代写代考 NUMS 32 #endif – cscodehelp代写
Computer Graphics
Parallel Programming Project Notes
To a Better Grade
Copyright By cscodehelp代写 加微信 cscodehelp
project.notes.pptx
mjb – June 16, 2021
These notes are here to:
Why Are These Notes Here?
1. Help you setup and run your projects
2. Help you get the data you collect in the right format for submission
3. Help you get a better grade by doing all of this correctly! better grade!
better grade! better grade! better grade! better grade!
Computer Graphics
mjb – June 16, 2021
Project Notes, I
• Feel free to run your projects on whatever systems you have access to.
• If you don’t have access to your own systems, then you can use what we have at OSU. On-campus users will have access to Windows and Linux systems here.
Ecampus users will have remote access to our Linux systems, such as flip.
• Most of the projects will require timing to determine performance. Use the OpenMP timing functions. They give decent answers, and this will make the timing consistent across projects and across people. The OpenMP call:
double prec = omp_get_wtick( );
tells you the precision of the clock in seconds. I get 10-9 seconds on the systems I’ve been using. (I really doubt that this is true.) The OpenMP call:
double time0 = omp_get_wtime( );
samples the clock right now. It gives you wall-clock time in seconds. In parallel computing, memory latency and thread-idle time are part of the equation, so wall clock time is what you want.
Computer Graphics
mjb – June 16, 2021
How We Doing Timing
In this class, we don’t want to just implement – we want to characterize performance. What speed-ups do we see, and why do we see them? How do we generalize that to other types of problems? What insights does this give us?
So, as part of your project assignments, you will be doing a lot of timing to determine program speed-ups.
#include
double time0 = omp_get_wtime( ); // seconds …
double time1 = omp_get_wtime( ); // seconds
fprintf( stderr, “Elapsed time = %10.2lf microseconds
”, 1000000. * ( time1 – time0 ) );
Computer Graphics
mjb – June 16, 2021
%10.2lf is a good way to print doubles (“long float”)
5 This way of timing measures wall-clock time, which is really what we want to know in a
parallel environment, not CPU time.
However, this puts you at the mercy of the other users on the system. If you are on one of our public systems (e.g., flip), I advise you to check the system load to see how much off your wall-clock time measurement will be due to the competition from other users. Use the Linux uptime command::
flip01 34% uptime
11:13:37 up 96 days, 11:52, 23 users, load average: 3.56, 3.08, 2.82
These three numbers represent total CPU load averages for the last 1, 5, and 15 minutes respectively. If the CPU load average is greater than the number of CPUs, then each CPU is over-burdened.
Clearly you want these numbers, especially the 1-minute one, to be as small as possible when you run your test. If they are “big”, you might want to ssh to other systems (flip01, flip02, flip03, …) to see if you can find a better place to run, or try again later.
How Reliable is the Timing?
Computer Graphics
mjb – June 16, 2021
How Reliable is the Timing? A Useful Trick!
I like to check the consistency of the timing by computing both peak speed and average speed and seeing how close they are:
double maxmflops = 0.; double summflops = 0.;
for( int t = 0; t < NUMTRIES; t++ ) {
double time0 = omp_get_wtime( );
#pragma omp parallel for
for( int i = 0; i < ARRAYSIZE; i++ ) {
C[ i ] = A[ i ] * B[ i ]; }
double time1 = omp_get_wtime( );
double mflops = (double)ARRAYSIZE/(time1-time0)/1000000.; summflops += mflops;
if( mflops > maxmflops )
maxmflops = mflops;
printf( “Average Performance = %8.2lf MFLOPS
”, summflops/(double)NUMTRIES );
This is a reliable result:
Peak Performance = 1183.31 MFLOPS CoAmpvueterraGgraephPices rformance = 1141.41 MFLOPS
Peak Performance = %8.2lf MFLOPS
”, maxmflops );
You should record the peak performance value. This gives you as close to the best answer that you will get. But, compare that with the average performance value. That will tell you how reliable that peak value is.
This is an unreliable result:
Peak Performance = 627.39 MFLOPS
Average Performance = 294.86 MFLOPS
mjb – June 16, 202
Project Notes, II
If you are on Linux and have access to the Intel compiler, icpc, don’t use it unless we tell you to! (icpc is so good that it often does optimizations that undermine the very things you are testing.)
Useg++. Thecompilationsequencesare:
On Linux, the typical compile sequence for files that use OpenMP is:
g++ -o proj proj.cpp -O3 -lm -fopenmp
icpc -o proj proj.cpp -O3 -lm -openmp -align -qopt-report=3 -qopt-report-phase=vec
Note that OpenMP should always be included because we are using OpenMP calls for timing.
Note that the second character in the 3-character sequence “-lm” is an ell, i.e., a lower-case L. This is how you link in the math library.
