def compile(self, sig):
# Use counter to track recursion compilation depth
with self._compiling_counter:
# XXX this is mostly duplicated from Dispatcher.
flags = self.flags
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
# (e.g. if another thread compiled it before we got the lock)
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing.entry_point
self._pre_compile(args, return_type, flags)
# Clone IR to avoid (some of the) mutation in the rewrite pass
cloned_func_ir = self.func_ir.copy()
cres = compiler.compile_ir(typingctx=self.typingctx,
targetctx=self.targetctx,
func_ir=cloned_func_ir,
args=args, return_type=return_type,
flags=flags, locals=self.locals,
lifted=(),
lifted_from=self.lifted_from,
is_lifted_loop=True,)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def _frontend_looplift(self, state):
"""
Loop lifting analysis and transformation
"""
loop_flags = state.flags.copy()
outer_flags = state.flags.copy()
# Do not recursively loop lift
outer_flags.unset('enable_looplift')
loop_flags.unset('enable_looplift')
if not state.flags.enable_pyobject_looplift:
loop_flags.unset('enable_pyobject')
loop_flags.unset('enable_ssa')
main, loops = transforms.loop_lifting(state.func_ir,
typingctx=state.typingctx,
targetctx=state.targetctx,
locals=state.locals,
flags=loop_flags)
if loops:
# Some loops were extracted
if config.DEBUG_FRONTEND or config.DEBUG:
for loop in loops:
print("Lifting loop", loop.get_source_location())
from numba.core.compiler import compile_ir
cres = compile_ir(state.typingctx,
state.targetctx,
main,
state.args,
state.return_type,
outer_flags,
state.locals,
lifted=tuple(loops),
lifted_from=None,
is_lifted_loop=True)
return cres
def compile(self, sig):
with ExitStack() as scope:
cres = None
def cb_compiler(dur):
if cres is not None:
self._callback_add_compiler_timer(dur, cres)
def cb_llvm(dur):
if cres is not None:
self._callback_add_llvm_timer(dur, cres)
scope.enter_context(
ev.install_timer("numba:compiler_lock", cb_compiler))
scope.enter_context(ev.install_timer("numba:llvm_lock", cb_llvm))
scope.enter_context(global_compiler_lock)
# Use counter to track recursion compilation depth
with self._compiling_counter:
# XXX this is mostly duplicated from Dispatcher.
flags = self.flags
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
# (e.g. if another thread compiled it before we got the lock)
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing.entry_point
self._pre_compile(args, return_type, flags)
# Clone IR to avoid (some of the) mutation in the rewrite pass
cloned_func_ir = self.func_ir.copy()
ev_details = dict(
dispatcher=self,
args=args,
return_type=return_type,
)
with ev.trigger_event("numba:compile", data=ev_details):
cres = compiler.compile_ir(
typingctx=self.typingctx,
targetctx=self.targetctx,
func_ir=cloned_func_ir,
args=args,
return_type=return_type,
flags=flags,
locals=self.locals,
lifted=(),
lifted_from=self.lifted_from,
is_lifted_loop=True,
)
# Check typing error if object mode is used
if (cres.typing_error is not None
and not flags.enable_pyobject):
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def compile_ir(self, the_ir, args=(), return_type=None):
typingctx = self.typingctx
targetctx = self.targetctx
flags = self.flags
# Register the contexts in case for nested @jit or @overload calls
with cpu_target.nested_context(typingctx, targetctx):
return compile_ir(typingctx, targetctx, the_ir, args,
return_type, flags, locals={})
def _stencil_wrapper(self, result, sigret, return_type, typemap, calltypes,
*args):
