Performance of large llvm::ConstantDataArrays

I’m running into some pretty bad performance in llc.exe when compiling some large neural networks into code that contains some very large llvm::ConstantDataArrays, some are { size=102,760,448 }. There’s a small about of actual code for processing the network, but the assembly is mostly global data.

I’m finding that llc.exe memory spikes up around 30 gigabytes and the job takes 20-30 minutes compiling from bitcode. When I looked into it I found that every single floating point number is loaded into ConstantFP object where the float is parsed into exponent, mantissa and stored in an integer part is stored in a heap allocated array, then these are emitted into MCDataFragments where again more heap allocated data, the float appears to be stored in SmallVectorImpl. On top of this I see a lot of MCFillFragments added because of long double padding.

All up the code I’m compiling ends up with 276 million MCFragments, which just take a super long time in each phase of compiling (loading from bitcode, emitting, layout and writing). With a peak working set of 30 gigabytes each float is taking around 108 bytes!

Is there a more efficient way to do this? Or is there any plan in the works to handle global data more efficiently in llc ?

Have you considered just writing out binary data directly and using i.e.
.incbin for including it?

Joerg

I'm running into some pretty bad performance in llc.exe when compiling
some large neural networks into code that contains some very large llvm::ConstantDataArrays,
some are { size=102,760,448 }. There's a small about of actual code for
processing the network, but the assembly is mostly global data.

I'm finding that llc.exe memory spikes up around 30 gigabytes and the job
takes 20-30 minutes compiling from bitcode. When I looked into it I found
that every single floating point number is loaded into ConstantFP object
where the float is parsed into exponent, mantissa and stored in an integer
part is stored in a heap allocated array, then these are emitted into
MCDataFragments where again more heap allocated data, the float appears to
be stored in SmallVectorImpl<char>. On top of this I see a lot of
MCFillFragments added because of long double padding.

All up the code I'm compiling ends up with 276 million MCFragments, which
just take a super long time in each phase of compiling (loading from
bitcode, emitting, layout and writing). With a peak working set of 30
gigabytes each float is taking around 108 bytes!

Is there a more efficient way to do this? Or is there any plan in the
works to handle global data more efficiently in llc ?

Maybe try putting the blob of floating point numbers in a string / i8 array?

-- Sean Silva

Thanks, I’ve considered that, but I am using LLVM to target multiple platforms, so if I do that I’m worried I need to consider the cross-platform floating point memory layouts … unless LLVM can help me with creating the correct binary blob…

If you don't care about cross-compiling, it is as simple as a scanf &
fwrite loop in a helper program. With cross-compiling, it becomes a lot
more tricky.

Joerg

I’m running into some pretty bad performance in llc.exe when compiling some large neural networks into code that contains some very large llvm::ConstantDataArrays, some are { size=102,760,448 }. There’s a small about of actual code for processing the network, but the assembly is mostly global data.

Yes, llvm’s representation of constant arrays is insanity for cases like this. Your case is bad, but just imagine the cost of a large char initialization: even though each byte is stored as a ConstantInt, the bloat isn’t huge because they are uniqued. The real problem comes from each entry in the ConstantArray being stored as an operand list. An operand in the operand list consumes something like 3-4 words per operand to maintain the uselist and a bunch of other nonsense that isn’t right for this.

IMO, there is a relatively easy solution for this. Introduce a new subclass of ConstantData which represents a blob of data that gets emitted to the .o file, stored in a reasonable native format. I think it would be fine to limit this to only representing arrays of primitive types (e.g. array of float, array of bytes, etc) since this keeps the API to the type simple (the type models an array, so it should have array element members only), and things that want to get the elements of the array out can have them returned as ConstantInt’s (or whatever). I’d name this something like “ConstantArrayBlob”.

There are cases this wouldn’t cover well, e.g. an array of small structs, but I think that is ok, and it could be feature crept to support that over time. The next trick is adding the corresponding special case to Clang to not generate the ConstantArray and the ConstantFP/Int members when given a candidate initialization. This can be done as a secondary optimization after the basic mechanics are in place.

-Chris

I'm running into some pretty bad performance in llc.exe when compiling
some large neural networks into code that contains some very large llvm::ConstantDataArrays,
some are { size=102,760,448 }. There's a small about of actual code for
processing the network, but the assembly is mostly global data.

Yes, llvm’s representation of constant arrays is insanity for cases like
this. Your case is bad, but just imagine the cost of a large char
initialization: even though each byte is stored as a ConstantInt, the bloat
isn’t huge because they are uniqued. The real problem comes from each
entry in the ConstantArray being stored as an operand list. An operand in
the operand list consumes something like 3-4 words per operand to maintain
the uselist and a bunch of other nonsense that isn’t right for this.

IMO, there is a relatively easy solution for this. Introduce a new
subclass of ConstantData which represents a blob of data that gets emitted
to the .o file, stored in a reasonable native format. I think it would be
fine to limit this to only representing arrays of primitive types (e.g.
array of float, array of bytes, etc) since this keeps the API to the type
simple (the type models an array, so it should have array element members
only), and things that want to get the elements of the array out can have
them returned as ConstantInt’s (or whatever). I’d name this something like
“ConstantArrayBlob”.

What's the relationship between ConstantDataArray and ConstantArray?

The former's doxygen says "An array constant whose element type is a simple
1/2/4/8-byte integer or float/double, and whose elements are just simple
data values (i.e. ConstantInt/ConstantFP). This Constant node has no
operands because it stores all of the elements of the constant as densely
packed data, instead of as Value*'s." so I assumed that it was a dense
representation and it seemed reasonable that an i8 typed one of them would
basically operate as a "ConstantArrayBlob". (but I guess if MC still
creates one fragment per element that will still be a memory hog regardless
of the IR's representation)

-- Sean Silva

Ah, it looks like ConstantDataArray is exactly what I was advocating for. Does Clang generate it from an array of doubles? Maybe that is all that is missing.

Yeah, MC should totally be fixed. That’s crazy!

-Chris

Well I have a great test case if someone wants to help show me where/how to fix this in MC.

Well I have a great test case if someone wants to help show me where/how to fix this in MC.

Hi Chris

Looks like its probably almost all there. In AsmPrinter.cpp emitGlobalConstantDataSequential we do:

if (CDS->isString())
return AP.OutStreamer->EmitBytes(CDS->getAsString());

After that, either all the time, or perhaps if the array isn’t a small number of integers/floats, you can emit it as a byte sequence with:

AP.OutStreamer->EmitBinaryData(CDS->getAsString());

Cheers,
Pete

MC should not be creating new fragments when outputting non
instructions, it should just append to the current one.

See MCObjectStreamer::getOrCreateDataFragment, something around it is
not working.

Cheers,
Rafael

Chris Lovett via llvm-dev <llvm-dev@lists.llvm.org> writes:

Ah, another suggestion to make it easier to isolate MC issues: benchmark
running llvm-mc instead of llc.

Cheers,
Rafael

Chris Lovett via llvm-dev <llvm-dev@lists.llvm.org> writes: