WEBVTT 00:00.000 --> 00:02.680 The following is a conversation with Chris Latner. 00:02.680 --> 00:04.560 Currently, he's a senior director 00:04.560 --> 00:08.400 at Google working on several projects, including CPU, GPU, 00:08.400 --> 00:12.040 TPU accelerators for TensorFlow, Swift for TensorFlow, 00:12.040 --> 00:14.400 and all kinds of machine learning compiler magic 00:14.400 --> 00:16.360 going on behind the scenes. 00:16.360 --> 00:18.440 He's one of the top experts in the world 00:18.440 --> 00:21.160 on compiler technologies, which means he deeply 00:21.160 --> 00:25.560 understands the intricacies of how hardware and software come 00:25.560 --> 00:27.920 together to create efficient code. 00:27.920 --> 00:31.400 He created the LLVM compiler infrastructure project 00:31.400 --> 00:33.360 and the Clang compiler. 00:33.360 --> 00:36.000 He led major engineering efforts at Apple, 00:36.000 --> 00:39.000 including the creation of the Swift programming language. 00:39.000 --> 00:41.720 He also briefly spent time at Tesla 00:41.720 --> 00:44.280 as vice president of Autopilot software 00:44.280 --> 00:46.760 during the transition from Autopilot hardware 1 00:46.760 --> 00:49.600 to hardware 2, when Tesla essentially 00:49.600 --> 00:52.640 started from scratch to build an in house software 00:52.640 --> 00:54.800 infrastructure for Autopilot. 00:54.800 --> 00:58.040 I could have easily talked to Chris for many more hours. 00:58.040 --> 01:01.200 Compiling code down across the levels of abstraction 01:01.200 --> 01:04.160 is one of the most fundamental and fascinating aspects 01:04.160 --> 01:06.640 of what computers do, and he is one of the world 01:06.640 --> 01:08.560 experts in this process. 01:08.560 --> 01:12.880 It's rigorous science, and it's messy, beautiful art. 01:12.880 --> 01:15.920 This conversation is part of the Artificial Intelligence 01:15.920 --> 01:16.760 podcast. 01:16.760 --> 01:19.440 If you enjoy it, subscribe on YouTube, iTunes, 01:19.440 --> 01:22.760 or simply connect with me on Twitter at Lex Friedman, 01:22.760 --> 01:24.680 spelled F R I D. 01:24.680 --> 01:29.360 And now, here's my conversation with Chris Ladner. 01:29.360 --> 01:33.160 What was the first program you've ever written? 01:33.160 --> 01:34.120 My first program. 01:34.120 --> 01:35.360 Back, and when was it? 01:35.360 --> 01:39.080 I think I started as a kid, and my parents 01:39.080 --> 01:41.560 got a basic programming book. 01:41.560 --> 01:44.200 And so when I started, it was typing out programs 01:44.200 --> 01:46.880 from a book, and seeing how they worked, 01:46.880 --> 01:49.680 and then typing them in wrong, and trying 01:49.680 --> 01:51.680 to figure out why they were not working right, 01:51.680 --> 01:52.960 that kind of stuff. 01:52.960 --> 01:54.880 So BASIC, what was the first language 01:54.880 --> 01:58.360 that you remember yourself maybe falling in love with, 01:58.360 --> 01:59.720 like really connecting with? 01:59.720 --> 02:00.400 I don't know. 02:00.400 --> 02:02.680 I mean, I feel like I've learned a lot along the way, 02:02.680 --> 02:05.800 and each of them have a different special thing 02:05.800 --> 02:06.640 about them. 02:06.640 --> 02:09.720 So I started in BASIC, and then went like GW BASIC, 02:09.720 --> 02:11.440 which was the thing back in the DOS days, 02:11.440 --> 02:15.280 and then upgraded to QBASIC, and eventually QuickBASIC, 02:15.280 --> 02:18.200 which are all slightly more fancy versions of Microsoft 02:18.200 --> 02:19.440 BASIC. 02:19.440 --> 02:21.360 Made the jump to Pascal, and started 02:21.360 --> 02:23.920 doing machine language programming and assembly 02:23.920 --> 02:25.280 in Pascal, which was really cool. 02:25.280 --> 02:28.080 Turbo Pascal was amazing for its day. 02:28.080 --> 02:31.600 Eventually got into C, C++, and then kind of did 02:31.600 --> 02:33.400 lots of other weird things. 02:33.400 --> 02:37.080 I feel like you took the dark path, which is the, 02:37.080 --> 02:39.480 you could have gone Lisp. 02:39.480 --> 02:40.000 Yeah. 02:40.000 --> 02:41.680 You could have gone higher level sort 02:41.680 --> 02:44.600 of functional philosophical hippie route. 02:44.600 --> 02:48.080 Instead, you went into like the dark arts of the C. 02:48.080 --> 02:49.720 It was straight into the machine. 02:49.720 --> 02:50.680 Straight to the machine. 02:50.680 --> 02:53.880 So I started with BASIC, Pascal, and then Assembly, 02:53.880 --> 02:55.320 and then wrote a lot of Assembly. 02:55.320 --> 03:00.080 And I eventually did Smalltalk and other things like that. 03:00.080 --> 03:01.880 But that was not the starting point. 03:01.880 --> 03:05.080 But so what is this journey to C? 03:05.080 --> 03:06.320 Is that in high school? 03:06.320 --> 03:07.560 Is that in college? 03:07.560 --> 03:09.320 That was in high school, yeah. 03:09.320 --> 03:13.720 And then that was really about trying 03:13.720 --> 03:16.240 to be able to do more powerful things than what Pascal could 03:16.240 --> 03:18.960 do, and also to learn a different world. 03:18.960 --> 03:20.760 So he was really confusing to me with pointers 03:20.760 --> 03:23.000 and the syntax and everything, and it took a while. 03:23.000 --> 03:28.800 But Pascal's much more principled in various ways. 03:28.800 --> 03:33.400 C is more, I mean, it has its historical roots, 03:33.400 --> 03:35.520 but it's not as easy to learn. 03:35.520 --> 03:39.880 With pointers, there's this memory management thing 03:39.880 --> 03:41.680 that you have to become conscious of. 03:41.680 --> 03:43.880 Is that the first time you start to understand 03:43.880 --> 03:46.520 that there's resources that you're supposed to manage? 03:46.520 --> 03:48.480 Well, so you have that in Pascal as well. 03:48.480 --> 03:51.440 But in Pascal, like the caret instead of the star, 03:51.440 --> 03:53.160 there's some small differences like that. 03:53.160 --> 03:55.680 But it's not about pointer arithmetic. 03:55.680 --> 03:58.760 And in C, you end up thinking about how things get 03:58.760 --> 04:00.840 laid out in memory a lot more. 04:00.840 --> 04:04.160 And so in Pascal, you have allocating and deallocating 04:04.160 --> 04:07.560 and owning the memory, but just the programs are simpler, 04:07.560 --> 04:10.080 and you don't have to. 04:10.080 --> 04:12.640 Well, for example, Pascal has a string type. 04:12.640 --> 04:14.040 And so you can think about a string 04:14.040 --> 04:15.880 instead of an array of characters 04:15.880 --> 04:17.720 which are consecutive in memory. 04:17.720 --> 04:20.400 So it's a little bit of a higher level abstraction. 04:20.400 --> 04:22.800 So let's get into it. 04:22.800 --> 04:25.560 Let's talk about LLVM, C lang, and compilers. 04:25.560 --> 04:26.560 Sure. 04:26.560 --> 04:32.160 So can you tell me first what LLVM and C lang are? 04:32.160 --> 04:33.960 And how is it that you find yourself 04:33.960 --> 04:35.720 the creator and lead developer, one 04:35.720 --> 04:39.400 of the most powerful compiler optimization systems 04:39.400 --> 04:40.080 in use today? 04:40.080 --> 04:40.580 Sure. 04:40.580 --> 04:43.320 So I guess they're different things. 04:43.320 --> 04:47.080 So let's start with what is a compiler? 04:47.080 --> 04:48.840 Is that a good place to start? 04:48.840 --> 04:50.200 What are the phases of a compiler? 04:50.200 --> 04:50.920 Where are the parts? 04:50.920 --> 04:51.600 Yeah, what is it? 04:51.600 --> 04:53.400 So what is even a compiler used for? 04:53.400 --> 04:57.880 So the way I look at this is you have a two sided problem of you 04:57.880 --> 05:00.120 have humans that need to write code. 05:00.120 --> 05:01.880 And then you have machines that need to run 05:01.880 --> 05:03.400 the program that the human wrote. 05:03.400 --> 05:05.280 And for lots of reasons, the humans 05:05.280 --> 05:07.040 don't want to be writing in binary 05:07.040 --> 05:09.080 and want to think about every piece of hardware. 05:09.080 --> 05:12.100 And so at the same time that you have lots of humans, 05:12.100 --> 05:14.800 you also have lots of kinds of hardware. 05:14.800 --> 05:17.400 And so compilers are the art of allowing 05:17.400 --> 05:19.240 humans to think at a level of abstraction 05:19.240 --> 05:20.920 that they want to think about. 05:20.920 --> 05:23.600 And then get that program, get the thing that they wrote, 05:23.600 --> 05:26.080 to run on a specific piece of hardware. 05:26.080 --> 05:29.480 And the interesting and exciting part of all this 05:29.480 --> 05:32.080 is that there's now lots of different kinds of hardware, 05:32.080 --> 05:35.780 chips like x86 and PowerPC and ARM and things like that. 05:35.780 --> 05:37.320 But also high performance accelerators 05:37.320 --> 05:38.900 for machine learning and other things like that 05:38.900 --> 05:41.520 are also just different kinds of hardware, GPUs. 05:41.520 --> 05:42.940 These are new kinds of hardware. 05:42.940 --> 05:45.640 And at the same time, on the programming side of it, 05:45.640 --> 05:48.680 you have basic, you have C, you have JavaScript, 05:48.680 --> 05:50.560 you have Python, you have Swift. 05:50.560 --> 05:52.840 You have lots of other languages 05:52.840 --> 05:55.200 that are all trying to talk to the human in a different way 05:55.200 --> 05:58.320 to make them more expressive and capable and powerful. 05:58.320 --> 06:01.500 And so compilers are the thing 06:01.500 --> 06:03.460 that goes from one to the other. 06:03.460 --> 06:05.200 End to end, from the very beginning to the very end. 06:05.200 --> 06:06.040 End to end. 06:06.040 --> 06:08.120 And so you go from what the human wrote 06:08.120 --> 06:11.600 and programming languages end up being about 06:11.600 --> 06:14.560 expressing intent, not just for the compiler 06:14.560 --> 06:17.980 and the hardware, but the programming language's job 06:17.980 --> 06:20.920 is really to capture an expression 06:20.920 --> 06:22.680 of what the programmer wanted 06:22.680 --> 06:25.120 that then can be maintained and adapted 06:25.120 --> 06:27.120 and evolved by other humans, 06:27.120 --> 06:29.720 as well as interpreted by the compiler. 06:29.720 --> 06:31.560 So when you look at this problem, 06:31.560 --> 06:34.200 you have, on the one hand, humans, which are complicated. 06:34.200 --> 06:36.760 And you have hardware, which is complicated. 06:36.760 --> 06:39.900 And so compilers typically work in multiple phases. 06:39.900 --> 06:42.760 And so the software engineering challenge 06:42.760 --> 06:45.000 that you have here is try to get maximum reuse 06:45.000 --> 06:47.140 out of the amount of code that you write, 06:47.140 --> 06:49.800 because these compilers are very complicated. 06:49.800 --> 06:51.240 And so the way it typically works out 06:51.240 --> 06:54.480 is that you have something called a front end or a parser 06:54.480 --> 06:56.640 that is language specific. 06:56.640 --> 06:59.500 And so you'll have a C parser, and that's what Clang is, 07:00.400 --> 07:03.480 or C++ or JavaScript or Python or whatever. 07:03.480 --> 07:05.000 That's the front end. 07:05.000 --> 07:07.120 Then you'll have a middle part, 07:07.120 --> 07:09.020 which is often the optimizer. 07:09.020 --> 07:11.120 And then you'll have a late part, 07:11.120 --> 07:13.320 which is hardware specific. 07:13.320 --> 07:15.020 And so compilers end up, 07:15.020 --> 07:16.680 there's many different layers often, 07:16.680 --> 07:20.860 but these three big groups are very common in compilers. 07:20.860 --> 07:22.200 And what LLVM is trying to do 07:22.200 --> 07:25.360 is trying to standardize that middle and last part. 07:25.360 --> 07:27.880 And so one of the cool things about LLVM 07:27.880 --> 07:29.740 is that there are a lot of different languages 07:29.740 --> 07:31.080 that compile through to it. 07:31.080 --> 07:35.600 And so things like Swift, but also Julia, Rust, 07:35.600 --> 07:39.140 Clang for C, C++, Subjective C, 07:39.140 --> 07:40.940 like these are all very different languages 07:40.940 --> 07:43.780 and they can all use the same optimization infrastructure, 07:43.780 --> 07:45.340 which gets better performance, 07:45.340 --> 07:47.240 and the same code generation infrastructure 07:47.240 --> 07:48.780 for hardware support. 07:48.780 --> 07:52.240 And so LLVM is really that layer that is common, 07:52.240 --> 07:55.580 that all these different specific compilers can use. 07:55.580 --> 07:59.300 And is it a standard, like a specification, 07:59.300 --> 08:01.140 or is it literally an implementation? 08:01.140 --> 08:02.140 It's an implementation. 08:02.140 --> 08:05.900 And so I think there's a couple of different ways 08:05.900 --> 08:06.740 of looking at it, right? 08:06.740 --> 08:09.700 Because it depends on which angle you're looking at it from. 08:09.700 --> 08:12.660 LLVM ends up being a bunch of code, okay? 08:12.660 --> 08:14.460 So it's a bunch of code that people reuse 08:14.460 --> 08:16.540 and they build compilers with. 08:16.540 --> 08:18.060 We call it a compiler infrastructure 08:18.060 --> 08:20.060 because it's kind of the underlying platform 08:20.060 --> 08:22.580 that you build a concrete compiler on top of. 08:22.580 --> 08:23.740 But it's also a community. 08:23.740 --> 08:26.820 And the LLVM community is hundreds of people 08:26.820 --> 08:27.980 that all collaborate. 08:27.980 --> 08:30.620 And one of the most fascinating things about LLVM 08:30.620 --> 08:34.260 over the course of time is that we've managed somehow 08:34.260 --> 08:37.060 to successfully get harsh competitors 08:37.060 --> 08:39.060 in the commercial space to collaborate 08:39.060 --> 08:41.120 on shared infrastructure. 08:41.120 --> 08:43.900 And so you have Google and Apple, 08:43.900 --> 08:45.860 you have AMD and Intel, 08:45.860 --> 08:48.860 you have Nvidia and AMD on the graphics side, 08:48.860 --> 08:52.620 you have Cray and everybody else doing these things. 08:52.620 --> 08:55.420 And all these companies are collaborating together 08:55.420 --> 08:58.520 to make that shared infrastructure really, really great. 