def test1(self):
typingctx = typing.Context()
targetctx = cpu.CPUContext(typingctx)
test_ir = compiler.run_frontend(test_will_propagate)
#print("Num blocks = ", len(test_ir.blocks))
#print(test_ir.dump())
with cpu_target.nested_context(typingctx, targetctx):
typingctx.refresh()
targetctx.refresh()
args = (types.int64, types.int64, types.int64)
typemap, return_type, calltypes = compiler.type_inference_stage(typingctx, test_ir, args, None)
#print("typemap = ", typemap)
#print("return_type = ", return_type)
type_annotation = type_annotations.TypeAnnotation(
func_ir=test_ir,
typemap=typemap,
calltypes=calltypes,
lifted=(),
lifted_from=None,
args=args,
return_type=return_type,
html_output=config.HTML)
remove_dels(test_ir.blocks)
in_cps, out_cps = copy_propagate(test_ir.blocks, typemap)
apply_copy_propagate(test_ir.blocks, in_cps, get_name_var_table(test_ir.blocks), typemap, calltypes)
remove_dead(test_ir.blocks, test_ir.arg_names, test_ir)
self.assertFalse(findLhsAssign(test_ir, "x"))
def test1(self):
typingctx = typing.Context()
targetctx = cpu.CPUContext(typingctx)
test_ir = compiler.run_frontend(test_will_propagate)
#print("Num blocks = ", len(test_ir.blocks))
#print(test_ir.dump())
with cpu_target.nested_context(typingctx, targetctx):
typingctx.refresh()
targetctx.refresh()
args = (types.int64, types.int64, types.int64)
typemap, return_type, calltypes = compiler.type_inference_stage(typingctx, test_ir, args, None)
#print("typemap = ", typemap)
#print("return_type = ", return_type)
type_annotation = type_annotations.TypeAnnotation(
func_ir=test_ir,
typemap=typemap,
calltypes=calltypes,
lifted=(),
lifted_from=None,
args=args,
return_type=return_type,
html_output=config.HTML)
remove_dels(test_ir.blocks)
in_cps, out_cps = copy_propagate(test_ir.blocks, typemap)
apply_copy_propagate(test_ir.blocks, in_cps, get_name_var_table(test_ir.blocks), typemap, calltypes)
remove_dead(test_ir.blocks, test_ir.arg_names, test_ir)
self.assertFalse(findLhsAssign(test_ir, "x"))
def run(self):
"""Run inline closure call pass.
"""
modified = False
work_list = list(self.func_ir.blocks.items())
debug_print = _make_debug_print("InlineClosureCallPass")
debug_print("START")
while work_list:
label, block = work_list.pop()
for i, instr in enumerate(block.body):
if isinstance(instr, ir.Assign):
lhs = instr.target
expr = instr.value
if isinstance(expr, ir.Expr) and expr.op == 'call':
call_name = guard(find_callname, self.func_ir, expr)
func_def = guard(get_definition, self.func_ir, expr.func)
if guard(self._inline_reduction,
work_list, block, i, expr, call_name):
modified = True
break # because block structure changed
if guard(self._inline_closure,
work_list, block, i, func_def):
modified = True
break # because block structure changed
if guard(self._inline_stencil,
instr, call_name, func_def):
modified = True
if enable_inline_arraycall:
# Identify loop structure
if modified:
# Need to do some cleanups if closure inlining kicked in
merge_adjacent_blocks(self.func_ir.blocks)
cfg = compute_cfg_from_blocks(self.func_ir.blocks)
debug_print("start inline arraycall")
_debug_dump(cfg)
loops = cfg.loops()
sized_loops = [(k, len(loops[k].body)) for k in loops.keys()]
visited = []
# We go over all loops, bigger loops first (outer first)
for k, s in sorted(sized_loops, key=lambda tup: tup[1], reverse=True):
visited.append(k)
if guard(_inline_arraycall, self.func_ir, cfg, visited, loops[k],
self.parallel_options.comprehension):
modified = True
if modified:
_fix_nested_array(self.func_ir)
if modified:
remove_dels(self.func_ir.blocks)
# repeat dead code elimintation until nothing can be further
# removed
while (remove_dead(self.func_ir.blocks, self.func_ir.arg_names,
self.func_ir)):
pass
self.func_ir.blocks = rename_labels(self.func_ir.blocks)
debug_print("END")
def run(self):
"""Run inline closure call pass.
