Changes between Version 41 and Version 42 of Next-GenOpenMPTransformation


Ignore:
Timestamp:
10/30/19 11:18:43 (7 years ago)
Author:
wuwenhao
Comment:

--

Legend:

Unmodified
Added
Removed
Modified

  • Next-GenOpenMPTransformation

    v41 v42  
    2525  * `shared`
    2626 * `num_threads(`n`)`
    27  * `reduction(op:list)`
     27 * `reduction(op:list)` ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#Case1:parallelreduction detail])
    2828
    2929* `sections` ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#Translatingsections detail])
    … …  
    4646 * `firstprivate(list)`
    4747 * `lastprivate(list)` . . . . .[Unsupported]
    48  * `reduction`
     48 * `reduction` ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#Case2:forreduction detail])
    4949 * `schedule` . . . . . . . . . . . . [Unsupported]
    5050 * `collapse`
    … …  
    8282* `master` ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#Translatingmaster detail])
    8383
    84 * `barrier`
    85 
    8684* `critical` . . . . . . . . . . . . . . . . [Unsupported]
    8785 * `[name]`
     86
     87* `barrier`
    8888
    8989* `atomic` . . . . . . . . . . . . . . . . . . [In-progress]
    … …  
    9595 * `seq_cst` . . . . . . . . . . . . . [Unsupported]
    9696
     97* `ordered` . . . . . . . . . . . . . . . . . .[Unsupported]
    9798
    9899== OpenMP Types ==
    99 * `omp_lock_t` . . . . . . . . . . . . . . .[Unsupported]
     100* `omp_lock_t` . . . . . . . . . . . . . . . [Unsupported]
    100101
    101102== OpenMP Functions ==
    … …  
    576577([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    577578
    578 === Translating `reduction` clause ===
    579 
     579=== Translating `master` ===
     580
     581{{{
     582#pragma omp master
     583S
     584}}}
     585
     586=>
     587
     588{{{
     589if (tid == 0) {
     590  translate(S);
     591}
     592}}}
     593
     594([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     595
     596=== Translating `barrier` ===
     597
     598{{{
     599#pragma omp barrier
     600}}}
     601
     602=>
     603
     604{{{
     605$omp_barrier_and_flush(team);
     606}}}
     607
     608([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     609
     610=== Translating `critical` ===
     611
     612Basically, use a lock for each critical name, plus one for the "no name".  All threads must obtain lock to enter the critical section, then release it.
     613I.e., if there are critical sections name a, b, and c, there should be global root-scope variables of boolean type named `_critical_noname`, `_critical_a`, etc.
     614
     615{{{
     616#pragma omp critical(a)
     617S
     618}}}
     619
     620=>
     621
     622{{{
     623...
     624_Bool _critical_a = $false;
     625.
     626.
     627.
     628$when (!_critical_a) _critical_a=$true;
     629translate(S);
     630_critical_a=$false;
     631}}}
     632
     633([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     634
     635=== Translating `atomic` ===
     636
     637In general, reads and writes to shared variables will be processed using the protocols described above.  However if the operation occurs within an omp atomic construct, it is translated differently.
     638
     639TODO: need to look up the rules on the different flavors of atomics.
     640
     641If sequentially consistent atomic...
     642
     643If non-sequentially consistent atomic...
     644
     645([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     646
     647===  Translating`ordered` ===
     648
     649This can only be used inside and OMP `for` loop in which the pragma used the `ordered` clause.  (Check that.)  It indicates that the specified region must be executed in iteration order.
     650
     651In this case the system function must return an int iterator in which the ints occur in loop order.
     652
     653{{{
     654#pragma omp for ordered
     655for (i=a; i<b; i++) {
     656  ...
     657  #pragma omp ordered
     658  S1
     659  ...
     660  #pragma omp ordered
     661  S2
     662  ...
     663}
     664}}}
     665
     666=>
     667
     668{{{
     669{
     670  $domain loop_domain = {a..b};