08:58.520 --> 09:01.380 And they do this not out of the goodness of their heart, 09:01.380 --> 09:03.420 but they do it because it's in their commercial interest 09:03.420 --> 09:05.140 of having really great infrastructure 09:05.140 --> 09:06.740 that they can build on top of 09:06.740 --> 09:09.080 and facing the reality that it's so expensive 09:09.080 --> 09:11.160 that no one company, even the big companies, 09:11.160 --> 09:14.580 no one company really wants to implement it all themselves. 09:14.580 --> 09:16.100 Expensive or difficult? 09:16.100 --> 09:16.940 Both. 09:16.940 --> 09:20.540 That's a great point because it's also about the skill sets. 09:20.540 --> 09:25.540 And the skill sets are very hard to find. 09:26.020 --> 09:27.980 How big is the LLVM? 09:27.980 --> 09:30.780 It always seems like with open source projects, 09:30.780 --> 09:33.500 the kind, an LLVM is open source? 09:33.500 --> 09:34.420 Yes, it's open source. 09:34.420 --> 09:38.660 It's about, it's 19 years old now, so it's fairly old. 09:38.660 --> 09:40.940 It seems like the magic often happens 09:40.940 --> 09:43.020 within a very small circle of people. 09:43.020 --> 09:43.860 Yes. 09:43.860 --> 09:46.060 At least their early birth and whatever. 09:46.060 --> 09:49.660 Yes, so the LLVM came from a university project, 09:49.660 --> 09:51.540 and so I was at the University of Illinois. 09:51.540 --> 09:53.900 And there it was myself, my advisor, 09:53.900 --> 09:57.500 and then a team of two or three research students 09:57.500 --> 09:58.380 in the research group, 09:58.380 --> 10:02.100 and we built many of the core pieces initially. 10:02.100 --> 10:03.740 I then graduated and went to Apple, 10:03.740 --> 10:06.480 and at Apple brought it to the products, 10:06.480 --> 10:09.340 first in the OpenGL graphics stack, 10:09.340 --> 10:11.580 but eventually to the C compiler realm, 10:11.580 --> 10:12.780 and eventually built Clang, 10:12.780 --> 10:14.640 and eventually built Swift and these things. 10:14.640 --> 10:16.380 Along the way, building a team of people 10:16.380 --> 10:18.620 that are really amazing compiler engineers 10:18.620 --> 10:20.060 that helped build a lot of that. 10:20.060 --> 10:21.860 And so as it was gaining momentum 10:21.860 --> 10:24.780 and as Apple was using it, being open source and public 10:24.780 --> 10:26.440 and encouraging contribution, 10:26.440 --> 10:28.780 many others, for example, at Google, 10:28.780 --> 10:30.220 came in and started contributing. 10:30.220 --> 10:33.740 And in some cases, Google effectively owns Clang now 10:33.740 --> 10:35.540 because it cares so much about C++ 10:35.540 --> 10:37.340 and the evolution of that ecosystem, 10:37.340 --> 10:41.420 and so it's investing a lot in the C++ world 10:41.420 --> 10:42.980 and the tooling and things like that. 10:42.980 --> 10:47.860 And so likewise, NVIDIA cares a lot about CUDA. 10:47.860 --> 10:50.780 And so CUDA uses Clang and uses LLVM 10:50.780 --> 10:54.060 for graphics and GPGPU. 10:54.060 --> 10:58.940 And so when you first started as a master's project, 10:58.940 --> 11:02.980 I guess, did you think it was gonna go as far as it went? 11:02.980 --> 11:06.340 Were you crazy ambitious about it? 11:06.340 --> 11:07.180 No. 11:07.180 --> 11:09.840 It seems like a really difficult undertaking, a brave one. 11:09.840 --> 11:11.380 Yeah, no, no, no, it was nothing like that. 11:11.380 --> 11:13.740 So my goal when I went to the University of Illinois 11:13.740 --> 11:17.540 was to get in and out with a non thesis masters in a year 11:17.540 --> 11:18.720 and get back to work. 11:18.720 --> 11:22.200 So I was not planning to stay for five years 11:22.200 --> 11:24.460 and build this massive infrastructure. 11:24.460 --> 11:27.380 I got nerd sniped into staying. 11:27.380 --> 11:29.580 And a lot of it was because LLVM was fun 11:29.580 --> 11:30.900 and I was building cool stuff 11:30.900 --> 11:33.420 and learning really interesting things 11:33.420 --> 11:36.900 and facing both software engineering challenges, 11:36.900 --> 11:38.540 but also learning how to work in a team 11:38.540 --> 11:40.100 and things like that. 11:40.100 --> 11:43.620 I had worked at many companies as interns before that, 11:43.620 --> 11:45.860 but it was really a different thing 11:45.860 --> 11:48.060 to have a team of people that are working together 11:48.060 --> 11:50.460 and try and collaborate in version control. 11:50.460 --> 11:52.420 And it was just a little bit different. 11:52.420 --> 11:54.060 Like I said, I just talked to Don Knuth 11:54.060 --> 11:56.860 and he believes that 2% of the world population 11:56.860 --> 11:58.820 have something weird with their brain, 11:58.820 --> 12:01.100 that they're geeks, they understand computers, 12:01.100 --> 12:02.580 they're connected with computers. 12:02.580 --> 12:04.380 He put it at exactly 2%. 12:04.380 --> 12:05.540 Okay, so. 12:05.540 --> 12:06.580 He's a specific guy. 12:06.580 --> 12:08.780 It's very specific. 12:08.780 --> 12:10.180 Well, he says, I can't prove it, 12:10.180 --> 12:11.780 but it's very empirically there. 12:13.180 --> 12:14.500 Is there something that attracts you 12:14.500 --> 12:16.940 to the idea of optimizing code? 12:16.940 --> 12:19.180 And he seems like that's one of the biggest, 12:19.180 --> 12:20.900 coolest things about LLVM. 12:20.900 --> 12:22.500 Yeah, that's one of the major things it does. 12:22.500 --> 12:26.460 So I got into that because of a person, actually. 12:26.460 --> 12:28.220 So when I was in my undergraduate, 12:28.220 --> 12:32.060 I had an advisor, or a professor named Steve Vegdahl. 12:32.060 --> 12:35.740 And he, I went to this little tiny private school. 12:35.740 --> 12:38.300 There were like seven or nine people 12:38.300 --> 12:40.340 in my computer science department, 12:40.340 --> 12:43.100 students in my class. 12:43.100 --> 12:47.460 So it was a very tiny, very small school. 12:47.460 --> 12:49.940 It was kind of a wart on the side of the math department 12:49.940 --> 12:51.260 kind of a thing at the time. 12:51.260 --> 12:53.820 I think it's evolved a lot in the many years since then. 12:53.820 --> 12:58.300 But Steve Vegdahl was a compiler guy. 12:58.300 --> 12:59.580 And he was super passionate. 12:59.580 --> 13:02.740 And his passion rubbed off on me. 13:02.740 --> 13:04.460 And one of the things I like about compilers 13:04.460 --> 13:09.100 is that they're large, complicated software pieces. 13:09.100 --> 13:12.940 And so one of the culminating classes 13:12.940 --> 13:14.540 that many computer science departments, 13:14.540 --> 13:16.700 at least at the time, did was to say 13:16.700 --> 13:18.380 that you would take algorithms and data structures 13:18.380 --> 13:19.460 and all these core classes. 13:19.460 --> 13:21.740 But then the compilers class was one of the last classes 13:21.740 --> 13:24.380 you take because it pulls everything together. 13:24.380 --> 13:26.980 And then you work on one piece of code 13:26.980 --> 13:28.700 over the entire semester. 13:28.700 --> 13:32.180 And so you keep building on your own work, 13:32.180 --> 13:33.460 which is really interesting. 13:33.460 --> 13:36.060 And it's also very challenging because in many classes, 13:36.060 --> 13:38.380 if you don't get a project done, you just forget about it 13:38.380 --> 13:41.300 and move on to the next one and get your B or whatever it is. 13:41.300 --> 13:43.860 But here you have to live with the decisions you make 13:43.860 --> 13:45.220 and continue to reinvest in it. 13:45.220 --> 13:48.500 And I really like that. 13:48.500 --> 13:50.700 And so I did an extra study project 13:50.700 --> 13:52.420 with him the following semester. 13:52.420 --> 13:53.940 And he was just really great. 13:53.940 --> 13:56.860 And he was also a great mentor in a lot of ways. 13:56.860 --> 13:59.500 And so from him and from his advice, 13:59.500 --> 14:01.380 he encouraged me to go to graduate school. 14:01.380 --> 14:03.420 I wasn't super excited about going to grad school. 14:03.420 --> 14:05.540 I wanted the master's degree, but I 14:05.540 --> 14:08.940 didn't want to be an academic. 14:08.940 --> 14:11.100 But like I said, I kind of got tricked into saying 14:11.100 --> 14:12.180 and was having a lot of fun. 14:12.180 --> 14:14.540 And I definitely do not regret it. 14:14.540 --> 14:17.940 What aspects of compilers were the things you connected with? 14:17.940 --> 14:22.100 So LLVM, there's also the other part 14:22.100 --> 14:24.940 that's really interesting if you're interested in languages 14:24.940 --> 14:29.620 is parsing and just analyzing the language, 14:29.620 --> 14:31.220 breaking it down, parsing, and so on. 14:31.220 --> 14:32.580 Was that interesting to you, or were you 14:32.580 --> 14:34.060 more interested in optimization? 14:34.060 --> 14:37.420 For me, it was more so I'm not really a math person. 14:37.420 --> 14:38.180 I could do math. 14:38.180 --> 14:41.540 I understand some bits of it when I get into it. 14:41.540 --> 14:43.940 But math is never the thing that attracted me. 14:43.940 --> 14:46.100 And so a lot of the parser part of the compiler 14:46.100 --> 14:47.820 has a lot of good formal theories 14:47.820 --> 14:50.060 that Don, for example, knows quite well. 14:50.060 --> 14:51.540 I'm still waiting for his book on that. 14:54.740 --> 14:57.900 But I just like building a thing and seeing what it could do 14:57.900 --> 15:00.740 and exploring and getting it to do more things 15:00.740 --> 15:04.020 and then setting new goals and reaching for them. 15:04.020 --> 15:09.580 And in the case of LLVM, when I started working on that, 15:09.580 --> 15:13.420 my research advisor that I was working for was a compiler guy. 15:13.420 --> 15:15.620 And so he and I specifically found each other 15:15.620 --> 15:16.940 because we were both interested in compilers. 15:16.940 --> 15:19.500 And so I started working with him and taking his class. 15:19.500 --> 15:21.580 And a lot of LLVM initially was, it's 15:21.580 --> 15:24.380 fun implementing all the standard algorithms and all 15:24.380 --> 15:26.380 the things that people had been talking about 15:26.380 --> 15:27.220 and were well known. 15:27.220 --> 15:30.620 And they were in the curricula for advanced studies 15:30.620 --> 15:31.340 and compilers. 15:31.340 --> 15:34.580 And so just being able to build that was really fun. 15:34.580 --> 15:37.660 And I was learning a lot by, instead of reading about it, 15:37.660 --> 15:38.660 just building. 15:38.660 --> 15:40.220 And so I enjoyed that. 15:40.220 --> 15:42.820 So you said compilers are these complicated systems. 15:42.820 --> 15:46.180 Can you even just with language try 15:46.180 --> 15:52.220 to describe how you turn a C++ program into code? 15:52.220 --> 15:53.460 Like, what are the hard parts? 15:53.460 --> 15:54.620 Why is it so hard? 15:54.620 --> 15:57.020 So I'll give you examples of the hard parts along the way. 15:57.020 --> 16:01.060 So C++ is a very complicated programming language. 16:01.060 --> 16:03.500 It's something like 1,400 pages in the spec. 16:03.500 --> 16:06.060 So C++ by itself is crazy complicated. 16:06.060 --> 16:07.140 Can we just pause? 16:07.140 --> 16:09.140 What makes the language complicated in terms 16:09.140 --> 16:12.340 of what's syntactically? 16:12.340 --> 16:14.300 So it's what they call syntax. 16:14.300 --> 16:16.700 So the actual how the characters are arranged, yes. 16:16.700 --> 16:20.020 It's also semantics, how it behaves. 16:20.020 --> 16:21.900 It's also, in the case of C++, there's 16:21.900 --> 16:23.380 a huge amount of history. 16:23.380 --> 16:26.700 C++ is built on top of C. You play that forward. 16:26.700 --> 16:29.860 And then a bunch of suboptimal, in some cases, decisions 16:29.860 --> 16:31.620 were made, and they compound. 16:31.620 --> 16:33.380 And then more and more and more things 16:33.380 --> 16:36.980 keep getting added to C++, and it will probably never stop. 16:36.980 --> 16:38.540 But the language is very complicated 16:38.540 --> 16:39.540 from that perspective. 16:39.540 --> 16:41.200 And so the interactions between subsystems 16:41.200 --> 16:42.420 is very complicated. 16:42.420 --> 16:43.580 There's just a lot there. 16:43.580 --> 16:45.660 And when you talk about the front end, 16:45.660 --> 16:47.060 one of the major challenges, which 16:47.060 --> 16:51.140 clang as a project, the C, C++ compiler that I built, 16:51.140 --> 16:54.480 I and many people built, one of the challenges we took on 16:54.480 --> 16:57.780 was we looked at GCC. 16:57.780 --> 17:02.540 GCC, at the time, was a really good industry standardized 17:02.540 --> 17:05.260 compiler that had really consolidated 17:05.260 --> 17:08.340 a lot of the other compilers in the world and was a standard. 17:08.340 --> 17:10.620 But it wasn't really great for research. 17:10.620 --> 17:12.580 The design was very difficult to work with. 17:12.580 --> 17:16.620 And it was full of global variables and other things 17:16.620 --> 17:18.540 that made it very difficult to reuse in ways 17:18.540 --> 17:20.420 that it wasn't originally designed for. 17:20.420 --> 17:22.740 And so with clang, one of the things that we wanted to do 17:22.740 --> 17:25.500 is push forward on better user interface, 17:25.500 --> 17:28.060 so make error messages that are just better than GCC's. 17:28.060 --> 17:29.580 And that's actually hard, because you 17:29.580 --> 17:32.780 have to do a lot of bookkeeping in an efficient way 17:32.780 --> 17:33.700 to be able to do that. 17:33.700 --> 17:35.180 We want to make compile time better. 17:35.180 --> 17:37.500 And so compile time is about making it efficient, 17:37.500 --> 17:38.900 which is also really hard when you're keeping 17:38.900 --> 17:40.540 track of extra information. 17:40.540 --> 17:43.380 We wanted to make new tools available, 17:43.380 --> 17:46.380 so refactoring tools and other analysis tools 17:46.380 --> 17:50.540 that GCC never supported, also leveraging the extra information 17:50.540 --> 17:54.060 we kept, but enabling those new classes of tools 17:54.060 --> 17:55.940 that then get built into IDEs. 17:55.940 --> 17:59.380 And so that's been one of the areas that clang has really 17:59.380 --> 18:01.300 helped push the world forward in, 18:01.300 --> 18:05.060 is in the tooling for C and C++ and things like that. 18:05.060 --> 18:07.500 But C++ and the front end piece is complicated. 