1.c.-OpenMP-directives.md

@@ -10,6 +10,50 @@ NOTE: This wiki does not contain all information, please also read the [specific
 
 ## Parallel directives
 
+These directives can be used for invoking GPU parallelization.
+
+**target**  
+Declares portion of code to be executed on GPU.
+
+**teams**  
+With teams, the code runs on GPU using multiple "teams".
+
+- Invokes fork/join model
+- No collective synchronization! (i.e. barriers)
+- Must be within "omp target"
+
+**distribute**  
+Iterations of the loop below are partitioned across "teams".
+Without, only thread 0 of each team would be used.
+
+- Loop partitioning (like omp do/far)
+- Iterations are partitioned across "teams"
+- NO implied barrier at end of loop!
+- Best practice to combine with teams (target teams distribute)
+
+**simd**  
+Optimize internal functions with SIMD instructions if possible. This directive gets ignored if this is not possible, thus it can always be used.  
+Maps the work below to GPU threads within the "teams" blocks.  
+
+- Used for two-level GPU parallelism
+- "teams" maps to GPU threadblocks
+- "simd" maps to GPU threads within the teams
+- According to Oak Ridge, always use in combination with parallel do.
+
+! NOTE - parallel and simd are inconsistent across implementations !
+
+- CCE-Classic maps "simd" to GPU threads and skips "parallel for"
+- Clang maps "parallel for" to GPU threads and skips "simd"
+- This will change for CCE16, where "parallel" will have the function that "simd" has in CCE15.
+
+**parallel do**  
+Maps the kernel to the available threads.
+Most usefull in combination with "teams" blocks.
+Causes the work done in a loop inside a parallel region to be divided among threads.  
+
+**loop**  
+Needs to be bounded to the teams with bind(teams) or is done implicitely when called inside a "teams" region.
+Basically does a *parallel do simd* internally? But it is newer.
 
 ## Memory directives.