In this course, you will learn to design the computer architecture of complex modern microprocessors. All the features of this course are available for free. It does not offer a certificate upon completion.
Vector Software and Compiler Optimizations
Loading...
来自 Princeton University 的课程
Computer Architecture
300 个评分
从本节课中
Vector Processors and GPUs
This lecture covers the vector processor and optimizations for vector processors.
与讲师见面
David WentzlaffAssociate Professor
Electrical Engineering
So, it's an important aspect of all this vector work of how do you compile for
this. Well,
Thankfully, we actually have compilers that can do automatic vectorization.
And one of the challenges here, if you look at this element wise multiply is, you
have a loop that's running and another loop that's running and your compiler
needs to figure out that it can merge those loops and run them at the same time.
And, compilers actually have gotten pretty sophisticated.
If you look at the, the, the Craig compiler now, it can basically do outer
loop parallelism, it can do certain types of parallelism with loop carry
dependencies and vectorize all this. But it requires some pretty deep compiler
analysis. This especially works well for things like
Fortran codes where you don't have random pointers pointing in different places.
C codes get a little bit hard. So, what if you don't want to execute the
same code in all the elements of your vector?
Well, that could be a problem. So, here we have a piece of code which
loops over some big vector, this is C code. And, it checks to see whether the
value is greater than zero. And only if it's greater than zero does it
do this next operation. So, there's been extensions to vector
processors that have allowed effectively predicates or masked operations on a per
element basis of the, of the vector. So, the way you would do this is you would
actually load the entire vector, set a mask register where you have a one or a
zero which is the result of this comparison on an element to element basis,
And then, do the operation. And you can basically put this together with these bit
by bit comparisons and have slightly different control flow for the different
elements within a vector. And, just sort of, showing the
implementation of this, if we looked at how to actually implement masking, one way
to do it is, you actually do every operation.
So, let's say, you're doing multiply and your vector length is 64.
You do all 64 but you just disable the right to the register file on, the ones
that have the mask bit turned off. Or, you could have a much more fancy
implementation which takes out the work that doesn't have to be done.
But, the control on this is, is quite a bit harder.
And, I would say, that this is probably more common, just the simple
implementation. And the, the,
This is, this is harder largely because, if you have the resources anyway, say if
you have multiple lanes, it might just make sense to go execute a sort of a null
operation later. Some other things that are pretty common
in vectors is you want to have reductions. What I mean by reduction is let's say, you
have this array and you want to add all the elements in the array into a variable.
There's a sort of a vector to scalar operation You can't really do this on what
we discussed so far. You can't do a vector operation which will
actually operate on all of these values and, and try to do something useful with
it. But, what you can do is you can try and do
some software tricks. So, one of the software tricks is, you
take a whole vector, and instead, call it two vectors.
Sort of, cut it in half, and then overlap them and do parallel adds.
And then, you take the results of that. You take, it was someplace else in there,
And you take those two parts and you overlap them, you do adds.
So, you could do lots of parallel adds and effectively build a reduction operation,
by building a tree of adds. So, if we have our vector here, we would
cut it in half and add this part with this part, and then the result would be half
the size. If we cut in half we had this, part of that part, the result is half the
size and cut again, we do, we keep doing adds. So, we can use our vector arithmetic
to effectively do a reduction. So we're about out of time here.
Talk about scatter gather, this isn't that deep.
The implementation of this can be very hard though.
Um,, A of d of i. So, we want to index base off
a index of the vector. This is called gather.
Scatter is the other direction when you're doing store with a double lead-in, a, a,
a, a, index of a index. And, in the instruction set in your book,
they actually have an instruction to do this.
Lvi here, Well, what that basically does is it takes
each element of vector D here, indexes into vector C, and then that is that
result. Problems with this is, of course, your
memory layout is not going to be all nicely laid out in memory.
You're going to be sort of jumping around in memory.
Let's, let's stop here for today, and we'll talk a little bit more about vectors
and GPUs next time.