18:07.500 --> 18:09.000 And you have to build syntax trees. 18:09.000 --> 18:11.340 And you have to check every rule in the spec. 18:11.340 --> 18:14.020 And you have to turn that back into an error message 18:14.020 --> 18:16.020 to the human that the human can understand 18:16.020 --> 18:17.820 when they do something wrong. 18:17.820 --> 18:20.740 But then you start doing what's called lowering, 18:20.740 --> 18:23.060 so going from C++ and the way that it represents 18:23.060 --> 18:24.980 code down to the machine. 18:24.980 --> 18:27.380 And when you do that, there's many different phases 18:27.380 --> 18:29.660 you go through. 18:29.660 --> 18:33.020 Often, there are, I think LLVM has something like 150 18:33.020 --> 18:36.260 different what are called passes in the compiler 18:36.260 --> 18:38.780 that the code passes through. 18:38.780 --> 18:41.860 And these get organized in very complicated ways, 18:41.860 --> 18:44.360 which affect the generated code and the performance 18:44.360 --> 18:45.980 and compile time and many other things. 18:45.980 --> 18:47.300 What are they passing through? 18:47.300 --> 18:53.980 So after you do the clang parsing, what's the graph? 18:53.980 --> 18:54.900 What does it look like? 18:54.900 --> 18:56.100 What's the data structure here? 18:56.100 --> 18:59.060 Yeah, so in the parser, it's usually a tree. 18:59.060 --> 19:01.100 And it's called an abstract syntax tree. 19:01.100 --> 19:04.580 And so the idea is you have a node for the plus 19:04.580 --> 19:06.820 that the human wrote in their code. 19:06.820 --> 19:09.020 Or the function call, you'll have a node for call 19:09.020 --> 19:11.900 with the function that they call and the arguments they pass, 19:11.900 --> 19:14.460 things like that. 19:14.460 --> 19:16.620 This then gets lowered into what's 19:16.620 --> 19:18.620 called an intermediate representation. 19:18.620 --> 19:22.100 And intermediate representations are like LLVM has one. 19:22.100 --> 19:26.940 And there, it's what's called a control flow graph. 19:26.940 --> 19:31.220 And so you represent each operation in the program 19:31.220 --> 19:34.480 as a very simple, like this is going to add two numbers. 19:34.480 --> 19:35.980 This is going to multiply two things. 19:35.980 --> 19:37.460 Maybe we'll do a call. 19:37.460 --> 19:40.260 But then they get put in what are called blocks. 19:40.260 --> 19:43.580 And so you get blocks of these straight line operations, 19:43.580 --> 19:45.340 where instead of being nested like in a tree, 19:45.340 --> 19:46.900 it's straight line operations. 19:46.900 --> 19:49.780 And so there's a sequence and an ordering to these operations. 19:49.780 --> 19:51.820 So within the block or outside the block? 19:51.820 --> 19:52.980 That's within the block. 19:52.980 --> 19:54.980 And so it's a straight line sequence of operations 19:54.980 --> 19:55.740 within the block. 19:55.740 --> 19:58.980 And then you have branches, like conditional branches, 19:58.980 --> 20:00.140 between blocks. 20:00.140 --> 20:04.860 And so when you write a loop, for example, in a syntax tree, 20:04.860 --> 20:08.060 you would have a for node, like for a for statement 20:08.060 --> 20:10.540 in a C like language, you'd have a for node. 20:10.540 --> 20:12.200 And you have a pointer to the expression 20:12.200 --> 20:14.080 for the initializer, a pointer to the expression 20:14.080 --> 20:16.040 for the increment, a pointer to the expression 20:16.040 --> 20:18.900 for the comparison, a pointer to the body. 20:18.900 --> 20:21.060 And these are all nested underneath it. 20:21.060 --> 20:22.900 In a control flow graph, you get a block 20:22.900 --> 20:26.820 for the code that runs before the loop, so the initializer 20:26.820 --> 20:27.620 code. 20:27.620 --> 20:30.340 And you have a block for the body of the loop. 20:30.340 --> 20:33.780 And so the body of the loop code goes in there, 20:33.780 --> 20:35.660 but also the increment and other things like that. 20:35.660 --> 20:37.860 And then you have a branch that goes back to the top 20:37.860 --> 20:39.900 and a comparison and a branch that goes out. 20:39.900 --> 20:43.820 And so it's more of an assembly level kind of representation. 20:43.820 --> 20:46.060 But the nice thing about this level of representation 20:46.060 --> 20:48.700 is it's much more language independent. 20:48.700 --> 20:51.900 And so there's lots of different kinds of languages 20:51.900 --> 20:54.540 with different kinds of, you know, 20:54.540 --> 20:56.840 JavaScript has a lot of different ideas of what 20:56.840 --> 20:58.180 is false, for example. 20:58.180 --> 21:00.780 And all that can stay in the front end. 21:00.780 --> 21:04.220 But then that middle part can be shared across all those. 21:04.220 --> 21:07.540 How close is that intermediate representation 21:07.540 --> 21:10.620 to neural networks, for example? 21:10.620 --> 21:13.540 Are they, because everything you describe 21:13.540 --> 21:16.100 is a kind of echoes of a neural network graph. 21:16.100 --> 21:18.940 Are they neighbors or what? 21:18.940 --> 21:20.980 They're quite different in details, 21:20.980 --> 21:22.520 but they're very similar in idea. 21:22.520 --> 21:24.320 So one of the things that neural networks do 21:24.320 --> 21:26.900 is they learn representations for data 21:26.900 --> 21:29.140 at different levels of abstraction. 21:29.140 --> 21:33.940 And then they transform those through layers, right? 21:33.940 --> 21:35.660 So the compiler does very similar things. 21:35.660 --> 21:37.320 But one of the things the compiler does 21:37.320 --> 21:40.660 is it has relatively few different representations. 21:40.660 --> 21:43.100 Where a neural network often, as you get deeper, for example, 21:43.100 --> 21:44.820 you get many different representations 21:44.820 --> 21:47.380 in each layer or set of ops. 21:47.380 --> 21:50.260 It's transforming between these different representations. 21:50.260 --> 21:53.100 In a compiler, often you get one representation 21:53.100 --> 21:55.240 and they do many transformations to it. 21:55.240 --> 21:59.540 And these transformations are often applied iteratively. 21:59.540 --> 22:02.940 And for programmers, there's familiar types of things. 22:02.940 --> 22:06.180 For example, trying to find expressions inside of a loop 22:06.180 --> 22:08.540 and pulling them out of a loop so they execute for times. 22:08.540 --> 22:10.740 Or find redundant computation. 22:10.740 --> 22:15.380 Or find constant folding or other simplifications, 22:15.380 --> 22:19.060 turning two times x into x shift left by one. 22:19.060 --> 22:21.980 And things like this are all the examples 22:21.980 --> 22:23.340 of the things that happen. 22:23.340 --> 22:26.180 But compilers end up getting a lot of theorem proving 22:26.180 --> 22:27.760 and other kinds of algorithms that 22:27.760 --> 22:30.100 try to find higher level properties of the program that 22:30.100 --> 22:32.280 then can be used by the optimizer. 22:32.280 --> 22:32.780 Cool. 22:32.780 --> 22:38.140 So what's the biggest bang for the buck with optimization? 22:38.140 --> 22:38.640 Today? 22:38.640 --> 22:39.140 Yeah. 22:39.140 --> 22:40.900 Well, no, not even today. 22:40.900 --> 22:42.900 At the very beginning, the 80s, I don't know. 22:42.900 --> 22:44.300 Yeah, so for the 80s, a lot of it 22:44.300 --> 22:46.420 was things like register allocation. 22:46.420 --> 22:50.460 So the idea of in a modern microprocessor, 22:50.460 --> 22:51.880 what you'll end up having is you'll 22:51.880 --> 22:54.340 end up having memory, which is relatively slow. 22:54.340 --> 22:57.060 And then you have registers that are relatively fast. 22:57.060 --> 23:00.340 But registers, you don't have very many of them. 23:00.340 --> 23:02.600 And so when you're writing a bunch of code, 23:02.600 --> 23:04.180 you're just saying, compute this, 23:04.180 --> 23:05.940 put in a temporary variable, compute this, compute this, 23:05.940 --> 23:07.780 compute this, put in a temporary variable. 23:07.780 --> 23:08.220 I have a loop. 23:08.220 --> 23:09.780 I have some other stuff going on. 23:09.780 --> 23:11.660 Well, now you're running on an x86, 23:11.660 --> 23:13.900 like a desktop PC or something. 23:13.900 --> 23:16.860 Well, it only has, in some cases, some modes, 23:16.860 --> 23:18.700 eight registers. 23:18.700 --> 23:21.620 And so now the compiler has to choose what values get 23:21.620 --> 23:24.820 put in what registers at what points in the program. 23:24.820 --> 23:26.580 And this is actually a really big deal. 23:26.580 --> 23:29.500 So if you think about, you have a loop, an inner loop 23:29.500 --> 23:31.620 that executes millions of times maybe. 23:31.620 --> 23:33.620 If you're doing loads and stores inside that loop, 23:33.620 --> 23:35.040 then it's going to be really slow. 23:35.040 --> 23:37.740 But if you can somehow fit all the values inside that loop 23:37.740 --> 23:40.180 in registers, now it's really fast. 23:40.180 --> 23:43.020 And so getting that right requires a lot of work, 23:43.020 --> 23:44.940 because there's many different ways to do that. 23:44.940 --> 23:46.980 And often what the compiler ends up doing 23:46.980 --> 23:48.840 is it ends up thinking about things 23:48.840 --> 23:52.020 in a different representation than what the human wrote. 23:52.020 --> 23:53.340 You wrote into x. 23:53.340 --> 23:56.820 Well, the compiler thinks about that as four different values, 23:56.820 --> 23:59.280 each which have different lifetimes across the function 23:59.280 --> 24:00.420 that it's in. 24:00.420 --> 24:03.180 And each of those could be put in a register or memory 24:03.180 --> 24:06.140 or different memory or maybe in some parts of the code 24:06.140 --> 24:08.360 recomputed instead of stored and reloaded. 24:08.360 --> 24:10.700 And there are many of these different kinds of techniques 24:10.700 --> 24:11.460 that can be used. 24:11.460 --> 24:15.780 So it's adding almost like a time dimension to it's 24:15.780 --> 24:18.300 trying to optimize across time. 24:18.300 --> 24:20.340 So it's considering when you're programming, 24:20.340 --> 24:21.860 you're not thinking in that way. 24:21.860 --> 24:23.220 Yeah, absolutely. 24:23.220 --> 24:27.100 And so the RISC era made things. 24:27.100 --> 24:32.020 So RISC chips, R I S C. The RISC chips, 24:32.020 --> 24:33.740 as opposed to CISC chips. 24:33.740 --> 24:36.700 The RISC chips made things more complicated for the compiler, 24:36.700 --> 24:40.660 because what they ended up doing is ending up 24:40.660 --> 24:42.500 adding pipelines to the processor, where 24:42.500 --> 24:45.020 the processor can do more than one thing at a time. 24:45.020 --> 24:47.740 But this means that the order of operations matters a lot. 24:47.740 --> 24:50.260 So one of the classical compiler techniques that you use 24:50.260 --> 24:51.940 is called scheduling. 24:51.940 --> 24:54.220 And so moving the instructions around 24:54.220 --> 24:57.740 so that the processor can keep its pipelines full instead 24:57.740 --> 24:59.220 of stalling and getting blocked. 24:59.220 --> 25:01.180 And so there's a lot of things like that that 25:01.180 --> 25:03.620 are kind of bread and butter compiler techniques 25:03.620 --> 25:06.220 that have been studied a lot over the course of decades now. 25:06.220 --> 25:08.540 But the engineering side of making them real 25:08.540 --> 25:10.580 is also still quite hard. 25:10.580 --> 25:12.460 And you talk about machine learning. 25:12.460 --> 25:14.420 This is a huge opportunity for machine learning, 25:14.420 --> 25:17.620 because many of these algorithms are full of these 25:17.620 --> 25:19.300 hokey, hand rolled heuristics, which 25:19.300 --> 25:21.820 work well on specific benchmarks that don't generalize, 25:21.820 --> 25:23.940 and full of magic numbers. 25:23.940 --> 25:26.620 And I hear there's some techniques that 25:26.620 --> 25:28.060 are good at handling that. 25:28.060 --> 25:32.220 So what would be the, if you were to apply machine learning 25:32.220 --> 25:34.740 to this, what's the thing you're trying to optimize? 25:34.740 --> 25:39.100 Is it ultimately the running time? 25:39.100 --> 25:41.180 You can pick your metric, and there's running time, 25:41.180 --> 25:43.900 there's memory use, there's lots of different things 25:43.900 --> 25:44.940 that you can optimize for. 25:44.940 --> 25:47.220 Code size is another one that some people care about 25:47.220 --> 25:48.860 in the embedded space. 25:48.860 --> 25:51.700 Is this like the thinking into the future, 25:51.700 --> 25:54.500 or has somebody actually been crazy enough 25:54.500 --> 25:58.060 to try to have machine learning based parameter 25:58.060 --> 26:01.060 tuning for the optimization of compilers? 26:01.060 --> 26:04.860 So this is something that is, I would say, research right now. 26:04.860 --> 26:06.820 There are a lot of research systems 26:06.820 --> 26:09.100 that have been applying search in various forms. 26:09.100 --> 26:11.460 And using reinforcement learning is one form, 26:11.460 --> 26:14.460 but also brute force search has been tried for quite a while. 26:14.460 --> 26:18.180 And usually, these are in small problem spaces. 26:18.180 --> 26:21.900 So find the optimal way to code generate a matrix 26:21.900 --> 26:24.460 multiply for a GPU, something like that, 26:24.460 --> 26:28.580 where you say, there, there's a lot of design space of, 26:28.580 --> 26:29.900 do you unroll loops a lot? 26:29.900 --> 26:32.660 Do you execute multiple things in parallel? 26:32.660 --> 26:35.340 And there's many different confounding factors here 26:35.340 --> 26:38.100 because graphics cards have different numbers of threads 26:38.100 --> 26:41.020 and registers and execution ports and memory bandwidth 26:41.020 --> 26:42.740 and many different constraints that interact 26:42.740 --> 26:44.460 in nonlinear ways. 26:44.460 --> 26:46.500 And so search is very powerful for that. 26:46.500 --> 26:49.820 And it gets used in certain ways, 26:49.820 --> 26:51.220 but it's not very structured. 