Updated timer.hpp and timer.f90 to be consistent. Fleshed out time_ad…

…d gpu exercise.

OpenMP_cpu/adv_intro/2_timers/README_timer.md

@@ -1,47 +1,56 @@
-# Class Timer  -- collects time measurements and reports them later.
+# Class Timer_Collector  -- collects time measurements and reports them later.
 
 ## BACKGROUND
 Fortran and CPP Timer classes have been created for measuring performance
 of code blocks demarked by start() and stop() class member functions.
 For each block a literal character string argument for start("my label")
 is used as a label when the times are reported with print().
-
-See timer.hpp (CPP) and timer.f90 (F90)
-
+```
+      Declare Timer_collector as timer
+               CPP                                F90
+      timer.start("my label")             call timer%start("my label")
+      <something to be timed>             <something to be timed> 
+      timer.stop()                        call timer%stop()
+      ...                                 ...
+      timer.print()                       call timer%print()
+
+      See timer.hpp (CPP) and timer.f90 (F90)
+
+```
 ## Exercises
 
 1.) 
-    Look over time_print(.cpp|.f90).  The todo's guide you
+    Look over time_print(.cpp|.f90).  The TODOs guide you
     through instrumenting the code for timing 
 
     a sleep routine, and
 
-    -- That's just to make sure the timer is working.---
+    -- That is just to make sure the timer is working.---
 
     a write to stdout (CPP: printf, F90: print*,)
 
     TODO 1:
-    Include the Timer class file (CPP: timer.hpp, F90: timer.f90)
+    Include the Timer_collector class file (CPP: timer.hpp, F90: timer.f90)
     at the beginning of the code.
 
     TODO 2:
     Instantiate the class in main as timer:
-    CPP:  Just use "Timer time;"
+    CPP:  Just use "Timer_Collector timer;"
     F90:  Use the module defined in timer.f90 (mod_timer)
-          and define timer as type(cls_timer).
+          and define "timer" as type(Timer_Collector).
 
     TODO 3:
     Time the sleep and print statements:
-    Put      time.start("<label>") and      time.stop() before and after
-    Put call time%start("<label>") and call time%stop() before and after
+    CPP: Put      time.start("my label") and      time.stop() before and after
+    F90: Put call time%start("my label") and call time%stop() before and after
 
     TODO 4:
     Report the times at the end.
 
     Is the timer reporting reasonable numbers?
     What are the units?
+    How accurate is the timer?
     How long did the print statement take?
-    What is the resolution of the timer.
+    (Note, the print time is just to write to an io buffer.)
     How much time is use in calling the timer?
     How would you measure that?
-    (TODO, fortran needs to report more digits)

OpenMP_cpu/adv_intro/2_timers/solutions/ans_output

@@ -1,11 +1,19 @@
 CPP code
 
+Ft1 solutions $ ./a.out
  Hello Helsinki.
 
 Action           ::    time/s     Time resolution = 1e-06
 ------
- 1 sec sleep  ::    1.00009
- printf io    ::    5.7e-05
-
+ 1 sec sleep  ::     1.000986
+ printf io    ::     0.000023
 
 Fortran code
+
+  Hello Helsinki.
+
+Action           ::   time/s     Time resolution =  1.0E-06
+------
+ 1 sec sleep     ::     1.000825
+ printf io       ::     0.000171
+

OpenMP_cpu/adv_intro/2_timers/solutions/ans_time_print.cpp

@@ -5,8 +5,8 @@
 #include "timer.hpp"
 int main(){
 
-// TODO 1a: Instantiate a Timer class as timer.
-   Timer timer;
+// TODO 1a: Instantiate  a Timer_Collector class as timer.
+   Timer_Collector timer;
 
 // TODO 1b: Time a 1 second sleep
    timer.start(" 1 sec sleep ");

OpenMP_cpu/adv_intro/2_timers/solutions/ans_time_print.f90

@@ -4,9 +4,9 @@
 program main
 
 !! TODO 1a: Use the module in timer.f90 
-!!          and instantiate a type(cls_timer) class
-   use mod_timer
-   type(cls_timer) timer
+!!          and instantiate Timer_Collector (type) class as timer
+   use mod_Timer
+   type(Timer_Collector) timer
 
 !! TODO 1b: Time a 1 second sleep
    call timer%start(" 1 sec sleep ");

OpenMP_cpu/adv_intro/2_timers/time_print.cpp

@@ -5,7 +5,7 @@
 // #include ...
 int main(){
 
-// TODO 1a: Instantiate a Timer class as timer.
+// TODO 1a: Instantiate a Timer_Collector class as timer.
 // 
 
