Related
I'm looking to optimize C++ code (mainly some for loops) using the NEON capability of computing 4 or 8 array elements at a time. Is there some kind of library or set of functions that can be used in C++ environment?
I use Eclipse IDE in Linux Gentoo to write C++ code.
UPDATE
After reading the answers I did some tests with the software. I compiled my project with the following flags:
-O3 -mcpu=cortex-a9 -ftree-vectorize -mfloat-abi=hard -mfpu=neon
Keep in mind that this project includes extensive libraries such as open frameworks, OpenCV, and OpenNI, and everything was compiled with these flags.
To compile for the ARM board we use a Linaro toolchain cross-compiler, and GCC's version is 4.8.3.
Would you expect this to improve the performance of the project? Because we experienced no changes at all, which is rather weird considering all the answers I read here.
Another question: all the for cycles have an apparent number of iterations, but many of them iterate through custom data types (structs or classes). Can GCC optimize these cycles even though they iterate through custom data types?
EDIT:
From your update, you may misunderstand what the NEON processor does. It is an SIMD (Single Instruction, Multiple Data) vector processor. That means that it is very good at performing an instruction (say "multiply by 4") to several pieces of data at the same time. It also loves to do things like "add all these numbers together" or "add each element of these two lists of numbers to create a third list of numbers." So if you problem looks like those things the NEON processor is going to be huge help.
To get that benefit, you must put your data in very specific formats so that the vector processor can load multiple data simultaneously, process it in parallel, and then write it back out simultaneously. You need to organize things such that the math avoids most conditionals (because looking at the results too soon means a roundtrip to the NEON). Vector programming is a different way of thinking about your program. It's all about pipeline management.
Now, for many very common kinds of problems, the compiler automatically can work all of this out. But it's still about working with numbers, and numbers in particular formats. For example, you almost always need to get all of your numbers into a contiguous block in memory. If you're dealing with fields inside of structs and classes, the NEON can't really help you. It's not a general-purpose "do stuff in parallel" engine. It's an SIMD processor for doing parallel math.
For very high-performance systems, data format is everything. You don't take arbitrary data formats (structs, classes, etc.) and try to make them fast. You figure out the data format that will let you do the most parallel work, and you write your code around that. You make your data contiguous. You avoid memory allocation at all costs. But this isn't really something a simple StackOverflow question can address. High-performance programming is a whole skill set and a different way of thinking about things. It isn't something you get by finding the right compiler flag. As you've found, the defaults are pretty good already.
The real question you should be asking is whether you could reorganize your data so that you can use more of OpenCV. OpenCV already has lots of optimized parallel operations that will almost certainly make good use of the NEON. As much as possible, you want to keep your data in the format that OpenCV works in. That's likely where you're going to get your biggest improvements.
My experience is that it is certainly possible to hand-write NEON assembly that will beat clang and gcc (at least from a couple of years ago, though the compiler certainly continues to improve). Having excellent ARM optimization is not the same as NEON optimization. As #Mats notes, the compiler will generally do an excellent job at obvious cases, but does not always handle every case ideally, and it is certainly possible for even a lightly skilled developer to sometimes beat it, sometimes dramatically. (#wallyk is also correct that hand-tuning assembly is best saved for last; but it can still be very powerful.)
That said, given your statement "Assembly, for which I have absolutely no background, and can't possibly afford to learn at this point," then no, you should not even bother. Without first at least understanding the basics (and a few non-basics) of assembly (and specifically vectorized NEON assembly), there is no point in second-guessing the compiler. Step one of beating the compiler is knowing the target.
If you are willing to learn the target, my favorite introduction is Whirlwind Tour of ARM Assembly. That, plus some other references (below), were enough to let me beat the compiler by 2-3x in my particular problems. On the other hand, they were insufficient enough that when I showed my code to an experienced NEON developer, he looked at it for about three seconds and said "you have a halt right there." Really good assembly is hard, but half-decent assembly can still be better than optimized C++. (Again, every year this gets less true as the compiler writers get better, but it can still be true.)
ARM Assembly language
A few things iOS developers ought to know about the ARM architecture (iPhone-focused, but the principles are the same for all uses.)
ARM NEON support in the ARM compiler
Coding for NEON
One side note, my experience with NEON intrinsics is that they are seldom worth the trouble. If you're going to beat the compiler, you're going to need to actually write full assembly. Most of the time, whatever intrinsic you would have used, the compiler already knew about. Where you get your power is more often in restructuring your loops to best manage your pipeline (and intrinsics don't help there). It's possible this has improved over the last couple of years, but I would expect the improving vector optimizer to outpace the value of intrinsics more than the other way around.
Here's a "mee too" with some blog posts from ARM. FIRST, start with the following to get the background information, including 32-bit ARM (ARMV7 and below), Aarch32 (ARMv8 32-bit ARM) and Aarch64 (ARMv8 64-bit ARM):
ARM NEON programming quick reference
Second, checkout the Coding for NEON series. Its a nice introduction with pictures so things like interleaved loads make sense with a glance.