26:51.220 --> 26:52.620 This is something that we need, 26:52.620 --> 26:54.500 we as an industry need to fix. 26:54.500 --> 26:59.220 So you said 80s, but like, so have there been like big jumps 26:59.220 --> 27:01.260 in improvement and optimization? 27:01.260 --> 27:02.340 Yeah. 27:02.340 --> 27:05.300 Yeah, since then, what's the coolest thing? 27:05.300 --> 27:07.100 It's largely been driven by hardware. 27:07.100 --> 27:09.860 So, well, it's hardware and software. 27:09.860 --> 27:13.700 So in the mid nineties, Java totally changed the world, 27:13.700 --> 27:14.540 right? 27:14.540 --> 27:17.540 And I'm still amazed by how much change was introduced 27:17.540 --> 27:19.340 by the way or in a good way. 27:19.340 --> 27:22.420 So like reflecting back, Java introduced things like, 27:22.420 --> 27:25.860 all at once introduced things like JIT compilation. 27:25.860 --> 27:27.780 None of these were novel, but it pulled it together 27:27.780 --> 27:30.580 and made it mainstream and made people invest in it. 27:30.580 --> 27:33.620 JIT compilation, garbage collection, portable code, 27:33.620 --> 27:36.620 safe code, like memory safe code, 27:36.620 --> 27:41.380 like a very dynamic dispatch execution model. 27:41.380 --> 27:42.620 Like many of these things, 27:42.620 --> 27:44.060 which had been done in research systems 27:44.060 --> 27:46.900 and had been done in small ways in various places, 27:46.900 --> 27:47.980 really came to the forefront, 27:47.980 --> 27:49.740 really changed how things worked 27:49.740 --> 27:51.980 and therefore changed the way people thought 27:51.980 --> 27:53.060 about the problem. 27:53.060 --> 27:56.300 JavaScript was another major world change 27:56.300 --> 27:57.740 based on the way it works. 27:59.300 --> 28:01.300 But also on the hardware side of things, 28:01.300 --> 28:06.300 multi core and vector instructions really change 28:06.660 --> 28:08.380 the problem space and are very, 28:09.460 --> 28:10.820 they don't remove any of the problems 28:10.820 --> 28:12.380 that compilers faced in the past, 28:12.380 --> 28:14.540 but they add new kinds of problems 28:14.540 --> 28:16.380 of how do you find enough work 28:16.380 --> 28:20.020 to keep a four wide vector busy, right? 28:20.020 --> 28:22.660 Or if you're doing a matrix multiplication, 28:22.660 --> 28:25.860 how do you do different columns out of that matrix 28:25.860 --> 28:26.700 at the same time? 28:26.700 --> 28:30.140 And how do you maximally utilize the arithmetic compute 28:30.140 --> 28:31.460 that one core has? 28:31.460 --> 28:33.500 And then how do you take it to multiple cores? 28:33.500 --> 28:35.780 How did the whole virtual machine thing change 28:35.780 --> 28:38.020 the compilation pipeline? 28:38.020 --> 28:40.460 Yeah, so what the Java virtual machine does 28:40.460 --> 28:44.180 is it splits, just like I was talking about before, 28:44.180 --> 28:46.300 where you have a front end that parses the code, 28:46.300 --> 28:48.020 and then you have an intermediate representation 28:48.020 --> 28:49.460 that gets transformed. 28:49.460 --> 28:51.020 What Java did was they said, 28:51.020 --> 28:53.100 we will parse the code and then compile to 28:53.100 --> 28:55.500 what's known as Java byte code. 28:55.500 --> 28:58.580 And that byte code is now a portable code representation 28:58.580 --> 29:02.420 that is industry standard and locked down and can't change. 29:02.420 --> 29:05.100 And then the back part of the compiler 29:05.100 --> 29:07.300 that does optimization and code generation 29:07.300 --> 29:09.460 can now be built by different vendors. 29:09.460 --> 29:10.300 Okay. 29:10.300 --> 29:13.020 And Java byte code can be shipped around across the wire. 29:13.020 --> 29:15.860 It's memory safe and relatively trusted. 29:16.860 --> 29:18.660 And because of that, it can run in the browser. 29:18.660 --> 29:20.540 And that's why it runs in the browser, right? 29:20.540 --> 29:22.980 And so that way you can be in, 29:22.980 --> 29:25.020 again, back in the day, you would write a Java applet 29:25.020 --> 29:29.300 and as a web developer, you'd build this mini app 29:29.300 --> 29:30.860 that would run on a webpage. 29:30.860 --> 29:33.620 Well, a user of that is running a web browser 29:33.620 --> 29:34.460 on their computer. 29:34.460 --> 29:37.860 You download that Java byte code, which can be trusted, 29:37.860 --> 29:41.060 and then you do all the compiler stuff on your machine 29:41.060 --> 29:42.460 so that you know that you trust that. 29:42.460 --> 29:44.060 Now, is that a good idea or a bad idea? 29:44.060 --> 29:44.900 It's a great idea. 29:44.900 --> 29:46.240 I mean, it's a great idea for certain problems. 29:46.240 --> 29:49.540 And I'm very much a believer that technology is itself 29:49.540 --> 29:50.520 neither good nor bad. 29:50.520 --> 29:51.620 It's how you apply it. 29:52.940 --> 29:54.660 You know, this would be a very, very bad thing 29:54.660 --> 29:56.980 for very low levels of the software stack. 29:56.980 --> 30:00.300 But in terms of solving some of these software portability 30:00.300 --> 30:02.820 and transparency, or portability problems, 30:02.820 --> 30:04.240 I think it's been really good. 30:04.240 --> 30:06.600 Now, Java ultimately didn't win out on the desktop. 30:06.600 --> 30:09.420 And like, there are good reasons for that. 30:09.420 --> 30:13.220 But it's been very successful on servers and in many places, 30:13.220 --> 30:16.300 it's been a very successful thing over decades. 30:16.300 --> 30:21.300 So what has been LLVMs and C langs improvements 30:21.300 --> 30:26.300 and optimization that throughout its history, 30:28.640 --> 30:31.080 what are some moments we had set back 30:31.080 --> 30:33.280 and really proud of what's been accomplished? 30:33.280 --> 30:36.160 Yeah, I think that the interesting thing about LLVM 30:36.160 --> 30:40.120 is not the innovations and compiler research. 30:40.120 --> 30:41.900 It has very good implementations 30:41.900 --> 30:44.000 of various important algorithms, no doubt. 30:44.880 --> 30:48.280 And a lot of really smart people have worked on it. 30:48.280 --> 30:50.560 But I think that the thing that's most profound about LLVM 30:50.560 --> 30:53.840 is that through standardization, it made things possible 30:53.840 --> 30:56.200 that otherwise wouldn't have happened, okay? 30:56.200 --> 30:59.120 And so interesting things that have happened with LLVM, 30:59.120 --> 31:01.260 for example, Sony has picked up LLVM 31:01.260 --> 31:03.920 and used it to do all the graphics compilation 31:03.920 --> 31:06.080 in their movie production pipeline. 31:06.080 --> 31:07.920 And so now they're able to have better special effects 31:07.920 --> 31:09.660 because of LLVM. 31:09.660 --> 31:11.180 That's kind of cool. 31:11.180 --> 31:13.000 That's not what it was designed for, right? 31:13.000 --> 31:15.480 But that's the sign of good infrastructure 31:15.480 --> 31:18.800 when it can be used in ways it was never designed for 31:18.800 --> 31:20.960 because it has good layering and software engineering 31:20.960 --> 31:23.440 and it's composable and things like that. 31:23.440 --> 31:26.120 Which is where, as you said, it differs from GCC. 31:26.120 --> 31:28.240 Yes, GCC is also great in various ways, 31:28.240 --> 31:31.800 but it's not as good as infrastructure technology. 31:31.800 --> 31:36.160 It's really a C compiler, or it's a Fortran compiler. 31:36.160 --> 31:38.920 It's not infrastructure in the same way. 31:38.920 --> 31:41.560 Now you can tell I don't know what I'm talking about 31:41.560 --> 31:44.500 because I keep saying C lang. 31:44.500 --> 31:48.080 You can always tell when a person has clues, 31:48.080 --> 31:49.400 by the way, to pronounce something. 31:49.400 --> 31:52.580 I don't think, have I ever used C lang? 31:52.580 --> 31:54.120 Entirely possible, have you? 31:54.120 --> 31:58.200 Well, so you've used code, it's generated probably. 31:58.200 --> 32:01.760 So C lang and LLVM are used to compile 32:01.760 --> 32:05.240 all the apps on the iPhone effectively and the OSs. 32:05.240 --> 32:09.380 It compiles Google's production server applications. 32:10.560 --> 32:14.840 It's used to build GameCube games and PlayStation 4 32:14.840 --> 32:16.680 and things like that. 32:16.680 --> 32:20.120 So as a user, I have, but just everything I've done 32:20.120 --> 32:22.120 that I experienced with Linux has been, 32:22.120 --> 32:23.560 I believe, always GCC. 32:23.560 --> 32:26.520 Yeah, I think Linux still defaults to GCC. 32:26.520 --> 32:27.800 And is there a reason for that? 32:27.800 --> 32:29.440 Or is it because, I mean, is there a reason for that? 32:29.440 --> 32:32.040 It's a combination of technical and social reasons. 32:32.040 --> 32:35.960 Many Linux developers do use C lang, 32:35.960 --> 32:39.720 but the distributions, for lots of reasons, 32:40.560 --> 32:44.240 use GCC historically, and they've not switched, yeah. 32:44.240 --> 32:46.640 Because it's just anecdotally online, 32:46.640 --> 32:50.640 it seems that LLVM has either reached the level of GCC 32:50.640 --> 32:53.520 or superseded on different features or whatever. 32:53.520 --> 32:55.200 The way I would say it is that they're so close, 32:55.200 --> 32:56.040 it doesn't matter. 32:56.040 --> 32:56.860 Yeah, exactly. 32:56.860 --> 32:58.160 Like, they're slightly better in some ways, 32:58.160 --> 32:59.160 slightly worse than otherwise, 32:59.160 --> 33:03.280 but it doesn't actually really matter anymore, that level. 33:03.280 --> 33:06.280 So in terms of optimization breakthroughs, 33:06.280 --> 33:09.160 it's just been solid incremental work. 33:09.160 --> 33:12.520 Yeah, yeah, which describes a lot of compilers. 33:12.520 --> 33:15.000 The hard thing about compilers, in my experience, 33:15.000 --> 33:17.440 is the engineering, the software engineering, 33:17.440 --> 33:20.160 making it so that you can have hundreds of people 33:20.160 --> 33:23.600 collaborating on really detailed, low level work 33:23.600 --> 33:25.400 and scaling that. 33:25.400 --> 33:27.880 And that's really hard. 33:27.880 --> 33:30.680 And that's one of the things I think LLVM has done well. 33:32.160 --> 33:34.200 And that kind of goes back to the original design goals 33:34.200 --> 33:37.200 with it to be modular and things like that. 33:37.200 --> 33:38.880 And incidentally, I don't want to take all the credit 33:38.880 --> 33:39.720 for this, right? 33:39.720 --> 33:41.760 I mean, some of the best parts about LLVM 33:41.760 --> 33:43.600 is that it was designed to be modular. 33:43.600 --> 33:45.600 And when I started, I would write, for example, 33:45.600 --> 33:48.500 a register allocator, and then somebody much smarter than me 33:48.500 --> 33:50.720 would come in and pull it out and replace it 33:50.720 --> 33:52.680 with something else that they would come up with. 33:52.680 --> 33:55.200 And because it's modular, they were able to do that. 33:55.200 --> 33:58.280 And that's one of the challenges with GCC, for example, 33:58.280 --> 34:01.280 is replacing subsystems is incredibly difficult. 34:01.280 --> 34:04.680 It can be done, but it wasn't designed for that. 34:04.680 --> 34:06.080 And that's one of the reasons that LLVM's been 34:06.080 --> 34:08.760 very successful in the research world as well. 34:08.760 --> 34:12.960 But in a community sense, Guido van Rossum, right, 34:12.960 --> 34:17.960 from Python, just retired from, what is it? 34:18.480 --> 34:20.500 Benevolent Dictator for Life, right? 34:20.500 --> 34:24.720 So in managing this community of brilliant compiler folks, 34:24.720 --> 34:28.660 is there, did it, for a time at least, 34:28.660 --> 34:31.480 fall on you to approve things? 34:31.480 --> 34:34.240 Oh yeah, so I mean, I still have something like 34:34.240 --> 34:37.980 an order of magnitude more patches in LLVM 34:37.980 --> 34:42.760 than anybody else, and many of those I wrote myself. 34:42.760 --> 34:47.760 But you still write, I mean, you're still close to the, 34:47.880 --> 34:49.480 to the, I don't know what the expression is, 34:49.480 --> 34:51.000 to the metal, you still write code. 34:51.000 --> 34:52.220 Yeah, I still write code. 34:52.220 --> 34:54.240 Not as much as I was able to in grad school, 34:54.240 --> 34:56.760 but that's an important part of my identity. 34:56.760 --> 34:58.880 But the way that LLVM has worked over time 34:58.880 --> 35:01.360 is that when I was a grad student, I could do all the work 35:01.360 --> 35:04.120 and steer everything and review every patch 35:04.120 --> 35:05.800 and make sure everything was done 35:05.800 --> 35:09.040 exactly the way my opinionated sense 35:09.040 --> 35:11.760 felt like it should be done, and that was fine. 35:11.760 --> 35:14.300 But as things scale, you can't do that, right? 35:14.300 --> 35:17.100 And so what ends up happening is LLVM 35:17.100 --> 35:20.520 has a hierarchical system of what's called code owners. 35:20.520 --> 35:22.880 These code owners are given the responsibility 35:22.880 --> 35:24.880 not to do all the work, 35:24.880 --> 35:26.640 not necessarily to review all the patches, 35:26.640 --> 35:28.800 but to make sure that the patches do get reviewed 35:28.800 --> 35:30.320 and make sure that the right thing's happening 35:30.320 --> 35:32.160 architecturally in their area. 35:32.160 --> 35:36.720 And so what you'll see is you'll see that, for example, 35:36.720 --> 35:38.560 hardware manufacturers end up owning 35:38.560 --> 35:43.560 the hardware specific parts of their hardware. 35:43.600 --> 35:44.520 That's very common. 35:45.520 --> 35:47.720 Leaders in the community that have done really good work 35:47.720 --> 35:50.880 naturally become the de facto owner of something. 35:50.880 --> 35:53.400 And then usually somebody else is like, 35:53.400 --> 35:55.520 how about we make them the official code owner? 35:55.520 --> 35:58.600 And then we'll have somebody to make sure 35:58.600 --> 36:00.320 that all the patches get reviewed in a timely manner. 36:00.320 --> 36:02.080 And then everybody's like, yes, that's obvious. 36:02.080 --> 36:03.240 And then it happens, right? 36:03.240 --> 36:06.080 And usually this is a very organic thing, which is great. 