In this episode, Conor and Bryce record live from C++ Under the Sea! We interview Ray and Paul from NVIDIA, talk about Parrot, scans and more! Link to Episode 260 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast: Twitter Conor Hoekstra: Twitter | BlueSky | Mastodon Bryce Adelstein Lelbach: Twitter About the Guests: Ray is a Senior Systems Software Engineer at NVIDIA since 2022. Studied Software Engineering at the University of Amsterdam. Founded the Dutch C++ Meetup in 2013 and co-organizes C++ Under the Sea since 2023. He has been programming for more than 25 years, his journey began on his father's Panasonic CF-2700 MSX--and has been hooked ever since. He is also 'the listener' of ADSP the podcast. Paul Grosse-Bley was first introduced to parallel programming with C+MPI at a student exchange to UmeΓ₯ (Sweden) in 2017 while studying Physics. In the following years he learned more about MPI, OpenMP, OpenACC, Thrust/parSTL and CUDA C++. After finishing his Master's degree in Physics at Heidelberg University (Germany) in 2021, he became a PhD candidate in Computational Science and Engineering researching the acceleration of iterative solvers in sparse linear algebra while being head-tutor for a course on GPU Algorithm Design. He learned using Thrust in 2019 shortly before learning C++ and became enamored with parallel algorithms which led to numerous answers on StackOverflow, contributions on GitHub, his NVIDIA internship in the summer of 2025 and full position starting in February of 2026. Show Notes Date Recorded: 2025-10-10 Date Released: 2025-11-14 NVIDIA BCM (Base Command Manager) C++11 std::ignore C++20 std::bind_front Parrot Parrot on GitHub Parrot Youtube Video: 1 Problem, 7 Libraries (on the GPU) thrust::inclusive_scan Single-pass Parallel Prefix Scan with Decoupled Look-back by Duane Merrill & Michael Garland Prefix Sums and Their Applications by Guy Blelloch Parallel Prefix Sum (Scan) with CUDA NVIDIA ON-Demand Videos A Faster Radix Sort Implementation Intro Song Info Miss You by Sarah Jansen https://soundcloud.com/sarahjansenmusic Creative Commons - Attribution 3.0 Unported - CC BY 3.0 Free Download / Stream: http://bit.ly/l-miss-you Music promoted by Audio Library
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Ray BurgemeestreHi, I'm Ray. I'm one of the co-organizers of this conference. For me, it was the best one so far.
ConorWhoa, whoa, whoa, whoa, whoa. If we're gonna throw the word princess around, it only is applied to you, buddy. Has been released at C under the sea. It's called Parrot. It's the best CUDA C experience out there.
BryceTemplate parameter, a non-type template parameter that's like an ant. Like you have a function that takes a scan?
ConorWe got a couple scans. Decoupled look back scan, not totally. I assume you've read the Dwayne Merrill paper. Did we originally say that? Absolutely, yeah.
BryceYou worked at NVIDIA for like two more days? Paul doesn't work in NVIDIA anymore. Yeah, he doesn't work in NVIDIA anymore.
ConorThis is the scan man. That is what you should be known as. Bryce, the scan man, Adelstein Lelbeck. Yes.
Scatman (Clip)I'm a scan man!
ConorWelcome to ADSP the podcast, episode 260, recorded on October 10th, 2025. My name is Connor. Today with my co-host Bryce, we record once again live from the Netherlands. The second time and second year in a row, we attended C under the sea. In this episode, we'll be talking to two different Nvidia's Ray and Paul about the new parrot library and also about scans. Here we go, folks, we'll redo that. C live from Breda, Netherlands, C Under the Sea 2025, second edition.
ConorIt's not the inaugural one, folks. We recorded here.
BryceBut but but this year, we are here as founding sponsors. Well, well, we're here for NVIDIA as founding sponsors. We had a booth, we had a bunch of NVIDIA people here.
ConorTherefore, we are founding sponsors of C under the sea. Day two just wrapped. We're about to talk to Sean, probably a couple other folks. And we were gonna record last night. Today is October 10th.
