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 _create_gufunc_for_parfor_body(
lowerer,
parfor,
typemap,
typingctx,
targetctx,
flags,
loop_ranges,
locals,
has_aliases,
index_var_typ,
races,
):
"""
Takes a parfor and creates a gufunc function for its body. There
are two parts to this function:
1) Code to iterate across the iteration space as defined by
the schedule.
2) The parfor body that does the work for a single point in
the iteration space.
Part 1 is created as Python text for simplicity with a sentinel
assignment to mark the point in the IR where the parfor body
should be added. This Python text is 'exec'ed into existence and its
IR retrieved with run_frontend. The IR is scanned for the sentinel
assignment where that basic block is split and the IR for the parfor
body inserted.
"""
loc = parfor.init_block.loc
# The parfor body and the main function body share ir.Var nodes.
# We have to do some replacements of Var names in the parfor body
# to make them legal parameter names. If we don't copy then the
# Vars in the main function also would incorrectly change their name.
loop_body = copy.copy(parfor.loop_body)
remove_dels(loop_body)
parfor_dim = len(parfor.loop_nests)
loop_indices = [l.index_variable.name for l in parfor.loop_nests]
# Get all the parfor params.
parfor_params = parfor.params
for start, stop, step in loop_ranges:
if isinstance(start, ir.Var):
parfor_params.add(start.name)
if isinstance(stop, ir.Var):
parfor_params.add(stop.name)
# Get just the outputs of the parfor.
parfor_outputs = numba.parfors.parfor.get_parfor_outputs(
parfor, parfor_params)
# Get all parfor reduction vars, and operators.
typemap = lowerer.fndesc.typemap
parfor_redvars, parfor_reddict = numba.parfors.parfor.get_parfor_reductions(
lowerer.func_ir, parfor, parfor_params, lowerer.fndesc.calltypes)
has_reduction = False if len(parfor_redvars) == 0 else True
if has_reduction:
_create_gufunc_for_reduction_parfor()
# Compute just the parfor inputs as a set difference.
parfor_inputs = sorted(list(set(parfor_params) - set(parfor_outputs)))
for race in races:
msg = ("Variable %s used in parallel loop may be written "
"to simultaneously by multiple workers and may result "
"in non-deterministic or unintended results." % race)
warnings.warn(NumbaParallelSafetyWarning(msg, loc))
replace_var_with_array(races, loop_body, typemap, lowerer.fndesc.calltypes)
if config.DEBUG_ARRAY_OPT >= 1:
print("parfor_params = ", parfor_params, type(parfor_params))
print("parfor_outputs = ", parfor_outputs, type(parfor_outputs))
print("parfor_inputs = ", parfor_inputs, type(parfor_inputs))
# Reorder all the params so that inputs go first then outputs.
parfor_params = parfor_inputs + parfor_outputs
def addrspace_from(params, def_addr):
addrspaces = []
for p in params:
if isinstance(to_scalar_from_0d(typemap[p]), types.npytypes.Array):
addrspaces.append(def_addr)
else:
addrspaces.append(None)
return addrspaces
addrspaces = addrspace_from(parfor_params, address_space.GLOBAL)
if config.DEBUG_ARRAY_OPT >= 1:
print("parfor_params = ", parfor_params, type(parfor_params))
print("loop_indices = ", loop_indices, type(loop_indices))
print("loop_body = ", loop_body, type(loop_body))
_print_body(loop_body)
# Some Var are not legal parameter names so create a dict of
# potentially illegal param name to guaranteed legal name.
param_dict = legalize_names_with_typemap(parfor_params, typemap)
if config.DEBUG_ARRAY_OPT >= 1:
print("param_dict = ", sorted(param_dict.items()), type(param_dict))
# Some loop_indices are not legal parameter names so create a dict
# of potentially illegal loop index to guaranteed legal name.
ind_dict = legalize_names_with_typemap(loop_indices, typemap)
# Compute a new list of legal loop index names.
legal_loop_indices = [ind_dict[v] for v in loop_indices]
if config.DEBUG_ARRAY_OPT >= 1:
print("ind_dict = ", sorted(ind_dict.items()), type(ind_dict))
print(
"legal_loop_indices = ",
legal_loop_indices,
type(legal_loop_indices),
)
for pd in parfor_params:
print("pd = ", pd)
print("pd type = ", typemap[pd], type(typemap[pd]))