36:06.080 --> 36:08.740 And so I'm nominally the top of that stack still, 36:08.740 --> 36:11.560 but I don't spend a lot of time reviewing patches. 36:11.560 --> 36:16.520 What I do is I help negotiate a lot of the technical 36:16.520 --> 36:18.040 disagreements that end up happening 36:18.040 --> 36:19.660 and making sure that the community as a whole 36:19.660 --> 36:22.040 makes progress and is moving in the right direction 36:22.040 --> 36:23.920 and doing that. 36:23.920 --> 36:28.240 So we also started a nonprofit six years ago, 36:28.240 --> 36:30.840 seven years ago, time's gone away. 36:30.840 --> 36:34.600 And the LLVM Foundation nonprofit helps oversee 36:34.600 --> 36:36.440 all the business sides of things and make sure 36:36.440 --> 36:38.800 that the events that the LLVM community has 36:38.800 --> 36:41.600 are funded and set up and run correctly 36:41.600 --> 36:42.800 and stuff like that. 36:42.800 --> 36:45.160 But the foundation is very much stays out 36:45.160 --> 36:49.060 of the technical side of where the project is going. 36:49.060 --> 36:52.160 Right, so it sounds like a lot of it is just organic. 36:53.160 --> 36:55.680 Yeah, well, LLVM is almost 20 years old, 36:55.680 --> 36:56.600 which is hard to believe. 36:56.600 --> 36:59.720 Somebody pointed out to me recently that LLVM 36:59.720 --> 37:04.600 is now older than GCC was when LLVM started, right? 37:04.600 --> 37:06.860 So time has a way of getting away from you. 37:06.860 --> 37:10.400 But the good thing about that is it has a really robust, 37:10.400 --> 37:13.520 really amazing community of people that are 37:13.520 --> 37:15.460 in their professional lives, spread across lots 37:15.460 --> 37:17.720 of different companies, but it's a community 37:17.720 --> 37:21.120 of people that are interested in similar kinds of problems 37:21.120 --> 37:23.680 and have been working together effectively for years 37:23.680 --> 37:26.460 and have a lot of trust and respect for each other. 37:26.460 --> 37:29.240 And even if they don't always agree that we're able 37:29.240 --> 37:31.200 to find a path forward. 37:31.200 --> 37:34.480 So then in a slightly different flavor of effort, 37:34.480 --> 37:38.120 you started at Apple in 2005 with the task 37:38.120 --> 37:41.800 of making, I guess, LLVM production ready. 37:41.800 --> 37:44.640 And then eventually 2013 through 2017, 37:44.640 --> 37:48.360 leading the entire developer tools department. 37:48.360 --> 37:52.960 We're talking about LLVM, Xcode, Objective C to Swift. 37:53.920 --> 37:58.580 So in a quick overview of your time there, 37:58.580 --> 37:59.600 what were the challenges? 37:59.600 --> 38:03.240 First of all, leading such a huge group of developers, 38:03.240 --> 38:06.540 what was the big motivator, dream, mission 38:06.540 --> 38:11.400 behind creating Swift, the early birth of it 38:11.400 --> 38:13.400 from Objective C and so on, and Xcode, 38:13.400 --> 38:14.240 what are some challenges? 38:14.240 --> 38:15.900 So these are different questions. 38:15.900 --> 38:19.720 Yeah, I know, but I wanna talk about the other stuff too. 38:19.720 --> 38:21.240 I'll stay on the technical side, 38:21.240 --> 38:24.480 then we can talk about the big team pieces, if that's okay. 38:24.480 --> 38:29.060 So it's to really oversimplify many years of hard work. 38:29.060 --> 38:32.440 LLVM started, joined Apple, became a thing, 38:32.440 --> 38:34.600 became successful and became deployed. 38:34.600 --> 38:35.960 But then there's a question about 38:35.960 --> 38:38.880 how do we actually parse the source code? 38:38.880 --> 38:40.320 So LLVM is that back part, 38:40.320 --> 38:42.320 the optimizer and the code generator. 38:42.320 --> 38:44.060 And LLVM was really good for Apple 38:44.060 --> 38:46.060 as it went through a couple of harder transitions. 38:46.060 --> 38:47.960 I joined right at the time of the Intel transition, 38:47.960 --> 38:51.820 for example, and 64 bit transitions, 38:51.820 --> 38:53.500 and then the transition to ARM with the iPhone. 38:53.500 --> 38:54.720 And so LLVM was very useful 38:54.720 --> 38:57.000 for some of these kinds of things. 38:57.000 --> 38:58.480 But at the same time, there's a lot of questions 38:58.480 --> 39:00.120 around developer experience. 39:00.120 --> 39:01.960 And so if you're a programmer pounding out 39:01.960 --> 39:03.460 at the time Objective C code, 39:04.480 --> 39:06.520 the error message you get, the compile time, 39:06.520 --> 39:09.760 the turnaround cycle, the tooling and the IDE, 39:09.760 --> 39:13.000 were not great, were not as good as they could be. 39:13.000 --> 39:18.000 And so, as I occasionally do, I'm like, 39:18.080 --> 39:20.720 well, okay, how hard is it to write a C compiler? 39:20.720 --> 39:22.560 And so I'm not gonna commit to anybody, 39:22.560 --> 39:25.320 I'm not gonna tell anybody, I'm just gonna just do it 39:25.320 --> 39:27.480 nights and weekends and start working on it. 39:27.480 --> 39:29.740 And then I built up in C, 39:29.740 --> 39:31.160 there's this thing called the preprocessor, 39:31.160 --> 39:33.040 which people don't like, 39:33.040 --> 39:35.480 but it's actually really hard and complicated 39:35.480 --> 39:37.700 and includes a bunch of really weird things 39:37.700 --> 39:39.280 like trigraphs and other stuff like that 39:39.280 --> 39:40.960 that are really nasty, 39:40.960 --> 39:44.080 and it's the crux of a bunch of the performance issues 39:44.080 --> 39:45.640 in the compiler. 39:45.640 --> 39:46.640 Started working on the parser 39:46.640 --> 39:47.800 and kind of got to the point where I'm like, 39:47.800 --> 39:49.880 ah, you know what, we could actually do this. 39:49.880 --> 39:51.460 Everybody's saying that this is impossible to do, 39:51.460 --> 39:53.960 but it's actually just hard, it's not impossible. 39:53.960 --> 39:57.560 And eventually told my manager about it, 39:57.560 --> 39:59.220 and he's like, oh, wow, this is great, 39:59.220 --> 40:00.360 we do need to solve this problem. 40:00.360 --> 40:02.560 Oh, this is great, we can get you one other person 40:02.560 --> 40:04.440 to work with you on this, you know? 40:04.440 --> 40:08.360 And slowly a team is formed and it starts taking off. 40:08.360 --> 40:12.040 And C++, for example, huge, complicated language. 40:12.040 --> 40:14.360 People always assume that it's impossible to implement 40:14.360 --> 40:16.260 and it's very nearly impossible, 40:16.260 --> 40:18.720 but it's just really, really hard. 40:18.720 --> 40:20.840 And the way to get there is to build it 40:20.840 --> 40:22.480 one piece at a time incrementally. 40:22.480 --> 40:26.440 And that was only possible because we were lucky 40:26.440 --> 40:28.160 to hire some really exceptional engineers 40:28.160 --> 40:30.380 that knew various parts of it very well 40:30.380 --> 40:32.680 and could do great things. 40:32.680 --> 40:34.440 Swift was kind of a similar thing. 40:34.440 --> 40:39.160 So Swift came from, we were just finishing off 40:39.160 --> 40:42.600 the first version of C++ support in Clang. 40:42.600 --> 40:47.260 And C++ is a very formidable and very important language, 40:47.260 --> 40:49.280 but it's also ugly in lots of ways. 40:49.280 --> 40:52.320 And you can't influence C++ without thinking 40:52.320 --> 40:54.380 there has to be a better thing, right? 40:54.380 --> 40:56.120 And so I started working on Swift, again, 40:56.120 --> 40:58.560 with no hope or ambition that would go anywhere, 40:58.560 --> 41:00.800 just let's see what could be done, 41:00.800 --> 41:02.620 let's play around with this thing. 41:02.620 --> 41:06.700 It was me in my spare time, not telling anybody about it, 41:06.700 --> 41:09.420 kind of a thing, and it made some good progress. 41:09.420 --> 41:11.260 I'm like, actually, it would make sense to do this. 41:11.260 --> 41:14.800 At the same time, I started talking with the senior VP 41:14.800 --> 41:17.720 of software at the time, a guy named Bertrand Serlet. 41:17.720 --> 41:19.280 And Bertrand was very encouraging. 41:19.280 --> 41:22.080 He was like, well, let's have fun, let's talk about this. 41:22.080 --> 41:23.440 And he was a little bit of a language guy, 41:23.440 --> 41:26.160 and so he helped guide some of the early work 41:26.160 --> 41:30.420 and encouraged me and got things off the ground. 41:30.420 --> 41:34.280 And eventually told my manager and told other people, 41:34.280 --> 41:38.800 and it started making progress. 41:38.800 --> 41:40.960 The complicating thing with Swift 41:40.960 --> 41:43.880 was that the idea of doing a new language 41:43.880 --> 41:47.840 was not obvious to anybody, including myself. 41:47.840 --> 41:50.240 And the tone at the time was that the iPhone 41:50.240 --> 41:53.440 was successful because of Objective C. 41:53.440 --> 41:54.440 Oh, interesting. 41:54.440 --> 41:57.160 Not despite of or just because of. 41:57.160 --> 42:01.160 And you have to understand that at the time, 42:01.160 --> 42:05.400 Apple was hiring software people that loved Objective C. 42:05.400 --> 42:07.960 And it wasn't that they came despite Objective C. 42:07.960 --> 42:10.240 They loved Objective C, and that's why they got hired. 42:10.240 --> 42:13.080 And so you had a software team that the leadership, 42:13.080 --> 42:15.200 in many cases, went all the way back to Next, 42:15.200 --> 42:19.400 where Objective C really became real. 42:19.400 --> 42:23.240 And so they, quote unquote, grew up writing Objective C. 42:23.240 --> 42:25.720 And many of the individual engineers 42:25.720 --> 42:28.360 all were hired because they loved Objective C. 42:28.360 --> 42:30.560 And so this notion of, OK, let's do new language 42:30.560 --> 42:34.120 was kind of heretical in many ways. 42:34.120 --> 42:36.960 Meanwhile, my sense was that the outside community wasn't really 42:36.960 --> 42:38.560 in love with Objective C. Some people were, 42:38.560 --> 42:40.360 and some of the most outspoken people were. 42:40.360 --> 42:42.620 But other people were hitting challenges 42:42.620 --> 42:44.760 because it has very sharp corners 42:44.760 --> 42:46.840 and it's difficult to learn. 42:46.840 --> 42:50.160 And so one of the challenges of making Swift happen that 42:50.160 --> 42:57.720 was totally non technical is the social part of what do we do? 42:57.720 --> 43:00.320 If we do a new language, which at Apple, many things 43:00.320 --> 43:02.240 happen that don't ship. 43:02.240 --> 43:05.560 So if we ship it, what is the metrics of success? 43:05.560 --> 43:06.400 Why would we do this? 43:06.400 --> 43:08.060 Why wouldn't we make Objective C better? 43:08.060 --> 43:10.160 If Objective C has problems, let's file off 43:10.160 --> 43:12.160 those rough corners and edges. 43:12.160 --> 43:15.640 And one of the major things that became the reason to do this 43:15.640 --> 43:18.960 was this notion of safety, memory safety. 43:18.960 --> 43:23.240 And the way Objective C works is that a lot of the object system 43:23.240 --> 43:27.560 and everything else is built on top of pointers in C. 43:27.560 --> 43:29.960 Objective C is an extension on top of C. 43:29.960 --> 43:32.680 And so pointers are unsafe. 43:32.680 --> 43:34.640 And if you get rid of the pointers, 43:34.640 --> 43:36.480 it's not Objective C anymore. 43:36.480 --> 43:39.080 And so fundamentally, that was an issue 43:39.080 --> 43:42.200 that you could not fix safety or memory safety 43:42.200 --> 43:45.640 without fundamentally changing the language. 43:45.640 --> 43:49.920 And so once we got through that part of the mental process 43:49.920 --> 43:53.200 and the thought process, it became a design process 43:53.200 --> 43:55.400 of saying, OK, well, if we're going to do something new, 43:55.400 --> 43:56.280 what is good? 43:56.280 --> 43:57.400 How do we think about this? 43:57.400 --> 43:58.200 And what do we like? 43:58.200 --> 44:00.040 And what are we looking for? 44:00.040 --> 44:02.440 And that was a very different phase of it. 44:02.440 --> 44:05.960 So what are some design choices early on in Swift? 44:05.960 --> 44:10.120 Like we're talking about braces, are you 44:10.120 --> 44:13.240 making a typed language or not, all those kinds of things. 44:13.240 --> 44:16.040 Yeah, so some of those were obvious given the context. 44:16.040 --> 44:17.800 So a typed language, for example, 44:17.800 --> 44:19.200 Objective C is a typed language. 44:19.200 --> 44:22.480 And going with an untyped language 44:22.480 --> 44:24.320 wasn't really seriously considered. 44:24.320 --> 44:26.000 We wanted the performance, and we 44:26.000 --> 44:27.680 wanted refactoring tools and other things 44:27.680 --> 44:29.600 like that that go with typed languages. 44:29.600 --> 44:31.440 Quick, dumb question. 44:31.440 --> 44:34.600 Was it obvious, I think this would be a dumb question, 44:34.600 --> 44:36.360 but was it obvious that the language 44:36.360 --> 44:40.120 has to be a compiled language? 44:40.120 --> 44:42.080 Yes, that's not a dumb question. 44:42.080 --> 44:44.520 Earlier, I think late 90s, Apple had seriously 44:44.520 --> 44:49.000 considered moving its development experience to Java. 44:49.000 --> 44:53.160 But Swift started in 2010, which was several years 44:53.160 --> 44:53.880 after the iPhone. 44:53.880 --> 44:55.380 It was when the iPhone was definitely 44:55.380 --> 44:56.640 on an upward trajectory. 44:56.640 --> 44:58.760 And the iPhone was still extremely, 44:58.760 --> 45:01.800 and is still a bit memory constrained. 45:01.800 --> 45:04.440 And so being able to compile the code 45:04.440 --> 45:08.160 and then ship it and then having standalone code that 45:08.160 --> 45:11.320 is not JIT compiled is a very big deal 45:11.320 --> 45:15.200 and is very much part of the Apple value system. 45:15.200 --> 45:17.480 Now, JavaScript's also a thing. 45:17.480 --> 45:19.360 I mean, it's not that this is exclusive, 45:19.360 --> 45:21.640 and technologies are good depending 45:21.640 --> 45:23.880 on how they're applied. 45:23.880 --> 45:26.600 But in the design of Swift, saying, 45:26.600 --> 45:28.320 how can we make Objective C better? 