BryceBryce was too tired.
ConorBryce was too tired. Dinner ended at like 10 30. Admittedly, dinner did go on for a while.
BryceIf we had recorded dinner, there were some Sean stories that I have never heard before, and they were some top quality.
ConorLet us tease and make the listener feel bad, folks. Was it his best story?
BryceI don't know. I think the two or three stories may have been the best like storytelling I.
ConorAnd I hadn't heard I think two of the three that he told last night, and one of them I was just like, my god, I wonder. The Photoshop 4 story? Are we even allowed to allude to the Photoshop? I guess most people most people probably don't know what that means, but yes, the Photoshop 4 story. Uh it's up to Sean. We're not, it's totally, you know, there are a number of stories that have to be at the bar or uh in a non-recorded environment. Not necessarily at a bar, folks. But uh, anyways, we were gonna record last night.
ConorDinner went a little bit too long. Thank you to Floris Bob who organized our dinner.
BryceWe got we have these wonderful NVIDIA developer tags we've been given out in addition to the ADSB stickers. And I I got I took a couple extra ones. So if you find me at a C event for maybe the next six months, I'll probably have a couple.
ConorYep, we gave out a lot of stickers, folks. And uh should we do initial thoughts on C under the sea 2025? Or should we just do some initial thoughts?
BryceWhat did you think? Great conference or greatest conference ever?
ConorWell, it's not the greatest conference ever. You know, now you got me thinking, what's the greatest conference I've ever been to?
BryceHey Ray, Ray! Hey Ray, we got on. We were just we were just debating we were just gonna recap up to the first one. I was I was asking, what did I ask?
ConorYou asked, because I was like, oh, let's give initial thoughts on how the conference went, and he's like, was it the greatest conference ever? And I was like, well, no. I said was it great or the greatest conference ever? It was great. Not the greatest, and that doesn't mean it's like it's not a bad conference.
BryceAlright, we should introduce Ray.
ConorBut uh we got Ray, fellow NVIDIA folks. Yes, Ray. He contacts us via email both with our NVIDIA emails and my personal. I'm not sure if he emails you personally.
Ray BurgemeestreBut Ray, introduce yourself and hi, I'm Ray, I'm uh one of the co-organizers of this conference. Yeah, for for me it was the best one so far.
BryceSo And what do you what do you apparently was the best C under the C I think it was the best C under the C, yeah, yeah, for sure. And uh what do you do at NVIDIA?
Ray BurgemeestreI work uh in a team in in the Netherlands, which is also where I'm from. I work on uh cluster management software. It's called BCM, base command manager. And uh yeah, there I do a mix of C, Python, but also Kubernetes.
BryceOh Kubernetes, the scary word here. We talk about all the programming languages, but as soon as somebody says the DevOps words, we run for the hills, folks. You do? We run for the hills. I guess we do. Do you do you work with the people who do DCGM?
Ray BurgemeestreUh no, but uh my in oh yeah, my integration with my knowledge of DCGM is or for the perspective from our product is we we collect those metrics. That's uh okay, okay, okay.
BryceSo uh so how did this conference come to be, Ray? What was uh what was your vision or what was the organizer's vision?
Ray BurgemeestreYeah, this is uh spin-off from the meetup group that I've been running since 2020 uh thirteen already. And uh I met Hans who's the main organizer, and Andrew, who is also a co-organizer of various meetups at various uh of these meetups. And Hans had the idea to uh to organize a conference for a long time, and that's uh how it all started.
BryceAlright, alright, we're gonna each do two things that we learned at this conference, and you two are gonna go first while I try to remember what I learned.
Ray BurgemeestreOh, what I learned? Yeah, two things you learned. Uh a lot of things, so they they they they don't come to mind uh in an order of the. I I have never thought about uh generic programming applied to higher level things like uh Sean was discussing in his keynote. That was uh yeah, that I'm still processing that to be honest. And little things like new features in C23, like uh maybe uh I have not been keeping up well enough, but like STD ignore or uh bind left, that's kind of uh bind left, was it?
