class Flags(utils.ConfigOptions):
# These options are all false by default, but the defaults are
# different with the @jit decorator (see targets.options.TargetOptions).
OPTIONS = {
# Enable loop-lifting
'enable_looplift': False,
# Enable pyobject mode (in general)
'enable_pyobject': False,
# Enable pyobject mode inside lifted loops
'enable_pyobject_looplift': False,
# Enable SSA:
'enable_ssa': True,
# Force pyobject mode inside the whole function
'force_pyobject': False,
# Release GIL inside the native function
'release_gil': False,
'no_compile': False,
'debuginfo': False,
'boundscheck': False,
'forceinline': False,
'no_cpython_wrapper': False,
'no_cfunc_wrapper': False,
# Enable automatic parallel optimization, can be fine-tuned by taking
# a dictionary of sub-options instead of a boolean, see parfor.py for
# detail.
'auto_parallel': cpu.ParallelOptions(False),
'nrt': False,
'no_rewrites': False,
'error_model': 'python',
'fastmath': cpu.FastMathOptions(False),
'noalias': False,
'inline': cpu.InlineOptions('never'),
}
def _run_parfor(cls, test_func, args, swap_map=None):
# TODO: refactor this with get_optimized_numba_ir() where this is
# copied from
typingctx = typing.Context()
targetctx = cpu.CPUContext(typingctx)
test_ir = compiler.run_frontend(test_func)
options = cpu.ParallelOptions(True)
tp = MyPipeline(typingctx, targetctx, args, test_ir)
with cpu_target.nested_context(typingctx, targetctx):
typingctx.refresh()
targetctx.refresh()
inline_pass = inline_closurecall.InlineClosureCallPass(
tp.state.func_ir, options, typed=True
)
inline_pass.run()
rewrites.rewrite_registry.apply("before-inference", tp.state)
untyped_passes.ReconstructSSA().run_pass(tp.state)
(
tp.state.typemap,
tp.state.return_type,
tp.state.calltypes,
_
) = typed_passes.type_inference_stage(
tp.state.typingctx, tp.state.func_ir, tp.state.args, None
)
typed_passes.PreLowerStripPhis().run_pass(tp.state)
diagnostics = numba.parfors.parfor.ParforDiagnostics()
preparfor_pass = numba.parfors.parfor.PreParforPass(
tp.state.func_ir,
tp.state.typemap,
tp.state.calltypes,
tp.state.typingctx,
options,
swapped=diagnostics.replaced_fns,
replace_functions_map=swap_map,
)
preparfor_pass.run()
rewrites.rewrite_registry.apply("after-inference", tp.state)
return tp, options, diagnostics, preparfor_pass
def test_no_copy_usm_shared(capfd):
a = usmarray.ones(10, dtype=np.int64)
b = np.ones(10, dtype=np.int64)
# f = njit(fn)
flags = compiler.Flags()
flags.no_compile = True
flags.no_cpython_wrapper = True
flags.nrt = False
flags.auto_parallel = cpu.ParallelOptions(True)
typingctx = cpu_target.typing_context
targetctx = cpu_target.target_context
args = typingctx.resolve_argument_type(a)
try:
device = dpctl.SyclDevice("opencl:gpu:0")
except ValueError:
pytest.skip("Device not found")
with dppy.offload_to_sycl_device(device):
cres = compiler.compile_extra(
typingctx=typingctx,
targetctx=targetctx,
func=fn,
args=tuple([args]),
return_type=args,
flags=flags,
locals={},
pipeline_class=DPPYCompiler,
)
assert "DPCTLQueue_Memcpy" not in cres.library.get_llvm_str()
args = typingctx.resolve_argument_type(b)
cres = compiler.compile_extra(
typingctx=typingctx,
targetctx=targetctx,
func=fn,
args=tuple([args]),
return_type=args,
flags=flags,
locals={},
pipeline_class=DPPYCompiler,
)
assert "DPCTLQueue_Memcpy" in cres.library.get_llvm_str()
def run_frontend(func, inline_closures=False):
"""
Run the compiler frontend over the given Python function, and return
the function's canonical Numba IR.
