In my OpenCL kernel I need to use what should normally be a small array of 4 entries, but because of my concerns over how that array would be stored (probably in a much slower kind of memory than regular variables) I'm instead using 4 separate variables and a switch-case statement to access the correct one based on an index.
Is there a way to make a small array of 4 x float4 work as fast and seamlessly as 4 separate float4 variables?
Here's what I'm trying to do: my kernel is meant to generate a single float4 variable v by going through a list of operations to apply to v. It runs sequentially, with operation after operation in the list being applied to v, however in that list there can be sort of brackets/parentheses, which just like in arithmetic isolate a group of operations for them to be done in isolation before the result of that bracket being brought back in with the rest.
So if a bracket is being opened then I should temporarily store the value of v into let's say v0 (to represent the current value at the bracket depth of 0), then v can be reset to 0 and perform the operations inside the bracket, and if there's yet another bracket inside that bracket I'd put v into v1 and so on with v2 and v3 as we go deeper into nested brackets. This is so that I can for instance apply a multiplication inside a bracket that would only affect the other things created inside that bracket and not the rest.
And once a bracket closes I would retrieve e.g. v3 and add v to it, and in the end all brackets would close and v would represent the final desired value of the series of operations and be written to a global buffer. This is doable using switch-case statements to select the correct variable according to the current bracket depth, but this is quite absurd as this is what arrays are for. So I'm not sure what the best thing to do is.
From what I've seen, compilers will usually put small arrays declared in the private address space directly in registers. Of course, this is not a guarantee and there are probably different parameters that intervene in the activation of that optimization, such as:
Array size;
Register pressure;
Cost of spilling;
And others.
As is usual with optimizations, the only way to be sure is to verify what the compiler is doing by checking the generated assembly.
So if a bracket is being opened then I should temporarily store the value of v into let's say v0 (to represent the current value at the bracket depth of 0), then v can be reset to 0 and perform the operations inside the bracket, and if there's yet another bracket inside that bracket I'd put v into v1 and so on with v2 and v3 as we go deeper into nested brackets. This is so that I can for instance apply a multiplication inside a bracket that would only affect the other things created inside that bracket and not the rest.
I don't think that would help. The compiler optimizes across scopes anyway. Just do the straightforward thing and let the optimizer do its job. Then, if you notice suboptimal codegen, you may start thinking about an alternate solution, but not before.
Related
I have been trying to implement some code in Julia JuMP. The idea of my code is that I have a for loop inside my while loop that runs S times. In each of these loops I solve a subproblem and get some variables as well as opt=1 if the subproblem was optimal or opt=0 if it was not optimal. Depending on the value of opt, I have two types of constraints, either optimality cuts (if opt=1) or feasibility cuts (if opt=0). So the intention with my code is that I only add all of the optimality cuts if there are no feasibility cuts for s=1:S (i.e. we get opt=1 in every iteration from 1:S).
What I am looking for is a better way to save the values of ubar, vbar and wbar. Currently I am saving them one at a time with the for-loop, which is quite expensive.
So the problem is that my values of ubar,vbar and wbar are sparse axis arrays. I have tried to save them in other ways like making a 3d sparse axis array, which I could not get to work, since I couldn't figure out how to initialize it.
The below code works (with the correct code inserted inside my <>'s of course), but does not perform as well as I wish. So if there is some way to save the values of 2d sparse axis arrays more efficiently, I would love to know it! Thank you in advance!
ubar2=zeros(nV,nV,S)
vbar2=zeros(nV,nV,S)
wbar2=zeros(nV,nV,S)
while <some condition>
opts=0
for s=1:S
<solve a subproblem, get new ubar,vbar,wbar and opt=1 if optimal or 0 if not>
opts+=opt
if opt==1
# Add opt cut Constraints
for i=1:nV
for k=1:nV
if i!=k
ubar2[i,k,s]=ubar[i,k]
end
end
for j=i:nV
if links[i,j]==1
vbar2[i,j,s]=vbar[i,j]
wbar2[i,j,s]=wbar[i,j]
end
end
end
else
# Add feas cut Constraints
#constraint(mas, <constraint from ubar,vbar,wbar> <= 0)
break
end
if opts==S
for s=1:S
#constraint(mas, <constraint from ubar2,vbar2,wbar2> <= <some variable>)
end
end
end
A SparseAxisArray is simply a thin wrapper in top of a Dict.
