Shiver
a shiver of sharks, biting and tearing, until all is consumed
A multi-threaded work queue for functions compiled with llvmpy. Give Shiver a function whose last argument is an index (or multiple indices) and an iteration space (i.e. a number of iterations, a tuple of integers, or even slice objects with start/stop/step fields), and shiver does all the messy plumbing of running your code in parallel.
Example:
# we're going to fill this array with the numbers [0...9]
x = np.empty(10, dtype=int)
# compile an LLVM function which takes an array, and an index
fn1 = shiver.from_c("void fn1(long *x, long i) { x[i] = i;}")
# run fn_one_idx in parallel;
# - shiver will supply x's data pointer as a fixed argument to all threads
# - each worker thread will also get a subrange of the indices [0..9]
# - the numpy array 'x' will be passed in as the underlying pointer x.ctypes.data
shiver.parfor(fn1, niters=len(x), fixed_args = [x])
# if the function you compile returns a value,
# then shiver will collect those values into a result array
ident = shiver.from_c("long identity(long i) { return i; }")
y = shiver.parfor(ident, 10)
assert (y==x).all()
# let's do the same thing again, but here we'll explicitly convert the
# fixed array argument to a type LLVM understands
x_gv = GenericValue.pointer(x.ctypes.data)
shiver.parfor(fn1, niters=len(x), fixed_args = [x_gv])
# Now we'll build a function which takes two indices which range
# over all pairs of integers [0..9] and [0..20] and fills x with their products
fn2 = shiver.from_c("float mult(long i, long j) { return (float) i*j; }")
result_grid = shiver.parfor(fn2, (10,20)
assert result_grid.shape == (10,20)
FAQ
Can I use this library to run Python code in parallel?
Sorry, no. Shiver is only useful if you're already compiling LLVM code using llvmpy or if you want to use shiver's from_c helper to compile simple C functions. Shiver takes an LLVM function which constitutes the body of a loop (or a nesting of loops) and runs that code in parallel. It uses the Python threading API to split up your work and saves you from having to deal with pthreads.
You're using Python threads, doesn't that mean you're still stuck behind the GIL?
When Shiver calls into native code it releases the Global Interpreter Lock, allowing its threads to actually utilize all of your processors.
