There are a lot of questions online about allocating, copying, indexing, etc 2d and 3d arrays on CUDA. I'm getting a lot of conflicting answers so I'm attempting to compile past questions to see if I can ask the right ones.
First link: https://devtalk.nvidia.com/default/topic/392370/how-to-cudamalloc-two-dimensional-array-/
Problem: Allocating a 2d array of pointers
User solution: use mallocPitch
"Correct" inefficient solution: Use malloc and memcpy in a for loop for each row (Absurd overhead)
"More correct" solution: Squash it into a 1d array "professional opinion," one comment saying no one with an eye on performance uses 2d pointer structures on the gpu
Second link: https://devtalk.nvidia.com/default/topic/413905/passing-a-multidimensional-array-to-kernel-how-to-allocate-space-in-host-and-pass-to-device-/
Problem: Allocating space on host and passing it to device
Sub link: https://devtalk.nvidia.com/default/topic/398305/cuda-programming-and-performance/dynamically-allocate-array-of-structs/
Sub link solution: Coding pointer based structures on the GPU is a bad experience and highly inefficient, squash it into a 1d array.
Third link: Allocate 2D Array on Device Memory in CUDA
Problem: Allocating and transferring 2d arrays
User solution: use mallocPitch
Other solution: flatten it
Fourth link: How to use 2D Arrays in CUDA?
Problem: Allocate and traverse 2d arrays
Submitted solution: Does not show allocation
Other solution: squash it
There are a lot of other sources mostly saying the same thing but in multiple instances I see warnings about pointer structures on the GPU.
Many people claim the proper way to allocate an array of pointers is with a call to malloc and memcpy for each row yet the functions mallocPitch and memcpy2D exist. Are these functions somehow less efficient? Why wouldn't this be the default answer?
The other 'correct' answer for 2d arrays is to squash them into one array. Should I just get used to this as a fact of life? I'm very persnickety about my code and it feels inelegant to me.
Another solution I was considering was to max a matrix class that uses a 1d pointer array but I can't find a way to implement the double bracket operator.
Also according to this link: Copy an object to device?
and the sub link answer: cudaMemcpy segmentation fault
This gets a little iffy.
The classes I want to use CUDA with all have 2/3d arrays and wouldn't there be a lot of overhead in converting those to 1d arrays for CUDA?
I know I've asked a lot but in summary should I get used to squashed arrays as a fact of life or can I use the 2d allocate and copy functions without getting bad overhead like in the solution where alloc and cpy are called in a for loop?
Since your question compiles a list of other questions, I'll answer by compiling a list of other answers.
cudaMallocPitch/cudaMemcpy2D:
First, the cuda runtime API functions like cudaMallocPitch and cudaMemcpy2D do not actually involve either double-pointer allocations or 2D (doubly-subscripted) arrays. This is easy to confirm simply by looking at the documentation, and noting the types of parameters in the function prototypes. The src and dst parameters are single-pointer parameters. They could not be doubly-subscripted, or doubly dereferenced. For additional example usage, here is one of many questions on this. here is a fully worked example usage. Another example covering various concepts associated with cudaMallocPitch/cudaMemcpy2d usage is here. Instead the correct way to think about these is that they work with pitched allocations. Also, you cannot use cudaMemcpy2D to transfer data when the underlying allocation has been created using a set of malloc (or new, or similar) operations in a loop. That sort of host data allocation construction is particularly ill-suited to working with the data on the device.
general, dynamically allocated 2D case:
If you wish to learn how to use a dynamically allocated 2D array in a CUDA kernel (meaning you can use doubly-subscripted access, e.g. data[x][y]), then the cuda tag info page contains the "canonical" question for this, it is here. The answer given by talonmies there includes the proper mechanics, as well as appropriate caveats:
there is additional, non-trivial complexity
the access will generally be less efficient than 1D access, because data access requires dereferencing 2 pointers, instead of 1.