Ray BurgemeestreYeah, yeah, that those kind of trivial things come to mind that I yeah, I just simply didn't know about them. And yeah, yeah, I don't know, more things may come to mind.
BryceI learned so the coolest thing I saw, there was a talk from some of the Conan folks uh about uh using Conan to install CUDA and uh uh using Conan and CMake to install CUDA and work with CUDA. And the the thing that I saw that was very, very cool was uh they have this uh support for cross-building where you can do the build on your server, and then they've got this like little CMake tool where it will figure out what files it needs to copy over to the embedded device, copy them over there, and then run C test remotely so you don't have to figure that out.
BryceAnd they can do that because they're the package manager, they know what all the dependencies are, and I thought that was super cool, not just for people doing embedded development on Nvidia Jetson, but people doing embedded development in general. And uh yeah, that was what I learned. And uh Connor has abandoned us to go get some food because he's very hungry, uh, because he doesn't like certain, he doesn't like like creams or like foods with textures, he's a little bit picky about food, and so he's he had limited options at lunch.
Ray BurgemeestreI actually I actually also learned a lot from your talk that comes to mind now. So I'm not uh uh Yakuda expert at all, but I I've been playing around with it like a couple years ago. And I remember syntax like uh three um less than uh the character I yeah, yeah, yeah.
BryceThe Chevron launch, the triple Chevron kernel launch.
Ray BurgemeestreThose kind of things are way uh away for a long time, I'm sure, and so much more library support. It actually makes me want to play with it again, even though I don't have to uh yeah.
BryceYou should. I mean it's you know, people assume that everybody at Nvidia does CUDA, does GPU stuff, but actually there's a lot of people at Nvidia that do other sorts of things that does infrastructure stuff, does the core libraries, core software. What's it like in your team? Like what uh what sort of C dialects do you use? How modern is the C code that you use? Like what's it you said you said there's some Python, like you know, what's it like dealing with the C Python crossover?
Ray BurgemeestreYeah, we the the boundary between Python and C is uh through APIs. So yeah, our core product, the the the product that runs on old servers is uh written in C to have a reasonably smaller uh footprint. Um we use, I believe currently C 17 and probably soon we'll jump to 20. So it's a mix of modern and uh some older parts. So yeah, we we try to keep it up to date. But uh it's been around for a while, this product.
BryceSo I see Connor's coming back. So Connor.
ConorWell I uh that I did get some food and I I decided to change up the nugget that I had had four of earlier, and it ended up being a cheese one.
BryceOh, I was just telling people how picky, what a little princess you were with the food. It's all in there.
ConorWhoa, whoa, whoa, whoa, whoa. If we're gonna throw the word princess around, it only is applied to you, buddy. I'm a little bit particular, and every one of the sandwiches had cheese on them. They look good, just don't like cheese. It's not a problem with the conference, it's a problem with me. What was the question we were asking?
BryceWhat were the top two things you learned at the conference? I only gave one, but I don't know if people noticed that because my answer is so good.
ConorSo, I'm not sure if this counts as something I learned, but I will be saying it as my answer. And I'm trying to think right now. We've got part two of walking in Copenhagen, then there's gonna be one or two parts from the conversation that we had at NDC. And then did we record a final time afterwards? Yeah, we did with Vittorio and with JF. So that is at least one, two, three minimum episodes, maybe four, which brings us into November.
ConorSo while you are listening to this, listener, I am either getting married or on my honeymoon. And the point being is that it is much in the past, and my long-awaited library that we haven't been able to talk about, that I was bleeping Bryce out, has been released at C under the sea. It's called Parrot. It's the best CUDA C experience out there. And the graphic signifier for the library is a is a parrot. And Kuhn, who I told to follow me, but then he disappeared, said to me while I was at this conference, he asked, Does the parrot have a name?
ConorAnd I said, No, it does not have a name. He's like, Well, you should call the parrot Sean Parrot.
BryceAnd that is Sean's right next to us, and he heard Sean Parrot, and he just looked up.
