I need to accelerate many computations I am now doing with PyLab. I thought of using CUDA. The overall unit of computation (A) consists in doing several (thousands) entirely independent smaller computations (B). Each of them involves, at their initial stage, doing 40-41 independent, even smaller, computations (C). So parallel programming should really help. With PyLab the overall (A) takes 20 minutes and (B) takes some tenth of a second.
As a beginner in this realm, my question is what level I should parallelize the computation at, whether at (C) or at (B).
I should clarify that the (C) stage consists in taking a bunch of data (thousands of floats) which is shared between all the (C) processes, and doing various tasks, among which one of the most time consuming is linear regression, which is, too, parallelizable! The output of each procedure (C) is a single float. Each computation (B) consists basically in doing many times procedure (C) and doing a linear regression on the data that comes out. Its output, again, is a single float.
I'm not familiar with CUDA programming so I am basically asking what would be the wisest strategy to start with.
An important consideration when deciding how (and if) to convert your project to CUDA is what type of memory access patterns your code requires. The GPU runs threads in groups of 32, called warps, and to get the best performance, the threads in a warp should access the memory in some basic patterns, that are described in the CUDA Programming Guide (included with CUDA). In general, the more random the access patterns, the more likely the kernel is to become memory bound. In that case, the compute power in the GPU cannot be fully utilized.
The other main case when the compute power in the GPU cannot be fully utilized is if there is conditional logic and loops that causes the threads in a warp to run through different code paths, as the GPU has to run all the threads in the warp through each code path.
If you find that these points may cause issues for your code, you should also do some research to see if there are known alternative ways to implement your code to run better on the GPU (this is often the case).
If you see your question about at which level to parallelize the computation in the light of the above considerations, it may become clear which choice to make.
Related
Does it make sense to implement own branch prediction optimization in own VM interpreter or it is enough to run VM on hardware that already has branch prediction optimization support?
It could make sense in a limited sense.
For example, in a JIT complier, when generating assembly you may decide to lay out code based on the observed branch probabilities. This only needs a very simple type of predictor that knows the overall probability but doesn't need to recognize any patterns. If you did recognize patterns you could do more sophisticated optimizations, e.g. a loop with an embedded branch that alternates every iteration could be unrolled 2x and the body created efficient for the observed case.
For an interpreter it seems a bit less useful, but one can imagine some sophisticated designs that fuse some adjacent instructions together into a single operation for efficiency and this might benefit from branch prediction. Similarly an interpreter might benefit from recognizing loops.
Apparently you're talking about a VM that interprets bytecode, not hardware virtualization of a CPU.
Implement how? Branch prediction in CPUs is only needed because they're pipelined, and for speculative out-of-order execution.
None of those things make sense for interpreter software if it would create more work to implement. Software pipelining can be worth it for loops over arrays to hide load and ALU latency, especially on older in-order CPUs, but that doesn't increase the total number of instructions to be run. If you don't know for sure what needs to be done next, leave the speculation to hardware OoO exec.
Note that for a pure non-JITing interpreter, control dependencies in the guest code become data dependencies in the interpreter, while a sequence of different instructions in the guest creates a control dependency in the interpreter (to dispatch to handler functions). See How exactly R is affected by Branch Prediction?
You do potentially need to care about branch prediction in the CPU that will run your code. Recently (like Intel since Haswell), CPUs are finally not bad for that, using IT-TAGE predictors: Branch Prediction and the Performance of Interpreters - Don’t Trust Folklore.
You don't implement branch prediction in software, but for older CPUs it was worth tuning interpreters with hardware branch prediction in mind. X86 prefetching optimizations: "computed goto" threaded code has some links, especially an article by Darek Mihocka discussing how badly it sucks for older CPUs (current at the time it was written) to have one "grand central" dispatch branch, like a single switch that every instruction-handler function returns to. That means the entire pattern of which instruction tends to follow which other instruction has to be predicted for that single branch. Without something like IT-TAGE, the prediction state for a single branch is very limited.
Tuning for older CPUs can involve putting dispatch to the next instruction at the end of each handler function, instead of returning to a single dispatch loop. But again, that's not implementing branch prediction, that's tuning for it.
While GPUs speed math calculations there is a fixed overhead for moving a kernel out to the GPU for execution that is high.