"""
modified = False
work_list = list(self.func_ir.blocks.items())
_debug_print("START InlineClosureCall")
while work_list:
label, block = work_list.pop()
for i in range(len(block.body)):
instr = block.body[i]
if isinstance(instr, ir.Assign):
lhs = instr.target
expr = instr.value
if isinstance(expr, ir.Expr) and expr.op == 'call':
try:
func_def = self.func_ir.get_definition(expr.func)
except KeyError:
func_def = None
_debug_print("found call to ", expr.func, " def = ", func_def)
if isinstance(func_def, ir.Expr) and func_def.op == "make_function":
new_blocks = self.inline_closure_call(block, i, func_def)
for block in new_blocks:
work_list.append(block)
modified = True
# current block is modified, skip the rest
break
if modified:
remove_dels(self.func_ir.blocks)
# repeat dead code elimintation until nothing can be further removed
while (remove_dead(self.func_ir.blocks, self.func_ir.arg_names)):
pass
self.func_ir.blocks = rename_labels(self.func_ir.blocks)
def run(self):
"""Run inline closure call pass.
"""
modified = False
work_list = list(self.func_ir.blocks.items())
debug_print = _make_debug_print("InlineClosureCallPass")
debug_print("START")
while work_list:
label, block = work_list.pop()
for i, instr in enumerate(block.body):
if isinstance(instr, ir.Assign):
lhs = instr.target
expr = instr.value
if isinstance(expr, ir.Expr) and expr.op == 'call':
call_name = guard(find_callname, self.func_ir, expr)
func_def = guard(get_definition, self.func_ir, expr.func)
if guard(self._inline_reduction,
work_list, block, i, expr, call_name):
modified = True
break # because block structure changed
if guard(self._inline_closure,
work_list, block, i, func_def):
modified = True
break # because block structure changed
if guard(self._inline_stencil,
instr, call_name, func_def):
modified = True
if enable_inline_arraycall:
# Identify loop structure
if modified:
# Need to do some cleanups if closure inlining kicked in
merge_adjacent_blocks(self.func_ir.blocks)
cfg = compute_cfg_from_blocks(self.func_ir.blocks)
debug_print("start inline arraycall")
_debug_dump(cfg)
loops = cfg.loops()
sized_loops = [(k, len(loops[k].body)) for k in loops.keys()]
visited = []
# We go over all loops, bigger loops first (outer first)
for k, s in sorted(sized_loops, key=lambda tup: tup[1], reverse=True):
visited.append(k)
if guard(_inline_arraycall, self.func_ir, cfg, visited, loops[k],
self.swapped, self.parallel_options.comprehension):
modified = True
if modified:
_fix_nested_array(self.func_ir)
if modified:
remove_dels(self.func_ir.blocks)
# repeat dead code elimintation until nothing can be further
# removed
while (remove_dead(self.func_ir.blocks, self.func_ir.arg_names,
self.func_ir)):
pass
self.func_ir.blocks = rename_labels(self.func_ir.blocks)
debug_print("END")
def run(self):
"""Run inline closure call pass.