ConorI mean, he's fully approved, and uh the first pull request that I merged from an external contributor was uh from Kuhn, who uh he created a fork and he uh renamed the alt text on the image in the README to be Sean Parrot. So officially the name of the parrot is Sean Parrot, and I am grumbling.
BryceI feel like you flubbed the story a little bit because the first time he suggested it, he just said, Oh, you should call the parrot Sean.
ConorNo, no, so he did say Sean Parrot, but I only heard Sean, and it or like maybe I heard uh parent, and I was like, Well, I mean, it's a little on the nose, I get it, like he's inspired me down the algorithm journey, etc. But uh uh he told me a second time, and then that was when I was like, Oh, you're saying Sean Parrot. And then like everyone burst out laughing, and we were like, that's the best, that's the best suggestion or thing I've been told in here. So is it something I learned? Not really, but it is something that is official, it's a part of the lore. Another thing, the actual another second thing that I learned was uh your contribution in terms of adding the uh std integral constant in order to avoid the template issue with the softmax example, and that is exactly why softmax wasn't in a function, because I knew that there was probably a solution, whether it was just making it a runtime parameter.
ConorI didn't even know about the integral constant runtime parameter.
BryceWe should we should explain this trick. If you have a template parameter, a non-type template parameter that's like an int, like you have a function that takes a uh C function that takes a int as a template parameter, if it's for example, a uh a member function, um uh like if if stdduple had a m member function git, then if you were using the if you're using the um that template in a context where uh it's dependent, like if you were doing like, you know, uh, you know, you had some tuple that you took as a template parameter, and then you did like t dot git, where git was a template function that took an integral template parameter, you you have to put the word template after the dot so it's not ambiguous, um, which is awful.
BryceUh and so one trick to get around this is that you can use user-defined literals. So you can add a user-defined literal like underscore IC that will return an integral constant object. And then you can change your template to add an additional function parameter, not a template parameter, but a function parameter that is an integral constant from which your non-type template parameter can be deduced. And you can just default that um that uh function parameter so that your the existing API continues to work, but then the nice API that people can do is they can do, you know, dot git open parens like one underscore iC, close parens, instead of having to do the dot template git, you know, uh angle brackets one, close angle brackets, open paren, close paren.
BryceAnd so that was a thing I suggested to Connor, and Connor went off and vibed coded, vibe coded it very quickly.
ConorI guide coded it, guide coded it. I did look, guide, there's a difference. Vibecoding you don't look, guide coding you definitely check. But it was a brilliant suggestion. It's still a little bit ugly. I don't love the underscore IC, but it is an order of magnitude better than the, and I people might have not encountered this, but I had encountered this kind of code before where you have to do a dot, so your uh original function would be like your A dot B and then your uh chevron, you know, with your you know uh integer in my case, so it's angle one, end angle, and then your param your parens.
ConorInstead you have to go a dot template space b and then uh your uh angle, angle, and paren paren, which is just awful. And so much so that I didn't I I didn't put softmax in a function, and I was like, we'll deal with this, we'll kick this down the road. And now we're uh we're approaching uh Sean. He's been approached. He's been he's been uh he's a popular guy. Sean was being asked for his autograph today, and uh he signed uh the EOP Elements of Programming Book.
BryceHe did. I've been I've been asked to sign stuff, and it is an interesting experience.
ConorYep, so anyways, that's uh Sean Parrot, number one highlight of this conference for me. And I mean also too, at this point, actually we shouldn't talk about the former color of the parrot. I mean, actually we can talk about it. At one point, Sean Parrot was pink before before the parrot was Sean Parrot.
BryceI I filed an issue to uh that the parrot should be green.
ConorYeah, it should be the NVIDIA green. So we have changed that, and we just crossed a hundred stars. At this point, at this point in November, I'm hoping we've crossed a thousand. Probably not, but it's uh hopeful thinking. And also, too, I will have given a YouTube version of the talk that I present today uh at some point. It might be as soon as Monday, but it might be a week and a half later. Regardless though, it'll be in the past from when this is airing on our podcast. So link in the show notes if you haven't seen that already.