# Get the types of each parameter.
param_types = [to_scalar_from_0d(typemap[v]) for v in parfor_params]
param_types_addrspaces = copy.copy(param_types)
# Calculate types of args passed to gufunc.
func_arg_types = [typemap[v] for v in (parfor_inputs + parfor_outputs)]
assert len(param_types_addrspaces) == len(addrspaces)
for i in range(len(param_types_addrspaces)):
if addrspaces[i] is not None:
# Convert Numba's npytype.Array to DPPYArray data type. DPPYArray
# allows us to specify an address space for the data and other
# pointer arguments for the array.
param_types_addrspaces[i] = npytypes_array_to_dppy_array(
param_types_addrspaces[i], addrspaces[i])
def print_arg_with_addrspaces(args):
for a in args:
print(a, type(a))
if isinstance(a, types.npytypes.Array):
print("addrspace:", a.addrspace)
if config.DEBUG_ARRAY_OPT >= 1:
print_arg_with_addrspaces(param_types)
print("func_arg_types = ", func_arg_types, type(func_arg_types))
# Replace illegal parameter names in the loop body with legal ones.
replace_var_names(loop_body, param_dict)
# remember the name before legalizing as the actual arguments
parfor_args = parfor_params
# Change parfor_params to be legal names.
parfor_params = [param_dict[v] for v in parfor_params]
parfor_params_orig = parfor_params
parfor_params = []
ascontig = False
for pindex in range(len(parfor_params_orig)):
if (ascontig and pindex < len(parfor_inputs)
and isinstance(param_types[pindex], types.npytypes.Array)):
parfor_params.append(parfor_params_orig[pindex] + "param")
else:
parfor_params.append(parfor_params_orig[pindex])
# Change parfor body to replace illegal loop index vars with legal ones.
replace_var_names(loop_body, ind_dict)
loop_body_var_table = get_name_var_table(loop_body)
sentinel_name = get_unused_var_name("__sentinel__", loop_body_var_table)
if config.DEBUG_ARRAY_OPT >= 1:
print("legal parfor_params = ", parfor_params, type(parfor_params))
# Determine the unique names of the scheduling and gufunc functions.
gufunc_name = "__numba_parfor_gufunc_%s" % (parfor.id)
if config.DEBUG_ARRAY_OPT:
# print("sched_func_name ", type(sched_func_name), sched_func_name)
print("gufunc_name ", type(gufunc_name), gufunc_name)
gufunc_txt = ""
# Create the gufunc function.
gufunc_txt += "def " + gufunc_name
gufunc_txt += "(" + (", ".join(parfor_params)) + "):\n"
gufunc_txt += _schedule_loop(parfor_dim, legal_loop_indices, loop_ranges,
param_dict)
# Add the sentinel assignment so that we can find the loop body position
# in the IR.
gufunc_txt += " "
gufunc_txt += sentinel_name + " = 0\n"
# gufunc returns nothing
gufunc_txt += " return None\n"
if config.DEBUG_ARRAY_OPT:
print("gufunc_txt = ", type(gufunc_txt), "\n", gufunc_txt)
sys.stdout.flush()
# Force gufunc outline into existence.
globls = {"np": np, "numba": numba, "dppy": dppy}
locls = {}
exec(gufunc_txt, globls, locls)
gufunc_func = locls[gufunc_name]
if config.DEBUG_ARRAY_OPT:
print("gufunc_func = ", type(gufunc_func), "\n", gufunc_func)
# Get the IR for the gufunc outline.
gufunc_ir = compiler.run_frontend(gufunc_func)
if config.DEBUG_ARRAY_OPT:
print("gufunc_ir dump ", type(gufunc_ir))
gufunc_ir.dump()
print("loop_body dump ", type(loop_body))
_print_body(loop_body)
# rename all variables in gufunc_ir afresh
var_table = get_name_var_table(gufunc_ir.blocks)
new_var_dict = {}
reserved_names = ([sentinel_name] + list(param_dict.values()) +
legal_loop_indices)
for name, var in var_table.items():
if not (name in reserved_names):
new_var_dict[name] = mk_unique_var(name)
replace_var_names(gufunc_ir.blocks, new_var_dict)
if config.DEBUG_ARRAY_OPT:
print("gufunc_ir dump after renaming ")
gufunc_ir.dump()
prs_dict = {}
pss_dict = {}
pspmd_dict = {}
gufunc_param_types = param_types
if config.DEBUG_ARRAY_OPT:
print(
"gufunc_param_types = ",
type(gufunc_param_types),
"\n",
gufunc_param_types,
)