(note that allocating an array of objects, where the object(s) has an embedded pointer to a dynamic allocation, is essentially the same as the 2D array concept, and the example you linked in your question is a reasonable demonstration for that)
Also, here is a thrust method for building a general dynamically allocated 2D array.
flattening:
If you think you must use the general 2D method, then go ahead, it's not impossible (although sometimes people struggle with the process!) However, due to the added complexity and reduced efficiency, the canonical "advice" here is to "flatten" your storage method, and use "simulated" 2D access. Here is one of many examples of questions/answers discussing "flattening".
general, dynamically allocated 3D case:
As we extend this to 3 (or higher!) dimensions, the general case becomes overly complex to handle, IMO. The additional complexity should strongly motivate us to seek alternatives. The triply-subscripted general case involves 3 pointer accesses before the data is actually retrieved, so even less efficient. Here is a fully worked example (2nd code example).
special case: array width known at compile time:
Note that it should be considered a special case when the array dimension(s) (the width, in the case of a 2D array, or 2 of the 3 dimensions for a 3D array) is known at compile-time. In this case, with an appropriate auxiliary type definition, we can "instruct" the compiler how the indexing should be computed, and in this case we can use doubly-subscripted access with considerably less complexity than the general case, and there is no loss of efficiency due to pointer-chasing. Only one pointer need be dereferenced to retrieve the data (regardless of array dimensionality, if n-1 dimensions are known at compile time for a n-dimensional array). The first code example in the already-mentioned answer here (first code example) gives a fully worked example of that in the 3D case, and the answer here gives a 2D example of this special case.
doubly-subscripted host code, singly-subscripted device code:
Finally another methodology option allows us to easily mix 2D (doubly-subscripted) access in host code while using only 1D (singly-subscripted, perhaps with "simulated 2D" access) in device code. A worked example of that is here. By organizing the underlying allocation as a contiguous allocation, then building the pointer "tree", we can enable doubly-subscripted access on the host, and still easily pass the flat allocation to the device. Although the example does not show it, it would be possible to extend this method to create a doubly-subscripted access system on the device based off a flat allocation and a manually-created pointer "tree", however this would have approximately the same issues as the 2D general dynamically allocated method given above: it would involve double-pointer (double-dereference) access, so less efficient, and there is some complexity associated with building the pointer "tree", for use in device code (e.g. it would necessitate an additional cudaMemcpy operation, probably).
From the above methods, you'll need to choose one that fits your appetite and needs. There is not one single recommendation that fits every possible case.
Related
The information about Ada.Containers.Functional_Maps in the GNAT documentation is quite—let's say—abstruse.
First, it says this:
…these containers can still be used safely.
In the second paragraph, it seems to me that you cannot free the memory allocated for those objects once the program exits the context where they are created. I am understanding that you could run into a memory leak. Am I right?
They are also memory consuming, as the allocated memory is not reclaimed when the container is no longer referenced.
Read the next two sentences in the doc:
Thus, they should in general be used in ghost code and annotations, so that they can be removed from the final executable. The specification of this unit is compatible with SPARK 2014.
Because the specification of Ada.Containers.Functional_Maps is compatible with SPARK, it may help to examine it in the context of related SPARK Libraries with regard to proof, testing and annotation. In particular,
The functional maps, sets and vectors are unbounded collections of indefinite elements that are neither controlled nor limited. While they are inefficient with regard to memory, they are simple, immutable and useful "to model user defined data structures."
The functional containers can be used in Ghost Code, "parts of the code that are only meant for specification and verification", as suggested here. This related example illustrates a ghost function.
it seems to me that you cannot free the memory allocated for those
objects once the program exits the context where they are created. I
am understanding that you could run into a memory leak. Am I right?
There are some things that you can do in Ada to manage memory, I would be surprised if (for example) the usage of an instance inside a declare-block were not cleaned-up on the block's exit. — This is, in fact, how some surprisingly robust applications can get away without "dynamically-allocated" memory/values (it's actually heap-allocated, but that's pedantic).
This sort of granular control is really nice, as you can constrain things/usages to specific points. Combined with Ada's good facilities for presenting interfaces, this means that changing some structure to another can be less-painful than it otherwise might be.