ConorHello, listener. It is October 30th when I am recording this interruption. And you are potentially listening to this on Friday, November 14th, or a few days or a couple weeks after this. I have not released the YouTube video that I said I would. However, I have released a YouTube video that mentions Parrot. It's not fully about Parrot, but it compares Parrot to a couple other Python and C libraries. So link in the show notes for that. And if you do happen to be listening to this on a bit of a delay, two to three weeks after this is released on November 14th, that talk may have been released at that point.
ConorSo I will leave a unlinked link in the show notes and we'll link it as soon as it has been released. So just be sure to check the show notes for when that talk does become available. And uh hey Paul, you want to say hi?
BryceHey, Paul! Paul's another person who works at NVIDIA. There's a lot of NVIDIA.
ConorWe had dinner with at I don't remember the name. Wolfus Wolfusar? Wolfread. Paul knows.
BryceYou were you worked at NVIDIA for like two more days? Paul doesn't work at NVIDIA anymore. Yeah, he doesn't work at NVIDIA anymore. Yes.
ConorHe was an N is still employed until like midnight? Yeah.
BrycePaul is an intern, great intern we had over the summer. Today is his last day, but then he'll be coming back sometime early next year. Yeah, first.
ConorGive him a mic. Uh tell us tell us about your internship, tell us about yourself. Uh hopefully you're coming back full time.
Paul Grosse-BleyListener, I know the answer to that question. I'm just setting him up. Oh, it was great, great experience. Uh I was really looking for looking forward to it because uh I've been using the CCCL where I had my internship for a long time. Um before you were working here. Yeah, yeah.
ConorThat's how you get a job at NVIDIA, folks. You gotta be using it on the daily and become an expert. Uh yeah.
Paul Grosse-BleyUh I've been even teaching courses based on thrust teaching.
BryceUsing it or contributing, contributing to our major open source products. Great way to get notice if you want to get hired at NVIDIA. I mean, hiring very tight right now. Please, no emails.
ConorWork for me. Yeah, you are an existence proof. Uh so we're using it before, then you got hired as an intern. So you've been here for four months, eight months, three months? Uh 3.8. 3.8? Okay. And today was your is your last day. Exactly. Gotcha. And so you're going back, taking some time off, going back to school?
Paul Grosse-BleyUh yeah, finishing my uh uh my my school. Yeah. Whatever that means.
ConorHoly shit. The mother love, where'd you get those? I just for the listener. That little uh diversion there, we've uh acquired some more drink tokens. So uh expect the next uh however bit, it's gonna be fun. Sean's been uh cornered, but we're talking to Paul now. So you're finishing your school, taking a little bit of time off, and then when do you start? And you are you joining the same team? Yeah, yeah.
Paul Grosse-BleyIn February I will rejoin the CCCL team and keep working on parallel algorithms, which is basically my dream job. Well, so what parallel algorithms, what's your favorite?
ConorWe don't ask this question anymore. We've lost touch with our roots. What is your favorite algorithm? Parallel, serial, it doesn't matter. Uh go.
Paul Grosse-BleyI mean, uh the problem is I already know the answer as a follower of your podcast.
BryceUh wait, there's there is no, it's your favorite algorithm. How can you I think our original answer was reductions, but I think my answer has changed. My answer is definite. It's a personal thing.
Paul Grosse-BleyIt's certainly scam. Scan? Yeah. Absolutely. We got a couple scans.
ConorDecoupled look back scan, not really. I assume you read the Dwayne Merrill paper. Did we originally say?
BryceI think we originally said reduce. Or I originally said reduce, transform reduce is my favorite. You said transform reduce. Yes. But that's changed, folks. Now I'm a scan man. It's all about the scan. What what's what's your favorite application of scan?
ConorI'm gonna I'm gonna insert. Do you know the scat man? I'm the scat man. Blah, blah, blah, blah, blah, blah.
Scatman (Clip)I'm a scan man!