45:28.320 --> 45:29.760 Objective C is statically compiled, 45:29.760 --> 45:32.520 and that was the contiguous, natural thing to do. 45:32.520 --> 45:35.360 Just skip ahead a little bit, and we'll go right back. 45:35.360 --> 45:40.040 Just as a question, as you think about today in 2019 45:40.040 --> 45:42.400 in your work at Google, TensorFlow and so on, 45:42.400 --> 45:48.600 is, again, compilations, static compilation still 45:48.600 --> 45:49.460 the right thing? 45:49.460 --> 45:52.000 Yeah, so the funny thing after working 45:52.000 --> 45:55.880 on compilers for a really long time is that, 45:55.880 --> 45:59.040 and this is one of the things that LLVM has helped with, 45:59.040 --> 46:01.440 is that I don't look at compilations 46:01.440 --> 46:05.240 being static or dynamic or interpreted or not. 46:05.240 --> 46:07.680 This is a spectrum. 46:07.680 --> 46:09.140 And one of the cool things about Swift 46:09.140 --> 46:12.160 is that Swift is not just statically compiled. 46:12.160 --> 46:14.080 It's actually dynamically compiled as well, 46:14.080 --> 46:15.320 and it can also be interpreted. 46:15.320 --> 46:17.440 Though, nobody's actually done that. 46:17.440 --> 46:20.400 And so what ends up happening when 46:20.400 --> 46:24.080 you use Swift in a workbook, for example in Colab or in Jupyter, 46:24.080 --> 46:26.360 is it's actually dynamically compiling the statements 46:26.360 --> 46:28.160 as you execute them. 46:28.160 --> 46:32.840 And so this gets back to the software engineering problems, 46:32.840 --> 46:34.960 where if you layer the stack properly, 46:34.960 --> 46:37.320 you can actually completely change 46:37.320 --> 46:39.360 how and when things get compiled because you 46:39.360 --> 46:41.120 have the right abstractions there. 46:41.120 --> 46:44.800 And so the way that a Colab workbook works with Swift 46:44.800 --> 46:47.720 is that when you start typing into it, 46:47.720 --> 46:50.280 it creates a process, a Unix process. 46:50.280 --> 46:52.160 And then each line of code you type in, 46:52.160 --> 46:56.120 it compiles it through the Swift compiler, the front end part, 46:56.120 --> 46:58.360 and then sends it through the optimizer, 46:58.360 --> 47:01.120 JIT compiles machine code, and then 47:01.120 --> 47:03.800 injects it into that process. 47:03.800 --> 47:05.400 And so as you're typing new stuff, 47:05.400 --> 47:09.360 it's like squirting in new code and overwriting and replacing 47:09.360 --> 47:11.200 and updating code in place. 47:11.200 --> 47:13.680 And the fact that it can do this is not an accident. 47:13.680 --> 47:15.560 Swift was designed for this. 47:15.560 --> 47:18.120 But it's an important part of how the language was set up 47:18.120 --> 47:21.320 and how it's layered, and this is a nonobvious piece. 47:21.320 --> 47:23.160 And one of the things with Swift that 47:23.160 --> 47:25.880 was, for me, a very strong design point 47:25.880 --> 47:29.640 is to make it so that you can learn it very quickly. 47:29.640 --> 47:31.880 And so from a language design perspective, 47:31.880 --> 47:33.340 the thing that I always come back to 47:33.340 --> 47:36.440 is this UI principle of progressive disclosure 47:36.440 --> 47:37.960 of complexity. 47:37.960 --> 47:41.680 And so in Swift, you can start by saying print, quote, 47:41.680 --> 47:44.040 hello world, quote. 47:44.040 --> 47:47.160 And there's no slash n, just like Python, one line of code, 47:47.160 --> 47:51.520 no main, no header files, no public static class void, 47:51.520 --> 47:55.640 blah, blah, blah, string like Java has, one line of code. 47:55.640 --> 47:58.400 And you can teach that, and it works great. 47:58.400 --> 48:00.400 Then you can say, well, let's introduce variables. 48:00.400 --> 48:02.400 And so you can declare a variable with var. 48:02.400 --> 48:03.780 So var x equals 4. 48:03.780 --> 48:04.700 What is a variable? 48:04.700 --> 48:06.280 You can use x, x plus 1. 48:06.280 --> 48:07.600 This is what it means. 48:07.600 --> 48:09.520 Then you can say, well, how about control flow? 48:09.520 --> 48:10.860 Well, this is what an if statement is. 48:10.860 --> 48:12.280 This is what a for statement is. 48:12.280 --> 48:15.280 This is what a while statement is. 48:15.280 --> 48:17.280 Then you can say, let's introduce functions. 48:17.280 --> 48:20.020 And many languages like Python have 48:20.020 --> 48:22.820 had this kind of notion of let's introduce small things, 48:22.820 --> 48:24.400 and then you can add complexity. 48:24.400 --> 48:25.760 Then you can introduce classes. 48:25.760 --> 48:28.040 And then you can add generics, in the case of Swift. 48:28.040 --> 48:29.520 And then you can build in modules 48:29.520 --> 48:32.200 and build out in terms of the things that you're expressing. 48:32.200 --> 48:35.800 But this is not very typical for compiled languages. 48:35.800 --> 48:38.000 And so this was a very strong design point, 48:38.000 --> 48:40.960 and one of the reasons that Swift, in general, 48:40.960 --> 48:43.480 is designed with this factoring of complexity in mind 48:43.480 --> 48:46.440 so that the language can express powerful things. 48:46.440 --> 48:49.280 You can write firmware in Swift if you want to. 48:49.280 --> 48:51.900 But it has a very high level feel, 48:51.900 --> 48:55.200 which is really this perfect blend, because often you 48:55.200 --> 48:57.520 have very advanced library writers that 48:57.520 --> 49:00.520 want to be able to use the nitty gritty details. 49:00.520 --> 49:02.960 But then other people just want to use the libraries 49:02.960 --> 49:04.880 and work at a higher abstraction level. 49:04.880 --> 49:07.240 It's kind of cool that I saw that you can just 49:07.240 --> 49:09.240 interoperability. 49:09.240 --> 49:11.320 I don't think I pronounced that word enough. 49:11.320 --> 49:14.960 But you can just drag in Python. 49:14.960 --> 49:16.000 It's just strange. 49:16.000 --> 49:19.640 You can import, like I saw this in the demo. 49:19.640 --> 49:21.280 How do you make that happen? 49:21.280 --> 49:23.120 What's up with that? 49:23.120 --> 49:25.560 Is that as easy as it looks, or is it? 49:25.560 --> 49:27.000 Yes, as easy as it looks. 49:27.000 --> 49:29.600 That's not a stage magic hack or anything like that. 49:29.600 --> 49:31.400 I don't mean from the user perspective. 49:31.400 --> 49:34.120 I mean from the implementation perspective to make it happen. 49:34.120 --> 49:37.000 So it's easy once all the pieces are in place. 49:37.000 --> 49:39.280 The way it works, so if you think about a dynamically typed 49:39.280 --> 49:41.480 language like Python, you can think about it 49:41.480 --> 49:42.360 in two different ways. 49:42.360 --> 49:45.800 You can say it has no types, which 49:45.800 --> 49:47.480 is what most people would say. 49:47.480 --> 49:50.400 Or you can say it has one type. 49:50.400 --> 49:53.320 And you can say it has one type, and it's the Python object. 49:53.320 --> 49:55.000 And the Python object gets passed around. 49:55.000 --> 49:58.200 And because there's only one type, it's implicit. 49:58.200 --> 50:00.880 And so what happens with Swift and Python talking 50:00.880 --> 50:02.760 to each other, Swift has lots of types. 50:02.760 --> 50:05.840 It has arrays, and it has strings, and all classes, 50:05.840 --> 50:07.000 and that kind of stuff. 50:07.000 --> 50:11.120 But it now has a Python object type. 50:11.120 --> 50:12.720 So there is one Python object type. 50:12.720 --> 50:16.440 And so when you say import NumPy, what you get 50:16.440 --> 50:19.840 is a Python object, which is the NumPy module. 50:19.840 --> 50:21.960 And then you say np.array. 50:21.960 --> 50:24.960 It says, OK, hey, Python object, I have no idea what you are. 50:24.960 --> 50:27.280 Give me your array member. 50:27.280 --> 50:27.960 OK, cool. 50:27.960 --> 50:31.160 And it just uses dynamic stuff, talks to the Python interpreter, 50:31.160 --> 50:33.680 and says, hey, Python, what's the.array member 50:33.680 --> 50:35.720 in that Python object? 50:35.720 --> 50:37.400 It gives you back another Python object. 50:37.400 --> 50:40.040 And now you say parentheses for the call and the arguments 50:40.040 --> 50:40.920 you're going to pass. 50:40.920 --> 50:43.520 And so then it says, hey, a Python object 50:43.520 --> 50:47.840 that is the result of np.array, call with these arguments. 50:47.840 --> 50:50.320 Again, calling into the Python interpreter to do that work. 50:50.320 --> 50:53.680 And so right now, this is all really simple. 50:53.680 --> 50:55.960 And if you dive into the code, what you'll see 50:55.960 --> 50:58.440 is that the Python module in Swift 50:58.440 --> 51:01.360 is something like 1,200 lines of code or something. 51:01.360 --> 51:02.400 It's written in pure Swift. 51:02.400 --> 51:03.560 It's super simple. 51:03.560 --> 51:06.560 And it's built on top of the C interoperability 51:06.560 --> 51:09.520 because it just talks to the Python interpreter. 51:09.520 --> 51:11.080 But making that possible required 51:11.080 --> 51:13.480 us to add two major language features to Swift 51:13.480 --> 51:15.400 to be able to express these dynamic calls 51:15.400 --> 51:17.240 and the dynamic member lookups. 51:17.240 --> 51:19.480 And so what we've done over the last year 51:19.480 --> 51:23.960 is we've proposed, implement, standardized, and contributed 51:23.960 --> 51:26.160 new language features to the Swift language 51:26.160 --> 51:29.560 in order to make it so it is really trivial. 51:29.560 --> 51:31.320 And this is one of the things about Swift 51:31.320 --> 51:35.000 that is critical to the Swift for TensorFlow work, which 51:35.000 --> 51:37.200 is that we can actually add new language features. 51:37.200 --> 51:39.160 And the bar for adding those is high, 51:39.160 --> 51:42.280 but it's what makes it possible. 51:42.280 --> 51:45.240 So you're now at Google doing incredible work 51:45.240 --> 51:47.680 on several things, including TensorFlow. 51:47.680 --> 51:53.080 So TensorFlow 2.0 or whatever leading up to 2.0 has, 51:53.080 --> 51:56.840 by default, in 2.0, has eager execution. 51:56.840 --> 52:00.520 And yet, in order to make code optimized for GPU or TPU 52:00.520 --> 52:04.120 or some of these systems, computation 52:04.120 --> 52:06.000 needs to be converted to a graph. 52:06.000 --> 52:07.440 So what's that process like? 52:07.440 --> 52:08.960 What are the challenges there? 52:08.960 --> 52:11.720 Yeah, so I am tangentially involved in this. 52:11.720 --> 52:15.280 But the way that it works with Autograph 52:15.280 --> 52:21.600 is that you mark your function with a decorator. 52:21.600 --> 52:24.280 And when Python calls it, that decorator is invoked. 52:24.280 --> 52:28.240 And then it says, before I call this function, 52:28.240 --> 52:29.480 you can transform it. 52:29.480 --> 52:32.400 And so the way Autograph works is, as far as I understand, 52:32.400 --> 52:34.440 is it actually uses the Python parser 52:34.440 --> 52:37.160 to go parse that, turn it into a syntax tree, 52:37.160 --> 52:39.400 and now apply compiler techniques to, again, 52:39.400 --> 52:42.320 transform this down into TensorFlow graphs. 52:42.320 --> 52:44.920 And so you can think of it as saying, hey, 52:44.920 --> 52:45.880 I have an if statement. 52:45.880 --> 52:48.360 I'm going to create an if node in the graph, 52:48.360 --> 52:51.080 like you say tf.cond. 52:51.080 --> 52:53.040 You have a multiply. 52:53.040 --> 52:55.320 Well, I'll turn that into a multiply node in the graph. 52:55.320 --> 52:57.760 And it becomes this tree transformation. 52:57.760 --> 53:00.480 So where does the Swift for TensorFlow 53:00.480 --> 53:04.960 come in, which is parallels? 53:04.960 --> 53:06.960 For one, Swift is an interface. 53:06.960 --> 53:09.200 Like, Python is an interface to TensorFlow. 53:09.200 --> 53:11.760 But it seems like there's a lot more going on in just 53:11.760 --> 53:13.120 a different language interface. 53:13.120 --> 53:15.960 There's optimization methodology. 53:15.960 --> 53:17.920 So the TensorFlow world has a couple 53:17.920 --> 53:21.240 of different what I'd call front end technologies. 53:21.240 --> 53:25.240 And so Swift and Python and Go and Rust and Julia 53:25.240 --> 53:29.320 and all these things share the TensorFlow graphs 53:29.320 --> 53:32.760 and all the runtime and everything that's later. 53:32.760 --> 53:36.640 And so Swift for TensorFlow is merely another front end 53:36.640 --> 53:40.640 for TensorFlow, just like any of these other systems are. 53:40.640 --> 53:43.080 There's a major difference between, I would say, 53:43.080 --> 53:44.600 three camps of technologies here. 53:44.600 --> 53:46.880 There's Python, which is a special case, 53:46.880 --> 53:49.160 because the vast majority of the community effort 53:49.160 --> 53:51.120 is going to the Python interface. 53:51.120 --> 53:52.920 And Python has its own approaches 53:52.920 --> 53:54.480 for automatic differentiation. 53:54.480 --> 53:58.160 It has its own APIs and all this kind of stuff. 53:58.160 --> 54:00.320 There's Swift, which I'll talk about in a second. 54:00.320 --> 54:02.040 And then there's kind of everything else. 54:02.040 --> 54:05.400 And so the everything else are effectively language bindings. 54:05.400 --> 54:07.960 So they call into the TensorFlow runtime, 54:07.960 --> 54:10.920 but they usually don't have automatic differentiation 54:10.920 --> 54:14.560 or they usually don't provide anything other than APIs 54:14.560 --> 54:16.440 that call the C APIs in TensorFlow. 54:16.440 --> 54:18.360 And so they're kind of wrappers for that. 54:18.360 --> 54:19.840 Swift is really kind of special. 54:19.840 --> 54:22.760 And it's a very different approach. 54:22.760 --> 54:25.360 Swift for TensorFlow, that is, is a very different approach. 54:25.360 --> 54:26.880 Because there we're saying, let's 54:26.880 --> 54:28.400 look at all the problems that need 54:28.400 --> 54:34.080 to be solved in the full stack of the TensorFlow compilation 54:34.080 --> 54:35.680 process, if you think about it that way. 54:35.680 --> 54:38.200 Because TensorFlow is fundamentally a compiler. 54:38.200 --> 54:42.760 It takes models, and then it makes them go fast on hardware. 