ConstantIterator

[AntonReinhard]

• Add ConstantIterator class
• Yields the same value it was initialized with every time
• Add testing for the class

[SimeonEhrig]

@bernhardmgruber Can you please review it.

[SimeonEhrig]

If the whole point of this iterator is to just return the same value whenever dereferences, I see no need to store an index. Then all iterator functions that move the iterator become noops. When do you need the index of the iterator?

You can probably make all member functions of ConstantIterator constexpr and noexcept.

The index is required for begin and end pointer:

ConstantIterator c_begin(10);
ConstantIterator c_end = c_begin + 4;

for(; c_begin < c_end; ++c_begin){
   std::cout << *c_begin << " "; // 10 10 10 10
}

Do you have a better idea?

[bernhardmgruber]

The index is required for begin and end pointer:

ConstantIterator c_begin(10);
ConstantIterator c_end = c_begin + 4;

for(; c_begin < c_end; ++c_begin){
   std::cout << *c_begin << " "; // 10 10 10 10
}

No, makes perfect sense now :D Without the index you would have an infinite range.

[SimeonEhrig]

@AntonReinhard Can you please rebase to the current master. I removed the C++ 14 support. New default is C++ 17.

[AntonReinhard]

Ah, I knew you had to include for the spaceship operator: https://godbolt.org/z/soa55sP73 Apparently gcc somehow handles it without, but clang needs it, because the return type of the defaulted operator<=> is the std::strong_ordered type, which is defined in

[SimeonEhrig]

Two points:

• You need to rebase on the this commit, otherwise the constant iterator will not work on CPU: dbc5d3a1653c8ae930c643d192ade66c070d17ef
• We cannot use the CI for benchmarking. All resources are virtualized and shared. Therefor the results are not reproduceable.

Can you execute your results on a local system or Taurus? If not, we have a Quadro V100 in our dev system. I can execute the benchmarks and send you the results.

[SimeonEhrig]

By the way, I was aware about the wrong results on the serial backend in the reduce benchmark. I think, it is a precession problem of float and the many summations on a single value. But I had no time yet to investigate this.