected number of curves to find a factor of n digits:
35    40    45    50    55    60    65    70    75    80
178    1111    8108    67760    637802    6669522    7.6e+07    9.8e+08    1.3e+10    6.6e+11
Step 1 took 33016ms
********** Factor found in step 1: 42010916546178371843382678844466221253227
Found prime factor of 41 digits: 42010916546178371843382678844466221253227
Composite cofactor ((103*10^286-1)/(3*43*47*9283691*60137568222544788641951))/42010916546178371843382678844466221253227 has 214 digits
Peak memory usage: 4MB

GMP-ECM 7.0.4 [configured with GMP 6.1.2, GWNUM 29.8, --enable-asm-redc] [ECM]
Tuned for x86_64/k8/params.h
Due to incompatible licenses, this binary file must not be distributed.
Input number is (103*10^286-1)/(3*43*47*9283691*60137568222544788641951) (255 digits)
Found number: 103*10^286 + -1
[Thu Dec  3 11:33:38 2020]
Using MODMULN [mulredc:0, sqrredc:2]
Using B1=6500000, B2=14271500890, polynomial Dickson(12), sigma=0:8646756833320245055
dF=16384, k=5, d=158340, d2=11, i0=31
Expected number of curves to find a factor of n digits:
35    40    45    50    55    60    65    70    75    80
178    1111    8108    67760    637802    6669522    7.6e+07    9.8e+08    1.3e+10    6.6e+11
Using gwnum_ecmStage1(103, 10, 286, -1, 6500000, 1)
Step 1 took 27268ms
Using 30 small primes for NTT
Estimated memory usage: 81.07MB
Initializing tables of differences for F took 52ms
Computing roots of F took 720ms
Building F from its roots took 552ms
Computing 1/F took 300ms
Initializing table of differences for G took 44ms
Computing roots of G took 636ms
Building G from its roots took 576ms
Computing roots of G took 668ms
Building G from its roots took 608ms
Computing G * H took 168ms
Reducing  G * H mod F took 172ms
Computing roots of G took 664ms
Building G from its roots took 588ms
Computing G * H took 164ms
Reducing  G * H mod F took 168ms
Computing roots of G took 664ms
Building G from its roots took 584ms
Computing G * H took 160ms
Reducing  G * H mod F took 164ms
Computing roots of G took 652ms
Building G from its roots took 576ms
Computing G * H took 152ms
Reducing  G * H mod F took 160ms
Computing polyeval(F,G) took 1076ms
Computing product of all F(g_i) took 8ms
Step 2 took 10312ms
Expected time to find a factor of n digits:
35    40    45    50    55    60    65    70    75    80
1.86h    11.60h    3.53d    29.47d    277.41d    7.95y    90.42y    1164y    15695y    790907y
Peak memory usage: 139MB

Lowering the value of B1, to say 3e6, exhibits the same behavior except that
the factor is found in step 2 instead of step 1.
Any thoughts?
Note: I built the binaries myself, I patched the source code of ECM 7.0.4 to
include rev3084 (which fixes the bug reported here) and that is the only
change I made.

[Prime95]

GWNUM only handles 53-bit sigma values -- 8646756833320245055 was truncated.

[mathwiz]

Originally Posted by Prime95

GWNUM only handles 53-bit sigma values -- 8646756833320245055 was truncated.

Is the net result that GMP-ECM is less likely to find factors when linked with gwnum, and if so by how much? (what's the range of sigma values used?)

[VBCurtis]

No, it just uses a smaller range of sigma. Each individual sigma has the same chance to find a factor, if one is to be found. Doesn't matter if the possible-sigma pool is 1 to 2^50 or 1 to 2^60.

If the pool were 1 to 10^9, then running hundreds of thousands of curves would likely repeat some work- that's a waste, so the pool is much bigger than that. GWNUM has a higher chance to repeat a curve due to a smaller pool, but if you calculate the odds you'll see it's still very very unlikely.

[YuL]

Then when using a GMP-ECM linked with GWNUM the value of sigma
displayed is not the real one, I have to truncate it to a value < 2⁵³
to deduce the real one?

[Prime95]

I am not familiar with the GMP-ECM code. The problem (truncation) may only occur when specifying a sigma on the command line.