54:42.760 --> 54:43.880 That's what a compiler does. 54:43.880 --> 54:47.560 And it has a front end, it has an optimizer, 54:47.560 --> 54:49.320 and it has many back ends. 54:49.320 --> 54:51.680 And so if you think about it the right way, 54:51.680 --> 54:54.800 or if you look at it in a particular way, 54:54.800 --> 54:55.560 it is a compiler. 54:59.280 --> 55:02.120 And so Swift is merely another front end. 55:02.120 --> 55:05.560 But it's saying, and the design principle is saying, 55:05.560 --> 55:08.240 let's look at all the problems that we face as machine 55:08.240 --> 55:11.320 learning practitioners and what is the best possible way we 55:11.320 --> 55:13.840 can do that, given the fact that we can change literally 55:13.840 --> 55:15.920 anything in this entire stack. 55:15.920 --> 55:18.440 And Python, for example, where the vast majority 55:18.440 --> 55:22.600 of the engineering and effort has gone into, 55:22.600 --> 55:25.000 is constrained by being the best possible thing you 55:25.000 --> 55:27.320 can do with a Python library. 55:27.320 --> 55:29.320 There are no Python language features 55:29.320 --> 55:31.040 that are added because of machine learning 55:31.040 --> 55:32.600 that I'm aware of. 55:32.600 --> 55:34.640 They added a matrix multiplication operator 55:34.640 --> 55:38.320 with that, but that's as close as you get. 55:38.320 --> 55:41.460 And so with Swift, it's hard, but you 55:41.460 --> 55:43.800 can add language features to the language. 55:43.800 --> 55:46.040 And there's a community process for that. 55:46.040 --> 55:48.200 And so we look at these things and say, well, 55:48.200 --> 55:49.720 what is the right division of labor 55:49.720 --> 55:52.000 between the human programmer and the compiler? 55:52.000 --> 55:55.280 And Swift has a number of things that shift that balance. 55:55.280 --> 56:00.560 So because it has a type system, for example, 56:00.560 --> 56:02.680 that makes certain things possible for analysis 56:02.680 --> 56:05.560 of the code, and the compiler can automatically 56:05.560 --> 56:08.880 build graphs for you without you thinking about them. 56:08.880 --> 56:10.520 That's a big deal for a programmer. 56:10.520 --> 56:11.680 You just get free performance. 56:11.680 --> 56:14.400 You get clustering and fusion and optimization, 56:14.400 --> 56:17.040 things like that, without you as a programmer 56:17.040 --> 56:20.080 having to manually do it because the compiler can do it for you. 56:20.080 --> 56:22.240 Automatic differentiation is another big deal. 56:22.240 --> 56:25.960 And I think one of the key contributions of the Swift 56:25.960 --> 56:29.640 TensorFlow project is that there's 56:29.640 --> 56:32.120 this entire body of work on automatic differentiation 56:32.120 --> 56:34.120 that dates back to the Fortran days. 56:34.120 --> 56:36.400 People doing a tremendous amount of numerical computing 56:36.400 --> 56:39.360 in Fortran used to write these what they call source 56:39.360 --> 56:43.280 to source translators, where you take a bunch of code, 56:43.280 --> 56:46.640 shove it into a mini compiler, and it would push out 56:46.640 --> 56:48.080 more Fortran code. 56:48.080 --> 56:50.240 But it would generate the backwards passes 56:50.240 --> 56:53.000 for your functions for you, the derivatives. 56:53.000 --> 56:57.840 And so in that work in the 70s, a tremendous number 56:57.840 --> 57:01.160 of optimizations, a tremendous number of techniques 57:01.160 --> 57:02.920 for fixing numerical instability, 57:02.920 --> 57:05.080 and other kinds of problems were developed. 57:05.080 --> 57:07.600 But they're very difficult to port into a world 57:07.600 --> 57:11.280 where, in eager execution, you get an op by op at a time. 57:11.280 --> 57:13.280 You need to be able to look at an entire function 57:13.280 --> 57:15.720 and be able to reason about what's going on. 57:15.720 --> 57:18.720 And so when you have a language integrated automatic 57:18.720 --> 57:20.520 differentiation, which is one of the things 57:20.520 --> 57:22.760 that the Swift project is focusing on, 57:22.760 --> 57:24.680 you can open all these techniques 57:24.680 --> 57:28.640 and reuse them in familiar ways. 57:28.640 --> 57:30.120 But the language integration piece 57:30.120 --> 57:33.240 has a bunch of design room in it, and it's also complicated. 57:33.240 --> 57:35.680 The other piece of the puzzle here that's kind of interesting 57:35.680 --> 57:37.560 is TPUs at Google. 57:37.560 --> 57:40.200 So we're in a new world with deep learning. 57:40.200 --> 57:42.960 It constantly is changing, and I imagine, 57:42.960 --> 57:46.360 without disclosing anything, I imagine 57:46.360 --> 57:48.400 you're still innovating on the TPU front, too. 57:48.400 --> 57:49.040 Indeed. 57:49.040 --> 57:53.560 So how much interplay is there between software and hardware 57:53.560 --> 57:55.240 in trying to figure out how to together move 57:55.240 --> 57:56.680 towards an optimized solution? 57:56.680 --> 57:57.760 There's an incredible amount. 57:57.760 --> 57:59.480 So we're on our third generation of TPUs, 57:59.480 --> 58:04.640 which are now 100 petaflops in a very large liquid cooled box, 58:04.640 --> 58:07.720 virtual box with no cover. 58:07.720 --> 58:11.240 And as you might imagine, we're not out of ideas yet. 58:11.240 --> 58:14.360 The great thing about TPUs is that they're 58:14.360 --> 58:17.520 a perfect example of hardware software co design. 58:17.520 --> 58:19.800 And so it's about saying, what hardware 58:19.800 --> 58:23.240 do we build to solve certain classes of machine learning 58:23.240 --> 58:23.840 problems? 58:23.840 --> 58:26.480 Well, the algorithms are changing. 58:26.480 --> 58:30.360 The hardware takes some cases years to produce. 58:30.360 --> 58:32.760 And so you have to make bets and decide 58:32.760 --> 58:36.520 what is going to happen and what is the best way to spend 58:36.520 --> 58:39.920 the transistors to get the maximum performance per watt 58:39.920 --> 58:44.000 or area per cost or whatever it is that you're optimizing for. 58:44.000 --> 58:46.560 And so one of the amazing things about TPUs 58:46.560 --> 58:49.960 is this numeric format called bfloat16. 58:49.960 --> 58:54.120 bfloat16 is a compressed 16 bit floating point format, 58:54.120 --> 58:55.960 but it puts the bits in different places. 58:55.960 --> 58:58.960 And in numeric terms, it has a smaller mantissa 58:58.960 --> 59:00.400 and a larger exponent. 59:00.400 --> 59:02.960 That means that it's less precise, 59:02.960 --> 59:05.680 but it can represent larger ranges of values, 59:05.680 --> 59:07.280 which in the machine learning context 59:07.280 --> 59:09.960 is really important and useful because sometimes you 59:09.960 --> 59:13.920 have very small gradients you want to accumulate 59:13.920 --> 59:17.480 and very, very small numbers that 59:17.480 --> 59:20.520 are important to move things as you're learning. 59:20.520 --> 59:23.160 But sometimes you have very large magnitude numbers as well. 59:23.160 --> 59:26.880 And bfloat16 is not as precise. 59:26.880 --> 59:28.040 The mantissa is small. 59:28.040 --> 59:30.360 But it turns out the machine learning algorithms actually 59:30.360 --> 59:31.520 want to generalize. 59:31.520 --> 59:34.320 And so there's theories that this actually 59:34.320 --> 59:36.440 increases the ability for the network 59:36.440 --> 59:37.960 to generalize across data sets. 59:37.960 --> 59:41.160 And regardless of whether it's good or bad, 59:41.160 --> 59:43.680 it's much cheaper at the hardware level to implement 59:43.680 --> 59:48.080 because the area and time of a multiplier 59:48.080 --> 59:50.840 is n squared in the number of bits in the mantissa, 59:50.840 --> 59:53.320 but it's linear with size of the exponent. 59:53.320 --> 59:55.400 And you're connected to both efforts 59:55.400 --> 59:57.160 here both on the hardware and the software side? 59:57.160 --> 59:58.880 Yeah, and so that was a breakthrough 59:58.880 --> 1:00:01.440 coming from the research side and people 1:00:01.440 --> 1:00:06.000 working on optimizing network transport of weights 1:00:06.000 --> 1:00:08.240 across the network originally and trying 1:00:08.240 --> 1:00:10.160 to find ways to compress that. 1:00:10.160 --> 1:00:12.120 But then it got burned into silicon. 1:00:12.120 --> 1:00:14.560 And it's a key part of what makes TPU performance 1:00:14.560 --> 1:00:17.880 so amazing and great. 1:00:17.880 --> 1:00:20.680 Now, TPUs have many different aspects that are important. 1:00:20.680 --> 1:00:25.080 But the co design between the low level compiler bits 1:00:25.080 --> 1:00:27.360 and the software bits and the algorithms 1:00:27.360 --> 1:00:28.680 is all super important. 1:00:28.680 --> 1:00:32.880 And it's this amazing trifecta that only Google can do. 1:00:32.880 --> 1:00:34.240 Yeah, that's super exciting. 1:00:34.240 --> 1:00:39.800 So can you tell me about MLIR project, previously 1:00:39.800 --> 1:00:41.400 the secretive one? 1:00:41.400 --> 1:00:43.040 Yeah, so MLIR is a project that we 1:00:43.040 --> 1:00:47.000 announced at a compiler conference three weeks ago 1:00:47.000 --> 1:00:49.280 or something at the Compilers for Machine Learning 1:00:49.280 --> 1:00:50.920 conference. 1:00:50.920 --> 1:00:53.760 Basically, again, if you look at TensorFlow as a compiler stack, 1:00:53.760 --> 1:00:56.120 it has a number of compiler algorithms within it. 1:00:56.120 --> 1:00:57.660 It also has a number of compilers 1:00:57.660 --> 1:00:59.000 that get embedded into it. 1:00:59.000 --> 1:01:00.480 And they're made by different vendors. 1:01:00.480 --> 1:01:02.840 For example, Google has XLA, which 1:01:02.840 --> 1:01:04.680 is a great compiler system. 1:01:04.680 --> 1:01:06.480 NVIDIA has TensorRT. 1:01:06.480 --> 1:01:08.640 Intel has NGRAPH. 1:01:08.640 --> 1:01:10.840 There's a number of these different compiler systems. 1:01:10.840 --> 1:01:13.840 And they're very hardware specific. 1:01:13.840 --> 1:01:16.480 And they're trying to solve different parts of the problems. 1:01:16.480 --> 1:01:19.400 But they're all kind of similar in a sense of they 1:01:19.400 --> 1:01:20.880 want to integrate with TensorFlow. 1:01:20.880 --> 1:01:22.960 Now, TensorFlow has an optimizer. 1:01:22.960 --> 1:01:25.540 And it has these different code generation technologies 1:01:25.540 --> 1:01:26.440 built in. 1:01:26.440 --> 1:01:28.720 The idea of MLIR is to build a common infrastructure 1:01:28.720 --> 1:01:31.160 to support all these different subsystems. 1:01:31.160 --> 1:01:33.500 And initially, it's to be able to make it 1:01:33.500 --> 1:01:34.880 so that they all plug in together 1:01:34.880 --> 1:01:37.880 and they can share a lot more code and can be reusable. 1:01:37.880 --> 1:01:39.680 But over time, we hope that the industry 1:01:39.680 --> 1:01:42.480 will start collaborating and sharing code. 1:01:42.480 --> 1:01:45.320 And instead of reinventing the same things over and over again, 1:01:45.320 --> 1:01:49.280 that we can actually foster some of that working together 1:01:49.280 --> 1:01:51.560 to solve common problem energy that 1:01:51.560 --> 1:01:54.480 has been useful in the compiler field before. 1:01:54.480 --> 1:01:57.360 Beyond that, MLIR is some people have joked 1:01:57.360 --> 1:01:59.320 that it's kind of LLVM too. 1:01:59.320 --> 1:02:01.840 It learns a lot about what LLVM has been good 1:02:01.840 --> 1:02:04.360 and what LLVM has done wrong. 1:02:04.360 --> 1:02:06.880 And it's a chance to fix that. 1:02:06.880 --> 1:02:09.840 And also, there are challenges in the LLVM ecosystem as well, 1:02:09.840 --> 1:02:12.760 where LLVM is very good at the thing it was designed to do. 1:02:12.760 --> 1:02:15.560 But 20 years later, the world has changed. 1:02:15.560 --> 1:02:17.980 And people are trying to solve higher level problems. 1:02:17.980 --> 1:02:20.360 And we need some new technology. 1:02:20.360 --> 1:02:24.720 And what's the future of open source in this context? 1:02:24.720 --> 1:02:25.760 Very soon. 1:02:25.760 --> 1:02:27.480 So it is not yet open source. 1:02:27.480 --> 1:02:29.320 But it will be hopefully in the next couple months. 1:02:29.320 --> 1:02:31.040 So you still believe in the value of open source 1:02:31.040 --> 1:02:31.640 in these kinds of contexts? 1:02:31.640 --> 1:02:31.880 Oh, yeah. 1:02:31.880 --> 1:02:32.440 Absolutely. 1:02:32.440 --> 1:02:36.160 And I think that the TensorFlow community at large 1:02:36.160 --> 1:02:37.720 fully believes in open source. 1:02:37.720 --> 1:02:40.120 So I mean, there is a difference between Apple, 1:02:40.120 --> 1:02:42.480 where you were previously, and Google now, 1:02:42.480 --> 1:02:43.520 in spirit and culture. 1:02:43.520 --> 1:02:45.480 And I would say the open source in TensorFlow 1:02:45.480 --> 1:02:48.400 was a seminal moment in the history of software, 1:02:48.400 --> 1:02:51.680 because here's this large company releasing 1:02:51.680 --> 1:02:56.200 a very large code base that's open sourcing. 1:02:56.200 --> 1:02:58.520 What are your thoughts on that? 1:02:58.520 --> 1:03:00.840 Happy or not, were you to see that kind 1:03:00.840 --> 1:03:02.920 of degree of open sourcing? 1:03:02.920 --> 1:03:05.360 So between the two, I prefer the Google approach, 1:03:05.360 --> 1:03:07.800 if that's what you're saying. 1:03:07.800 --> 1:03:12.400 The Apple approach makes sense, given the historical context 1:03:12.400 --> 1:03:13.400 that Apple came from. 1:03:13.400 --> 1:03:15.760 But that's been 35 years ago. 1:03:15.760 --> 1:03:18.200 And I think that Apple is definitely adapting. 1:03:18.200 --> 1:03:20.280 And the way I look at it is that there's 1:03:20.280 --> 1:03:23.160 different kinds of concerns in the space. 1:03:23.160 --> 1:03:24.880 It is very rational for a business 1:03:24.880 --> 1:03:28.720 to care about making money. 1:03:28.720 --> 1:03:31.640 That fundamentally is what a business is about. 1:03:31.640 --> 1:03:34.880 But I think it's also incredibly realistic to say, 1:03:34.880 --> 1:03:36.360 it's not your string library that's 1:03:36.360 --> 1:03:38.080 the thing that's going to make you money. 1:03:38.080 --> 1:03:41.480 It's going to be the amazing UI product differentiating 1:03:41.480 --> 1:03:43.840 features and other things like that that you built on top 1:03:43.840 --> 1:03:45.280 of your string library. 