ARM NEON programming quick reference
Coding for NEON - Part 1: Load and Stores
Coding for NEON - Part 2: Dealing With Leftovers
Coding for NEON - Part 3: Matrix Multiplication
Coding for NEON - Part 4: Shifting Left and Right
Coding for NEON - Part 5: Rearranging Vectors
I also went on Amazon looking for some books on ARM assembly with a treatment of NEON. I could only find two, and neither book's treatment of NEON were impressive. They reduced to a single chapter with the obligatory Matrix example.
I believe ARM Intrinsics are a very good idea. The instrinsics allow you to write code for GCC, Clang and Visual C/C++ compilers. We have one code base that works for ARM Linux distros (like Linaro), some iOS devices (using -arch armv7) and Microsoft gadgets (like Windows Phone and Windows Store Apps).
If you have access to a reasonably modern GCC (GCC 4.8 and upwards) I would recommend giving intrinsics a go. The NEON intrinsics are a set of functions that the compiler knows about, which can be used from C or C++ programs to generate NEON/Advanced SIMD instructions. To gain access to them in your program, it is necessary to #include <arm_neon.h>. The verbose documentation of all available intrinsics is available at http://infocenter.arm.com/help/topic/com.arm.doc.ihi0073a/IHI0073A_arm_neon_intrinsics_ref.pdf , but you may find more user-friendly tutorials elsewhere online.
Advice on this site is generally against the NEON intrinsics, and certainly there are GCC versions which have done a poor job of implementing them, but recent versions do reasonably well (and if you spot bad code generation, please do raise it as a bug - https://gcc.gnu.org/bugzilla/ )
They are an easy way to program to the NEON/Advanced SIMD instruction set, and the performance you can achieve is often rather good. They are also "portable", in that when you move to an AArch64 system, a superset of the intrinsics you can use from ARMv7-A are available. They are also portable across implementations of the ARM architecture, which can vary in their performance characteristics, but which the compiler will model for performance tuning.
The principle benefit of the NEON intrinsics over hand-written assembly, is that the compiler can understand them when performing its various optimization passes. By contrast hand-written assembler is an opaque block to GCC, and will not be optimized. On the other hand, expert assembler programmers can often beat the compiler's register allocation policies, particularly when using the instructions which write to or read from to multiple consecutive registers.
In addition to Wally's answer - and probably should be a comment, but I couldn't make it short enough: ARM has a team of compiler developers whose entire role is to improve the parts of GCC and Clang/llvm that does code generation for ARM CPUs, including features that provides "auto-vectorization" - I have not looked deeply into it, but from my experience on x86 code generation, I'd expect for anything that is relatively easy to vectorize, the compiler should do a deecent job. Some code is hard for the compiler to understand when it can vectorize or not, and may need some "encouragement" - such as unrolling loops or marking conditions as "likely" or "unlikely", etc.
Disclaimer: I work for ARM, but have very little to do with the compilers or even CPUs, as I work for the group that does graphics (where I have some involvement with compilers for the GPUs in the OpenCL part of the GPU driver).
Edit:
Performance, and use of various instruction extensions is really depending on EXACTLY what the code is doing. I'd expect that libraries such as OpenCV is already doing a fair amount of clever stuff in their code (such as both handwritten assembler as compiler intrinsics and generally code that is designed to allow the compiler to already do a good job), so it may not really give you much improvement. I'm not a computer vision expert, so I can't really comment on exactly how much such work is done on OpenCV, but I'd certainly expect the "hottest" points of the code to have been fairly well optimised already.
Also, profile your application. Don't just fiddle with optimisation flags, measure it's performance and use a profiling tool (e.g. the Linux "perf" tool) to measure WHERE your code is spending time. Then see what can be done to that particular code. Is it possible to write a more parallel version of it? Can the compiler help, do you need to write assembler? Is there a different algorithm that does the same thing but in a better way, etc, etc...
Although tweaking compiler options CAN help, and often does, it can give tens of percent, where a change in algorithm can often lead to 10 times or 100 times faster code - assuming of course, your algorithm can be improved!
Understanding what part of your application is taking the time, however, is KEY. It's no point in changing things to make the code that takes 5% of the time 10% faster, when a change somewhere else could make a piece of code that is 30 or 60% of the total time 20% faster. Or optimise some math routine, when 80% of the time is spent on reading a file, where making the buffer twice the size would make it twice as fast...
Although a long time has passed since I submitted this question, I realize that it gathers some interest and I decided to tell what I ended up doing regarding this.
My main goal was to optimize a for-loop which was the bottleneck of the project. So, since I don't know anything about Assembly I decided to give NEON intrinsics a go. I ended up having a 40-50% gain in performance (in this loop alone), and a significant overall improvement in performance of the whole project.
The code does some math to transform a bunch of raw distance data into distance to a plane in millimetres. I use some constants (like _constant05, _fXtoZ) that are not defined here, but they are just constant values defined elsewhere.