ConorI'm gonna cut that, I'm gonna club that in. It's the scan man. That is what you should be known as. Bryce, the scan man, Adelstein Lelbeck. Yes.
Paul Grosse-BleyUh you uh sometimes, for example, you know uh how many elements you have in each row of your matrix, and then with the scan you can get the offset into memory for that.
BryceWe were just talking earlier today about applications of scans and uh and sparse linear algebra. You gotta hold this. What are you gonna do? You're you're talking it over there, like yeah, you can't. Well, it's your job. I'm I'm drinking here. I'm drinking my well everybody's routine. Everybody's drinking.
ConorAnd you're just anyway.
BryceOkay. But yeah, we were just talking earlier today about how to implement uh certain uh sparse uh linear algebra uh uh techniques in this uh this new Q-tile thing, and we were talking about the potential uses of scans there, and yeah, but my my uh my favorite recent scans have been these uh these uh like uh um monoidal transformation scans where you have a scan that takes both a transform iterator as the input and then a transform output iterator.
BryceUm and uh we that we I don't want to I don't want to spoil anything, but we talked about one application of that for the proportional tracker. I think we talked about that on the podcast. Yes, we did. But but we talked we talked in Copenhagen about run length decode, and you were convinced it couldn't be done in a single pass, and folks, let me tell you, uh, it's beautiful. It's beautiful the single pass run length d uh decoding, and we're gonna we're gonna look at it. We're gonna look at it in uh, I don't know, December at this point because we've got such a backlog of episodes.
BryceBut the next time we're like home and recording, we're gonna look at that. It's it's beautiful.
ConorWait, so you're talking about decoupled look back, aka Dwayne Merrill, I think Michael Garland's co-author on that paper. Yeah. And do you have other because when you said scan man, uh there's a little voice in my head being like, well, technically Mark Harris should be the scan man. If there's anybody that's the scan man, it's Mark Harris, who is an employee of NVIDIA. He's written many uh influential papers, one of them being Prefix Sums and Their Applications, I think is the title. Modulo, maybe a word or two. I gotta go get some of that food. Alright, Bryce is leaving to go get some food. Have you I assume you've read that if you're if you've read the decoupled look bad, that's a trickier paper than the others.
ConorDo you have papers or resources for people that want to become scan experts like yourself?
Paul Grosse-BleyMaybe less scan, but uh there's the the uh NVIDIA on demand uh website where you can see all the old talks from old GTCs. And I think there's, for example, a talk on how to implement Radix sort or optimize it. Uh that one was very very interesting.
BryceNot to put you on the spot, but how do you implement Radix sort?
ConorI mean, there's certainly scans involved. Be sure to check these show notes either in your podcast app or at adspthepodcast.com for links to anything we mentioned in today's episode, as well as a link to a GitHub discussion where you can leave thoughts, comments, and questions. Thanks for listening. We hope you enjoyed and have a great day.
BryceLow quality, high quality! That is the tagline of our podcast.
ConorThat's not the tagline. Our tagline is chaos with sprinkles of information. Live from C under the C 2025, folks. Is the audio levels correct? We don't know. Is this the last time we interviewed him? In person?
BryceNo, just in general.
ConorYou got cheese all over your face, bro.
BryceI mean, what else is new? No, you sound like my girlfriend. Don't uh don't tell me that.
ConorWhat do you mean the last time we interviewed him?
In this episode, Conor and Bryce chat with Zach Laine about C++26 Reflection vs CRTP, AI, kids & more. Link to Episode 313 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP:...
In this episode, Conor and Bryce chat with Zach Laine about the C++26 contracts fiasco. Link to Episode 312 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...
In this episode, Conor and Bryce talk with Zach Laine about his career and the career changes he has made. Link to Episode 311 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials...
In this episode, Conor and Bryce talk about Bryce's next chapter and where he is leaving NVIDIA for. Link to Episode 310 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The...
In this episode, Conor and Ben chat with Jason Turner about default construction in C++. Link to Episode 309 on Website Discuss this episode, leave a comment, or ask a question (on GitHub) Socials ADSP: The Podcast:...