# Overall approach:
# 1) Construct a string containing a function definition for the stencil function
# that will execute the stencil kernel. This function definition includes a
# unique stencil function name, the parameters to the stencil kernel, loop
# nests across the dimensions of the input array. Those loop nests use the
# computed stencil kernel size so as not to try to compute elements where
# elements outside the bounds of the input array would be needed.
# 2) The but of the loop nest in this new function is a special sentinel
# assignment.
# 3) Get the IR of this new function.
# 4) Split the block containing the sentinel assignment and remove the sentinel
# assignment. Insert the stencil kernel IR into the stencil function IR
# after label and variable renaming of the stencil kernel IR to prevent
# conflicts with the stencil function IR.
# 5) Compile the combined stencil function IR + stencil kernel IR into existence.
# Copy the kernel so that our changes for this callsite
# won't effect other callsites.
(kernel_copy,
copy_calltypes) = self.copy_ir_with_calltypes(self.kernel_ir,
calltypes)
# The stencil kernel body becomes the body of a loop, for which args aren't needed.
ir_utils.remove_args(kernel_copy.blocks)
first_arg = kernel_copy.arg_names[0]
in_cps, out_cps = ir_utils.copy_propagate(kernel_copy.blocks, typemap)
name_var_table = ir_utils.get_name_var_table(kernel_copy.blocks)
ir_utils.apply_copy_propagate(kernel_copy.blocks, in_cps,
name_var_table, typemap, copy_calltypes)
if "out" in name_var_table:
raise ValueError(
"Cannot use the reserved word 'out' in stencil kernels.")
sentinel_name = ir_utils.get_unused_var_name("__sentinel__",
name_var_table)
if config.DEBUG_ARRAY_OPT >= 1:
print("name_var_table", name_var_table, sentinel_name)
the_array = args[0]
if config.DEBUG_ARRAY_OPT >= 1:
print("_stencil_wrapper", return_type, return_type.dtype,
type(return_type.dtype), args)
ir_utils.dump_blocks(kernel_copy.blocks)
# We generate a Numba function to execute this stencil and here
# create the unique name of this function.
stencil_func_name = "__numba_stencil_%s_%s" % (hex(
id(the_array)).replace("-", "_"), self.id)
# We will put a loop nest in the generated function for each
# dimension in the input array. Here we create the name for
# the index variable for each dimension. index0, index1, ...
index_vars = []
for i in range(the_array.ndim):
index_var_name = ir_utils.get_unused_var_name(
"index" + str(i), name_var_table)
index_vars += [index_var_name]
# Create extra signature for out and neighborhood.
out_name = ir_utils.get_unused_var_name("out", name_var_table)
neighborhood_name = ir_utils.get_unused_var_name(
"neighborhood", name_var_table)
sig_extra = ""
if result is not None:
sig_extra += ", {}=None".format(out_name)
if "neighborhood" in dict(self.kws):
sig_extra += ", {}=None".format(neighborhood_name)
# Get a list of the standard indexed array names.
standard_indexed = self.options.get("standard_indexing", [])
if first_arg in standard_indexed:
raise ValueError("The first argument to a stencil kernel must "
"use relative indexing, not standard indexing.")
if len(set(standard_indexed) - set(kernel_copy.arg_names)) != 0:
raise ValueError("Standard indexing requested for an array name "
"not present in the stencil kernel definition.")
# Add index variables to getitems in the IR to transition the accesses
# in the kernel from relative to regular Python indexing. Returns the
# computed size of the stencil kernel and a list of the relatively indexed
# arrays.
kernel_size, relatively_indexed = self.add_indices_to_kernel(
kernel_copy, index_vars, the_array.ndim, self.neighborhood,
standard_indexed, typemap, copy_calltypes)
if self.neighborhood is None:
self.neighborhood = kernel_size
if config.DEBUG_ARRAY_OPT >= 1:
print("After add_indices_to_kernel")
ir_utils.dump_blocks(kernel_copy.blocks)
# The return in the stencil kernel becomes a setitem for that
# particular point in the iteration space.
ret_blocks = self.replace_return_with_setitem(kernel_copy.blocks,
index_vars, out_name)
if config.DEBUG_ARRAY_OPT >= 1:
print("After replace_return_with_setitem", ret_blocks)
ir_utils.dump_blocks(kernel_copy.blocks)
# Start to form the new function to execute the stencil kernel.
func_text = "def {}({}{}):\n".format(stencil_func_name,
",".join(kernel_copy.arg_names),
sig_extra)