It was defined such that when the user creates a container in a JuMP macro, whether he gets an Array, a DenseAxisArray or a SparseAxisArray, it behaves as close as possible to one another hence the user does not need to care about what he obtained for most operations.
For this reason we could not just create a Dict as it behaves differently as an array. For instance you cannot do getindex with multiple indices as x[2, 2].
Here you can use either a Dict or a SparseAxisArray, as you prefer.
Both of them have O(1) complexity for setting and getting new elements and a sparse storage which seems to be adequate for what you need.
If you choose SparseAxisArray, you can initialize it with
ubar2 = JuMP.Containers.SparseAxisArray(Dict{Tuple{Int,Int,Int},Float64}())
and set it with
ubar2[i,k,s]=ubar[i,k]
If you choose Dict, you can initialize it with
ubar2 = Dict{Tuple{Int,Int,Int},Float64}()
and set it with
ubar2[(i,k,s)]=ubar[i,k]
I'm working on a practice program for doing belief propagation stereo vision. The relevant aspect of that here is that I have a fairly long array representing every pixel in an image, and want to carry out an operation on every second entry in the array at each iteration of a for loop - first one half of the entries, and then at the next iteration the other half (this comes from an optimisation described by Felzenswalb & Huttenlocher in their 2006 paper 'Efficient belief propagation for early vision'.) So, you could see it as having an outer for loop which runs a number of times, and for each iteration of that loop I iterate over half of the entries in the array.
I would like to parallelise the operation of iterating over the array like this, since I believe it would be thread-safe to do so, and of course potentially faster. The operation involved updates values inside the data structures representing the neighbouring pixels, which are not themselves used in a given iteration of the outer loop. Originally I just iterated over the entire array in one go, which meant that it was fairly trivial to carry this out - all I needed to do was put .Parallel between Array and .iteri. Changing to operating on every second array entry is trickier, however.
To make the change from simply iterating over every entry, I from Array.iteri (fun i p -> ... to using for i in startIndex..2..(ArrayLength - 1) do, where startIndex is either 1 or 0 depending on which one I used last (controlled by toggling a boolean). This means though that I can't simply use the really nice .Parallel to make things run in parallel.
I haven't been able to find anything specific about how to implement a parallel for loop in .NET which has a step size greater than 1. The best I could find was a paragraph in an old MSDN document on parallel programming in .NET, but that paragraph only makes a vague statement about transforming an index inside a loop body. I do not understand what is meant there.
I looked at Parallel.For and Parallel.ForEach, as well as creating a custom partitioner, but none of those seemed to include options for changing the step size.
The other option that occurred to me was to use a sequence expression such as
let getOddOrEvenArrayEntries myarray oddOrEven =
seq {
let startingIndex =
if oddOrEven then
1
else
0
for i in startingIndex..2..(Array.length myarray- 1) do
yield (i, myarray.[i])
}
and then using PSeq.iteri from ParallelSeq, but I'm not sure whether it will work correctly with .NET Core 2.2. (Note that, currently at least, I need to know the index of the given element in the array, as it is used as the index into another array during the processing).
How can I go about iterating over every second element of an array in parallel? I.e. iterating over an array using a step size greater than 1?
You could try PSeq.mapi which provides not only a sequence item as a parameter but also the index of an item.
Here's a small example
let res = nums
|> PSeq.mapi(fun index item -> if index % 2 = 0 then item else item + 1)
You can also have a look over this sampling snippet. Just be sure to substitute Seq with PSeq
I am parallelizing a certain dynamic programming problem using AVX2/SSE instructions.
In the main iteration of my calculation, I calculate column in matrix where each cell is a structure of AVX2 registers (_m256i). I use values from the previous matrix column as input values for calculating the current column. Columns can be big, so what I do is I have an array of structures (on stack), where each structure has two _m256i elements.
Structure:
struct Cell {
_m256i first;
_m256i second;
};
An then I have array like this: Cell prevColumn [N]. N will tipically be few hundreds.
I know that _m256i basically represents an avx2 register, so I am wondering how should I think about this array, how does it behave, since N is much larger than 16 (which is number of avx registers)? Is it a good practice to create such an array, or is there some better approach that i should use when storing a lot of _m256i values that are going to be reused real soon?