As you can see, I'm doing the math for 4 elements at a time, talk about real parallelization :)
unsigned short* frameData = frame.ptr<unsigned short>(_depthLimits.y, _depthLimits.x);
unsigned short step = _runWidth - _actWidth; //because a ROI being processed, not the whole image
cv::Mat distToPlaneMat = cv::Mat::zeros(_runHeight, _runWidth, CV_32F);
float* fltPtr = distToPlaneMat.ptr<float>(_depthLimits.y, _depthLimits.x); //A pointer to the start of the data
for(unsigned short y = _depthLimits.y; y < _depthLimits.y + _depthLimits.height; y++)
{
for (unsigned short x = _depthLimits.x; x < _depthLimits.x + _depthLimits.width - 1; x +=4)
{
float32x4_t projX = {(float)x, (float)(x + 1), (float)(x + 2), (float)(x + 3)};
float32x4_t projY = {(float)y, (float)y, (float)y, (float)y};
framePixels = vld1_u16(frameData);
float32x4_t floatFramePixels = {(float)framePixels[0], (float)framePixels[1], (float)framePixels[2], (float)framePixels[3]};
float32x4_t fNormalizedY = vmlsq_f32(_constant05, projY, _yResInv);
float32x4_t auxfNormalizedX = vmulq_f32(projX, _xResInv);
float32x4_t fNormalizedX = vsubq_f32(auxfNormalizedX, _constant05);
float32x4_t realWorldX = vmulq_f32(fNormalizedX, floatFramePixels);
realWorldX = vmulq_f32(realWorldX, _fXtoZ);
float32x4_t realWorldY = vmulq_f32(fNormalizedY, floatFramePixels);
realWorldY = vmulq_f32(realWorldY, _fYtoZ);
float32x4_t realWorldZ = floatFramePixels;
realWorldX = vsubq_f32(realWorldX, _tlVecX);
realWorldY = vsubq_f32(realWorldY, _tlVecY);
realWorldZ = vsubq_f32(realWorldZ, _tlVecZ);
float32x4_t distAuxX, distAuxY, distAuxZ;
distAuxX = vmulq_f32(realWorldX, _xPlane);
distAuxY = vmulq_f32(realWorldY, _yPlane);
distAuxZ = vmulq_f32(realWorldZ, _zPlane);
float32x4_t distToPlane = vaddq_f32(distAuxX, distAuxY);
distToPlane = vaddq_f32(distToPlane, distAuxZ);
*fltPtr = (float) distToPlane[0];
*(fltPtr + 1) = (float) distToPlane[1];
*(fltPtr + 2) = (float) distToPlane[2];
*(fltPtr + 3) = (float) distToPlane[3];
frameData += 4;
fltPtr += 4;
}
frameData += step;
fltPtr += step;
}
If you don't want to mess with assembly code at all, then tweak the compiler flags to maximally optimize for speed. gcc given the proper ARM target should do this provided the number of loop iterations is apparent.
To check gcc code generation, request assembly output by adding the -S flag.
If after several tries (of reading the gcc documentation and tweaking flags) you still can't get it to produce the code you want, then take the assembly output and edit it to your satisfaction.
Beware of premature optimization. The proper development order is to get the code functional, then see if it needs optimization. Only when the code is stable does it makes sense to do so.
Play with some minimal assembly examples on QEMU to understand the instructions
The following setup does not have many examples yet, but it serves as a neat playground:
v7 examples
v8 examples
setup usage
The examples run on QEMU user mode, which dispenses extra hardware, and the GDB is working just fine.
The asserts are done through the C standard library.
You should be a able to easily extend that setup with new instructions as you learn them.
ARM intrinsincs in particular were asked at: Is there a good reference for ARM Neon intrinsics?
I am aware that C/C++ is a lower-level language and generates relatively optimized machine code when we compare with any other high-level language. But I guess there is pretty much more than that, which is also evident from the practice.
When I do simple calculations like montecarlo averaging of a Gaussian sample collection or so, I see there is not much of a difference between a C++ implementation or MATLAB implementation, sometimes in fact MATLAB performs a bit better in time.
When I move on to larger scale simulations with thousands of lines of code, slowly the real picture shows up. C++ simulations show superior performance like 100x better in time complexity than an equivalent MATLAB implementation.
The code in C++ most of the times, is pretty much serial and no hi-fi optimization is done explicitly. Whereas, as per my awareness, MATLAB inherently does a lot of optimization. This shows up for example when I try to generate a huge chunk of random samples, where as the equivalent in C++ using some library like IT++/GSL/Boost performs relatively slower (the algorithm used is the same namely mt19937).
My question is simply to know if there is a simpler tradeoff between MATLAB/C++ in performance. Is it just like what people say, "Whenever you can, C/C++ is the better"(The frequently experienced)?. In a different perspective, "What is MATLAB good for, other than comfort?"
