`get_fmem()` issues

[MicheleCancilla]

I was looking for windows incompatibilites in develop branch and I spotted some issues regarding the get_fmem() function that could be solved.

-

  unsigned long freemem = get_free_mem();

The get_free_mem() function returns an unsigned long which could cause integer overflow. I would move return type to unsigned long long.

• try{
    //ptr = new float[size];
    //ptr=(float *)malloc(size*sizeof(float));
    //ptr=(float *)aligned_alloc(64, size*sizeof(float));
    posix_memalign((void **)&ptr, 64, size*sizeof(float));
  }
  catch (std::bad_alloc& badAlloc){
    error=true;
  }

  posix_memalign() is not available on windows. _aligned_malloc() could be a windows replacement but requires a specific function for freeing memory (_aligned_free()) instead of free(). Two feasible choices are:

  1. Define eddl_malloc() and eddl_free() which contain call to posix_memalign/_aligned_malloc (and free/_aligned_free).
  2. Use malloc() under windows loosing the alignment feature.

  The try/catch is now useless since posix_memalign() does not throw any exceptions, but it returns an error code.

I can open a PR so you can test the changes.

[jonandergomez]

Thanks Michele, I'm going to solve it asap.

[jonandergomez]

Dear Michele,

I just pushed a commit into develop facing what you commented, but I have no the chance to test it in Windows, could you check it? I'm ready to fix any bug derived from this commit.

Regads,

Jon

[MicheleCancilla]

Dear Jon, the tests do not pass on windows, but I understand why. The TensorTestSuite.tensor_math_unary_log test fails throwing exception in eddl_malloc. I have verified that errno != 0 is true before calling _aligned_malloc and errno value is 33 (EDOM Math argument).

auto err_ = errno;
ptr = _aligned_malloc(size, alignment_block_size);
error = (nullptr == ptr || errno != 0);
err_ = errno;

I suggest only checking if allocation fails (doc):

ptr = _aligned_malloc(size, alignment_block_size);
error = (nullptr == ptr || errno == ENOMEM);

I get a different compilation error (a parenthesized type followed by an initializer list is a non-standard explicit type conversion syntax) on _5_nlp_textgeneration. You should remove parentheses around vtensor:

forward(decoder, vtensor{ word,treshape,state });

[CostantinoGrana]

For everybody: errno should be set to 0 before calling a function, if you plan to check its value. errno should be checked only in case of error in function. Check this: https://softwareengineering.stackexchange.com/a/209731/360023

[jonandergomez]

Thanks for the comment Costantino. And thanks for the suggestion Michele.

Let me solve it and I will push a commit.

Jon

[jonandergomez]

Hi again,

Commit following your indications pushed, also the compiling problem with the call to forward in the NLP example fixed.

Michele, could you please check in windows if both problems are solved?

Regards,

Jon

[MicheleCancilla]

No, it does not work 😂 There is an error in the errno condition error = (nullptr == ptr || errno != ENOMEM);. It should be error = (nullptr == ptr || errno == ENOMEM);.

[jonandergomez]

Sorry, my logic fails. I need a windows instance to do checks and avoid bothering you with this. Please, check it after the last commit and close the issue if finally it runs. Regards, Jon

[MicheleCancilla]

Yes, it works! Thank you Jon.