As an example of the above, I had a nested key-value map (a JSON object) that was being used to pass parameters around; the method for doing this changed and so I had a string of values (with common-rooted keys) coming in and a procedure that took JSON as input. Obviously what was needed was a "keys&values-to-JSON function, so inside the function I used the multiway-tree container where the leafs represented values and the internal-nodes the keys, the second step was to traverse the tree and create the JSON-object as needed - simple recursion and data-structure selection used to address the problem of adapting the textual key-value pairs of these nested parameters to JSON. — And because the usage of multi-way trees was exclusive to this function, I can be confident that the memory used by the intermediate tree-object I used is released on the function's exit.
I am looking for an Array-like type with the following properties:
stores elements on disk
elements can have composite type
elements are read into memory, not the whole array
it is possible to write individual elements without writing the whole array
supports setindex!, getindex, push!, pop!, shift!, unshift! and maybe vcat
is reasonably efficient
So far I have found the following leads:
https://docs.julialang.org/en/latest/stdlib/SharedArrays/
http://juliadb.org
https://github.com/JuliaIO/JLD.jl
The first one seems promising, but it seems the type of the elements has to be isbits (meaning a simple number, some structs but not, e.g., an Array{Float64,1}). And it's not clear if the whole array contents are loaded into memory.
If it does not exist yet, I will of course try to construct it myself.
NCDatasets.jl addresses part of the requirements:
stores elements on disk: yes
elements can have composite type: no (although some support for composite type is in NetCDF4, but not yet in NCDatasets.jl). Currently you can have only Arrays of basic types and Arrays of Vectors (of basic types).
elements are read into memory, not the whole array: yes
it is possible to write individual elements without writing the whole array supports setindex!, getindex, push!, pop!, shift!, unshift! and maybe vcat: just setindex!, getindex
is reasonably efficient: the efficency is reasonable for me :-)
The project making it yourself sounds very interesting. I think it would server certainly a gap in the current ecosystem.
Some storage technologies that might be good to have a look at are:
HDF5 (for storage, cross-platform and cross-language)
JLD2 (successor of JLD) https://github.com/simonster/JLD2.jl
rasdaman (a "database" for arrays) http://www.rasdaman.org/
possibly also BSON http://bsonspec.org/
Maybe you can also reach out to the JuliaIO group.
I have a code in which a one-dimensional array R is used which has 3N elements. You can think of it as the position vector of N particles, such that R=[r1x,r1y,r1z,r2x,r2y,...]. Note that the pattern should be defined as this for concise usage of the array.
In sections of the code, I need to perform some operations only on the x-coordinates. I am currently using something like this:
Rx => R(1:3N-2:3)
and Rx is subsequently used in the operations. This makes the access non-contiguous but I was wondering if I can hope for a way to vectorize the operations. Alternatively, one may use OMP with a loop over the particles. I want to get the expert's idea on this matter and particularly the best possible practice performance-wise.
You can't have your cake and eat it too. If you want to make strided access to non-contiguous array elements you're going to pay a price in performance. For small arrays, in which all the elements fit into cache, you'll probably never notice the price. For larger arrays you'll do a lot more data movement through cache than if you step through array elements one-by-one in memory-layout order. Using pointers to non-contiguous array sections doesn't magically alter these facts (as you seem to be aware).
So what you do is what Fortran programmers have always done, optimise the memory layout of your arrays for the most common access pattern. In your case many of us would have either a 3,x rank-2 array or a x,3 one depending on whether accessing all the x (or y or z) elements together was more frequent than accessing particle-by-particle.
Sometimes it's worth transposing an array prior to operations on elements in non-memory-layout order. Sometimes it's even worth holding the same data twice, once in one order, once in the other. But you're going to have to figure out which is the best solution for your program, we don't have all the facts necessary to provide a high-quality recommendation. If it matters to you, then it should matter enough for you to conduct some tests and develop a quantified view of the situation.
You pays your money and you makes your choice.
I'm wondering if there is any perf benchmark on raw objects vs pointers to objects.
I'm aware that it doesn't make sense to use pointers on reference types (e.g. maps) so please don't mention it.
I'm aware that you "must" use pointers if the data needs to be updated so please don't mention it.