By the way, I don't see coding efficiency parameter being significant here, thinking of the same programmer in both cases. And also, I think the other alternatives like python,R are not relevant here. But dependence on the specific libraries we use should be interesting.
[I am a phd student in Coding Theory in communication systems. I do simulations using matlab/C++ all the time, and have reasonable experience of coding few 10K's of lines in both cases]
I have been using Matlab and C++ for about 10 years. For every numerical algorithms implemented for my research, I always start from prototyping with Matlab and then translate the project to C++ to gain a 10x to 100x (I am not kidding) performance improvement. Of course, I am comparing optimized C++ code to the fully vectorized Matlab code. On average, the improvement is about 50x.
There are lot of subtleties behind both of the two programming languages, and the following are some misunderstandings:
Matlab is a script language but C++ is compiled
Matlab uses JIT compiler to translate your script to machine code, you can improve your speed at most by a factor 1.5 to 2 by using the compiler that Matlab provides.
Matlab code might be able to get fully vectorized but you have to optimize your code by hand in C++
Fully vectorized Matlab code can call libraries written in C++/C/Assembly (for example Intel MKL). But plain C++ code can be reasonably vectorized by modern compilers.
Toolboxes and routines that Matlab provides should be very well tuned and should have reasonable performance
No. Other than linear algebra routines, the performance is generally bad.
The reasons why you can gain 10x~100x performance in C++ comparing to vectorized Matlab code:
Calling external libraries (MKL) in Matlab costs time.
Memory in Matlab is dynamically allocated and freed. For example, small matrices multiplication:
A = B*C + D*E + F*G
requires Matlab to create 2 temporary matrices. And in C++, if you allocate your memory before hand, you create NONE. And now imagine you loop that statement for 1000 times. Another solution in C++ is provided by C++11 Rvalue reference. This is the one of the biggest improvement in C++, now C++ code can be as fast as plain C code.
If you want to do parallel processing, Matlab model is multi-process and the C++ way is multi-thread. If you have many small tasks needing to be parallelized, C++ provides linear gain up to many threads but you might have negative performance gain in Matlab.
Vectorization in C++ involves using intrinsics/assembly, and sometimes SIMD vectorization is only possible in C++.
In C++, it is possible for an experienced programmer to completely avoid L2 cache miss and even L1 cache miss, hence pushing CPU to its theoretical throughput limit. Performance of Matlab can lag behind C++ by a factor of 10x due to this reason alone.
In C++, computational intensive instructions sometimes can be grouped according to their latencies (code carefully in assembly or intrinsics) and dependencies (most of time is done automatically by compiler or CPU hardware), such that theoretical IPC (instructions per clock cycle) could be reached and CPU pipelines are filled.
However, development time in C++ is also a factor of 10x comparing to Matlab!
The reasons why you should use Matlab instead of C++:
Data visualization. I think my career can go on without C++ but I won't be able to survive without Matlab just because it can generate beautiful plots!
Low efficiency but mathematically robust build-in routines and toolboxes. Get the correct answer first and then talk about efficiency. People can make subtle mistakes in C++ (for example implicitly convert double to int) and get sort of correct results.
Express your ideas and present your code to your colleagues. Matlab code is much easier to read and much shorter than C++, and Matlab code can be correctly executed without compiler. I just refuse to read other people's C++ code. I don't even use C++ GNU scientific libraries because the code quality is not guaranteed. It is dangerous for a researcher/engineer to use a C++ library as a black box and take the accuracy as granted. Even for commercial C/C++ libraries, I remember Intel compiler had a sign error in its sin() function last year and numerical accuracy problems also occurred in MKL.
Debugging Matlab script with interactive console and workspace is a lot more efficient than C++ debugger. Finding an index calculation bug in Matlab could be done within minutes, but it could take hours in C++ figuring out why the program crashes randomly if boundary check is removed for the sake of speed.
Last but not the least:
Because once Matlab code is vectorized, there is not much left for a programmer to optimize, Matlab code performance is much less sensitive to the quality of the code comparing with C++ code. Therefore it is best to optimize computation algorithms in Matlab, and marginally better algorithms normally have marginally better performance in Matlab. On the other hand, algorithm test in C++ requires decent programmer to write algorithms optimized more or less in the same way, and to make sure the compiler does not optimize the algorithms differently.
My recent experience in C++ and Matlab:
I made several large Matlab data analysis tools in the past year and suffered from the slow speed of Matlab. But I was able to improve my Matlab program speed by 10x through the following techniques:
Run/profile the Matlab script, re-implement critical routines in C/C++ and compile with MEX. Critical routines are mostly likely logically simple but numerically heavy. This improves speed by 5x.
Simplify ".m" files shipped with Matlab tool boxes by commenting all unnecessary safety checks and output parameter computations. Please be reminded that the modified code cannot be distributed with the rest of the user scripts. This improves speed by another 2x (after C/C++ and MEX).
The improved code is ~98% in Matlab and ~2% in C++.