"""
modified = False
work_list = list(self.func_ir.blocks.items())
debug_print = _make_debug_print("InlineClosureCallPass")
debug_print("START")
while work_list:
label, block = work_list.pop()
for i in range(len(block.body)):
instr = block.body[i]
if isinstance(instr, ir.Assign):
lhs = instr.target
expr = instr.value
if isinstance(expr, ir.Expr) and expr.op == 'call':
func_def = guard(get_definition, self.func_ir, expr.func)
debug_print("found call to ", expr.func, " def = ", func_def)
if isinstance(func_def, ir.Expr) and func_def.op == "make_function":
new_blocks = self.inline_closure_call(block, i, func_def)
for block in new_blocks:
work_list.append(block)
modified = True
# current block is modified, skip the rest
break
if enable_inline_arraycall:
# Identify loop structure
if modified:
# Need to do some cleanups if closure inlining kicked in
merge_adjacent_blocks(self.func_ir)
cfg = compute_cfg_from_blocks(self.func_ir.blocks)
debug_print("start inline arraycall")
_debug_dump(cfg)
loops = cfg.loops()
sized_loops = [(k, len(loops[k].body)) for k in loops.keys()]
visited = []
# We go over all loops, bigger loops first (outer first)
for k, s in sorted(sized_loops, key=lambda tup: tup[1], reverse=True):
visited.append(k)
if guard(_inline_arraycall, self.func_ir, cfg, visited, loops[k],
self.flags.auto_parallel):
modified = True
if modified:
_fix_nested_array(self.func_ir)
if modified:
remove_dels(self.func_ir.blocks)
# repeat dead code elimintation until nothing can be further
# removed
while (remove_dead(self.func_ir.blocks, self.func_ir.arg_names)):
pass
self.func_ir.blocks = rename_labels(self.func_ir.blocks)
debug_print("END")
def run_pass(self, state):
parfor_pass = numba.parfor.ParforPass(
state.func_ir, state.type_annotation.typemap,
state.type_annotation.calltypes, state.return_type,
state.typingctx, state.flags.auto_parallel, state.flags,
state.parfor_diagnostics)
remove_dels(state.func_ir.blocks)
parfor_pass.array_analysis.run(state.func_ir.blocks)
parfor_pass._convert_loop(state.func_ir.blocks)
remove_dead(state.func_ir.blocks, state.func_ir.arg_names,
state.func_ir, state.type_annotation.typemap)
numba.parfor.get_parfor_params(state.func_ir.blocks,
parfor_pass.options.fusion,
parfor_pass.nested_fusion_info)
return True
def get_inner_ir(func):
# get untyped numba ir
f_ir = numba_compiler.run_frontend(func)
blocks = f_ir.blocks
remove_dels(blocks)
topo_order = find_topo_order(blocks)
first_block = blocks[topo_order[0]]
last_block = blocks[topo_order[-1]]
# remove arg nodes
new_first_body = []
for stmt in first_block.body:
if isinstance(stmt, ir.Assign) and isinstance(stmt.value, ir.Arg):
continue
new_first_body.append(stmt)
first_block.body = new_first_body
# rename all variables to avoid conflict, except args
var_table = get_name_var_table(blocks)
new_var_dict = {}
for name, var in var_table.items():
if not (name in f_ir.arg_names):
new_var_dict[name] = mk_unique_var(name)
replace_var_names(blocks, new_var_dict)
return blocks
def _create_gufunc_for_parfor_body(lowerer, parfor, typemap, typingctx,
targetctx, flags, locals):
'''
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.
'''
# TODO: need copy?
# 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)
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
# Get just the outputs of the parfor.
parfor_outputs = numba.parfor.get_parfor_outputs(parfor, parfor_params)
# Get all parfor reduction vars, and operators.
parfor_redvars, parfor_reddict = numba.parfor.get_parfor_reductions(
parfor, parfor_params, lowerer.fndesc.calltypes)
# Compute just the parfor inputs as a set difference.
parfor_inputs = sorted(
list(set(parfor_params) - set(parfor_outputs) - set(parfor_redvars)))
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))
print("parfor_redvars = ", parfor_redvars, " ", type(parfor_redvars))
# Reduction variables are represented as arrays, so they go under
# different names.
parfor_redarrs = []
for var in parfor_redvars:
arr = var + "_arr"
parfor_redarrs.append(arr)
typemap[arr] = types.npytypes.Array(typemap[var], 1, "C")