I'm using cupy and numba. THe first time I execute a function call that is using cupy's GPU version of numpy it is quite slow. But the second time it is fast.
I've realized I don't understand how the kernel, or GPU code, gets out to the GPU to run. Operationally I want to understand this better so that I can know when the things I do will accidentally create a slow step due to some kernel transfer. So I need some sorts of rules or rules of thumb understand the concept.
For example, if I multiply two cupy arrays that are stashed on the GPU already I might write C= A*B
At some point the cupy overload on * multiplication has to be coded out on the GPU, and it automagically needs will also get wrapped by loops that break it down into blocks and threads. So presumably this code is some kernel that gets transported out to the GPU. I'm guessing that the next time I call C*D that the GPU no longer needs to be taught what * means and so it will be fast.
But at some point I would imagine the GPU needs to clear out old code so * or other operations not being used at that moment might get flushed from memory, and so later on when the call for A*B happens again there's going to be a penalty in time to recompile it out on the GPU.
Or so I imagine. If I'm right how do I know when these kernels are going to stick around or disappear?
If I'm wrong and this isn't how it works or there is some other slow step (I'm assuming the data is already transported to arrays on the GPU) then what is this slow step and how does organize things so one pay it as little as possible?
I'm trying to avoid writing explicit numba thread managment kernels as one does in cuda++ but just use the standard numba #njit, #vectorize, #stencil decorators. Likewise in Cupy I want to just work at the level of the numpy syntax not dive into thread management.
I've read a lot of docs on this but they just refer to overheads of kernels, not when these get paid and how one controls that so I'm confused.
I don't have a full answer to this yet. But so far the biggest clue I've gotten has come from reading up on the currently undocumented function #cupy.fuse() which makes it more clear than the #numba.jit documents where the kernel launch costs are paid. I have not found the connection to Contexts yet as recommended by #talonmies.
see https://gist.github.com/unnonouno/877f314870d1e3a2f3f45d84de78d56c
The key example is this
c = cupy.arange(4)
##cupy.fuse()
def foo(x):
return x+x+x+x+x
foo(.) will be three times slower with #cupy.fuse() commented out because each "+" involves a kernel load and a kernel free. Fusion merges all the adds into a single kernel so those the launch and free are paid onces. FOr matricies less than 1 million in size on a typical 2018 GPU, the add() is so fast that the launch and free are the dominate times.
I wish I could find some documentation on #fuse. FOr example, does it unroll internal functions the way #jit does. Could I achieve that by stacking #jit and #fuse?
I'm still however largely in the dark about when the costs are getting paid in numba.
Just reading a bit about what the advantage of GPU is, and I want to verify I understand on a practical level. Lets say I have 10,000 arrays each containing a billion simple equations to run. On a cpu it would need to go through every single equation, 1 at a time, but with a GPU I could run all 10,000 arrays as as 10,000 different threads, all at the same time, so it would finish a ton faster...is this example spot on or have I misunderstood something?
I wouldn't call it spot on, but I think you're headed in the right direction. Mainly, a GPU is optimized for graphics-related calculations. This does not, however, mean that's all it is capable of.
Without knowing how much detail you want me to go into here, I can say at the very least the concept of running things in parallel is relevant. The GPU is very good at performing many tasks simultaneously in one go (known as running in parallel). CPUs can do this too, but the GPU is specifically optimized to handle much larger numbers of specific calculations with preset data.
For example, to render every pixel on your screen requires a calculation, and the GPU will attempt to do as many of these calculations as it can all at the same time. The more powerful the GPU, the more of these it can handle at once and the faster its clock speed. The end result is a higher-end GPU can run your OS and games in 4k resolution, whereas other cards (or integrated graphics) might only be able to handle 1080p or less.
There's a lot more to this as well, but I figured you weren't looking for the insanely technical explanation.
The bottom line is this: For running a single task on one piece of data, the CPU will normally be faster. A single CPU core is generally much faster than a single GPU core. However, they typically have many cores and for running a single task on many pieces of data (so you have to run it once for each), the GPU will usually be faster. But these are data-driven situations, and as such each situation should be assessed on an individual basis to determine which to use and how to use it.