Computer Graphics
mjb – June 16, 2021
Project Notes, III
• Most of these projects will require you to submit graphs. You can prepare the graphs any way you want, except for drawing them by hand. (The Excel Scatter-with-Smooth- Lines-and-Markers works well.) So that we can easily look at each other’s graphs, please follow the convention that up is faster. That is, do not plot seconds on the Y axis because then “up” would mean “slower”. Instead, plot something like Speedup or MFLOPS or frames-per-second.
• I expect the graphs to show off your scientific literacy — that is, I expect axes with numbers, labels, and units, If there are multiple curves on the same set of axes, I expect to be able to easily tell which curve goes with which quantity. After all, there is a reason this major is called Computer Science. Not doing this makes your project unacceptable for grading.
You lose points if you don’t do it this way.
Computer Graphics
mjb – June 16, 2021
Diverting to File
We all have a tendency to want to write our performance results out using printf (or cout) so that we can see them on the screen. That’s fine. But, then we want to get those results into a file. You could mess with file I/O, or you could use a divert.
If you are currently running your program like this:
and it prints to the screen via printf or cout, then running it like this: ./proj01 > output
will write your results into the file output.
( If you do it a second time, you will probably have to remove the previous output first. ) You can also divert the entire output of a looping script:
loop.csh > output
Computer Graphics
mjb – June 16, 2021
Importing into Excel – csv Files 10
csv stands for comma-separated values. It is a file format where you write your numbers out as text with commas between them. The great part is that Excel recognizes csv files and will read them in automatically.
Say you are using a printf that looks like this:
printf( “%2d threads ; %8d trials ; probability = %6.2f%% ; megatrials/sec = %6.2lf
”, NUMT, NUMTRIALS, 100.*currentProb, maxPerformance);
You probably did this because it looks really nice on your screen as you use this output to debug your program. But, now you want to change it to get the numbers into Excel quickly and painlessly. Comment out the old way and change it to this:
//printf( “%2d threads ; %8d trials ; probability = %6.2f%% ; megatrials/sec = %6.2lf
”, //NUMT, NUMTRIALS, 100.*currentProb, maxPerformance);
printf( “%2d, %8d, %6.2lf
”, NUMT, NUMTRIALS, maxPerformance );
This will now be printing just what you need into CSV format. You could divert it like this:
./proj01 > OUT.csv
loop.csh > OUT.csv
Which would then let you read the OUT.csv file right into Excel.
Computer Graphics
mjb – June 16, 2021
Computer Graphics
Importing into Excel – csv Files 11
mjb – June 16, 2021
12 In Excel, I have had the most success with creating graphs from tables that look like this:
where the 1,2,4,8 rows are holding the number of threads constant, and the 1, 10, 100, 1000, etc. columns are holding the dataset size constant. The cells are holding performance numbers, with higher numbers representing faster performance.
You will need to do some copying and pasting to get the linear format (from the previous page) into this 2D format, but it will be worth it!
Making Graphs
Computer Graphics
mjb – June 16, 2021
You can avoid all that copying and pasting by using an Excel feature called Pivot Tables! See the Pivot Table noteset.
Making Graphs
Then, sweep over the entire table
Computer Graphics
mjb – June 16, 2021
Making Graphs
Then, select Copy, and then insert it into one of the scatterplot options.
I like this one:
Computer Graphics
mjb – June 16, 2021
Here’s the Graph that Excel Produced from this Data
Computer Graphics
mjb – June 16, 2021
Adding Annotation
Number of Trials
Number of Cores
Computer Graphics
mjb – June 16, 2021
I like using PowerPoint for this, but you need to do it somewhere to get the proper graph into your project report.
MegaTrials Per Second
Transposing the Graph
To transpose the sense of the graph (which you also need to do), right-click on the border of the graph and then click on “Select Data”.
Then click on “Switch Row/Column”.
Computer Graphics
mjb – June 16, 2021
Here’s the Graph that Excel Produced from this Data
You might have to adjust the X axis number maximum and minimum.
Computer Graphics
mjb – June 16, 2021
Adding Annotation
Number of Cores
Number of Trials
Computer Graphics
mjb – June 16, 2021
I like using PowerPoint for this, but you need to do it somewhere to get the proper graph into your project report.
It’s the Same Data, but Each Graph Gives You a Different Insight into what the Data is Telling You
Computer Graphics
mjb – June 16, 2021
MegaTrials Per Second
This Data is actually a 3D Surface Plot –
The 2D Graphs are actually sets of 2D slices through the 3D Surface
Computer Graphics
mjb – June 16, 2021
Making Graphs
When we plot, we will all put execution performance on the Y axis (as opposed to putting elapsed time on the Y axis). Thus, as far as performance goes, up will mean “good”. So, for example:
Computer Graphics
mjb – June 16, 2021
Making Graphs
As you can tell, these performance measurements will be far more intelligible when examined as a graph than as raw numbers. Thus, you are expected to have access to a good automated graphing package. If you don’t have one or can’t get access to one – go get one!