# Reorder all the params so that inputs go first then outputs.
parfor_params = parfor_inputs + parfor_outputs + parfor_redarrs
if config.DEBUG_ARRAY_OPT == 1:
print("parfor_params = ", parfor_params, " ", type(parfor_params))
#print("loop_ranges = ", loop_ranges, " ", type(loop_ranges))
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(parfor_params + parfor_redvars)
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(loop_indices)
# 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 = [typemap[v] for v in parfor_params]
# if config.DEBUG_ARRAY_OPT==1:
# param_types_dict = { v:typemap[v] for v in parfor_params }
# print("param_types_dict = ", param_types_dict, " ", type(param_types_dict))
# print("param_types = ", param_types, " ", type(param_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]
# Change parfor body to replace illegal loop index vars with legal ones.
replace_var_names(loop_body, ind_dict)
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.
# sched_func_name = "__numba_parfor_sched_%s" % (hex(hash(parfor)).replace("-", "_"))
gufunc_name = "__numba_parfor_gufunc_%s" % (hex(hash(parfor)).replace(
"-", "_"))
if config.DEBUG_ARRAY_OPT:
# print("sched_func_name ", type(sched_func_name), " ", sched_func_name)
print("gufunc_name ", type(gufunc_name), " ", gufunc_name)
# Create the gufunc function.
gufunc_txt = "def " + gufunc_name + \
"(sched, " + (", ".join(parfor_params)) + "):\n"
# Add initialization of reduction variables
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[var] + \
"=" + param_dict[arr] + "[0]\n"
# For each dimension of the parfor, create a for loop in the generated gufunc function.
# Iterate across the proper values extracted from the schedule.
# The form of the schedule is start_dim0, start_dim1, ..., start_dimN, end_dim0,
# end_dim1, ..., end_dimN
for eachdim in range(parfor_dim):
for indent in range(eachdim + 1):
gufunc_txt += " "
sched_dim = eachdim
gufunc_txt += ("for " + legal_loop_indices[eachdim] +
" in range(sched[" + str(sched_dim) + "], sched[" +
str(sched_dim + parfor_dim) + "] + 1):\n")
# Add the sentinel assignment so that we can find the loop body position
# in the IR.
for indent in range(parfor_dim + 1):
gufunc_txt += " "
gufunc_txt += "__sentinel__ = 0\n"
# Add assignments of reduction variables (for returning the value)
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[arr] + \
"[0] = " + param_dict[var] + "\n"
gufunc_txt += " return None\n"
if config.DEBUG_ARRAY_OPT:
print("gufunc_txt = ", type(gufunc_txt), "\n", gufunc_txt)
# Force gufunc outline into existence.
exec(gufunc_txt)
gufunc_func = eval(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__"] + \
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()
gufunc_param_types = [numba.types.npytypes.Array(numba.intp, 1, "C")
] + 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())
# 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 config.DEBUG_ARRAY_OPT:
_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__":
# 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:
print("gufunc_ir last dump")
gufunc_ir.dump()
kernel_func = compiler.compile_ir(typingctx, targetctx, gufunc_ir,
gufunc_param_types, types.none, flags,
locals)
kernel_sig = signature(types.none, *gufunc_param_types)
if config.DEBUG_ARRAY_OPT:
print("kernel_sig = ", kernel_sig)
return kernel_func, parfor_args, kernel_sig
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 dimenions 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)
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
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.
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 lables 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 get_stencil_ir(sf, typingctx, args, scope, loc, input_dict, typemap,
calltypes):
"""get typed IR from stencil bytecode
"""
from numba.targets.cpu import CPUContext
from numba.targets.registry import cpu_target
from numba.annotations import type_annotations
from numba.compiler import type_inference_stage
# get untyped IR
stencil_func_ir = sf.kernel_ir.copy()
# copy the IR nodes to avoid changing IR in the StencilFunc object
stencil_blocks = copy.deepcopy(stencil_func_ir.blocks)
stencil_func_ir.blocks = stencil_blocks
name_var_table = ir_utils.get_name_var_table(stencil_func_ir.blocks)
if "out" in name_var_table:
raise ValueError("Cannot use the reserved word 'out' in stencil kernels.")