If inline_closures is Truthy then closure inlining will be run
"""
# XXX make this a dedicated Pipeline?
func_id = bytecode.FunctionIdentity.from_function(func)
interp = interpreter.Interpreter(func_id)
bc = bytecode.ByteCode(func_id=func_id)
func_ir = interp.interpret(bc)
if inline_closures:
inline_pass = InlineClosureCallPass(func_ir,
cpu.ParallelOptions(False), {},
False)
inline_pass.run()
post_proc = postproc.PostProcessor(func_ir)
post_proc.run()
return func_ir
def compile_parallel(self, func, arg_types):
fast_pflags = Flags()
fast_pflags.set('auto_parallel', cpu.ParallelOptions(True))
fast_pflags.set('nrt')
fast_pflags.set('fastmath', cpu.FastMathOptions(True))
return compile_isolated(func, arg_types, flags=fast_pflags).entry_point
def assert_prune(self, func, args_tys, prune, *args, **kwargs):
# This checks that the expected pruned branches have indeed been pruned.
# func is a python function to assess
# args_tys is the numba types arguments tuple
# prune arg is a list, one entry per branch. The value in the entry is
# encoded as follows:
# True: using constant inference only, the True branch will be pruned
# False: using constant inference only, the False branch will be pruned
# None: under no circumstances should this branch be pruned
# *args: the argument instances to pass to the function to check
# execution is still valid post transform
# **kwargs:
# - flags: compiler.Flags instance to pass to `compile_isolated`,
# permits use of e.g. object mode
func_ir = compile_to_ir(func)
before = func_ir.copy()
if self._DEBUG:
print("=" * 80)
print("before inline")
func_ir.dump()
# run closure inlining to ensure that nonlocals in closures are visible
inline_pass = InlineClosureCallPass(
func_ir,
cpu.ParallelOptions(False),
)
inline_pass.run()
# Remove all Dels, and re-run postproc
post_proc = postproc.PostProcessor(func_ir)
post_proc.run()
rewrite_semantic_constants(func_ir, args_tys)
if self._DEBUG:
print("=" * 80)
print("before prune")
func_ir.dump()
dead_branch_prune(func_ir, args_tys)
after = func_ir
if self._DEBUG:
print("after prune")
func_ir.dump()
before_branches = self.find_branches(before)
self.assertEqual(len(before_branches), len(prune))
# what is expected to be pruned
expect_removed = []
for idx, prune in enumerate(prune):
branch = before_branches[idx]
if prune is True:
expect_removed.append(branch.truebr)
elif prune is False:
expect_removed.append(branch.falsebr)
elif prune is None:
pass # nothing should be removed!
elif prune == 'both':
expect_removed.append(branch.falsebr)
expect_removed.append(branch.truebr)
else:
assert 0, "unreachable"
# compare labels
original_labels = set([_ for _ in before.blocks.keys()])
new_labels = set([_ for _ in after.blocks.keys()])
# assert that the new labels are precisely the original less the
# expected pruned labels
try:
self.assertEqual(new_labels, original_labels - set(expect_removed))
except AssertionError as e:
print("new_labels", sorted(new_labels))
print("original_labels", sorted(original_labels))
print("expect_removed", sorted(expect_removed))
raise e
supplied_flags = kwargs.pop('flags', False)
compiler_kws = {'flags': supplied_flags} if supplied_flags else {}
cres = compile_isolated(func, args_tys, **compiler_kws)
if args is None:
res = cres.entry_point()
expected = func()
else:
res = cres.entry_point(*args)
expected = func(*args)
self.assertEqual(res, expected)
def compile_parallel(self, func, arg_types):
fast_pflags = Flags()
fast_pflags.auto_parallel = cpu.ParallelOptions(True)
fast_pflags.nrt = True
fast_pflags.fastmath = cpu.FastMathOptions(True)
return compile_isolated(func, arg_types, flags=fast_pflags).entry_point
class Flags(TargetConfig):
enable_looplift = Option(
type=bool,
default=False,