Most of the answers/ docs that I've found basically rephrase the guidelines from the official documentation:
... If the receiver is large, a big struct for instance, it will be much cheaper to use a pointer receiver.
My question is simply what means "large" / "big"? Is a pointer on a string overkill ? what about a struct with two strings, what about a struct 3 string fields??
I think we deal with this use case quite often so it's a fair question to ask. Some advise to don't mind the performance issue but maybe some people want to use the right notation whenever they have to chance even if the performance gain is not signifiant. After all a pointer is not that expensive (i.e. one additional keystroke).
An example where it doesn't make sense to use a pointer is for reference types (slices, maps, and channels)
As mentioned in this thread:
The concept of a reference just means something that serves the purpose of referring you to something. It's not magical.
A pointer is a simple reference that tells you where to look.
A slice tells you where to start looking and how far.
Maps and channels also just tell you where to look, but the data they reference and the operations they support on it are more complex.
The point is that all the actually data is stored indirectly and all you're holding is information on how to access it.
As a result, in many cases you don't need to add another layer of indirection, unless you want a double indirection for some reason.
As twotwotwo details in "Pointers vs. values in parameters and return values", strings, interface values, and function values are also implemented with pointers.
As a consequence, you would rarely need a to use a pointer on those objects.
To quote the official golang documentation
...the consideration of efficiency. If the receiver is large, a big struct for instance, it will be much cheaper to use a pointer receiver.
It's very hard to give you exact conditions since there can be different performance goals. As a rule of thumb, by default, all objects larger than 128 bits should be passed by pointer. Possible exceptions of the rule:
you are writing latency sensitive server, so you want to minimise garbage collection pressure. In order to achieve that your Request struct has byte[8] field instead of pointer to Data struct which holds byte[8]. One allocation instead of two.
algorithm you are writing is more readable when you pass the struct and make a copy
etc.
There are lots of real-world reasons you'd want to do this. Ours is because we have a list of variable length data structures, and we want to be able to change the size of one of the elements without recopying them all.
Here's a few things I've tried:
Just have a lot of kernel arguments. Sure, sounds hacky, but works for small N. This is actually what we've been doing.
Do 1) with some sort of macro loop which extends the kernel args to the max size (which I think is device dependent). I don't really want to do this... it sounds bad.
Create some sort of list of structs which contain pointers, and fill it before your kernel invocation. I tried this, and I think it violates the spec. According to what I've seen on the nVidia forums, preserving the address of a device pointer beyond one kernel invocation is illegal. If anyone can point to where in the spec it says this, I'd love to know, because I can't find it. However, this definitely breaks on ATI hardware, as it moves the objects around.
Give up, store the variable sized objects in a big array, and write a clever algorithm to use empty space so the whole array must be reflowed less often. This will work, but is an inelegant, complicated design. Also, it requires lots of scary pointer arithmetic...
Does anyone else have other ideas? What about experiences trying to do this; is there a least hacky way? Why?
To 3:
OpenCL 1.1 spec page 193 says "Arguments to kernel functions in a program cannot be declared as a pointer to a pointer(s)."
Struct containing a pointer to pointer (pointer to a buffer object) might not be against strict reading of this sentence but it's within the spirit: No pointers to buffer objects may be passed as arguments from host code to kernel even if they're hidden inside a user defined struct.
I'd opt for option 5: Do not use variable size data structures. If you have any way of making them constant size by all means do it. It will make your life a whole lot easier. To be precise there is no 'variable size structure'. Every struct definition produces constant sized structs, so if the size has changed then the struct itself has changed and therefore requires another mem object. Every pointer passed to kernel function must have a single type.
In addition to sharpnelis answer option 5:
If the objects have similar size you could use unions on the biggest possible object size. But make sure you use explicit alignment. Pass a second buffer identifying the union used in each object in your variable-sized-objects-in-static-size-union buffer.
I reverted to this when using opencl lib code that only allowed one variable array of arbitrary type. I simply used cl_float2 to pass two floats. Since the cl_floatN types are implemented as unions - what works for the build in types will work for you as well.