# get typed IR with a dummy pipeline (similar to test_parfors.py)
targetctx = CPUContext(typingctx)
with cpu_target.nested_context(typingctx, targetctx):
tp = DummyPipeline(typingctx, targetctx, args, stencil_func_ir)
numba.rewrites.rewrite_registry.apply(
'before-inference', tp, tp.func_ir)
tp.typemap, tp.return_type, tp.calltypes = type_inference_stage(
tp.typingctx, tp.func_ir, tp.args, None)
type_annotations.TypeAnnotation(
func_ir=tp.func_ir,
typemap=tp.typemap,
calltypes=tp.calltypes,
lifted=(),
lifted_from=None,
args=tp.args,
return_type=tp.return_type,
html_output=numba.config.HTML)
# make block labels unique
stencil_blocks = ir_utils.add_offset_to_labels(stencil_blocks,
ir_utils.next_label())
min_label = min(stencil_blocks.keys())
max_label = max(stencil_blocks.keys())
ir_utils._max_label = max_label
if config.DEBUG_ARRAY_OPT == 1:
print("Initial stencil_blocks")
ir_utils.dump_blocks(stencil_blocks)
# rename variables,
var_dict = {}
for v, typ in tp.typemap.items():
new_var = ir.Var(scope, mk_unique_var(v), loc)
var_dict[v] = new_var
typemap[new_var.name] = typ # add new var type for overall function
ir_utils.replace_vars(stencil_blocks, var_dict)
if config.DEBUG_ARRAY_OPT == 1:
print("After replace_vars")
ir_utils.dump_blocks(stencil_blocks)
# add call types to overall function
for call, call_typ in tp.calltypes.items():
calltypes[call] = call_typ
arg_to_arr_dict = {}
# replace arg with arr
for block in stencil_blocks.values():
for stmt in block.body:
if isinstance(stmt, ir.Assign) and isinstance(stmt.value, ir.Arg):
if config.DEBUG_ARRAY_OPT == 1:
print("input_dict", input_dict, stmt.value.index,
stmt.value.name, stmt.value.index in input_dict)
arg_to_arr_dict[stmt.value.name] = input_dict[stmt.value.index].name
stmt.value = input_dict[stmt.value.index]
if config.DEBUG_ARRAY_OPT == 1:
print("arg_to_arr_dict", arg_to_arr_dict)
print("After replace arg with arr")
ir_utils.dump_blocks(stencil_blocks)
ir_utils.remove_dels(stencil_blocks)
stencil_func_ir.blocks = stencil_blocks
return stencil_func_ir, sf.get_return_type(args)[0], arg_to_arr_dict
def _create_gufunc_for_parfor_body(lowerer, parfor, typemap, typingctx,
targetctx, flags, 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
# Get just the outputs of the parfor.
parfor_outputs = numba.parfor.get_parfor_outputs(parfor, parfor_params)
# Get all parfor reduction vars, and operators.
parfor_redvars, parfor_reddict = numba.parfor.get_parfor_reductions(
parfor, parfor_params, lowerer.fndesc.calltypes)
# Compute just the parfor inputs as a set difference.
parfor_inputs = sorted(
list(set(parfor_params) - set(parfor_outputs) - set(parfor_redvars)))
races = races.difference(set(parfor_redvars))
for race in races:
warnings.warn_explicit(
"Variable %s used in parallel loop may be written "
"to simultaneously by multiple workers and may result "
"in non-deterministic or unintended results." % race,
ParallelSafetyWarning, loc.filename, loc.line)
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))
print("parfor_redvars = ", parfor_redvars, " ", type(parfor_redvars))
# Reduction variables are represented as arrays, so they go under
# different names.
parfor_redarrs = []
for var in parfor_redvars:
arr = var + "_arr"
parfor_redarrs.append(arr)
typemap[arr] = types.npytypes.Array(typemap[var], 1, "C")
# Reorder all the params so that inputs go first then outputs.
parfor_params = parfor_inputs + parfor_outputs + parfor_redarrs
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(parfor_params + parfor_redvars)
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(loop_indices)