Hypothetically speaking, if my scientific work was leading toward the development of functions/modules/subroutines (on a desktop), what would I need to know to incorporate it into a large-scale simulation to be run on a supercomputer (which might simulate molecules, fluids, reactions, and so on)?
My impression is that it has to do with taking advantage of certain libraries (e.g., BLAS, LAPLACK) where possible, revising algorithms (reducing iteration), profiling, parallelizing, considering memory-hard disk-processor use/access... I am aware of the adage, "want to optimize your code? don't do it", but if one were interested in learning about writing efficient code, what references might be available?
I think this question is language agnostic, but since many number-crunching packages for biomolecular simulation, climate modeling, etc. are written in some version of Fortran, this language would probably be my target of interest (and I have programmed rather extensively in Fortran 77).
Profiling is a must at any level of machinery. In common usage, I've found that scaling to larger and larger grids requires a better understanding of the grid software and the topology of the grid. In that sense, everything you learn about optimizing for one machine is still applicable, but understanding the grid software gets you additional mileage. Hadoop is one of the most popular and widespread grid systems, so learning about the scheduler options, interfaces (APIs and web interfaces), and other aspects of usage will help. Although you may not use Hadoop for a given supercomputer, it is one of the less painful methods for learning about distributed computing. For parallel computing, you may pursue MPI and other systems.
Additionally, learning to parallelize code on a single machine, across multiple cores or processors, is something you can begin learning on a desktop machine.
Recommendations:
Learn to optimize code on a single machine:
Learn profiling
Learn to use optimized libraries (after profiling: so that you see the speedup)
Be sure you know algorithms and data structures very well (*)
Learn to do embarrassingly parallel programming on multiple core machines.
Later: consider multithreaded programming. It's harder and may not pay off for your problem.
Learn about basic grid software for distributed processing
Learn about tools for parallel processing on a grid
Learn to program for alternative hardware, e.g. GPUs, various specialized computing systems.
This is language agnostic. I have had to learn the same sequence in multiple languages and multiple HPC systems. At each step, take a simpler route to learn some of the infrastructure and tools; e.g. learn multicore before multithreaded, distributed before parallel, so that you can see what fits for the hardware and problem, and what doesn't.
Some of the steps may be reordered depending on local computing practices, established codebases, and mentors. If you have a large GPU or MPI library in place, then, by all means, learn that rather than foist Hadoop onto your collaborators.
(*) The reason to know algorithms very well is that as soon as your code is running on a grid, others will see it. When it is hogging up the system, they will want to know what you're doing. If you are running a process that is polynomial and should be constant, you may find yourself mocked. Others with more domain expertise may help you find good approximations for NP-hard problems, but you should know that the concept exists.
Parallelization would be the key.
Since the problems you cited (e.g. CFD, multiphysics, mass transfer) are generally expressed as large-scale linear algebra problems, you need matrix routines that parallelize well. MPI is a standard for those types of problems.
Physics can influence as well. For example, it's possible to solve some elliptical problems efficiently using explicit dynamics and artificial mass and damping matricies.
3D multiphysics means coupled differential equations with varying time scales. You'll want a fine mesh to resolve details in both space and time, so the number of degrees of freedom will rise rapidly; time steps will be governed by the stability requirements of your problem.
If someone ever figures out how to run linear algebra as a map-reduce problem they'll have it knocked.
Hypothetically speaking, if my scientific work was leading toward the development of functions/modules/subroutines (on a desktop), what would I need to know to incorporate it into a large-scale simulation to be run on a supercomputer (which might simulate molecules, fluids, reactions, and so on)?
First, you would need to understand the problem. Not all problems can be solved in parallel (and I'm using the term parallel in as wide meaning as it can get). So, see how the problem is now solved. Can it be solved with some other method quicker. Can it be divided in independent parts ... and so on ...
Fortran is the language specialized for scientific computing, and during the recent years, along with the development of new language features, there has also been some very interesting development in terms of features that are aiming for this "market". The term "co-arrays" could be an interesting read.
But for now, I would suggest reading first into a book like Using OpenMP - OpenMP is a simpler model but the book (fortran examples inside) explains nicely the fundamentals. Message parsing interface (for friends, MPI :) is a larger model, and one of often used. Your next step from OpenMP should probably go in this direction. Books on the MPI programming are not rare.