I believe it is possible to improve the speed by another 2x (total 20x) if the entire tool is coded in C++, this is ~100x speed improvement of the computation routines. The hard drive I/O will then dominate the program run time.
Question for Mathworks engineers:
When Matlab code is fully vectorized, one of the performance limiting factor is the matrix indexing operation. For instance, a finite difference operation needs to be performed on Matrix A which has a dimension of 5000x5000:
B = A(:,2:end)-A(:,1:end-1)
The matrix indexing operation makes the Matlab code multiple times slower than the C++ code. Can the matrix indexing performance be improved?
In my experience (several years of Computer Vision and image processing in both languages) there is no simple answer to this question, as Matlab performance depends strongly (and much more than C++ performance) on your coding style.
Generally, Matlab wraps the classic C++ / Fortran based linear algebra libraries. So anything like x = A\b is going to be very fast. Also, Matlab does a good job in choosing the most efficient solver for these types of problems, so for x = A\b Matlab will look at the size of your matrices and chose the appropriate low-level routines.
Matlab also shines in data manipulation of large matrices if you "vectorize" your code, i.e. if you avoid for loops and use index arrays or boolean arrays to access your data. This stuff is highly optimised.
For other routines, some are written in Matlab code, while others point to a C/C++ implementation (e.g. the Delaunay stuff). You can check this yourself by typing edit some_routine.m. This opens the code and you see whether it is all Matlab or just a wrapper for something compiled.
Matlab, I think, is primarily for comfort - but comfort translates to coding time and ultimately money which is why Matlab is used in the industry. Also, it is easy to learn for engineers from other fields than computer science, with little training in programming.
As a PhD Student too, and a 10years long Matlab user, I'm glad to share my POV:
Matlab is a great tool for developing and prototyping algorithms, especially when dealing with GUIs, high-level analysis (Frequency Domain, LS Optimization etc.): fast coding, powerful syntaxis (think about [],{},: etc.).
As soon as your processing chain is more stable and defined and data dimensions grows move to C/C++.
The main Matlab limit rises when considering its language is script-like: as long as you avoid any cycle (using arrayfun, cellfun or other matrix procedures) performances are high since the called subroutine is again in C/C++.
Your question is difficult to answer. In general C++ is faster, but if make use of the well written algorithms of Matlab it can outperform C++. In some cases Matlab can parallelize your code which has to be done manually in many cases for C++. Mathlab can kind of export C++ code.
So my conclusion is, that you have to measure the performance of both programs to get an answer. But then you compare your two implementations and not Matlab and C++ in general.
Matlab does very well with linear algebra and array/matrix operations, since they seem to have been doing some extra optimizations on the underlying operations - if you want to beat Matlab there, you would need a similarly optimized BLAS/LAPACK library.
As an interpreted language, Matlab loses time whenever a Matlab function is called, due to internal overhead, which traditionally meant that Matlab loops were slow. This has been alleviated somewhat in recent years thanks to significant improvement in the JIT compiler (search for "performance" questions on Matlab on SO for examples). As a consequence of the function call overhead, all Matlab functions that have not been implemented in C/C++ behind the scenes (call edit functionName to see whether it's written in Matlab) risks being slower than a C/C++ counterpart.
Finally, Matlab attempts to be user friendly, and may do "unnecessary" input checking that can take time (due to function call overhead). For example, if you know that ismember gets sorted inputs, you can call ismembc directly (the behind-the-scene compiled function), saving quite a bit of time.
I think you can consider the difference in four folds at least.
Compiled vs Interpreted
Strongly-typed vs Dynamically-typed
Performance vs Fast-prototyping
Special strength
For 1-3 can be easily generalized into comparison between two family of programming languages.
For 4, MATLAB is optimized for matrix operations. So if you can vectorize more code in MATLAB, the performance can be drastically boosted. Conversely, if many loops are required, never hesitate to use C++ or create a mex file.
It is a difficult quesion after all.
I saw a 5.5x speed improvement when switching from MATLAB to C++. This was for a robot controller- lots of loops and ode solving. I spent many hours trying to optimize the MATLAB code, hardly any time optimizing the C++ (I'm sure it could have been 10x faster with a little more effort).
However, it was easy to add a GUI for the MATLAB code, so I still use it more often. Like others have said, it was nice to prototype first on MATLAB. That made the implementation on C++ much simpler.
Besides the speed of the final program, you should also take into account the total development time of your code, ie., not only the time to write, but also to debug, etc. Matlab (and its open-source counterpart, Octave) can be good for quick prototyping due to its visualisation capabilities.
If you're using straight C++ (ie. no matrix libraries), it may take you much longer to write C++ code that's equivalent to Matlab code (eg. there might be no point in spending 10 hours writing C++ code that only runs 10 seconds quicker, compared to a Matlab program that took 5 minutes to write).
However, there are dedicated C++ matrix libraries, such as Armadillo, which provide a Matlab-like API. This can be useful for writing performance critical code that can be called from Matlab, or for converting Matlab code into "real" programs.