Also, is there any aligning I should be doing with this structures? I read a lot about aligning, but I am still not sure how and when to do it exactly.
It's better to structure your code to do everything it can with a value before moving on. Small buffers that fit in L1 cache aren't going to be too bad for performance, but don't do that unless you need to.
I think it's more typical to write your code with buffers of int [] type, rather than __m256i type, but I'm not sure. Either way works, and should get the compile to generate efficient code. But the int [] way means less code has to be different for the SSE, AVX2, and AVX512 version. And it might make it easier to examine things with a debugger, to have your data in an array with a type that will get the data formatted nicely.
As I understand it, the load/store intrinsics are partly there as a cast between _m256i and int [], since AVX doesn't fault on unaligned, just slows down on cacheline boundaries. Assigning to / from an array of _m256i should work fine, and generate load/store instructions where needed, otherwise generate vector instructions with memory source operands. (for more compact code and fewer fused-domain uops.)
After compacting an array(putting required elements from an input array into an output array) by doing a scan operation, there might be some empty spaces left in the output(compacted) array in a contiguous form after the required elements are placed. Is there a way to free these empty spaces in the OpenCL kernel code itself without going back in the host(just for the sake of deleting)...?
for eg I have an input array of 100 elements with some no.s greater than 50 and some of them less than 50 and want to store the no.s more than 50 in a different array and do further processing only on those elements in that array, and I don't know the size of this output array since I don't know how many no.s are actually greater than 50(so I declare the size of this array to be 100)... then after performing a scan I get the output array with all elements more than 50... but there might be some continuous spaces empty in the output array after the storage of these elements... then how do we delete these spaces... Is there a way of doing this in the kernel code itself...? Or do we have to come back in the Host code for this...?
How do we deal with such compacted arrays to do further processing if we can't delete the remaining spaces in the kernel code itself and also if we don't want to go back in the host code..?
There is no simple solution to your problem I'm afraid.
What I think you might do, is to have a counter of the elements in each array. You can increment the counter first locally with atomic_inc() and then globally with atomic_add().
This way at the end of your kernel execution the total number of elements in each array will be present.
You can also use this atomic operation as an index for the array. This way you can write to the output without any "hole" in your array. However you will probably lose some speed due to abusing of atomic operations I'm afraid.
I have been playing with an implementation of lookandsay (OEIS A005150) in J. I have made two versions, both very simple, using while. type control structures. One recurs, the other loops. Because I am compulsive, I started running comparative timing on the versions.
look and say is the sequence 1 11 21 1211 111221 that s, one one, two ones, etc.
For early elements of the list (up to around 20) the looping version wins, but only by a tiny amount. Timings around 30 cause the recursive version to win, by a large enough amount that the recursive version might be preferred if the stack space were adequate to support it. I looked at why, and I believe that it has to do with handling intermediate results. The 30th number in the sequence has 5808 digits. (32nd number, 9898 digits, 34th, 16774.)
When you are doing the problem with recursion, you can hold the intermediate results in the recursive call, and the unstacking at the end builds the results so that there is minimal handling of the results.
In the list version, you need a variable to hold the result. Every loop iteration causes you to need to add two elements to the result.
The problem, as I see it, is that I can't find any way in J to modify an extant array without completely reassigning it. So I am saying
try. o =. o,e,(0&{y) catch. o =. e,(0&{y) end.
to put an element into o where o might not have a value when we start. That may be notably slower than
o =. i.0
.
.
.
o =. (,o),e,(0&{y)
The point is that the result gets the wrong shape without the ravels, or so it seems. It is inheriting a shape from i.0 somehow.
But even functions like } amend don't modify a list, they return a list that has a modification made to it, and if you want to save the list you need to assign it. As the size of the assigned list increases (as you walk the the number from the beginning to the end making the next number) the assignment seems to take more time and more time. This assignment is really the only thing I can see that would make element 32, 9898 digits, take less time in the recursive version while element 20 (408 digits) takes less time in the loopy version.
The recursive version builds the return with:
e,(0&{y),(,lookandsay e }. y)
The above line is both the return line from the function and the recursion, so the whole return vector gets built at once as the call gets to the end of the string and everything unstacks.
