Hello Everyone,

Hopefully this comment has the more legible line spacing, word grouping, bold, code section, etc.

From https://devblogs.microsoft.com/cppblog/simd-extension-to-c-openmp-in-visual-studio/

Hello, I have two topics to understand customer requirements. The easy statement describing OpenMP user requirements, and pleasing us 100%, is to fully implement support of OpenMP 4.5 and/or newer. If that functionality is too large to release in the next 30 to 60 days, then all of the clauses above are a good start:

simdlen(length) : specify the number of vector lanes
safelen(length) : specify the vector dependency distance
linear(list[ : linear-step]) : the linear mapping from loop induction variable to array subscription
aligned(list[ : alignment]): the alignment of data
private(list) : specify data privatization
lastprivate(list) : specify data privatization with final value from the last iteration
reduction(reduction-identifier : list) : specify customized reduction operations
collapse(n) : coalescing loop nest The company I work for will then verify functionality and use accordingly.

Regarding the user defined reduction details, the best performance I have so far with a single dimension array size range from as few as ten elements, to as large as one billion elements, is the combination of the parallel and simd directives in a single pragma:

#pragma omp parallel for simd  default(none) firstprivate(theMaxValue) shared(pArrayOfDoubles)  reduction(theReductionName : instanceNameOfUserStructure)


followed by a  normal looking for loop with a size_t loop control variable and  condition checking by i < theMaxValue

When I separate the parallel and simd into separate pragma directives on separate source code lines, and use scoping braces accordingly, the total time increases. If I use only parallel or only simd, then the total time increases more than the separated parallel simd combination.

In other words, supporting only one aspect of the pragma directive, for example simd, will likely not be as fast as possible. Supporting the combination of the parallel and simd directives in the single pragma line above is preferred.

Users need multiple OpenMP clauses to correctly determine which aspect(s) will have the best performance. It is understood that the performance test results will vary between array sizes and algorithm implementations. Plus supporting all of the various ways to use simd, parallel, and for allows users to test and select the directives and clauses that perform best.

Within the instanceNameOfUserStructure above are two data members. One of them happens to be a size_t, tracking the index position of the reduction search. The other data element is the value at that index position.

Implementing user defined reduction functionality this way, then functions within the deprecated Intel cilk library (e.g. Intel Cilk Plus is being deprecated) are present from OpenMP user defined reduction. Additional details on the user defined reduction functionality can be found in the book titled Using OpenMP - The Next Step by Ruud van der Pas, Eric Stotzer, and Christian Terboven, section 2.4.3, MIT Press, 2017, 978-0-262-53478-9 (use no dashes when searching the Internet).

Sincerely,

sim-mse

Developer Community

Get help from our community supported forum

[OpenMP] Support #pragma omp parallel for simd

Add comment

Related Ideas