Some Matlab code uses standard linear algebra fictions with multithreading built into it. So, it appears that they are faster than a sequential C code.
GPU Compute Programmers,
I have a C++ program which currently relies on the ACML (LAPACK) to invert and multiple fairly large matrices of single precision fp values (E.g. 4,000 x 4,000). These matrices are very sparse although they do not always fit nicely into a diagonal matrix so I cannot presently reduce them. The other thing about this program is I have to do this invert and multiply several times (serially) as part of a Newton Rapson. However, I have several thousand permutations which can be done in parallel, each with a small change to the matrix before again calculating and inverting the Jacobian. This is all single precision fp, and seems perfectly suited for the GPU. My question is this...
I suspect I will need to use the AMD Accelerated Parallel Processing Math Libraries (APPML) for OpenGL as that is the only thing (non-CUDA, I want to be GPU agnostic) I know of which is available with BLAS functionality. My problem is I do not see the LAPACK dgetrf and dgetri functions included in APPML (yes, these are fp64 but I don't need that precision). Would C++ AMP be a better alternative? I am very interested in HSA features of passing pointers rather than copying data as there is a lot of data in flight here and some calculations still are done on the CPU. I believe that copy overhead would kill me otherwise. Ultimately, performance is the key and I want to make the right architectural decisions to set myself up for the most performance I can wring out of HSA GPUs coming out over the next 6 months.
I am using VS 2013 Ultimate preview and would be able to take advantage of C++ AMP for these HSA capabilities. I just want to make sure I am making the right long term architectural decision now while my program is in its infancy. Here is a link and snippet from some interesting data I found on Anandtech:
http://anandtech.com/show/7118/windows-81-and-vs2013-bring-gpu-computing-updates-to-direct3d-and-c-amp-
C++ AMP, Microsoft's C++ extension for GPU computing, has also been updated with the upcoming VS2013. I think the biggest feature update is that C++ AMP programs will also gain a shared memory feature on APUs/SoCs where the compiler and runtime will be able to eliminate extra data copies between CPU and GPU. This feature will also be available only on Windows 8.1 and it is likely built on top of the "map default buffer" as Microsoft's AMP implementation uses Direct3D under the hood. C++ AMP also brings some other nice additions including enhanced texture support and better debugging abilities.
Any thoughts, additional questions or discussion would be greatly appreciated!
I am in a project to process an image using CUDA. The project is simply an addition or subtraction of the image.
May I ask your professional opinion, which is best and what would be the advantages and disadvantages of those two?
I appreciate everyone's opinions and/or suggestions since this project is very important to me.
General answer: It doesn't matter. Use the language you're more comfortable with.
Keep in mind, however, that pycuda is only a wrapper around the CUDA C interface, so it may not always be up-to-date, also it adds another potential source of bugs, …
Python is great at rapid prototyping, so I'd personally go for Python. You can always switch to C++ later if you need to.
If the rest of your pipeline is in Python, and you're using Numpy already to speed things up, pyCUDA is a good complement to accelerate expensive operations. However, depending on the size of your images and your program flow, you might not get too much of a speedup using pyCUDA. There is latency involved in passing the data back and forth across the PCI bus that is only made up for with large data sizes.
In your case (addition and subtraction), there are built-in operations in pyCUDA that you can use to your advantage. However, in my experience, using pyCUDA for something non-trivial requires knowing a lot about how CUDA works in the first place. For someone starting from no CUDA knowledge, pyCUDA might be a steep learning curve.
Take a look at openCV, it contains a lot of image processing functions and all the helpers to load/save/display images and operate cameras.
It also now supports CUDA, some of the image processing functions have been reimplemented in CUDA and it gives you a good framework to do your own.
Alex's answer is right. The amount of time consumed in the wrapper is minimal. Note that PyCUDA has some nice metaprogramming constructs for generating kernels which might be useful.
If all you're doing is adding or subtracting elements of an image, you probably shouldn't use CUDA for this at all. The amount of time it takes to transfer back and forth across the PCI-E bus will dwarf the amount of savings you get from parallelism.
Any time you deal with CUDA, it's useful to think about the CGMA ratio (computation to global memory access ratio). Your addition/subtraction is only 1 float point operation for 2 memory accesses (1 read and 1 write). This ends up being very lousy from a CUDA perspective.
I'm developing non-interactive cpu-bound application which does only computations, almost no IO. Currently it works too long and while I'm working on improving the algorithm, I also think if it can give any benefit to change language or platform. Currently it is C++ (no OOP so it is almost C) on windows compiled with Intel C++ compiler. Can switching to ASM help and how much? Can switching to Linux and GCC help?
Just to be thorough: the first thing to do is to gather profile data and the second thing to do is consider your algorithms. I'm sure you know that, but they've got to be #included into any performance-programming discussion.
To be direct about your question "Can switching to ASM help?" the answer is "If you don't know the answer to that, then probably not." Unless you're very familiar with the CPU architecture and its ins and outs, it's unlikely that you'll do a significantly better job than a good optimizing C/C++ compiler on your code.
