vdExp 32 bit vs 64 bit different results

[Mark_L_6]

Hi,

I'm new to MKL so please bare with me. We're moving a project from 32 to 64 bit and we've encountered some inconsistencies between 32 & 64 bit vdExp.

I have the following code

double value = 0.027771918023383080;

double result;

vdExp( 1, &value, &result );

In 32 bit mode, the value in result is 1.0281611526482297 whilst when I run it in 64 bit mode I get 1.0281611526482299.

Any ideas why this is? I've double checked and I'm definitely compiling with the correct libraries. I used the Intel MKL link advisor.

Regards

Mark.

[TimP]

Those results are so close that the function results may be identical. It may be a difference in the library used by your debugger or compiler.

[Mark_L_6]

Thanks for the response but it may seem like a tiny difference and I've replicated it across several other double values but it makes a big difference as we're already dealing with large numbers and when you multiply or use other maths with an incorrect, no, inconsistent value then it makes most of our unit tests fail and presents a burden on our porting effort.

I'd really like to get to the bottom of it if possible so we can come up with a strategy.

Btw, we've also noticed the same inconsistencies using the normal std::exp function albeit with different values.

[Mark_L_6]

It turns out, after reading a bit more on the MKL that if I use

mkl_cbwr_set( MKL_CBWR_SSE2);

which I gleaned from this intel documentation https://software.intel.com/en-us/node/528731

then I get consistent results.

[Jingwei_Z_Intel]

Are you running the ia32 and intel64 executables on the same CPU? I tried your test case with the latest Intel compiler (16.0.3 20160415) and I got the same result on most platforms I tried, for both 32bit and 64bit:

input  = 0.027771918023383080 = 0x3f9c703ddea16800
result = 1.028161152648229715 = 0x3ff073591bda44db

I see different result on Penryn processors:

input  = 0.027771918023383080 = 0x3f9c703ddea16800
result = 1.028161152648229937 = 0x3ff073591bda44dc

For vector math functions in MKL, the result should usually be the same for ia32 and intel64 on the same CPU. However, on different CPUs we cannot guarantee the same result, since we are optimizing for different instruction sets. For that, you should use conditional numerical reproducibility (CBWR) setting of MKL.

[Mark_L_6]

Thanks Jingwei, I found that out and you're right that using the CBWR does indeed produce correct results across our machines.

We are definitely using Intel x86/x64 machines for our testing. One of the machines we saw no problems with but the others all exhibited this issue. the one which didn't have the issue was an intel i7-3840QM, the others were intel i5's.

Anyway, I'm happy to report we're getting consistent results now.

Thanks for all your comments.