# Get loop ranges for each dimension, which could be either int
# or variable. In the latter case we'll use the extra neighborhood
# argument to the function.
ranges = []
for i in range(the_array.ndim):
if isinstance(kernel_size[i][0], int):
lo = kernel_size[i][0]
hi = kernel_size[i][1]
else:
lo = "{}[{}][0]".format(neighborhood_name, i)
hi = "{}[{}][1]".format(neighborhood_name, i)
ranges.append((lo, hi))
# If there are more than one relatively indexed arrays, add a call to
# a function that will raise an error if any of the relatively indexed
# arrays are of different size than the first input array.
if len(relatively_indexed) > 1:
func_text += " raise_if_incompatible_array_sizes(" + first_arg
for other_array in relatively_indexed:
if other_array != first_arg:
func_text += "," + other_array
func_text += ")\n"
# Get the shape of the first input array.
shape_name = ir_utils.get_unused_var_name("full_shape", name_var_table)
func_text += " {} = {}.shape\n".format(shape_name, first_arg)
# If we have to allocate the output array (the out argument was not used)
# then us numpy.full if the user specified a cval stencil decorator option
# or np.zeros if they didn't to allocate the array.
if result is None:
return_type_name = numpy_support.as_dtype(
return_type.dtype).type.__name__
if "cval" in self.options:
cval = self.options["cval"]
if return_type.dtype != typing.typeof.typeof(cval):
raise ValueError(
"cval type does not match stencil return type.")
out_init = "{} = np.full({}, {}, dtype=np.{})\n".format(
out_name, shape_name, cval, return_type_name)
else:
out_init = "{} = np.zeros({}, dtype=np.{})\n".format(
out_name, shape_name, return_type_name)
func_text += " " + out_init
else: # result is present, if cval is set then use it
if "cval" in self.options:
cval = self.options["cval"]
cval_ty = typing.typeof.typeof(cval)
if not self._typingctx.can_convert(cval_ty, return_type.dtype):
msg = "cval type does not match stencil return type."
raise ValueError(msg)
out_init = "{}[:] = {}\n".format(out_name, cval)
func_text += " " + out_init
offset = 1
# Add the loop nests to the new function.
for i in range(the_array.ndim):
for j in range(offset):
func_text += " "
# ranges[i][0] is the minimum index used in the i'th dimension
# but minimum's greater than 0 don't preclude any entry in the array.
# So, take the minimum of 0 and the minimum index found in the kernel
# and this will be a negative number (potentially -0). Then, we do
# unary - on that to get the positive offset in this dimension whose
# use is precluded.
# ranges[i][1] is the maximum of 0 and the observed maximum index
# in this dimension because negative maximums would not cause us to
# preclude any entry in the array from being used.
func_text += ("for {} in range(-min(0,{}),"
"{}[{}]-max(0,{})):\n").format(
index_vars[i], ranges[i][0], shape_name, i,
ranges[i][1])
offset += 1
for j in range(offset):
func_text += " "
# Put a sentinel in the code so we can locate it in the IR. We will
# remove this sentinel assignment and replace it with the IR for the
# stencil kernel body.
func_text += "{} = 0\n".format(sentinel_name)
func_text += " return {}\n".format(out_name)
if config.DEBUG_ARRAY_OPT >= 1:
print("new stencil func text")
print(func_text)
# Force the new stencil function into existence.
exec(func_text) in globals(), locals()
stencil_func = eval(stencil_func_name)
if sigret is not None:
pysig = utils.pysignature(stencil_func)
sigret.pysig = pysig
# Get the IR for the newly created stencil function.
from numba.core import compiler
stencil_ir = compiler.run_frontend(stencil_func)
ir_utils.remove_dels(stencil_ir.blocks)
# rename all variables in stencil_ir afresh
var_table = ir_utils.get_name_var_table(stencil_ir.blocks)
new_var_dict = {}
reserved_names = (
[sentinel_name, out_name, neighborhood_name, shape_name] +
kernel_copy.arg_names + index_vars)
for name, var in var_table.items():
if not name in reserved_names:
new_var_dict[name] = ir_utils.mk_unique_var(name)
ir_utils.replace_var_names(stencil_ir.blocks, new_var_dict)
