National Energy Research Scientific Computing Center (NERSC) Codee Analyzer Training, April 2022, 12 May 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: 06 Codee identification of defects in parallel code using MATMUL

Description

No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).

A

If you remember, we are going to use the matrix matrix multiplication example. The same implementation we use today and we are going to use this example to identify defects in the data transfers of our openmp or open sec codes. What this means is that in the implementation code, we have a part of the main section that is allocating double pointed data structure.

A

This means that it is using in this simply in this code, an auxiliary matrix that first allocates the array of pointers and then makes individual separate allocations of memory for each of the rows. So in the end, we have this double pointer data structure.

A

From the point of view of the computations, they look like plane, arrays, similar to fortran.

A

It seems that the data is consecutive in memory, but it is actually not so if we try to use, modify the matrix multiplication example. Just by adding this pragma pragma omp target teams distribute parallel. Target teams, distribute parallel 4 to make the data transfer and the parallelization of the computation on the gpu, and we need. We need to specify what to map from the cpu memory to jpm memory.

A

So here in an intuitive manner, we can sync that mapping pointers a b and c to the memory of the gpu and back from the memory of the gpu. The result in the in pointer see this should be the the specification and the implementation of openmp should be smart enough to navigate through all the data transfer it to the gpu memory and transfer it back. But again, this there is a hidden problem of tip copy here.

A

So we you will be seeing in this lab is that by following exactly the same process, you will get the performance, optimization report. You will get the entry level report you will get in the list of actions, a defect in this case defect number six, which we encourage you to go to the knowledge base and see how it is described and explained.

A

And then you compile and execute this code with the defect, and you see that the code will break will fail to execute because it is accessing a legal uh addresses on the in memory. It's not accessing the correct data and memory. Okay. So- and this is all we wanted to say, you have all the data in the same repository than today and yesterday, the same set of files, and you already have the source code and the launch script with this implementation.

A

So you just need to go through the usage of kodi, reproduce the steps in the slides and see how it identifies the defects, compile it run the code and see that the code breaks is, has an illegal execution.

B

And, and for today's exercises we actually have updated example codes. There are more lab lab6 lab eight and fortran much more details in the uh mirrors april. 2022 example codes so grab a new. um Do a new copy, so I'll put the link as s copy command in chat the new directory. So you get updated, exercises and rename the old one to day one.

B

Meanwhile, I think uh it's a good idea, maybe oldest, can do a demo for the defect. One.

C

For defect six, I I think you mean.

C

So first, let's see the code, but it's already shown a little slice.

C

C

So it's the regular man mood, but it has this fragma, this openmp pragma that it has an error in the map: um klaus klausel. um So let's analyze it with kodi and see what it reports.

C

The entry level report is always evaluation.

C

Which is the default flag? So if you don't put anything, it will use evolution anyway. So italian report evaluation.

C

And as default, we have multithreading and offloading disabled, but we haven't. We can enable them sorry, okay, for projects that use libraries and include paths that are not the current path. You can use dash dash to pass compiler arguments in this case. We need to pass dash, I and input that is the folder, where the input path is so now right. Now it's successfully analyzed and we have to enable the multituding and offloading with uh include text all the we have.

C

We have to follow the third suggestion we can copy basic command invocation, and now we have two offloading actions, uh two actions related with offloading. We can list them with dash dash actions.

C

And now the tool will tell us to use uh to focus the analysis on on a function or a loop, in this case, we're going to analyze the main function.

C

Sorry, there's no, my friendship. Yes,.

C

C

And here we have the defect: six for the line: 15.

C

missing, deep copies of non-contiguous arrays, a b and c and data transferred to accelerator device. So here cody is detecting the the error located at the pragma. So now we will try to compile it and we will see that we'll get an error.

C

Okay, let's do it manually.

C

And here we have the error, so it is compiling it correctly, but it has an execution time error.

C

But manuel, do you want to you want me to generate the correct practice.

C

Or here showing the pwd6 is fine.

A

I think it's not really needed, because that is that was covered in the lab2 yesterday. So here we want to point out that kt can detect openmp openscc programmers that seem to be correct, but for some reason for some problem hidden like deep copy, it will make our codes break. So I think it's enough helen. Do you think? Also it's enough.

B

Yeah, uh anyway, lots of questions. I would suggest uh dude as patched around this, but this. Otherwise, this one is running on the logging note, but um you'll see the same error on the gpu node as well.

A

So maybe we can use the remaining time before the break to know that ulysses is sharing his screen. Maybe we can start with a fortran demo, yes, and then.

C

A

Right before the breakup right.

C

After the breaking the.

A

C

So yesterday we analyzed pi in a c implementation of pi using kodi and now we're going to do the same, but with a fortran implementation. So let's see it with.

C

So it's the same c implementation, but with fortran with this subroutine calculate p, sequential implementation, simple and so the entry level tool again is pw report evaluation.

C

And like yesterday, two actions and multiplying and offloading disabled we can enable it with includes excel.