You mentioned also libraries - yes, some of those you mentioned are widely used. Others are also available. A person who does not know exactly where the problem in performance lies should IMHO never try to undertake the task of trying to rewrite library routines.
Also there are books on parallel algorithms, you might want to check out.
I think this question is language agnostic, but since many number-crunching packages for biomolecular simulation, climate modeling, etc. are written in some version of Fortran, this language would probably be my target of interest (and I have programmed rather extensively in Fortran 77).
In short it comes down to understanding the problem, learning where the problem in performance is, re-solving the whole problem again with a different approach, iterating a few times, and by that time you'll already know what you're doing and where you're stuck.
We're in a position similar to yours.
I'm most in agreement with #Iterator's answer, but I think there's more to say.
First of all, I believe in "profiling" by the random-pausing method, because I'm not really interested in measuring things (It's easy enough to do that) but in pinpointing code that is causing time waste, so I can fix it. It's like the difference between a floodlight and a laser.
For one example, we use LAPACK and BLAS. Now, in taking my stack samples, I saw a lot of the samples were in the routine that compares characters. This was called from a general routine that multiplies and scales matrices, and that was called from our code. The matrix-manipulating routine, in order to be flexible, has character arguments that tell it things like, if a matrix is lower-triangular or whatever. In fact, if the matrices are not very large, the routine can spend more than 50% of its time just classifying the problem. Of course, the next time it is called from the same place, it does the same thing all over again. In a case like that, a special routine should be written. When it is optimized by the compiler, it will be as fast as it reasonably can be, and will save all that classifying time.
For another example, we use a variety of ODE solvers. These are optimized to the nth degree of course. They work by calling user-provided routines to calculate derivatives and possibly a jacobian matrix. If those user-provided routines don't actually do much, samples will indeed show the program counter in the ODE solver itself. However, if the user-provided routines do much more, samples will find the lower end of the stack in those routines mostly, because they take longer, while the ODE code takes roughly the same time. So, optimization should be concentrated in the user-provided routines, not the ODE code.
Once you've done several of the kind of optimization that is pinpointed by stack sampling, which can speed things up by 1-2 orders of magnitude, then by all means exploit parallelism, MPI, etc. if the problem allows it.
I'm a CS master student, and next semester I will have to start working on my thesis. I've had trouble coming up with a thesis idea, but I decided it will be related to Computer Graphics as I'm passionate about game development and wish to work as a professional game programmer one day.
Unfortunately I'm kinda new to the field of 3D Computer Graphics, I took an undergraduate course on the subject and hope to take an advanced course next semester, and I'm already reading a variety of books and articles to learn more. Still, my supervisor thinks its better if I come up with a general thesis idea now and then spend time learning about it in preparation for doing my thesis proposal. My supervisor has supplied me with some good ideas but I'd rather do something more interesting on my own, which hopefully has to do with games and gives me more opportunities to learn more about the field. I don't care if it's already been done, for me the thesis is more of an opportunity to learn about things in depth and to do substantial work on my own.
I don't know much about GPU programming and I'm still learning about shaders and languages like CUDA. One idea I had is to program an entire game (or as much as possible) on the GPU, including all the game logic, AI, and tests. This is inspired by reading papers on GPGPU and questions like this one I don't know how feasible that is with my knowledge, and my supervisor doesn't know a lot about recent GPUs. I'm sure with time I will be able to answer this question on my own, but it'd be handy if I could know the answer in advance so I could also consider other ideas.
So, if you've got this far, my question: Using only shaders or something like CUDA, can you make a full, simple 3D game that exploits the raw power and parallelism of GPUs? Or am I missing some limitation or difference between GPUs and CPUs that will always make a large portion of my code bound to CPU? I've read about physics engines running on the GPU, so why not everything else?
DISCLAIMER: I've done a PhD, but have never supervised a student of my own, so take all of what I'm about to say with a grain of salt!
I think trying to force as much of a game as possible onto a GPU is a great way to start off your project, but eventually the point of your work should be: "There's this thing that's an important part of many games, but in it's present state doesn't fit well on a GPU: here is how I modified it so it would fit well".
For instance, fortran mentioned that AI algorithms are a problem because they tend to rely on recursion. True, but, this is not necessarily a deal-breaker: the art of converting recursive algorithms into an iterative form is looked upon favorably by the academic community, and would form a nice center-piece for your thesis.