doc="Enable loop-lifting",
)
enable_pyobject = Option(
type=bool,
default=False,
doc="Enable pyobject mode (in general)",
)
enable_pyobject_looplift = Option(
type=bool,
default=False,
doc="Enable pyobject mode inside lifted loops",
)
enable_ssa = Option(
type=bool,
default=True,
doc="Enable SSA",
)
force_pyobject = Option(
type=bool,
default=False,
doc="Force pyobject mode inside the whole function",
)
release_gil = Option(
type=bool,
default=False,
doc="Release GIL inside the native function",
)
no_compile = Option(
type=bool,
default=False,
doc="TODO",
)
debuginfo = Option(
type=bool,
default=False,
doc="TODO",
)
boundscheck = Option(
type=bool,
default=False,
doc="TODO",
)
forceinline = Option(
type=bool,
default=False,
doc="Force inlining of the function. Overrides _dbg_optnone.",
)
no_cpython_wrapper = Option(
type=bool,
default=False,
doc="TODO",
)
no_cfunc_wrapper = Option(
type=bool,
default=False,
doc="TODO",
)
auto_parallel = Option(
type=cpu.ParallelOptions,
default=cpu.ParallelOptions(False),
doc="""Enable automatic parallel optimization, can be fine-tuned by
taking a dictionary of sub-options instead of a boolean, see parfor.py for
detail""",
)
nrt = Option(
type=bool,
default=False,
doc="TODO",
)
no_rewrites = Option(
type=bool,
default=False,
doc="TODO",
)
error_model = Option(
type=str,
default="python",
doc="TODO",
)
fastmath = Option(
type=cpu.FastMathOptions,
default=cpu.FastMathOptions(False),
doc="TODO",
)
noalias = Option(
type=bool,
default=False,
doc="TODO",
)
inline = Option(
type=cpu.InlineOptions,
default=cpu.InlineOptions("never"),
doc="TODO",
)
# Defines a new target option for tracking the "target backend".
# This will be the XYZ in @jit(_target=XYZ).
target_backend = Option(
type=str,
default="cpu", # if not set, default to CPU
doc="backend")
dbg_extend_lifetimes = Option(
type=bool,
default=False,
doc=("Extend variable lifetime for debugging. "
"This automatically turns on with debug=True."),
)
dbg_optnone = Option(
type=bool,
default=False,
doc=("Disable optimization for debug. "
"Equivalent to adding optnone attribute in the LLVM Function."))
class Flags(TargetConfig):
enable_looplift = Option(
type=bool,
default=False,
doc="Enable loop-lifting",
)
enable_pyobject = Option(
type=bool,
default=False,
doc="Enable pyobject mode (in general)",
)
enable_pyobject_looplift = Option(
type=bool,
default=False,
doc="Enable pyobject mode inside lifted loops",
)
enable_ssa = Option(
type=bool,
default=True,
doc="Enable SSA",
)
force_pyobject = Option(
type=bool,
default=False,
doc="Force pyobject mode inside the whole function",
)
release_gil = Option(
type=bool,
default=False,
doc="Release GIL inside the native function",
)
no_compile = Option(
type=bool,
default=False,
doc="TODO",
)
debuginfo = Option(
type=bool,
default=False,
doc="TODO",
)
boundscheck = Option(
type=bool,
default=False,
doc="TODO",
)
forceinline = Option(
type=bool,
default=False,
doc="TODO",
)
no_cpython_wrapper = Option(
type=bool,
default=False,
doc="TODO",
)
no_cfunc_wrapper = Option(
type=bool,
default=False,
doc="TODO",
)
auto_parallel = Option(
type=cpu.ParallelOptions,
default=cpu.ParallelOptions(False),
doc="""Enable automatic parallel optimization, can be fine-tuned by
taking a dictionary of sub-options instead of a boolean, see parfor.py for
detail""",
)
nrt = Option(
type=bool,
default=False,
doc="TODO",
)
no_rewrites = Option(
type=bool,
default=False,
doc="TODO",
)
error_model = Option(
type=str,
default="python",
doc="TODO",
)
fastmath = Option(
type=cpu.FastMathOptions,
default=cpu.FastMathOptions(False),
doc="TODO",
)
noalias = Option(
type=bool,
default=False,
doc="TODO",
)
inline = Option(
type=cpu.InlineOptions,
default=cpu.InlineOptions("never"),
doc="TODO",
)