 // TODO 1b: Time a 1 second sleep

OpenMP_cpu/adv_intro/2_timers/time_print.f90

@@ -4,7 +4,7 @@
 program main
 
 !! TODO 1a: Use the module in timer.f90 
-!!          and instantiate a type(cls_timer) class
+!!          and instantiate Timer_Collector (type) class as timer
   !use  ...
   !type(...) timer
 

OpenMP_cpu/adv_intro/2_timers/timer.f90

@@ -4,7 +4,7 @@ module mod_timer
    private
    integer, parameter     :: m = 20
 
-   type, public :: cls_timer
+   type, public :: Timer_Collector
 
    private
    integer                         :: n = 0
@@ -19,13 +19,13 @@ module mod_timer
    procedure, public :: stop
    procedure, public :: print
 
-   endtype cls_timer
+   endtype Timer_Collector
 
    contains
 
    subroutine reset(this)
 
-   class(cls_timer) :: this
+   class(Timer_Collector) :: this
    this%n  = 0
    this%it = -1
    this%c  = 'undef'
@@ -35,7 +35,7 @@ subroutine reset(this)
 
    subroutine start(this, c)
 
-   class(cls_timer) :: this
+   class(Timer_Collector) :: this
    character(len=*) :: c
    character(len=16) :: c16
 
@@ -54,7 +54,7 @@ subroutine start(this, c)
 
    subroutine stop(this)
 
-   class(cls_timer) :: this
+   class(Timer_Collector) :: this
 
    call system_clock(this%it(2,this%n))
 
@@ -63,14 +63,14 @@ subroutine stop(this)
 
    subroutine print(this)
 
-   class(cls_timer) :: this
+   class(Timer_Collector) :: this
 
    write (*,*)
-   write (*,'(a,es7.1)') 'Action           ::   time/s     Time resolution = ', 1./real(this%itr)
+   write (*,'(a,es8.1)') 'Action           ::   time/s     Time resolution = ', 1./real(this%itr)
    write (*,'(a)')       '------'
    do i=1, this%n
-     write (*,'(a,a, f7.3)') this%c(i), ' :: ',               &
-                               (real(this%it(2,i) - this%it(1,i))) / real(this%itr)
+     write (*,'(a,a, f12.6)') this%c(i), ' :: ',               &  !TODO change .6 to .precision
+                              (real(this%it(2,i) - this%it(1,i))) / real(this%itr)
    enddo
 
    end subroutine

OpenMP_cpu/adv_intro/2_timers/timer.hpp

@@ -2,44 +2,47 @@
 #define SIMPLE_TIMER_H
 
 #include <iostream>
+#include <iomanip>    //setprecision
 #include <cmath>
 #include <string>
 #include <sys/time.h>
+using namespace std;
 
 /* simple timer that stores multiple times 
    and can print them afterwards.
    Does not support nested timing calls. 
 */
-class Timer
+class Timer_Collector
 {
 public:
-    Timer(): n(0) { }
-    void start(std::string label) 
+    Timer_Collector(): n(0) { }
+    void start(string label) 
     { 
         if (n < 20) { labels[n] = label; gettimeofday(&times[2*n], NULL); }
-        else { std::cerr << "No more timers, " << label << " will not be timed." << std::endl; }
+        else { cerr << "No more timers, " << label << " will not be timed." << endl; }
     }
 
     void stop() { gettimeofday(&times[2*n+1], NULL); n++;}
     void reset() { n=0; }
     void print();
 private:
-    std::string labels[20];
+    string labels[20];
     timeval      times[40];
     int n;
 };
 
-void Timer::print()
+void Timer_Collector::print()
 {
-    std::cout << std::endl;
-    std::cout << "Action           ::    time/s     Time resolution = " << 1.f/(float)CLOCKS_PER_SEC << std::endl;
-    std::cout << "------" << std::endl;
+    cout << endl;
+    cout << "Action           ::    time/s     Time resolution = " << 1.f/(float)CLOCKS_PER_SEC << endl;
+    cout << "------" << endl;
     for (int i=0; i < n; ++i)
     {
-        time_t seconds = times[2*i+1].tv_sec - times[2*i+0].tv_sec;
-        suseconds_t ms = times[2*i+1].tv_usec - times[2*i+0].tv_usec;
-        if (ms < 0) { ms += 1000000; seconds--; }
-        std::cout << labels[i] << " ::    " << (float)seconds + ms/1000000.f << std::endl;
+        time_t seconds = times[2*i+1].tv_sec  - times[2*i+0].tv_sec;
+        suseconds_t us = times[2*i+1].tv_usec - times[2*i+0].tv_usec;
+        if (us < 0) { us += 1000000; seconds--; }
+        double secs = (double)seconds + us/1000000.0f;
+        cout << labels[i] << " :: " << setiosflags(ios::fixed)<<setprecision(6)<< setw(12)<<secs<<endl;
     }
 }
 #endif