However, as a masters student, you haven't got much time so you would really need to identify the kernel of interest very quickly. I would not bother trying to get the whole game to actually fit onto the GPU as part of the outcome of your masters: I would treat it as an exercise just to see which part won't fit, and then focus on that part alone.
But be careful with your choice of supervisor. If your supervisor doesn't have any relevant experience, you should pick someone else who does.
I'm still waiting for a Gameboy Emulator that runs entirely on the GPU, which is just fed the game ROM itself and current user input and results in a texture displaying the game - maybe a second texture for sound output :)
The main problem is that you can't access persistent storage, user input or audio output from a GPU. These parts have to be on the CPU, by definition (even though cards with HDMI have audio output, but I think you can't control it from the GPU). Apart from that, you can already push large parts of the game code into the GPU, but I think it's not enough for a 3D game, since someone has to feed the 3D data into the GPU and tell it which shaders should apply to which part. You can't really randomly access data on the GPU or run arbitrary code, someone has to do the setup.
Some time ago, you would just setup a texture with the source data, a render target for the result data, and a pixel shader that would do the transformation. Then you rendered a quad with the shader to the render target, which would perform the calculations, and then read the texture back (or use it for further rendering). Today, things have been made simpler by the fourth and fifth generation of shaders (Shader Model 4.0 and whatever is in DirectX 11), so you can have larger shaders and access memory more easily. But still they have to be setup from the outside, and I don't know how things are today regarding keeping data between frames. In worst case, the CPU has to read back from the GPU and push again to retain game data, which is always a slow thing to do. But if you can really get to a point where a single generic setup/rendering cycle would be sufficient for your game to run, you could say that the game runs on the GPU. The code would be quite different from normal game code, though. Most of the performance of GPUs comes from the fact that they execute the same program in hundreds or even thousands of parallel shading units, and you can't just write a shader that can draw an image to a certain position. A pixel shader always runs, by definition, on one pixel, and the other shaders can do things on arbitrary coordinates, but they don't deal with pixels. It won't be easy, I guess.
I'd suggest just trying out the points I said. The most important is retaining state between frames, in my opinion, because if you can't retain all data, all is impossible.
First, Im not a computer engineer so my assumptions cannot even be a grain of salt, maybe nano scale.
Artificial intelligence? No problem.There are countless neural network examples running in parallel in google. Example: http://www.heatonresearch.com/encog
Pathfinding? You just try some parallel pathfinding algorithms that are already on internet. Just one of them: https://graphics.tudelft.nl/Publications-new/2012/BB12a/BB12a.pdf
Drawing? Use interoperability of dx or gl with cuda or cl so drawing doesnt cross pci-e lane. Can even do raytracing at corners so no z-fighting anymore, even going pure raytraced screen is doable with mainstream gpu using a low depth limit.
Physics? The easiest part, just iterate a simple Euler or Verlet integration and frequently stability checks if order of error is big.
Map/terrain generation? You just need a Mersenne-twister and a triangulator.
Save game? Sure, you can compress the data parallelly before writing to a buffer. Then a scheduler writes that data piece by piece to HDD through DMA so no lag.
Recursion? Write your own stack algorithm using main vram, not local memory so other kernels can run in wavefronts and GPU occupation is better.
Too much integer needed? You can cast to a float then do 50-100 calcs using all cores then cast the result back to integer.
Too much branching? Compute both cases if they are simple, so every core is in line and finish in sync. If not, then you can just put a branch predictor of yourself so the next time, it predicts better than the hardware(could it be?) with your own genuine algorithm.
Too much memory needed? You can add another GPU to system and open DMA channel or a CF/SLI for faster communication.
Hardest part in my opinion is the object oriented design since it is very weird and hardware dependent to build pseudo objects in gpu. Objects should be represented in host(cpu) memory but they must be separated over many arrays in gpu to be efficient. Example objects in host memory: orc1xy_orc2xy_orc3xy. Example objects in gpu memory: orc1_x__orc2_x__ ... orc1_y__orc2_y__ ...
The answer has already been chosen 6 years ago but for those interested to the actual question, Shadertoy, a live-coding WebGL platform, recently added the "multipass" feature allowing preservation of state.