In APL I thought that one could say something on the order of:
a[1+rho a] <- new element
But when I try this in NARS2000 I find that it causes an index error. I don't have access to any other APL, I might be remembering this idiom from APL Plus, I doubt it worked this way in APL\360 or APL\1130. I might be misremembering it completely.
I can find no way to do that in J. It might be that there is no way to do that, but the next thought is to pre-allocate an array that could hold results, and to change individual entries. I see no way to do that either - that is, J does not seem to support the APL idiom:
a<- iota 5
a[3] <- -1
Is this one of those side effect things that is disallowed because of language purity?
Does the interpreter recognize a=. a,foo or some of its variants as a thing that it should fastpath to a[>:#a]=.foo internally?
This is the recursive version, just for the heck of it. I have tried a bunch of different versions and I believe that the longer the program, the slower, and generally, the more complex, the slower. Generally, the program can be chained so that if you want the nth number you can do lookandsay^: n ] y. I have tried a number of optimizations, but the problem I have is that I can't tell what environment I am sending my output into. If I could tell that I was sending it to the next iteration of the program I would send it as an array of digits rather than as a big number.
I also suspect that if I could figure out how to make a tacit version of the code, it would run faster, based on my finding that when I add something to the code that should make it shorter, it runs longer.
lookandsay=: 3 : 0
if. 0 = # ,y do. return. end. NB. return on empty argument
if. 1 ~: ##$ y do. NB. convert rank 0 argument to list of digits
y =. (10&#.^:_1) x: y
f =. 1
assert. 1 = ##$ y NB. the converted argument must be rank 1
else.
NB. yw =. y
f =. 0
end.
NB. e should be a count of the digits that match the leading digit.
e=.+/*./\y=0&{y
if. f do.
o=. e,(0&{y),(,lookandsay e }. y)
assert. e = 0&{ o
10&#. x: o
return.
else.
e,(0&{y),(,lookandsay e }. y)
return.
end.
)
I was interested in the characteristics of the numbers produced. I found that if you start with a 1, the numerals never get higher than 3. If you start with a numeral higher than 3, it will survive as a singleton, and you can also get a number into the generated numbers by starting with something like 888888888 which will generate a number with one 9 in it and a single 8 at the end of the number. But other than the singletons, no digit gets higher than 3.
Edit:
I did some more measuring. I had originally written the program to accept either a vector or a scalar, the idea being that internally I'd work with a vector. I had thought about passing a vector from one layer of code to the other, and I still might using a left argument to control code. With I pass the top level a vector the code runs enormously faster, so my guess is that most of the cpu is being eaten by converting very long numbers from vectors to digits. The recursive routine always passes down a vector when it recurs which might be why it is almost as fast as the loop.
That does not change my question.
I have an answer for this which I can't post for three hours. I will post it then, please don't do a ton of research to answer it.
assignments like
arr=. 'z' 15} arr
are executed in place. (See JWiki article for other supported in-place operations)
Interpreter determines that only small portion of arr is updated and does not create entire new list to reassign.
What happens in your case is not that array is being reassigned, but that it grows many times in small increments, causing memory allocation and reallocation.
If you preallocate (by assigning it some large chunk of data), then you can modify it with } without too much penalty.
After I asked this question, to be honest, I lost track of this web site.
Yes, the answer is that the language has no form that means "update in place, but if you use two forms
x =: x , most anything
or
x =: most anything } x
then the interpreter recognizes those as special and does update in place unless it can't. There are a number of other specials recognized by the interpreter, like:
199(1000&|#^)199
That combined operation is modular exponentiation. It never calculates the whole exponentiation, as
199(1000&|^)199
would - that just ends as _ without the #.
So it is worth reading the article on specials. I will mark someone else's answer up.
The link that sverre provided above ( http://www.jsoftware.com/jwiki/Essays/In-Place%20Operations ) shows the various operations that support modifying an existing array rather than creating a new one. They include:
myarray=: myarray,'blah'
If you are interested in a tacit version of the lookandsay sequence see this submission to RosettaCode:
las=: ,#((# , {.);.1~ 1 , 2 ~:/\ ])&.(10x&#.inv)#]^:(1+i.#[)
5 las 1
11 21 1211 111221 312211