def chunkwise_kernel():
ast, env = compile(source)
#Array = ca.carray(xrange(25000), rootdir='example1', mode='w',
#dtype='int32', cparams=ca.cparams(clevel=0))
Array = open('example1', mode='w')
c = Array.data.ca
ctx = Context(env)
for i in range(c.nchunks):
chunk = c.chunks[i]
# read only access
#x = c.chunks[0][:]
# write access
x = view(chunk)
size = x.strides[0]
args = (x, size)
execute(ctx, args, fname='main')
# this does a _save() behind the scenes
c.chunks[i] = chunk
ctx.destroy()
rts = Runtime(1,2,3)
rts.join()
print Array
def chunked_eval(blz_expr, chunk_size=32768):
operands, code = blz_expr.gen_blir()
total_size = _dimension(operands)
temps = [_temp_for(i, chunk_size) for i in operands]
temp_op = [
i for i in zip(temps, operands) if isinstance(i[1], blaze.Array)
]
offset = 0
accum = 0.0
_, env = blir.compile(code)
ctx = blir.Context(env)
while offset < total_size:
curr_chunk_size = min(total_size - offset, chunk_size)
slice_chunk = slice(0, curr_chunk_size)
slice_src = slice(offset, offset + curr_chunk_size)
for temp, op in temp_op:
temp[slice_chunk] = op[slice_src]
accum += blir.execute(ctx,
args=temps + [curr_chunk_size],
fname='main')
offset = slice_src.stop
ctx.destroy()
return accum
def test_cgen2_add():
with namesupply():
krn = ElementwiseKernel(
[
(IN , VectorArg((300,), 'array[int]')),
(IN , VectorArg((300,), 'array[int]')),
(OUT , VectorArg((300,), 'array[int]')),
],
'_out0[i0] = _in0[i0] + _in1[i0]',
)
krn.verify()
ast, env = krn.compile()
ctx = Context(env)
a = np.array(xrange(300), dtype='int32')
b = np.array(xrange(300), dtype='int32')
c = np.empty_like(b)
execute(ctx, args=(a,b,c), fname='kernel0', timing=False)
assert np.allclose(c, a + b)
def chunked_eval(blz_expr, chunk_size=32768):
operands, code = blz_expr.gen_blir()
total_size = _dimension(operands)
temps = [_temp_for(i, chunk_size) for i in operands]
temp_op = [i for i in zip(temps, operands) if isinstance(i[1], blaze.Array)]
offset = 0
accum = 0.0
_, env = blir.compile(code)
ctx = blir.Context(env)
while offset < total_size:
curr_chunk_size = min(total_size - offset, chunk_size)
slice_chunk = slice(0, curr_chunk_size)
slice_src = slice(offset, offset + curr_chunk_size)
for temp, op in temp_op:
temp[slice_chunk] = op[slice_src]
accum += blir.execute(ctx, args=temps + [curr_chunk_size], fname="main")
offset = slice_src.stop
ctx.destroy()
return accum
for i in range(n) {
for j in range(n) {
x[i,j] = i+j;
}
}
}
"""
N = 15
ast, env = compile(source)
arr = np.eye(N, dtype='int32')
args = (arr, N)
ctx = Context(env)
execute(ctx, args, timing=True)
ctx.destroy()
print arr
#------------------------------------------------------------------------
# Vector Dot Product
#------------------------------------------------------------------------
N = 50000
A = np.arange(N, dtype='double')
B = np.arange(N, dtype='double')
source = open('samples/blir/dot.bl')
ast, env = compile(source.read())
for i in range(n) {
for j in range(n) {
x[i,j] = i+j;
}
}
}
"""
N = 15
ast, env = compile(source)
arr = np.eye(N, dtype='int32')
args = (arr, N)
ctx = Context(env)
execute(ctx, args, timing=True)
ctx.destroy()
print arr
#------------------------------------------------------------------------
# Vector Dot Product
#------------------------------------------------------------------------
N = 50000
A = np.arange(N, dtype='double')
B = np.arange(N, dtype='double')
source = open('samples/blir/dot.bl')
ast, env = compile(source.read())