def test_uniquify_instruction_ids():
i1 = lp.Assignment("b", 1, id=None)
i2 = lp.Assignment("b", 1, id=None)
i3 = lp.Assignment("b", 1, id=lp.UniqueName("b"))
i4 = lp.Assignment("b", 1, id=lp.UniqueName("b"))
knl = lp.make_kernel("{[i]: i = 1}", []).copy(instructions=[i1, i2, i3, i4])
from loopy.transform.instruction import uniquify_instruction_ids
knl = uniquify_instruction_ids(knl)
insn_ids = {insn.id for insn in knl.instructions}
assert len(insn_ids) == 4
assert all(isinstance(id, str) for id in insn_ids)
def test_uniquify_instruction_ids():
i1 = lp.Assignment("b", 1, id=None)
i2 = lp.Assignment("b", 1, id=None)
i3 = lp.Assignment("b", 1, id=lp.UniqueName("b"))
i4 = lp.Assignment("b", 1, id=lp.UniqueName("b"))
prog = lp.make_kernel("{[i]: i = 1}", [], name="lpy_knl")
new_root_kernel = prog["lpy_knl"].copy(instructions=[i1, i2, i3, i4])
prog = prog.with_kernel(new_root_kernel)
from loopy.transform.instruction import uniquify_instruction_ids
prog = uniquify_instruction_ids(prog)
insn_ids = {insn.id for insn in prog["lpy_knl"].instructions}
assert len(insn_ids) == 4
assert all(isinstance(id, str) for id in insn_ids)
def get_kernel(self):
if self.src_expansion is not self.tgt_expansion:
raise RuntimeError("%s requires that the source "
"and target expansion are the same object"
% type(self).__name__)
ncoeffs = len(self.src_expansion)
# To clarify terminology:
#
# isrc_box -> The index in a list of (in this case, source) boxes
# src_ibox -> The (global) box number for the (in this case, source) box
#
# (same for itgt_box, tgt_ibox)
loopy_insns = [
insn.copy(
predicates=insn.predicates | frozenset(["is_src_box_valid"]),
id=lp.UniqueName("compute_coeff"))
for insn in self.get_translation_loopy_insns()]
from sumpy.tools import gather_loopy_arguments
loopy_knl = lp.make_kernel(
[
"{[itgt_box]: 0<=itgt_box<ntgt_boxes}",
"{[isrc_box]: 0<=isrc_box<nchildren}",
"{[idim]: 0<=idim<dim}",
],
["""
for itgt_box
<> tgt_ibox = target_boxes[itgt_box]
<> tgt_center[idim] = centers[idim, tgt_ibox] \
for isrc_box
<> src_ibox = box_child_ids[isrc_box,tgt_ibox] \
{id=read_src_ibox}
<> is_src_box_valid = src_ibox != 0
if is_src_box_valid
<> src_center[idim] = centers[idim, src_ibox] {dup=idim}
<> d[idim] = tgt_center[idim] - src_center[idim] \
{dup=idim}
"""] + ["""
<> src_coeff{i} = \
src_expansions[src_ibox - src_base_ibox, {i}] \
{{id_prefix=read_coeff,dep=read_src_ibox}}
""".format(i=i) for i in range(ncoeffs)] + [
] + loopy_insns + ["""
tgt_expansions[tgt_ibox - tgt_base_ibox, {i}] = \
tgt_expansions[tgt_ibox - tgt_base_ibox, {i}] \
+ coeff{i} \
{{id_prefix=write_expn,dep=compute_coeff*,
nosync=read_coeff*}}
""".format(i=i) for i in range(ncoeffs)] + ["""
end
end
end
"""],
[
lp.GlobalArg("target_boxes", None, shape=lp.auto,
offset=lp.auto),
lp.GlobalArg("centers", None, shape="dim, aligned_nboxes"),
lp.ValueArg("src_rscale,tgt_rscale", None),
lp.GlobalArg("box_child_ids", None,
shape="nchildren, aligned_nboxes"),
lp.GlobalArg("tgt_expansions", None,
shape=("ntgt_level_boxes", ncoeffs), offset=lp.auto),
lp.GlobalArg("src_expansions", None,
shape=("nsrc_level_boxes", ncoeffs), offset=lp.auto),
lp.ValueArg("src_base_ibox,tgt_base_ibox", np.int32),
lp.ValueArg("ntgt_level_boxes,nsrc_level_boxes", np.int32),
lp.ValueArg("aligned_nboxes", np.int32),
"..."
] + gather_loopy_arguments([self.src_expansion, self.tgt_expansion]),
name=self.name,
assumptions="ntgt_boxes>=1",
silenced_warnings="write_race(write_expn*)",
fixed_parameters=dict(dim=self.dim, nchildren=2**self.dim),
lang_version=MOST_RECENT_LANGUAGE_VERSION)
for expn in [self.src_expansion, self.tgt_expansion]:
loopy_knl = expn.prepare_loopy_kernel(loopy_knl)
loopy_knl = lp.tag_inames(loopy_knl, "idim*:unr")
return loopy_knl