# 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 = [typemap[v] for v in parfor_params]
# if config.DEBUG_ARRAY_OPT==1:
# param_types_dict = { v:typemap[v] for v in parfor_params }
# print("param_types_dict = ", param_types_dict, " ", type(param_types_dict))
# print("param_types = ", param_types, " ", type(param_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.
# sched_func_name = "__numba_parfor_sched_%s" % (hex(hash(parfor)).replace("-", "_"))
gufunc_name = "__numba_parfor_gufunc_%s" % (hex(hash(parfor)).replace(
"-", "_"))
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 + \
"(sched, " + (", ".join(parfor_params)) + "):\n"
for pindex in range(len(parfor_inputs)):
if ascontig and isinstance(param_types[pindex], types.npytypes.Array):
gufunc_txt += (" " + parfor_params_orig[pindex] +
" = np.ascontiguousarray(" + parfor_params[pindex] +
")\n")
# Add initialization of reduction variables
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[var] + \
"=" + param_dict[arr] + "[0]\n"
# For each dimension of the parfor, create a for loop in the generated gufunc function.
# Iterate across the proper values extracted from the schedule.
# The form of the schedule is start_dim0, start_dim1, ..., start_dimN, end_dim0,
# end_dim1, ..., end_dimN
for eachdim in range(parfor_dim):
for indent in range(eachdim + 1):
gufunc_txt += " "
sched_dim = eachdim
gufunc_txt += ("for " + legal_loop_indices[eachdim] +
" in range(sched[" + str(sched_dim) + "], sched[" +
str(sched_dim + parfor_dim) + "] + np.uint8(1)):\n")
if config.DEBUG_ARRAY_OPT_RUNTIME:
for indent in range(parfor_dim + 1):
gufunc_txt += " "
gufunc_txt += "print("
for eachdim in range(parfor_dim):
gufunc_txt += "\"" + legal_loop_indices[
eachdim] + "\"," + legal_loop_indices[eachdim] + ","
gufunc_txt += ")\n"
# Add the sentinel assignment so that we can find the loop body position
# in the IR.
for indent in range(parfor_dim + 1):
gufunc_txt += " "
gufunc_txt += sentinel_name + " = 0\n"
# Add assignments of reduction variables (for returning the value)
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[arr] + \
"[0] = " + param_dict[var] + "\n"
gufunc_txt += " return None\n"
if config.DEBUG_ARRAY_OPT:
print("gufunc_txt = ", type(gufunc_txt), "\n", gufunc_txt)
# Force gufunc outline into existence.
globls = {"np": np}
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()
gufunc_param_types = [numba.types.npytypes.Array(index_var_typ, 1, "C")
] + 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:
for label, block in loop_body.items():
new_block = block.copy()
new_block.clear()
loc = block.loc
scope = block.scope
for inst in block.body:
new_block.append(inst)
# Append print after assignment
if isinstance(inst, ir.Assign):
# Only apply to numbers
if typemap[inst.target.name] not in types.number_domain:
continue
# Make constant string
strval = "{} =".format(inst.target.name)
strconsttyp = types.Const(strval)
lhs = ir.Var(scope, mk_unique_var("str_const"), loc)
assign_lhs = ir.Assign(value=ir.Const(value=strval,
loc=loc),
target=lhs,
loc=loc)
typemap[lhs.name] = strconsttyp
new_block.append(assign_lhs)
# Make print node
print_node = ir.Print(args=[lhs, inst.target],
vararg=None,
loc=loc)
new_block.append(print_node)
sig = numba.typing.signature(types.none, typemap[lhs.name],
typemap[inst.target.name])
lowerer.fndesc.calltypes[print_node] = sig
loop_body[label] = new_block
if config.DEBUG_ARRAY_OPT:
print("parfor loop body")
_print_body(loop_body)
wrapped_blocks = wrap_loop_body(loop_body)
hoisted = hoist(parfor_params, loop_body, typemap, wrapped_blocks)
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)
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:
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
kernel_func = compiler.compile_ir(typingctx, targetctx, gufunc_ir,
gufunc_param_types, types.none, flags,
locals)
flags.noalias = old_alias
kernel_sig = signature(types.none, *gufunc_param_types)
if config.DEBUG_ARRAY_OPT:
print("kernel_sig = ", kernel_sig)
return kernel_func, parfor_args, kernel_sig
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 dimenions 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)
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:
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.dtype)
else:
out_init ="{} = np.zeros({}, dtype=np.{})\n".format(
out_name, shape_name, return_type.dtype)
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.
