Interleaved computation with communication in halo exchange #881
Conversation
| return; | ||
| } | ||
|
|
||
| assert(aligned(dr)); |
There was a problem hiding this comment.
makes no sense checking this for one range, not aligned can be two or more ranges, one is always aligned
|
|
||
| assert(aligned(dr)); | ||
|
|
||
| for (auto &s : dr::ranges::segments(dr)) { |
There was a problem hiding this comment.
probably you wanted local_segments
| void fence() { backend.fence(); } | ||
| void fence() { backend_.fence(); } | ||
|
|
||
| backend_type& backend(const std::size_t segment_index) { return backend_; } |
There was a problem hiding this comment.
add __attribute__((unused)) to these functoins, but I'm unsure this function is really needed
|
|
||
| static constexpr std::size_t DUAL_SEGMENTS_PER_PROC = 2; | ||
|
|
||
| class DualMpiBackend { |
There was a problem hiding this comment.
what is the difference between DualMPiBackend and MpiBackend types? if none, please use one type
There was a problem hiding this comment.
It's a leftover from when I was experimenting with changing some code in the backend, thanks for pointing it out
|
|
||
| distribution distribution_; | ||
| std::size_t size_; | ||
| std::vector<dual_dv_segment<dual_distributed_vector>> segments_; |
There was a problem hiding this comment.
if these 2 lines are the only differences between ordinary and dual vector, then let's pass different template parameters and make segment type a template paramter
There was a problem hiding this comment.
init() is different, also there are multiple pointers to local segments (std::vector<T *> datas_) and there are additional member functions that return the local segment currently suited for computation. Although I agree that the two classes might be merged, as long as it's still not fully developed I'd rather keep it split and merge them a bit later when it fully works.
| for (std::size_t i = 0; i < DUAL_SEGMENTS_PER_PROC; i++) { | ||
| if (size_ > 0) { | ||
| datas_.push_back(static_cast<T *>( | ||
| backends_[i].allocate(data_size_ * sizeof(value_type)))); |
There was a problem hiding this comment.
do I understand correctly that you allocate twice as many memory as normal distributed_vector of the same size?
if so - this is unacceptable
There was a problem hiding this comment.
I think this isn't the case -- note the added multiplication by DUAL_SEGMENTS_PER_PROC:
std::size_t segment_count = comm_size * DUAL_SEGMENTS_PER_PROC;
This results in segment_size_ being smaller, so the memory footprint should only come from more halos.
No description provided.