     671  $domain(1) my_iters = ($domain(1))$omp_arrive_loop(team, FOR_LOC++, loop_domain, STRATEGY);
     672  int order1=a, order2=a;
     673
     674  $for (int i : my_iters) {
     675     ...
     676    $when (order1==i) {
     677      translate(S1);
     678      order1++;
     679    }
     680      ...
     681    $when (order2==i) {
     682      translate(S2);
     683      order2++;
     684    }
     685      ...
     686  }
     687}
     688}}}
     689
     690([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     691
     692=== Translating `nowait` ===
     693
     694Just leave out the `$omp_barrier_and_flush` at the end of the translated construct.
     695
     696([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
     697
     698=== Translating `reduction` ===
    580699==== Case 1: parallel reduction ====
    581700{{{
    … …  
    664783
    665784([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    666 
    667 === Translating `master` ===
    668 
    669 {{{
    670 #pragma omp master
    671 S
    672 }}}
    673 
    674 =>
    675 
    676 {{{
    677 if (tid == 0) {
    678   translate(S);
    679 }
    680 }}}
    681 
    682 ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    683 
    684 === Translating `barrier` ===
    685 
    686 {{{
    687 #pragma omp barrier
    688 }}}
    689 
    690 =>
    691 
    692 {{{
    693 $omp_barrier_and_flush(team);
    694 }}}
    695 
    696 ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    697 
    698 === Translating `critical` ===
    699 
    700 Basically, use a lock for each critical name, plus one for the "no name".  All threads must obtain lock to enter the critical section, then release it.
    701 I.e., if there are critical sections name a, b, and c, there should be global root-scope variables of boolean type named `_critical_noname`, `_critical_a`, etc.
    702 
    703 {{{
    704 #pragma omp critical(a)
    705 S
    706 }}}
    707 
    708 =>
    709 
    710 {{{
    711 ...
    712 _Bool _critical_a = $false;
    713 .
    714 .
    715 .
    716 $when (!_critical_a) _critical_a=$true;
    717 translate(S);
    718 _critical_a=$false;
    719 }}}
    720 
    721 ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    722 
    723 === Translating `atomic` ===
    724 
    725 In general, reads and writes to shared variables will be processed using the protocols described above.  However if the operation occurs within an omp atomic construct, it is translated differently.
    726 
    727 TODO: need to look up the rules on the different flavors of atomics.
    728 
    729 If sequentially consistent atomic...
    730 
    731 If non-sequentially consistent atomic...
    732 
    733 ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    734 
    735 ===  Translating`ordered` ===
    736 
    737 This can only be used inside and OMP `for` loop in which the pragma used the `ordered` clause.  (Check that.)  It indicates that the specified region must be executed in iteration order.
    738 
    739 In this case the system function must return an int iterator in which the ints occur in loop order.
    740 
    741 {{{
    742 #pragma omp for ordered
    743 for (i=a; i<b; i++) {
    744   ...
    745   #pragma omp ordered
    746   S1
    747   ...
    748   #pragma omp ordered
    749   S2
    750   ...
    751 }
    752 }}}
    753 
    754 =>
    755 
    756 {{{
    757 {
    758   $domain loop_domain = {a..b};
    759   $domain(1) my_iters = ($domain(1))$omp_arrive_loop(team, FOR_LOC++, loop_domain, STRATEGY);
    760   int order1=a, order2=a;
    761 
    762   $for (int i : my_iters) {
    763      ...
    764     $when (order1==i) {
    765       translate(S1);
    766       order1++;
    767     }
    768       ...
    769     $when (order2==i) {
    770       translate(S2);
    771       order2++;
    772     }
    773       ...
    774   }
    775 }
    776 }}}
    777 
    778 ([https://vsl.cis.udel.edu/trac/civl/wiki/Next-GenOpenMPTransformation#OpenMPConstructs back to top])
    779 
    780 === Translating `nowait` ===
    781 
    782 Just leave out the `$omp_barrier_and_flush` at the end of the translated construct.