C

Now for actions when multithing and one of loading, we will see them using vw report dash dash actions.

C

And we have to focus the analysis on a function. In this case we say this calculate b.

C

C

And here we have the three opportunities opportunity for multi-threading, cd, vectorization and offloading, and the cmd vectorization opportunity is associated with this remark. 11 at least telling us feedback from the vectorization cost model that cody has. But now we are, we will focusing on offloading opportunity. We will um increase the level of detail of the analysis using the flag dash dash level two, as it is suggested in the first session, and here we can see suggestions made for the opportunity three.

C

We have two beta directors invocations such suggested and one for openmp for offloading and one for openacc.

C

We are going to involve both, but we are going to change the last flag, so we generate the the output in another five, in this case, pi r, p off.

C

And we're going to do the same for the open, acc.

C

Yes, and now, uh as fortran support is experimental for kodi. For now we detected a bug and the generated code for openmp offloading, but we can easily fix it manually in this case, vim.

C

By rb, basically, we need to change the four, uh the four key, that is for the the c, the bracket for c, and we have to change it with do and we have to delete the schedule static. That is failing with np fortran, compiler and.

C

I want now, let's see the acc version generated by peter directives in this case. It's okay and we don't have to change anything.

C

Here are the pragmas otherwise, directives of openhcc for offloading.

C

Okay and now we can run it.

C

Okay, here it is so the sequential version executed at three seconds: the openmp first loading uh 0.17 seconds and the open acc version 0.14 seconds and the results are okay. So this is how we analyze a fortran code, in this case pi using kodi.

C

I am another, we wanna add something.

A

uh No, no, I think it's pretty clear that the same workflow we have proposed since yesterday for all the codes also works for fortune, which is what, as a user, we might expect calen. Would you like to point something or repeat something.

B

Something, let's say that fortune support is experimental, as you can see, when you do the pw directives. There's a first line says warning fortune. Support is experimental. That's why, um once this is finished, you have to make some changes to the generated code. One of them is apparel. Four has to change the power of do another. One is related, possibly a compiler bug that schedule auto. um This was not supported for 4chan, but I was for um not scheduled, auto schedule.

B

Anything schedule auto has to be changed to schedule static for for the c code, but the same schedule doesn't work for the fortran compiler, so that's potential compiler bug, but otherwise the 4chan support will be upcoming in much more mature in the future releases. I hope, because we have lots of fortune users at nursk and oak ridge.

A

Yes, indeed, we will also, probably at the end of our sliding presentation, we will show you very briefly the the current roadmap we have today, so that you can see how we are somehow prioritizing fortran for this yeah.

B

But this year this is very encouraging because um for for your development workers, there was lots of dependencies for fortran um fortune dependencies that for um for your tour development, work, the the rvm and flan stuff like that yeah. um So there is also a madmo unfortune code that people can play with it as well.

A

We can also see it in a fast demo in okay and then we can continue with the presentation of the final lab and then three hands on until the end take off a break and yet and they can decide if they want to work and we will be assisting or they prefer to make a break so lisa. Yes, yes,.

C

So let's see the code uh matthew fortune.

C

So it's this is our routine. That is calculating the matrix multiplication, so calculate my mole and we use the entry level tool to make the first report bw.

C

And now we have um eight actions, but we want to see again the offloading action, so we have to enable them with includes excel.

C

Two offloading opportunities: let's see them with pw report actions, actions and remember that in dash dash actions mode, you have to focus the analysis on on a specific function or loop. In this case, we said is calculate math move.

C

C

And well here we have three opportunities: sorry, two opportunities, one for multi-threading, one for offloading and a remark from the vectorization cost model. Again the study stating that the scene is not a single opportunity due to strided memory, access in the loop body. But again we will focus on opportunities for offloading.

C

We will increase the level of detail again with dash dash level two.

C

And here we have the pw directive, suggestions, one for openmp offload and another one for openacc we're going to invoke both of them first dash o, instead of in place and.

C

It generated it correctly and now, let's do the same for open, acc.

C

Everything fine.

C

Let me check the script.

C

Yes, I choose the right name. So now, let's see the the culturated automatically, but with w directives.

C

Here we can see the pragma for offloading. This is an openmp blackmass.

C

Remember that we have to change this four for a two and we had to remove this schedule static manually. So.

C

C

Let's see the open acc version.

C

And here they are the pragmas for offloading with open acc generated automatically and again they are they work fine, so we don't have to change them manually and now, let's run it.

C

C

Okay, I forgot to make the ch mode classic. We have to give the execution permission to this. The scripts.

C

C

Okay, here we have the execution time this. This removes our failing because we didn't have the binaries created before, but it just ignored this first three lines had the execution time for the serial version, the execution time for the openmp version and the time for the openacc.

C

We can see clearly an improvement.

C

So manuela wanna add something.

A

No, I think it is quite clear the same workflow again and.

A

And the it and remark that we are really interested in providing support for fortran as soon as possible. So we are working very very actively on that.