1:03:45.280 --> 1:03:48.280 And so keeping your string library 1:03:48.280 --> 1:03:50.360 proprietary and secret and things 1:03:50.360 --> 1:03:54.760 like that is maybe not the important thing anymore. 1:03:54.760 --> 1:03:57.720 Where before, platforms were different. 1:03:57.720 --> 1:04:01.520 And even 15 years ago, things were a little bit different. 1:04:01.520 --> 1:04:02.920 But the world is changing. 1:04:02.920 --> 1:04:04.840 So Google strikes a very good balance, 1:04:04.840 --> 1:04:05.340 I think. 1:04:05.340 --> 1:04:09.040 And I think that TensorFlow being open source really 1:04:09.040 --> 1:04:12.000 changed the entire machine learning field 1:04:12.000 --> 1:04:14.080 and caused a revolution in its own right. 1:04:14.080 --> 1:04:17.560 And so I think it's amazingly forward looking 1:04:17.560 --> 1:04:20.880 because I could have imagined, and I wasn't at Google 1:04:20.880 --> 1:04:23.160 at the time, but I could imagine a different context 1:04:23.160 --> 1:04:25.520 and different world where a company says, 1:04:25.520 --> 1:04:27.640 machine learning is critical to what we're doing. 1:04:27.640 --> 1:04:29.640 We're not going to give it to other people. 1:04:29.640 --> 1:04:35.560 And so that decision is a profoundly brilliant insight 1:04:35.560 --> 1:04:37.480 that I think has really led to the world being 1:04:37.480 --> 1:04:40.120 better and better for Google as well. 1:04:40.120 --> 1:04:42.200 And has all kinds of ripple effects. 1:04:42.200 --> 1:04:45.160 I think it is really, I mean, you 1:04:45.160 --> 1:04:48.800 can't understate Google deciding how profound that 1:04:48.800 --> 1:04:49.840 is for software. 1:04:49.840 --> 1:04:50.880 It's awesome. 1:04:50.880 --> 1:04:54.900 Well, and again, I can understand the concern 1:04:54.900 --> 1:04:58.440 about if we release our machine learning software, 1:04:58.440 --> 1:05:00.000 our competitors could go faster. 1:05:00.000 --> 1:05:02.500 But on the other hand, I think that open sourcing TensorFlow 1:05:02.500 --> 1:05:03.960 has been fantastic for Google. 1:05:03.960 --> 1:05:09.120 And I'm sure that decision was very nonobvious at the time, 1:05:09.120 --> 1:05:11.480 but I think it's worked out very well. 1:05:11.480 --> 1:05:13.240 So let's try this real quick. 1:05:13.240 --> 1:05:15.640 You were at Tesla for five months 1:05:15.640 --> 1:05:17.640 as the VP of autopilot software. 1:05:17.640 --> 1:05:20.520 You led the team during the transition from H hardware 1:05:20.520 --> 1:05:22.360 one to hardware two. 1:05:22.360 --> 1:05:23.520 I have a couple of questions. 1:05:23.520 --> 1:05:26.320 So one, first of all, to me, that's 1:05:26.320 --> 1:05:33.000 one of the bravest engineering decisions undertaking really 1:05:33.000 --> 1:05:36.040 ever in the automotive industry to me, software wise, 1:05:36.040 --> 1:05:37.440 starting from scratch. 1:05:37.440 --> 1:05:39.200 It's a really brave engineering decision. 1:05:39.200 --> 1:05:42.600 So my one question there is, what was that like? 1:05:42.600 --> 1:05:43.920 What was the challenge of that? 1:05:43.920 --> 1:05:45.720 Do you mean the career decision of jumping 1:05:45.720 --> 1:05:48.800 from a comfortable good job into the unknown, or? 1:05:48.800 --> 1:05:51.480 That combined, so at the individual level, 1:05:51.480 --> 1:05:54.560 you making that decision. 1:05:54.560 --> 1:05:57.960 And then when you show up, it's a really hard engineering 1:05:57.960 --> 1:05:58.760 problem. 1:05:58.760 --> 1:06:03.560 So you could just stay, maybe slow down, 1:06:03.560 --> 1:06:06.680 say hardware one, or those kinds of decisions. 1:06:06.680 --> 1:06:10.160 Just taking it full on, let's do this from scratch. 1:06:10.160 --> 1:06:11.080 What was that like? 1:06:11.080 --> 1:06:12.640 Well, so I mean, I don't think Tesla 1:06:12.640 --> 1:06:16.080 has a culture of taking things slow and seeing how it goes. 1:06:16.080 --> 1:06:18.080 And one of the things that attracted me about Tesla 1:06:18.080 --> 1:06:20.020 is it's very much a gung ho, let's change the world, 1:06:20.020 --> 1:06:21.520 let's figure it out kind of a place. 1:06:21.520 --> 1:06:25.640 And so I have a huge amount of respect for that. 1:06:25.640 --> 1:06:28.680 Tesla has done very smart things with hardware one 1:06:28.680 --> 1:06:29.400 in particular. 1:06:29.400 --> 1:06:32.200 And the hardware one design was originally 1:06:32.200 --> 1:06:36.560 designed to be very simple automation features 1:06:36.560 --> 1:06:39.360 in the car for like traffic aware cruise control and things 1:06:39.360 --> 1:06:39.840 like that. 1:06:39.840 --> 1:06:42.920 And the fact that they were able to effectively feature creep 1:06:42.920 --> 1:06:47.720 it into lane holding and a very useful driver assistance 1:06:47.720 --> 1:06:50.120 feature is pretty astounding, particularly given 1:06:50.120 --> 1:06:52.560 the details of the hardware. 1:06:52.560 --> 1:06:54.640 Hardware two built on that in a lot of ways. 1:06:54.640 --> 1:06:56.180 And the challenge there was that they 1:06:56.180 --> 1:07:00.040 were transitioning from a third party provided vision stack 1:07:00.040 --> 1:07:01.720 to an in house built vision stack. 1:07:01.720 --> 1:07:05.680 And so for the first step, which I mostly helped with, 1:07:05.680 --> 1:07:08.480 was getting onto that new vision stack. 1:07:08.480 --> 1:07:10.800 And that was very challenging. 1:07:10.800 --> 1:07:14.000 And it was time critical for various reasons, 1:07:14.000 --> 1:07:14.960 and it was a big leap. 1:07:14.960 --> 1:07:16.640 But it was fortunate that it built 1:07:16.640 --> 1:07:18.800 on a lot of the knowledge and expertise and the team 1:07:18.800 --> 1:07:22.920 that had built hardware one's driver assistance features. 1:07:22.920 --> 1:07:25.360 So you spoke in a collected and kind way 1:07:25.360 --> 1:07:28.960 about your time at Tesla, but it was ultimately not a good fit. 1:07:28.960 --> 1:07:31.840 Elon Musk, we've talked on this podcast, 1:07:31.840 --> 1:07:33.880 several guests to the course, Elon Musk 1:07:33.880 --> 1:07:36.880 continues to do some of the most bold and innovative engineering 1:07:36.880 --> 1:07:39.560 work in the world, at times at the cost 1:07:39.560 --> 1:07:41.280 some of the members of the Tesla team. 1:07:41.280 --> 1:07:45.080 What did you learn about working in this chaotic world 1:07:45.080 --> 1:07:46.720 with Elon? 1:07:46.720 --> 1:07:50.560 Yeah, so I guess I would say that when I was at Tesla, 1:07:50.560 --> 1:07:54.440 I experienced and saw the highest degree of turnover 1:07:54.440 --> 1:07:58.240 I'd ever seen in a company, which was a bit of a shock. 1:07:58.240 --> 1:08:00.520 But one of the things I learned and I came to respect 1:08:00.520 --> 1:08:03.760 is that Elon's able to attract amazing talent because he 1:08:03.760 --> 1:08:05.660 has a very clear vision of the future, 1:08:05.660 --> 1:08:07.200 and he can get people to buy into it 1:08:07.200 --> 1:08:09.840 because they want that future to happen. 1:08:09.840 --> 1:08:11.840 And the power of vision is something 1:08:11.840 --> 1:08:14.240 that I have a tremendous amount of respect for. 1:08:14.240 --> 1:08:17.040 And I think that Elon is fairly singular 1:08:17.040 --> 1:08:20.120 in the world in terms of the things 1:08:20.120 --> 1:08:22.360 he's able to get people to believe in. 1:08:22.360 --> 1:08:27.360 And there are many people that stand in the street corner 1:08:27.360 --> 1:08:30.200 and say, ah, we're going to go to Mars, right? 1:08:30.200 --> 1:08:31.600 But then there are a few people that 1:08:31.600 --> 1:08:35.200 can get others to buy into it and believe and build the path 1:08:35.200 --> 1:08:36.160 and make it happen. 1:08:36.160 --> 1:08:39.120 And so I respect that. 1:08:39.120 --> 1:08:41.880 I don't respect all of his methods, 1:08:41.880 --> 1:08:45.000 but I have a huge amount of respect for that. 1:08:45.000 --> 1:08:46.920 You've mentioned in a few places, 1:08:46.920 --> 1:08:50.440 including in this context, working hard. 1:08:50.440 --> 1:08:52.000 What does it mean to work hard? 1:08:52.000 --> 1:08:53.520 And when you look back at your life, 1:08:53.520 --> 1:08:57.080 what were some of the most brutal periods 1:08:57.080 --> 1:09:00.760 of having to really put everything 1:09:00.760 --> 1:09:03.360 you have into something? 1:09:03.360 --> 1:09:05.040 Yeah, good question. 1:09:05.040 --> 1:09:07.440 So working hard can be defined a lot of different ways, 1:09:07.440 --> 1:09:12.480 so a lot of hours, and so that is true. 1:09:12.480 --> 1:09:14.520 The thing to me that's the hardest 1:09:14.520 --> 1:09:18.760 is both being short term focused on delivering and executing 1:09:18.760 --> 1:09:21.120 and making a thing happen while also thinking 1:09:21.120 --> 1:09:24.400 about the longer term and trying to balance that. 1:09:24.400 --> 1:09:28.520 Because if you are myopically focused on solving a task 1:09:28.520 --> 1:09:31.240 and getting that done and only think 1:09:31.240 --> 1:09:32.600 about that incremental next step, 1:09:32.600 --> 1:09:36.440 you will miss the next big hill you should jump over to. 1:09:36.440 --> 1:09:39.600 And so I've been really fortunate that I've 1:09:39.600 --> 1:09:42.120 been able to kind of oscillate between the two. 1:09:42.120 --> 1:09:45.480 And historically at Apple, for example, that 1:09:45.480 --> 1:09:47.920 was made possible because I was able to work with some really 1:09:47.920 --> 1:09:50.360 amazing people and build up teams and leadership 1:09:50.360 --> 1:09:55.280 structures and allow them to grow in their careers 1:09:55.280 --> 1:09:58.280 and take on responsibility, thereby freeing up 1:09:58.280 --> 1:10:02.960 me to be a little bit crazy and thinking about the next thing. 1:10:02.960 --> 1:10:04.640 And so it's a lot of that. 1:10:04.640 --> 1:10:06.760 But it's also about with experience, 1:10:06.760 --> 1:10:10.080 you make connections that other people don't necessarily make. 1:10:10.080 --> 1:10:12.880 And so I think that's a big part as well. 1:10:12.880 --> 1:10:16.000 But the bedrock is just a lot of hours. 1:10:16.000 --> 1:10:19.600 And that's OK with me. 1:10:19.600 --> 1:10:21.480 There's different theories on work life balance. 1:10:21.480 --> 1:10:25.200 And my theory for myself, which I do not project onto the team, 1:10:25.200 --> 1:10:28.520 but my theory for myself is that I 1:10:28.520 --> 1:10:30.400 want to love what I'm doing and work really hard. 1:10:30.400 --> 1:10:35.000 And my purpose, I feel like, and my goal is to change the world 1:10:35.000 --> 1:10:36.280 and make it a better place. 1:10:36.280 --> 1:10:40.000 And that's what I'm really motivated to do. 1:10:40.000 --> 1:10:44.760 So last question, LLVM logo is a dragon. 1:10:44.760 --> 1:10:47.880 You explain that this is because dragons have connotations 1:10:47.880 --> 1:10:50.320 of power, speed, intelligence. 1:10:50.320 --> 1:10:53.320 It can also be sleek, elegant, and modular, 1:10:53.320 --> 1:10:56.280 though you remove the modular part. 1:10:56.280 --> 1:10:58.920 What is your favorite dragon related character 1:10:58.920 --> 1:11:01.440 from fiction, video, or movies? 1:11:01.440 --> 1:11:03.840 So those are all very kind ways of explaining it. 1:11:03.840 --> 1:11:06.200 Do you want to know the real reason it's a dragon? 1:11:06.200 --> 1:11:07.000 Yeah. 1:11:07.000 --> 1:11:07.920 Is that better? 1:11:07.920 --> 1:11:11.040 So there is a seminal book on compiler design 1:11:11.040 --> 1:11:12.520 called The Dragon Book. 1:11:12.520 --> 1:11:16.320 And so this is a really old now book on compilers. 1:11:16.320 --> 1:11:22.080 And so the dragon logo for LLVM came about because at Apple, 1:11:22.080 --> 1:11:24.720 we kept talking about LLVM related technologies 1:11:24.720 --> 1:11:26.960 and there's no logo to put on a slide. 1:11:26.960 --> 1:11:28.480 And so we're like, what do we do? 1:11:28.480 --> 1:11:30.480 And somebody's like, well, what kind of logo 1:11:30.480 --> 1:11:32.160 should a compiler technology have? 1:11:32.160 --> 1:11:33.360 And I'm like, I don't know. 1:11:33.360 --> 1:11:37.320 I mean, the dragon is the best thing that we've got. 1:11:37.320 --> 1:11:41.520 And Apple somehow magically came up with the logo. 1:11:41.520 --> 1:11:42.680 And it was a great thing. 1:11:42.680 --> 1:11:44.520 And the whole community rallied around it. 1:11:44.520 --> 1:11:46.760 And then it got better as other graphic designers 1:11:46.760 --> 1:11:47.360 got involved. 1:11:47.360 --> 1:11:49.360 But that's originally where it came from. 1:11:49.360 --> 1:11:50.160 The story. 1:11:50.160 --> 1:11:51.960 Is there dragons from fiction that you 1:11:51.960 --> 1:11:57.240 connect with, that Game of Thrones, Lord of the Rings, 1:11:57.240 --> 1:11:58.080 that kind of thing? 1:11:58.080 --> 1:11:59.200 Lord of the Rings is great. 1:11:59.200 --> 1:12:00.760 I also like role playing games and things 1:12:00.760 --> 1:12:02.240 like computer role playing games. 1:12:02.240 --> 1:12:04.280 And so dragons often show up in there. 1:12:04.280 --> 1:12:07.160 But really, it comes back to the book. 1:12:07.160 --> 1:12:09.960 Oh, no, we need a thing. 1:12:09.960 --> 1:12:13.720 And hilariously, one of the funny things about LLVM 1:12:13.720 --> 1:12:19.520 is that my wife, who's amazing, runs the LLVM Foundation. 1:12:19.520 --> 1:12:21.080 And she goes to Grace Hopper and is 1:12:21.080 --> 1:12:23.360 trying to get more women involved in the. 1:12:23.360 --> 1:12:24.640 She's also a compiler engineer. 1:12:24.640 --> 1:12:26.080 So she's trying to get other women 1:12:26.080 --> 1:12:28.020 to get interested in compilers and things like this. 1:12:28.020 --> 1:12:30.000 And so she hands out the stickers. 1:12:30.000 --> 1:12:34.320 And people like the LLVM sticker because of Game of Thrones. 1:12:34.320 --> 1:12:36.880 And so sometimes culture has this helpful effect 1:12:36.880 --> 1:12:39.960 to get the next generation of compiler engineers 1:12:39.960 --> 1:12:42.400 engaged with the cause. 1:12:42.400 --> 1:12:43.320 OK, awesome. 1:12:43.320 --> 1:12:44.800 Chris, thanks so much for talking with us. 1:12:44.800 --> 1:13:05.920 It's been great talking with you.