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 lables 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 run(self):
"""Run inline closure call pass.
"""
modified = False
work_list = list(self.func_ir.blocks.items())
debug_print = _make_debug_print("InlineClosureCallPass")
debug_print("START")
while work_list:
label, block = work_list.pop()
for i in range(len(block.body)):
instr = block.body[i]
if isinstance(instr, ir.Assign):
lhs = instr.target
expr = instr.value
if isinstance(expr, ir.Expr) and expr.op == 'call':
# inline reduce() when parallel is off
if not self.flags.auto_parallel:
call_name = guard(find_callname, self.func_ir, expr)
if (call_name == ('reduce', 'builtin')
or call_name == ('reduce', 'functools')):
if len(expr.args) != 3:
raise TypeError("invalid reduce call, "
"three arguments including initial "
"value required")
check_reduce_func(self.func_ir, expr.args[0])
def reduce_func(f, A, v):
s = v
it = iter(A)
for a in it:
s = f(s, a)
return s
new_blocks = inline_closure_call(self.func_ir,
self.func_ir.func_id.func.__globals__,
block, i, reduce_func)
for block in new_blocks:
work_list.append(block)
modified = True
# current block is modified, skip the rest
break
func_def = guard(get_definition, self.func_ir, expr.func)
debug_print("found call to ", expr.func, " def = ", func_def)
if isinstance(func_def, ir.Expr) and func_def.op == "make_function":
new_blocks = inline_closure_call(self.func_ir,
self.func_ir.func_id.func.__globals__,
block, i, func_def)
for block in new_blocks:
work_list.append(block)
modified = True
# current block is modified, skip the rest
break
if enable_inline_arraycall:
# Identify loop structure
if modified:
# Need to do some cleanups if closure inlining kicked in
merge_adjacent_blocks(self.func_ir)
cfg = compute_cfg_from_blocks(self.func_ir.blocks)
debug_print("start inline arraycall")
_debug_dump(cfg)
loops = cfg.loops()
sized_loops = [(k, len(loops[k].body)) for k in loops.keys()]
visited = []
# We go over all loops, bigger loops first (outer first)
for k, s in sorted(sized_loops, key=lambda tup: tup[1], reverse=True):
visited.append(k)
if guard(_inline_arraycall, self.func_ir, cfg, visited, loops[k],
self.flags.auto_parallel):
modified = True
if modified:
_fix_nested_array(self.func_ir)
if modified:
remove_dels(self.func_ir.blocks)
# repeat dead code elimintation until nothing can be further
# removed
while (remove_dead(self.func_ir.blocks, self.func_ir.arg_names)):
pass
self.func_ir.blocks = rename_labels(self.func_ir.blocks)
debug_print("END")
def _create_gufunc_for_parfor_body(
lowerer,
parfor,
typemap,
typingctx,
targetctx,
flags,
locals,
has_aliases,
index_var_typ):
'''
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.