Here's a live demo of the Bricks game running on Gpu.
I don't care if it's already been
done, for me the thesis is more of an
opportunity to learn about things in
depth and to do substantial work on my
own.
Then your idea of what a thesis is is completely wrong. A thesis must be an original research. --> edit: I was thinking about a PhD thesis, not a master thesis ^_^
About your question, the GPU's instruction sets and capabilities are very specific to vector floating point operations. The game logic usually does little floating point, and much logic (branches and decision trees).
If you take a look to the CUDA wikipedia page you will see:
It uses a recursion-free,
function-pointer-free subset of the C
language
So forget about implementing any AI algorithms there, that are essentially recursive (like A* for pathfinding). Maybe you could simulate the recursion with stacks, but if it's not allowed explicitly it should be for a reason. Not having function pointers also limits somewhat the ability to use dispatch tables for handling the different actions depending on state of the game (you could use again chained if-else constructions, but something smells bad there).
Those limitations in the language reflect that the underlying HW is mostly thought to do streaming processing tasks. Of course there are workarounds (stacks, chained if-else), and you could theoretically implement almost any algorithm there, but they will probably make the performance suck a lot.
The other point is about handling the IO, as already mentioned up there, this is a task for the main CPU (because it is the one that executes the OS).
It is viable to do a masters thesis on a subject and with tools that you are, when you begin, unfamiliar. However, its a big chance to take!
Of course a masters thesis should be fun. But ultimately, its imperative that you pass with distinction and that might mean tackling a difficult subject that you have already mastered.
Equally important is your supervisor. Its imperative that you tackle some problem they show an interest in - that they are themselves familiar with - so that they can become interested in helping you get a great grade.
You've had lots of hobby time for scratching itches, you'll have lots more hobby time in the future too no doubt. But master thesis time is not the time for hobbies unfortunately.
Whilst GPUs today have got some immense computational power, they are, regardless of things like CUDA and OpenCL limited to a restricted set of uses, whereas the CPU is more suited towards computing general things, with extensions like SSE to speed up specific common tasks. If I'm not mistaken, some GPUs have the inability to do a division of two floating point integers in hardware. Certainly things have improved greatly compared to 5 years ago.
It'd be impossible to develop a game to run entirely in a GPU - it would need the CPU at some stage to execute something, however making a GPU perform more than just the graphics (and physics even) of a game would certainly be interesting, with the catch that game developers for PC have the biggest issue of having to contend with a variety of machine specification, and thus have to restrict themselves to incorporating backwards compatibility, complicating things. The architecture of a system will be a crucial issue - for example the Playstation 3 has the ability to do multi gigabytes a second of throughput between the CPU and RAM, GPU and Video RAM, however the CPU accessing GPU memory peaks out just past 12MiB/s.
The approach you may be looking for is called "GPGPU" for "General Purpose GPU". Good starting points may be:
http://en.wikipedia.org/wiki/GPGPU
http://gpgpu.org/
Rumors about spectacular successes in this approach have been around for a few years now, but I suspect that this will become everyday practice in a few years (unless CPU architectures change a lot, and make it obsolete).
The key here is parallelism: if you have a problem where you need a large number of parallel processing units. Thus, maybe neural networks or genetic algorithms may be a good range of problems to attack with the power of a GPU. Maybe also looking for vulnerabilities in cryptographic hashes (cracking the DES on a GPU would make a nice thesis, I imagine :)). But problems requiring high-speed serial processing don't seem so much suited for the GPU. So emulating a GameBoy may be out of scope. (But emulating a cluster of low-power machines might be considered.)
I would think a project dealing with a game architecture that targets multiple core CPUs and GPUs would be interesting. I think this is still an area where a lot of work is being done. In order to take advantage of current and future computer hardware, new game architectures are going to be needed. I went to GDC 2008 and there were ome talks related to this. Gamebryo had an interesting approach where they create threads for processing computations. You can designate the number of cores you want to use so that if you don't starve out other libraries that might be multi-core. I imagine the computations could be targeted to GPUs as well.
Other approaches included targeting different systems for different cores so that computations could be done in parallel. For instance, the first split a talk suggested was to put the renderer on its own core and the rest of the game on another. There are other more complex techniques but it all basically boils down to how do you get the data around to the different cores.