Hand-drawn graphs, whether analog or digital, will not be accepted for your assignments.
You will also need a word processor, with a way to import your tables
and graphs, and with a way to turn that document into PDF.
Computer Graphics
mjb – June 16, 2021
Setting up Your Benchmarks to run from Scripts: #1 — the #define Approach
There are always advantages to not hardcoding constants into the middle of your program and, instead, setting them with a #define at the top where you can find that value and change it easily, like this:
Then, in the C or C++ program, all you have to do is use NUMT to set the number of threads, like this:
omp_set_num_threads( NUMT );
But, the use of the #ifndef/#endif construct has other advantages. It lets you either run this as a standalone program or run many occurrences of the program from a script.
Computer Graphics
mjb – June 16, 2021
#include
#ifndef NUMT
#define NUMT 8 #endif
#ifndef NUMS
#define NUMS 32 #endif
Setting up Your Benchmarks to run from Scripts: #1 — the #define Approach
In our project assignments, you will run benchmarks, that is, you will try your application using several different combinations of parameters. Setting these combinations by hand inside your program one-by-one is a time-consuming pain.
Your time is more valuable than that. Try doing it from a script.
In most C and C++ compilers, there is some mechanism to set a #define from outside the
program. Many of them use the –D construct on the command line:
Then, in the C or C++ program, all you have to do is use NUMT. For example:
omp_set_num_threads( NUMT );
This lets you automatically run your program 5 times with 1, 2, 4, 6, and 8 threads.
#!/bin/csh
#number of threads: foreach t (12468)
echo NUMT = $t
g++ -DNUMT=$t prog.cpp -o prog -lm -fopenmp ./prog
Computer Graphics
mjb – June 16, 2021
Setting up Your Benchmarks to run from Scripts: #1 — the #define Approach
You can also test multiple parameters from the same script by nesting the loops. This one is done using C Shell (csh):
#!/bin/csh
# number of threads: foreach t (12468)
echo NUMT = $t
# number of subdivisions:
foreachs(2 4 8 16 32 64 128 256 512 1024 2048 3072 4096)
echo NUMS = $s
g++ -DNUMS=$s -DNUMT=$t prog.cpp -o prog -lm -fopenmp ./prog
Computer Graphics
mjb – June 16, 2021
Or, in bash (Bourne-again Shell) …
#!/bin/bash
# number of threads: for t in 12468
echo NUMT = $t
# number of subdivisions:
forsin 2 4 8 16 32 64 128 256 512 1024 2048 3072 4096 do
echo NUMS = $s
g++ -DNUMS=$s -DNUMT=$t prog.cpp -o prog -lm -fopenmp ./prog
Computer Graphics
http://teespring.com
mjb – June 16, 2021
Or, in Python…
for t in[1,2,4,6,8]:
print “NUMT = %d” % t
for s in [ 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 3072, 4096 ]:
print “NUMS = %d” % s
cmd = “g++ -DNUMS=%d -DNUMT=%d prog.cpp -o prog -lm -fopenmp” % ( s, t ) os.system( cmd )
cmd = “./prog”
os.system( cmd )
Computer Graphics
mjb – June 16, 2021
Setting up Your Benchmarks to run from Scripts: #2 — the Command Line Arguments Approach
Instead of this: Do this:
#include
#ifndef NUMT
#define NUMT 8 #endif
#ifndef NUMS
#define NUMS 32 #endif
Then, in the C or C++ program, all you have to do is use NUMT to set the number of threads, like this:
omp_set_num_threads( NUMT );
But, the use of the #ifndef/#endif construct has other advantages. It lets you either run this as a standalone program or run many occurrences of the program from a script.
Computer Graphics
mjb – June 16, 2021
#include
int NUMT = 8; int NUMS = 32
argc and argv
When you write in C or C++, your main program, which is really a special function call, looks like this:
int main( int argc, char *argv[ ] ) {
These arguments describe what was entered on the command line used to run the program.
The argc is the number of arguments (the arg count)
The argv is a list of argc character strings that were typed (the arg vector).
The name of the program counts as the 0th argv (i.e., argv[0])
So, for example, when you type
in a shell, the ls program sees argc and argv filled like this: argc = 2
argv[0] = “ls” argv[1] = “-l”
Computer Graphics
mjb – June 16, 2021
argc and argv
So, if NUMT and NUMS are global int variables:
int NUMT = 2; int NUMS = 32;
and you want to set them from the command line, like this:
./prog 1 64
Then, inside your main program, you would say this: if( argc >= 2 )
NUMT = atoi( argv[1] );
if( argc >= 3 )
NUMS = atoi( argv[2] );
The if-statements guarantee that nothing bad happens if you forget to type values on the command line.
