[mlpack] Question about Armadillo

[Kirill Mishchenko]

> I believe this is a result of strange inference by the compiler.  You
> should double-check my explanation here, I'm not 100% sure that it is
> correct, but my gut feeling is that it is…

It looks like there is some other reason. I got the same error when I tried the following

template<typename T>
void g(const T&)
{
  const T& ref = arma::join_rows(arma::zeros(2).t(), arma::zeros(2).t());
  arma::mat m = ref;
}

BOOST_AUTO_TEST_CASE(RefTest4)
{
  g(arma::join_rows(arma::zeros(2).t(), arma::zeros(2).t()));
}

Best regards,

Kirill Mishchenko



> On 27 Jun 2017, at 19:41, Ryan Curtin <ryan at ratml.org> wrote:
> 
> On Tue, Jun 27, 2017 at 09:10:06AM +0500, Kirill Mishchenko wrote:
>> Hi mlpack community!
>> 
>> I’m not sure it is a right place to ask a question that I want to ask,
>> but, anyway, maybe somebody will be eager to answer. Otherwise, just
>> ignore this message.
>> 
>> I have encountered some unexpected behaviour with the Armadillo
>> library. The following code snippet works as expected:
>> 
>> template<typename T>
>> void f(const T& arg)
>> {
>>  arma::mat m = arg;
>> }
>> 
>> BOOST_AUTO_TEST_CASE(RefTest)
>> {
>>  f(arma::join_rows(arma::zeros(2).t(), arma::zeros(2).t()));
>> }
>> 
>> But if I run the following
>> 
>> BOOST_AUTO_TEST_CASE(RefTest2)
>> {
>>  const auto& ref = arma::join_rows(arma::zeros(2).t(), arma::zeros(2).t());
>>  arma::mat m = ref;
>> }
>> 
>> I get this error:
>> 
>> unknown location:0: fatal error: in "ArmaExtendTest/RefTest2": memory
>> access violation at address: 0x00000000: no mapping at fault address
>> 
>> Surprisingly, if run the following code snippet
>> 
>> BOOST_AUTO_TEST_CASE(RefTest3)
>> {
>>  const auto& ref = arma::join_rows(arma::zeros(2), arma::zeros(2));
>>  arma::mat m = ref;
>> }
>> 
>> everything is ok again.
>> 
>> Are there any thoughts why I observe such behaviour?
> 
> I believe this is a result of strange inference by the compiler.  You
> should double-check my explanation here, I'm not 100% sure that it is
> correct, but my gut feeling is that it is...
> 
> Armadillo does lazy evaluation at compile time, so if you do, e.g., mat
> + mat, the type returned is not 'mat' but instead a type that represents
> that two matrices should be added.  If you're interested I actually just
> gave a talk on this yesterday, here are the slides:
> 
> http://www.ratml.org/presentation.pdf <http://www.ratml.org/presentation.pdf>
> 
> Anyway, I think that the C++ compiler will infer strange things for
> 'auto' types, and as a result if you try to use auto with Armadillo, you
> can end up with bizarre results, and it's often better to avoid it if
> possible.
> 
> Hope this is helpful... let me know if I can clarify anything.
> 
> -- 
> Ryan Curtin    | "If it's something that can be stopped, then just try to stop it!"
> ryan at ratml.org <mailto:ryan at ratml.org> |   - Skull Kid

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20170627/e3bfc4d2/attachment-0001.html>