OpenMP_gpu/adv/time_add/README_time_add.md

@@ -1,10 +1,45 @@
-# This exercise is incomplete.
+# Time Offload kernel for vector add, and data motion
+
+## Background
+
+By putting timers around the "target data" directive"
+and immediately after the target data region begind,,
+the "target" construct, and the termination of the target
+data region, one can measure the time required to 
+
+```
+   (allocate device data and) move data "to" the device, 
+
+   execute and kernel, 
+
+   and move data "from" the device (and deallocate).
+```
 
 ## Exercise
-Only C/C++ code is available at this time
-1.) Experiment with the target ... directives 
-    executing the addition.
-    Some different setting are in comments.
+    STATUS:  Makefile works for CPP only at this time.
 
+1.) Add the start and stop timers around the appropriate
+    areas, and run with the default (present) "target"
+    construct to offload the add function.  Make sue it works.
+    See previous examples (basic/worksharing) for including a timer:
+    TODO 1:
     Note, the data motions due to the target data
-    construct are measured separately.
+    construct are measured separately at the beginning and end
+    of the target data region.
+
+    TODO 2:  2a target ENCLOSING function, 2b constructs IN function
+    Change 1<<25 to 1<<28 for C/C++; 2**25 to 2**28 for Fortran for timing.
+
+    Note: For the "add" procedure, the target and the target teams can be
+    hoisted outside the function.  The different forms are labeled with
+    F1, F2, F3, and F4.  Experiment with (time) the differents forms,
+    and determine which one works best with defaults, and then experiment
+    with the number of teams and thread limit for the optimal version.
+    Which one has the highest performance.
+
+    TODO 3:
+    Now try running the Loop version (which doesn't
+    call the function) L1, L2, L3, with various different 
+    "target" directives (number of teams and thread limit).
+
+    Which one has the highest performance.

OpenMP_gpu/adv/time_add/add.F90

@@ -0,0 +1,66 @@
+include "timer.f90"
+
+subroutine add(n, x, y)
+  implicit none
+  integer :: n,i
+  real    :: x(n),y(n)
+
+    !! TODO 2b:
+!F1 !$omp target teams distribute parallel for num_teams(108) num_threads(256) !!1 .4s
+!F2 !$omp        teams distribute parallel for num_teams(108) num_threads(256) !!2 mins
+!F3 !$omp              distribute parallel for                num_threads(256) !!2 .1s
+    !$omp              distribute parallel for !!             num_threads(256) !!2 .1s
+    do i=1,n;  y(i)=x(i)+y(i);  enddo
+
+end subroutine 
+
+program main
+  use mod_Timer
+  implicit none
+
+  type(Timer_Collector) timer
+
+ !integer,parameter :: N=2**28 !! use 28 for timing, 1024*1024*2**8
+  integer,parameter :: N=2**25 !! use 25 for code validation,  1024*1024*2**5
+  real    :: x(N), y(N), error
+  integer :: i
+
+
+  do i=1,n;  x(i)=1.0e0; y(i)=1.0e0; enddo  !init x and y on host
+
+  call timer%start(" Data TO   & Alloc  "); !! TODO 1
+    !$omp target data map(tofrom:y) map(to:x) !allocate and move data
+  call timer%stop();                        !! TODO 1
+
+  call timer%start(" Add on GPU         "); !! TODO 1
+
+    !! TODO 3:
+    !!    Time Loop on Host
+    !!#pragma omp target teams distribute parallel for num_teams(108) num_threads(256)
+    !!do i=1,n;  y(i)=x(i)+y(i);  enddo
+
+    !! TODO 2a:
+    !!    Time function call with "target" "target teams and num_teams/thread_limit clauses
+    !F1 -- no construct here                      !!1 .4s
+    !F2 #pragma !$omp target                      !!3  mins
+    !F3 #pragma !$omp target teams num_teams(108) !!2 .1s
+    !F4 #pragma !$omp target teams thread_limit(256) !! num_teams(108) !!2 .008s
+    !$omp target teams thread_limit(512) !! num_teams(108) !!2 
+    call add(N,x,y);
+
+  call timer%stop();                        !! TODO 1
+
+  call timer%start(" Data FROM & Dealloc"); !! TODO 1
+    !$omp end target data 
+  call timer%stop();                        !! TODO 1
+
+  do i=1,N
+     error = max( 0.0e0, abs(y(i)-2.0e0))
+     if ( error > 0.0000001 ) &
+        print*, " error e10-6 exceeded at ",i," val= ",error
+  end do
+  print*, "Max error = ", error, " for count of ", N
+
+  call timer%print();                       !! TODO 1
+
+end program main