def test_ast_build_expr():
pa = isl.pw_aff("[n] -> { [n + 1] }")
build = isl.ast_build.from_context(pa.domain())
op = build.expr_from(pa)
assert (type(op) == isl.ast_expr_op_add)
assert (op.n_arg() == 2)
def test_ast_build_expr():
pa = isl.pw_aff("[n] -> { [n + 1] }")
build = isl.ast_build.from_context(pa.domain())
op = build.expr_from(pa)
assert(type(op) == isl.ast_expr_op_add)
assert(op.n_arg() == 2)
def test_return_string():
context = isl.set("[n] -> { : }")
build = isl.ast_build.from_context(context)
pw_aff = isl.pw_aff("[n] -> { [n] }")
set = isl.set("[n] -> { : n >= 0 }")
expr = build.expr_from(pw_aff)
expected_string = "n"
assert (expected_string == expr.to_C_str())
expr = build.expr_from(set)
expected_string = "n >= 0"
assert (expected_string == expr.to_C_str())
def test_return_string():
context = isl.set("[n] -> { : }")
build = isl.ast_build.from_context(context)
pw_aff = isl.pw_aff("[n] -> { [n] }")
set = isl.set("[n] -> { : n >= 0 }")
expr = build.expr_from(pw_aff)
expected_string = "n"
assert(expected_string == expr.to_C_str())
expr = build.expr_from(set)
expected_string = "n >= 0"
assert(expected_string == expr.to_C_str())
def cuda_tile(tree, tile_size, permutation=None):
assert tree.parallel_tilable()
box_size, lowers, strides = tree.outermost_band_box()
n = len(box_size)
tile_size = tile_size[:n]
real_tile_size = [tile_size[i] * strides[i] for i in range(n)]
filled_box_size = [
-(-box_size[i] // (real_tile_size[i])) * real_tile_size[i]
for i in range(n)
]
fake_args = ['i%d' % i for i in range(n)]
thread_fake_constraints = [
f'({i} mod {stride}) = (({lower}) mod {stride})'
f' and 0 <= {i} - ({lower}) < {size}' for i, lower, stride, size in
zip(fake_args, lowers, strides, filled_box_size)
]
thread_fake_named_tuple = f'_thread[{", ".join(fake_args)}]'
thread_fake_statement = isl.union_set(
f'{{ {thread_fake_named_tuple} : {" and ".join(thread_fake_constraints)} }}'
).coalesce()
block_fake_constraints = [
f'({i} mod {stride}) = (({lower}) mod {stride})'
f' and 0 <= {i} - ({lower}) < {size}'
f' and ({i} mod {rt_size}) = (({lower}) mod {rt_size})'
for i, lower, stride, size, rt_size in zip(
fake_args, lowers, strides, filled_box_size, real_tile_size)
]
block_fake_named_tuple = f'_block[{", ".join(fake_args)}]'
block_fake_statement = isl.union_set(
f'{{ {block_fake_named_tuple} : {" and ".join(block_fake_constraints)} }}'
).coalesce()
old_domain = tree.domain()
tree.add_to_domain(thread_fake_statement)
tree.add_to_domain(block_fake_statement)
band = tree.outermost_band()
for i in range(n):
s = band.schedule.at(i).union_add(
isl.pw_aff(
f'{{ {thread_fake_named_tuple} -> [({fake_args[i]})] }}'))
band.schedule = band.schedule.set_at(i, s.coalesce())
s = band.schedule.at(i).union_add(
isl.pw_aff(
f'{{ {block_fake_named_tuple} -> [({fake_args[i]})] }}'))
band.schedule = band.schedule.set_at(i, s.coalesce())
fake_branch = SequenceNode()
fake_branch.add_child(FilterNode(filter='{%s}' % thread_fake_named_tuple))
fake_branch.add_child(FilterNode(filter='{%s}' % block_fake_named_tuple))
kernel_branch = FilterNode(filter=old_domain)
if band.child:
kernel_branch.child = band.child
fake_branch.add_child(kernel_branch)
band.child = fake_branch
if permutation is not None:
band.permute(*permutation)
band.tile(*real_tile_size)
band.insert_before(MarkNode('bind=blockIdx'))
child = band.child
child.insert_before(MarkNode('bind=threadIdx'))
kernel = child.child
kernel.insert_before(MarkNode('clear(bind)'))