The atoi function converts a string into an integer (“ascii-to-integer”). If you ever need it, there is also an atof function for floating-point.
Computer Graphics
mjb – June 16, 2021
32 Also, remember, since NUMS is a variable, it needs to be declared as shared in the
#pragma omp line:
#pragma omp parallel for default(none) shared(NUMS,xcs,ycs,rs,tn) reduction(+:numHits)
NUMT does not need to be declared in this way because it is not used in the for-loop that has the #pragma omp in front of it.
shared( ) in the #pragma omp Line
Computer Graphics
mjb – June 16, 2021
Setting up Your Benchmarks to run from Scripts: #2 — the Command Line Arguments Approach
In our project assignments, you will run benchmarks, that is, you will try your application using several different combinations of parameters. Setting these combinations by hand inside your program one-by-one is a time-consuming pain.
Your time is more valuable than that. Try doing it from a script.
#!/bin/csh
g++ prog.cpp -o prog -lm –fopenmp
#number of threads: foreach t (12468)
echo NUMT = $t
./prog $t end
Then, in the C or C++ program, all you have to do is use NUMT. For example:
omp_set_num_threads( NUMT );
This lets you automatically run your program 5 times with 1, 2, 4, 6, and 8 threads.
Computer Graphics
mjb – June 16, 2021
Setting up Your Benchmarks to run from Scripts: #2 — the Command Line Arguments Approach
You can also test multiple parameters from the same script by nesting the loops. This one is done using C Shell (csh):
#!/bin/csh
g++ prog.cpp -o prog -lm -fopenmp
# number of threads: foreach t (12468)
echo NUMT = $t
# number of subdivisions:
foreachs(2 4 8 16 32 64 128 256 512 1024 2048 3072 4096)
echo NUMS = $s
./prog $t $s end
Computer Graphics
mjb – June 16, 2021
Or, in bash (Bourne-again Shell) …
#!/bin/bash
g++ prog.cpp -o prog -lm -fopenmp
# number of threads: for t in 12468
echo NUMT = $t
# number of subdivisions:
forsin 2 4 8 16 32 64 128 256 512 1024 2048 3072 4096 do
echo NUMS = $s
./prog $t $s done
Computer Graphics
http://teespring.com
mjb – June 16, 2021
Or, in Python…
cmd = “g++ prog.cpp -o prog -lm -fopenmp” os.system( cmd )
for t in[1,2,4,6,8]:
print “NUMT = %d” % t
for s in [ 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 3072, 4096 ]:
print “NUMS = %d” % s cmd=”./prog %d %d” % (s,t)
os.system( cmd )
Computer Graphics
mjb – June 16, 2021
I Don’t Recommend That You Put These Loops in the Program!
I know what you’re thinking.
You’re thinking:
Those scripts are a pain, and I’ve never done them before. So, I’ll just build the iterations through all the parameters into for-loops in the program.
Don’t! I have seen evidence that the first time OpenMP does anything, it also does some one-time setups. This will mess up your timing because your first test will seem slower than it should be and the others will seem faster than they should be.
I recommend you run the program separately for each combination of parameters. (The script code in the previous pages shows that.)
Computer Graphics
mjb – June 16, 2021
Computing Performance
When computing performance, be sure that the numerator is amount of work done and the denominator is the amount of time it took to do that work. For example, in the Bezier surface example, computing one height is the work done at each node and you have NUMS*NUMS total nodes, so (NUMS*NUMS)/dt is one good way to measure performance.
NUMS, NUMS*NUMS, 1./dt, and NUMS/dt are not good ways to measure performance as they don’t reflect the true amount of work done per time.
If you are using ridiculously high values for NUMS, the quantity NUMS*NUMS might overflow a normal 32-bit int. You can use a long int, or just float each one separately. Instead of (float)(NUMS*NUMS)/dt, you could say (float)NUMS*(float)NUMS/dt
If you are squaring a size number, and are using signed ints, the largest NUMS you can use is:
2,147, 483, 647 46, 340
If you are squaring a size number, and are using unsigned ints, the largest NUMS you can use is:
4, 294, 967, 295 65, 535
Computer Graphics
mjb – June 16, 2021
Project Turn-in Procedures
Your project turnins will all be electronic.
Your project turnins will be done at http://engr.oregonstate.edu/teach and will consist of:
1. Source files of everything (.cpp, .cl, .cuda)
2. A Linux or Windows executable file, if needed.
3. A report in PDF format. You can .zip everything else if you want, but please leave
the PDF as a separate file.
Electronic submissions are due at 23:59:59 on the listed due date.
Your PDF repor
程序代写 CS代考 加微信: cscodehelp QQ: 2235208643 Email: kyit630461@163.com
