ippsLn_64f_A53 differences between Xcode (clang 10) and VS 2019

[Vitor]

Hi,

I’m working on a cross-platform C++ app that makes lots of math computations. I have some test cases to assess the correctness of the algorithms and everything works well except on some edge cases. 

Even though both platforms use IEEE 754 to represent doubles and both codebases have the same rounding mode, sometimes the results for the Ln function differ. For example, for the value “0.99843853339480171” ippsLn_64f_A53 gives:

-0.0015626869647103901 on MacOs

-0.0015626869647103903 on Windows

Code was compiled without any kind of optimizations. All other math functions give the same value on both platforms except the Ln family functions. Any suggestion?

Thanks

Note: using IPP version 2019.0.1

[Gennady_F_Intel]

If you will truncate to a double floating point - will you see the same results?

[Vitor]

Hi Gennady, the result is already truncate to a double floating point. The difference between the OSs is the least significant bit of the mantissa.

[Gennady_F_Intel]

Hi Vitor,

It looks like the issue. We need to check the problem on our side. Do you see the problem with some specific CPU types of everywhere?

[Vitor]

Hi Gennady, both CPUs are Intel models:

Mac: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
Windows: Intel(R) Core(TM) i5-8265U CPU @ 1.60GHz

I didn't try in other CPU types.

[Gennady_F_Intel]

as IPP uses MKL VML implementation as a backend, you may try to get the bit-to-bit output with CNR mode enabled. e.x. export MKL_CBWR = AVX2

[Vitor]

Hi Gennady, thanks for the insight about CNR mode. Although I’m using IPP (without linking with MKL libraries), I have changed the code to link with MKL but the results between platforms still have a 1-bit off on the mantissa.

This simple test case fails:

    assert(mkl_cbwr_set(MKL_CBWR_COMPATIBLE) == MKL_CBWR_SUCCESS);
    double source = 0.99843853339480171;
    Ipp64f out{0};
    IppStatus status = ippsLn_64f_A53(&source, &out, 1);
    assert(status == ippStsNoErr);
    double result = static_cast<double>(out);

    uint64_t bin = *((uint64_t*)&result); 
    std::cout << std::hex << bin << "\n";
    assert(result == -0.0015626869647103901);

On Mac the output is: bf599a625a1179c7

On Windows the assert fails and the output is: bf599a625a1179c8

[Gennady_F_Intel]

Hi Vitor,

I must admit it was my fault. Actually MKL VML is used a backend from IPP vector math functions, but IPP didn't check this environment variables.

regarding the original problem - the differences you obtained on difference OS is within +- 1 ulp and therefore it aligns with committed accuracy.  MKL and IPP never accepted to provide bit-to-bit output across the different OS.