gufunc_stub_last_label = max(gufunc_ir.blocks.keys()) + 1
# Add gufunc stub last label to each parfor.loop_body label to prevent
# label conflicts.
loop_body = add_offset_to_labels(loop_body, gufunc_stub_last_label)
# new label for splitting sentinel block
new_label = max(loop_body.keys()) + 1
# If enabled, add a print statement after every assignment.
if config.DEBUG_ARRAY_OPT_RUNTIME:
_dbgprint_after_each_array_assignments(lowerer, loop_body, typemap)
if config.DEBUG_ARRAY_OPT:
print("parfor loop body")
_print_body(loop_body)
wrapped_blocks = wrap_loop_body(loop_body)
# hoisted, not_hoisted = hoist(parfor_params, loop_body,
# typemap, wrapped_blocks)
setitems = set()
find_setitems_body(setitems, loop_body, typemap)
hoisted = []
not_hoisted = []
start_block = gufunc_ir.blocks[min(gufunc_ir.blocks.keys())]
start_block.body = start_block.body[:-1] + hoisted + [start_block.body[-1]]
unwrap_loop_body(loop_body)
# store hoisted into diagnostics
diagnostics = lowerer.metadata["parfor_diagnostics"]
diagnostics.hoist_info[parfor.id] = {
"hoisted": hoisted,
"not_hoisted": not_hoisted,
}
lowerer.metadata["parfor_diagnostics"].extra_info[str(parfor.id)] = str(
dpctl.get_current_queue().get_sycl_device().name)
if config.DEBUG_ARRAY_OPT:
print("After hoisting")
_print_body(loop_body)
# Search all the block in the gufunc outline for the sentinel assignment.
for label, block in gufunc_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 the sentinel.
block.body = block.body[i + 1:]
# But the current block gets a new label.
body_first_label = min(loop_body.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 loop_body.items():
gufunc_ir.blocks[l] = b
body_last_label = max(loop_body.keys())
gufunc_ir.blocks[new_label] = block
gufunc_ir.blocks[label] = prev_block
# Add a jump from the last parfor body block to the block
# containing statements after the sentinel.
gufunc_ir.blocks[body_last_label].append(
ir.Jump(new_label, loc))
break
else:
continue
break
if config.DEBUG_ARRAY_OPT:
print("gufunc_ir last dump before renaming")
gufunc_ir.dump()
gufunc_ir.blocks = rename_labels(gufunc_ir.blocks)
remove_dels(gufunc_ir.blocks)
if config.DEBUG_ARRAY_OPT:
sys.stdout.flush()
if config.DEBUG_ARRAY_OPT:
print("gufunc_ir last dump")
gufunc_ir.dump()
print("flags", flags)
print("typemap", typemap)
old_alias = flags.noalias
if not has_aliases:
if config.DEBUG_ARRAY_OPT:
print("No aliases found so adding noalias flag.")
flags.noalias = True
remove_dead(gufunc_ir.blocks, gufunc_ir.arg_names, gufunc_ir, typemap)
if config.DEBUG_ARRAY_OPT:
print("gufunc_ir after remove dead")
gufunc_ir.dump()
kernel_sig = signature(types.none, *gufunc_param_types)
if config.DEBUG_ARRAY_OPT:
sys.stdout.flush()
if config.DEBUG_ARRAY_OPT:
print("before DUFunc inlining".center(80, "-"))
gufunc_ir.dump()
# Inlining all DUFuncs
dufunc_inliner(
gufunc_ir,
lowerer.fndesc.calltypes,
typemap,
lowerer.context.typing_context,
lowerer.context,
)
if config.DEBUG_ARRAY_OPT:
print("after DUFunc inline".center(80, "-"))
gufunc_ir.dump()
kernel_func = dppy.compiler.compile_kernel_parfor(
dpctl.get_current_queue(),
gufunc_ir,
gufunc_param_types,
param_types_addrspaces,
debug=flags.debuginfo,
)
flags.noalias = old_alias
if config.DEBUG_ARRAY_OPT:
print("kernel_sig = ", kernel_sig)
return kernel_func, parfor_args, kernel_sig, func_arg_types, setitems