[Numpy-discussion] Back to numexpr (original) (raw)

Tim Hochberg tim.hochberg at cox.net
Tue Jun 13 15:49:45 EDT 2006

Francesc Altet wrote:

A Dimarts 13 Juny 2006 20:46, Tim Hochberg va escriure:

Uh, I'm afraid that yes. In PyTables, int64, while being a bit bizarre for some users (specially in 32-bit platforms), is a type with the same rights than the others and we would like to give support for it in numexpr. In fact, Ivan Vilata already has implemented this suport in our local copy of numexpr, so perhaps (I say perhaps because we are in the middle of a big project now and are a bit scarce of time resources) we can provide the patch against the latest version of David for your consideration. With this we can solve the problem with int64 support in 32-bit platforms (although addmittedly, the VM gets a bit more complicated, I really think that this is worth the effort)

In addition to complexity, I worry that we'll overflow the code cache at some point and slow everything down. To be honest I have no idea at what point that is likely to happen, but I know they worry about it with the Python interpreter mainloop. That's true. I didn't think about this :-/ Also, it becomes much, much slower to compile past a certain number of case statements under VC7, not sure why. That's mostly my problem though. No, this is a general problem (I'd say much more in GCC, because the optimizer runs so slooooow). However, this should only affect to poor developers, not users and besides, we should find a solution for int64 in 32-bit platforms. Yeah. This is just me whining. Under VC7, there is a very sudden change when adding more cases where compile times go from seconds to minutes. I think we're already past that now anyway, so slowing that down more isn't going to hurt me. Overflowing the cache is the real thing I worry about.

One idea that might be worth trying for int64 is to special case them using functions. That is using OPFUNCLL and OPFUNCLLL and some casting opcodes. This could support int64 with relatively few new opcodes. There's obviously some exta overhead introduced here by the function call. How much this matters is probably a function of how well the compiler / hardware supports int64 to begin with.

Mmm, in my experience int64 operations are reasonable well supported by modern 32-bit processors (IIRC they normally take twice of the time than int32 ops). The problem with using a long for representing ints in numexpr is that we have the duality of being represented differently in 32/64-bit platforms and that could a headache in the long term (int64 support in 32-bit platforms is only one issue, but there should be more). IMHO, it is much better to assign the role for ints in numexpr to a unique datatype, and this should be int64, for the sake of wide int64 support, but also for future (and present!) 64-bit processors. The problem would be that operations with 32-bit ints in 32-bit processors can be slowed-down by a factor 2x (or more, because there is a casting now), but in exchange, whe have full portable code and int64 support. This certainly makes things simpler. I think that this would be fine with me since I mostly use float and complex, so the speed issue wouldn't hit me much. But that's 'cause I'm selfish that way ;-)

In case we consider entering this way, we have two options here: keep VM simple and advertise that int32 arithmetic in numexpr in 32-bit platforms will be sub-optimal, or, as we already have done, add the proper machinery to support both integer separately (at the expense of making the VM more complex). Or perhaps David can come with a better solution (vmgen from gforth? no idea what this is, but the name sounds sexy;-)

Yeah!

That brings up another point. We probably don't want to have casting opcodes from/to everything. Given that there are 8 types on the table now, if we support every casting opcode we're going to have 56(?) opcodes just for casting. I imagine what we'll have to do is write a cast from int16 to float as OPCASTIi; OPCASTFI; trading an extra step in these cases for keeping the number of casting opcodes under control. Once again, int64 is problematic since you lose precision casting to int. I guess in this case you could get by with being able to cast back and forth to float and int. No need to cast directly to booleans, etc as two stage casting should suffice for this.

Well, we already thought about this. Not only you can't safely cast an int64 to an int32 without loosing precistion, but what is worse, you can't even cast it to any other commonly available datatype (casting to a float64 will also loose precision). And, although you can afford loosing precision when dealing with floating data in some scenarios (but not certainly with a general-purpose library like numexpr tries to be), it is by any means unacceptable loosing 'precision' in ints. So, to my mind, the only solution is completely avoiding casting int64 to any type. I forgot that the various OP_CAST_xy opcodes only do safe casting. That makes the number of potential casts much less, so I guess this is not as big a deal as I thought. I'm still not sure, for instance, if we need boolean to int16, int32, int64, float32, float64, complex64 and complex128. It wouldn't kill us, but it's probably overkill.

-tim

More information about the NumPy-Discussion mailing list