The next point to make is that significant speed-ups in your code (aside from algorithmic improvements) will almost certainly come from parallelism, not linear increases. Desktop machines can now throw 4 or 8 cores at a task, which has much more performance potential than a slightly better code generator. Since you're comfortable with C/C++, OpenMP is pretty much a no-brainer; it's very easy to use to parallelize your loops (obviously, you have to watch loop-carried dependencies, but it's definitely "the simplest parallelism that could possibly work").
Having said all that, code generation quality does vary between C/C++ compilers. The Intel C++ compiler is well-regarded for its optimization quality and has full support not just for OpenMP but for other technologies such as the Threading Building Blocks.
Moving into the question of what programming languages might be even better than C++, the answer would be "programming languages that actively promote / facilitate concepts of parallelism and concurrent programming." Erlang is the belle of the ball in that regard, and is a "hot" language right now and most people interested in performance programming are paying at least some attention to it, so if you want to improve your skills in that area, you might want to check it out.
It's always algorithm, rarely language. Here's my clue: "while I'm working on improving the algorithm".
Tweaking may not be enough.
Consider radical changes to the algorithm. You've got to eliminate processing, not make the processing go faster. The culprit is often "search" -- looping through data looking for something. Find ways to eliminate search. If you can't eliminate it, replace linear search with some kind of tree search or a hash map of some kind.
Switching to ASM is not going to help much, unless you're very good at it and/or have a specific critical path routine which you know you can do better. As several people have remarked, modern compilers are just better in most cases at taking advantages of caching/etc. than anyone can do by hand.
I'd suggest:
Try a different compiler, and/or different optimization options
Run a code coverage/analysis utility, and figure out where the critical paths are, and work on optimizing those in the code
C++ should be able to give you very near the best possible performance from the code, so I wouldn't recommend switching the language. Depending on the app, you may be able to get better performance on multi code/processor systems using multiple thread, as another suggestion.
While just switching to asm won't give any benefits, since the Intel C++ Compiler is likely better at optimizing than you, you can try one of the following options:
Try a compiler that will parallelize your code, like the VectorC compiler.
Try to switch to asm with heavy use of MMX, 3DNow!, SSE or whatever fits your needs (and your CPU). This will give more of a benefit than pure asm.
You can also try GPGPU, i.e. execute large parts of your algorithm on a GPU instead of a CPU. Depending on your algorithm, it can be dramatically faster.
Edit: I also second the profile approach. I recommend AQTime, which supports the Intel C++ compiler.
Personally I'd look at languages which allow you to take advantage of parallelism most easily, unless it's a thoroughly non-parallelisable situation. Being able to bolt on some extra cores and get (if possible!) near-linear improvement may well be a lot more cost-effective than squeezing the extra few percent of efficiency out.
When it comes to parallelisation, I believe functional languages are often regarded as the best way to go, or you could look at OpenMP for C/C++. (Personally, as a managed language guy, I'd be looking at libraries for Java/.NET, but I quite understand that not everyone has the same preferences!)
Try Fortran 77 - when it comes to computations still nothing beats the granddaddy of programming languages. Also, try it with OpenMP to take advantage of multiple cores.
Hand optimizing your ASM code compared to what C++ can do for you is rarely cost effective.
If you've done anything you can to the algorithm from a traditional algorithmic view, and you've also eliminated excesses, then you may either be SOL, or you can consider optimizing your program from a hardware point of view.
For example, any time you follow a pointer around the heap you are paying a huge cost due to cache misses, possibly paging, etc., which all affect branching predictions. Most programmers (even C gurus) tend to look at the CPU from the functional standpoint rather than what happens behind the scenes. Sometimes reorganizing memory, for example by "flattening" or manually allocating memory to fit on the same page can obtain ENORMOUS speedups. I managed to get 2X speedups on graph traversals just by flattening my structures.
These are not things that your compiler will do for you since they are based on your high-level understanding of the program.
As lobrien said, you haven't given us any information to tell you if hand-optimized ASM code would help... which means the answer is probably, "not yet."
Have you run your code with a profiler?
Do you know if the code is slow because of memory constraints or processor constraints?
Are you using all your available cores?
Have you identified any algorithms you're using that aren't O(1)? Can you get them to O(1)? If not, why not?
If you've done all that, how much control do you have over the environment your program is running in? (presumably a lot if you're thinking of switching operating systems) Can you disable other processes, give your process highest priority, etc? What about just finding a machine with a faster processor, more cores, or more memory (depending on what you're constrained on)
And on and on.
If you've already done all that and more, it's certainly possible you'll get to a point where you think, "I wonder if these few lines of code right here could be optimized better than the assembly that I'm looking at in the debugger right now?" And at that point you can ask specifically.
Good luck! You're solving a problem that's fun to solve.
Sometimes you can find libraries that have optimized implementations of the algorithms you care about. Often times they will have done the multithreading for you.
For example switching from LINPACK to LAPACK got us a 10x speed increase in LU factorization/solve with a good BLAS library.
First, figure out if you can change the algorithm, as S.Lott suggested.
Assuming the algorithm choice is correct, you might look a the memory access patterns, if you have a lot of data you are processing. For a lot of number crunching applications these days, they're bound by the memory bus, not by the ALU(s). I recently optimized some code that was of the form:
// Assume N is a big number
for (int i=0; i<N; i++) {
myArray[i] = dosomething(i);
}
for (int i=0; i<N; i++) {
myArray[i] = somethingElse(myArray[i]);
}
...
and converted it to look like:
for (int i=0; i<N; i++) {
double tmp = dosomething(i);
tmp = somethingElse(tmp);
...
myArray[i] = tmp;
}
...
In this particular case, this yielded about a 2x speedup.
As Oregonghost already hinted - The VectorC compiler might help. It does not really parallelize the code though, instead you can use it to leverage on extended command sets like mmx or sse. I used it for the most time-critical parts in a software rendering engine and it resulted in a speedup of about 150%-200% on most processors.
For an alternative approach, you could look into Distributed Computing which sounds like it could suit your needs.
If you're sticking with C++ on the intel compiler, take a look at the compiler intrinsics (full reference here). I know that VC++ has similar functionality, and I'm sure you can do the same thing with gcc. These can let you take full advantage of the parallelism built into your CPU. You can use the MMX, SSE and SSE2 instructions to improve performance to a degree. Like others have said, you're probably best looking at the algorithm first.
I suggest you rethink your algorithm, or maybe even better, your approach. On the other hand maybe what you are trying to calculate just takes a lot of computing time. Have you considered to make it distributed so it can run in a cluster of some sort? If you want to focus on pure code optimization by introducing Assembler for your inner loops then often that can be very beneficial (if you know what you're doing).
For modern processors, learning ASM will take you a long time. Further, with all the different versions of SSE around, your code will end up very processor dependant.
I do quite a lot of CPU-bound work, and have found that the difference between intel's C++ compiler and g++ usually isn't that big (at most 15% or so), and there is no measurable difference between Mac OS X, Windows and Linux.
You are going to have to optimise your code and improve your algorithm by hand. There is no "magic fairy dust" which can make existing code that much faster I'm afraid.
If you haven't yet, and you care about performance, you MUST run your code through a good profiler (personally, I like kcachegrind & valgrind on Linux, or Shark on Mac OS X. I don't know what is good for windows I'm afraid).
Based on my past experience, there is a very good chance you'll find some method is taking 95% of your CPU time, and some simple change or addition of caching will make a massive improvement to your performance. On a similar note, if some method is only taking 1% of your CPU time, no amount of optimising is going to gain you anything.
The 2 obvious answers to "CPU-bound" are:
1. Use more CPU (core)s
2. Use something else.
Using 2 threads instead of 1 will cut the time spent by up to 50%. In comparision, C++ to ASM rarely gives you 5% (and for novice ASM programmers, it's often -5%!). Some problems scale well, and may benefit from 8 or 16 cores. That kind of hardware is still pretty mainstream, so see if your problems fall in that category.
The other solution is to throw more specialized hardware at the task. This could be the vector unit of your CPU - considering Windows=x86/x64, that's going to be a flavor of SSE. Another kind of vector hardware is the modern GPU. The GPU also has its own memory bus, which is quite speedy.
First get the lead out. Then if it's as fast as it can possibly be without going to ASM, so be it. But thinking you have to go to ASM assumes you know what's making it slow, and I'll bet a donut that you're guessing.
If you feel you have optimized your code to a point there is no improvement, increase your CPU's. This can be done on different platforms. One I develop with is Appistry. A few links:
http://www.appistry.com/resource-library/index.html
and you can download the product free from here:
http://www.appistry.com/developers/
I work for Appistry and we have done many installations for tasks that were cpu bound by spreading work out over 10's or 100's of machines.
Hope this helps,
-Brett
Probable small help:
Optimization of 64-bit programs
AMD64 (EM64T) architecture
Debugging and optimization of multi-thread OpenMP-programs
Introduction into the problems of developing parallel programs
Development of Resource-intensive Applications in Visual C++
Linux
Switching to Linux can help, if you strip it down to only the parts you actually need.
CrowdProcess has about 2000 workers you can use to compute your algorithm. The API is extremely simple and we've been observing speedups close to the number of workers. Also you can write Javascript which should make you more productive than C++ or ASM.
So if you're in between C++ or ASM, I'd say you should first use all your CPU cores, then if it's not enough, CrowdProcess should be an interesting platform.
Disclaimer: I built CrowdProcess.
It is hard to produce ASM code that is faster than naive C or C++ code. In most cases if you do this job really well, you probably gain not much than few percents and getting like 10% speedup is considered great success but in most cases it is just impossible.
Compilers are capable of understanding how to compile efficiently. You should profile in order to figure out where to optimize.