stencil_stub_last_label = max(stencil_ir.blocks.keys()) + 1
# Shift labels in the kernel copy so they are guaranteed unique
# and don't conflict with any labels in the stencil_ir.
kernel_copy.blocks = ir_utils.add_offset_to_labels(
kernel_copy.blocks, stencil_stub_last_label)
new_label = max(kernel_copy.blocks.keys()) + 1
# Adjust ret_blocks to account for addition of the offset.
ret_blocks = [x + stencil_stub_last_label for x in ret_blocks]
if config.DEBUG_ARRAY_OPT >= 1:
print("ret_blocks w/ offsets", ret_blocks, stencil_stub_last_label)
print("before replace sentinel stencil_ir")
ir_utils.dump_blocks(stencil_ir.blocks)
print("before replace sentinel kernel_copy")
ir_utils.dump_blocks(kernel_copy.blocks)
# Search all the block in the stencil outline for the sentinel.
for label, block in stencil_ir.blocks.items():
for i, inst in enumerate(block.body):
if (isinstance(inst, ir.Assign)
and inst.target.name == sentinel_name):
# We found the sentinel assignment.
loc = inst.loc
scope = block.scope
# split block across __sentinel__
# A new block is allocated for the statements prior to the
# sentinel but the new block maintains the current block
# label.
prev_block = ir.Block(scope, loc)
prev_block.body = block.body[:i]
# The current block is used for statements after sentinel.
block.body = block.body[i + 1:]
# But the current block gets a new label.
body_first_label = min(kernel_copy.blocks.keys())
# The previous block jumps to the minimum labelled block of
# the parfor body.
prev_block.append(ir.Jump(body_first_label, loc))
# Add all the parfor loop body blocks to the gufunc
# function's IR.
for (l, b) in kernel_copy.blocks.items():
stencil_ir.blocks[l] = b
stencil_ir.blocks[new_label] = block
stencil_ir.blocks[label] = prev_block
# Add a jump from all the blocks that previously contained
# a return in the stencil kernel to the block
# containing statements after the sentinel.
for ret_block in ret_blocks:
stencil_ir.blocks[ret_block].append(
ir.Jump(new_label, loc))
break
else:
continue
break
stencil_ir.blocks = ir_utils.rename_labels(stencil_ir.blocks)
ir_utils.remove_dels(stencil_ir.blocks)
assert (isinstance(the_array, types.Type))
array_types = args
new_stencil_param_types = list(array_types)
if config.DEBUG_ARRAY_OPT >= 1:
print("new_stencil_param_types", new_stencil_param_types)
ir_utils.dump_blocks(stencil_ir.blocks)
# Compile the combined stencil function with the replaced loop
# body in it.
new_func = compiler.compile_ir(self._typingctx, self._targetctx,
stencil_ir, new_stencil_param_types,
None, compiler.DEFAULT_FLAGS, {})
return new_func
def compile_with_dppy(pyfunc, return_type, args, is_kernel, debug=None):
"""
Compiles with Numba_dppy's pipeline and returns the compiled result.
Args:
pyfunc: The Python function to be compiled.
return_type: The Numba type of the return value.
args: The list of arguments sent to the Python function.
is_kernel (bool): Indicates whether the function is decorated
with @dppy.kernel or not.
debug (bool): Flag to turn debug mode ON/OFF.
Returns:
cres: Compiled result.
Raises:
TypeError: @dppy.kernel does not allow users to return any
value. TypeError is raised when users do.
"""
# First compilation will trigger the initialization of the OpenCL backend.
from .descriptor import dppy_target
typingctx = dppy_target.typing_context
targetctx = dppy_target.target_context
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.debuginfo = config.DEBUGINFO_DEFAULT
flags.no_compile = True
flags.no_cpython_wrapper = True
flags.nrt = False
if debug is not None:
flags.debuginfo = debug
# Run compilation pipeline
if isinstance(pyfunc, FunctionType):
cres = compiler.compile_extra(
typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={},
pipeline_class=DPPYCompiler,
)
elif isinstance(pyfunc, ir.FunctionIR):
cres = compiler.compile_ir(
typingctx=typingctx,
targetctx=targetctx,
func_ir=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={},
pipeline_class=DPPYCompiler,
)
else:
assert 0
if is_kernel:
assert_no_return(cres.signature.return_type)
# Linking depending libraries
library = cres.library
library.finalize()
return cres
