How To Multiply Big Numbers Faster?

I was experimenting with multiplying large numbers in python. For my purpose I was trying to evaluate. 2*odd*(4*odd^2 + 7)/3 - 6*odd^2 - 3 Now my question boils down to how to mu

Solution 1:

First of all, 4*odd^2 + 7 is always going to be 2 mod 3. So you don't need to check.

But you don't want to reorder the operations, anyway. Consider what happens to integer arithmetic of 4*5/3 vs 4/3*5. The first result is 6. While the 2nd result is 5. With larger numbers you'd get even more loss of information if you move division to an earlier position in the order of operations.

Another point: you never want to do x**2. ** uses C's pow() so it will have a floating point result. Just use x*x instead. It'll be faster and more accurate.

As for the example of n*(n+1)/2, n*(n+1) will always be divisible by 2. And all you do is shift a large number by 1 bit to the right and you know that the bit you lose is 0.

It seems like you are generally worried about dividing large numbers by small ones. But (as in the 1st example that I gave), if you do the division too early, you'll lose information and get less accurate answers. The calculation won't be faster, either. You might need tricks like this if you divide large numbers by large numbers, but if you are dividing large numbers by small numbers (and you are using integer arithmetic), the result will be only a few bits shorter than the large number you started with. It will neither be faster, nor more accurate, to "divide early" or to use floating-point arithmetic.

Solution 2:

In general you should specialised libraries for specialised jobs. In your case dealing with large numbers and getting optimal speed might require this. Have a look at this this blog which deals with big numbers and makes speed comparisons between pure Python and using the GNU Multiple Precision Arithmetic Library from Python, which results in a big performance improvement.

Solution 3:

The core of your question appears to be a variation of "how fast are numeric operations in python?" and the answer to that depends on what kind of number you're working on, and what kind of machine you are using.

There are 4 kinds of numbers in Python: https://docs.python.org/2/library/stdtypes.html#numeric-types-int-float-long-complex

The question of what kind of machine you are using affects whether or not floating point math will be fast or slow. For the most part, all numeric work will be eclipsed by the slowness of memory, but there are some cases where the number set you are working and instructions will all fit in cache and not be bounded by memory access time.

So, the next question is, since an "int" is really usually a 32-bit C long in python, and a "long" is anything larger than that, are your ints really large enough to cross the 4B barrier into longs? From the sound of it, yes, so let's look at the implementation of the Python longobject: https://github.com/python/cpython/blob/master/Objects/longobject.c

So, it's just a bunch of int mults. Admittedly, it's a linear chain of them, so, how long is that chain?

>>>k = 10**(10**6) + 1>>>k.bit_length()/32
103810

Ok, the int chain is longer than the few hundred cycles each hit to main memory takes, so you might be hitting the point where you are actually CPU bound...

until you look at how big that number actually is.

>>>k.bit_length()/(8*1024)
405

405KB? How big are caches? http://www.intel.com/content/www/us/en/processors/core/6th-gen-core-family-mobile-u-y-processor-lines-datasheet-vol-1.html

So, L1, 64B? L2 is probably 256KB? L3 is the one that's probably around 4M. Even then - how many of these 405KB numbers are being kept around during the process? How much memory does python need to run? Are we thrashing the cache? How expensive is that?

Approximate cost to access various caches and main memory?

This is as far as I'm digging into this hole for now, but signs are pointing to slowness from cache exhaustion. You are using big numbers. Big is more than a math concept, it pushes up against the physical reality of memory. The remaining profiling of "how many temp copies is python keeping around?" and "how is the interpreter using the rest of memory?" are interesting questions, and beyond the scope of this answer.