'''
# 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
# Get just the outputs of the parfor.
parfor_outputs = numba.parfor.get_parfor_outputs(parfor, parfor_params)
# Get all parfor reduction vars, and operators.
parfor_redvars, parfor_reddict = numba.parfor.get_parfor_reductions(
parfor, parfor_params, lowerer.fndesc.calltypes)
# Compute just the parfor inputs as a set difference.
parfor_inputs = sorted(
list(
set(parfor_params) -
set(parfor_outputs) -
set(parfor_redvars)))
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))
print("parfor_redvars = ", parfor_redvars, " ", type(parfor_redvars))
# Reduction variables are represented as arrays, so they go under
# different names.
parfor_redarrs = []
for var in parfor_redvars:
arr = var + "_arr"
parfor_redarrs.append(arr)
typemap[arr] = types.npytypes.Array(typemap[var], 1, "C")
# Reorder all the params so that inputs go first then outputs.
parfor_params = parfor_inputs + parfor_outputs + parfor_redarrs
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(parfor_params + parfor_redvars)
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(loop_indices)
# 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 = [typemap[v] for v in parfor_params]
# if config.DEBUG_ARRAY_OPT==1:
# param_types_dict = { v:typemap[v] for v in parfor_params }
# print("param_types_dict = ", param_types_dict, " ", type(param_types_dict))
# print("param_types = ", param_types, " ", type(param_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.
# sched_func_name = "__numba_parfor_sched_%s" % (hex(hash(parfor)).replace("-", "_"))
gufunc_name = "__numba_parfor_gufunc_%s" % (
hex(hash(parfor)).replace("-", "_"))
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 + \
"(sched, " + (", ".join(parfor_params)) + "):\n"
for pindex in range(len(parfor_inputs)):
if ascontig and isinstance(param_types[pindex], types.npytypes.Array):
gufunc_txt += (" " + parfor_params_orig[pindex]
+ " = np.ascontiguousarray(" + parfor_params[pindex] + ")\n")
# Add initialization of reduction variables
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[var] + \
"=" + param_dict[arr] + "[0]\n"
# For each dimension of the parfor, create a for loop in the generated gufunc function.
# Iterate across the proper values extracted from the schedule.
# The form of the schedule is start_dim0, start_dim1, ..., start_dimN, end_dim0,
# end_dim1, ..., end_dimN
for eachdim in range(parfor_dim):
for indent in range(eachdim + 1):
gufunc_txt += " "
sched_dim = eachdim
gufunc_txt += ("for " +
legal_loop_indices[eachdim] +
" in range(sched[" +
str(sched_dim) +
"], sched[" +
str(sched_dim +
parfor_dim) +
"] + np.uint8(1)):\n")
if config.DEBUG_ARRAY_OPT_RUNTIME:
for indent in range(parfor_dim + 1):
gufunc_txt += " "
gufunc_txt += "print("
for eachdim in range(parfor_dim):
gufunc_txt += "\"" + legal_loop_indices[eachdim] + "\"," + legal_loop_indices[eachdim] + ","
gufunc_txt += ")\n"
# Add the sentinel assignment so that we can find the loop body position
# in the IR.
for indent in range(parfor_dim + 1):
gufunc_txt += " "
gufunc_txt += sentinel_name + " = 0\n"
# Add assignments of reduction variables (for returning the value)
for arr, var in zip(parfor_redarrs, parfor_redvars):
gufunc_txt += " " + param_dict[arr] + \
"[0] = " + param_dict[var] + "\n"
gufunc_txt += " return None\n"
if config.DEBUG_ARRAY_OPT:
print("gufunc_txt = ", type(gufunc_txt), "\n", gufunc_txt)
# Force gufunc outline into existence.
globls = {"np": np}
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()
gufunc_param_types = [
numba.types.npytypes.Array(
index_var_typ, 1, "C")] + 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:
for label, block in loop_body.items():
new_block = block.copy()
new_block.clear()
loc = block.loc
scope = block.scope
for inst in block.body:
new_block.append(inst)
# Append print after assignment
if isinstance(inst, ir.Assign):
# Only apply to numbers
if typemap[inst.target.name] not in types.number_domain:
continue
# Make constant string
strval = "{} =".format(inst.target.name)
strconsttyp = types.Const(strval)
lhs = ir.Var(scope, mk_unique_var("str_const"), loc)
assign_lhs = ir.Assign(value=ir.Const(value=strval, loc=loc),
target=lhs, loc=loc)
typemap[lhs.name] = strconsttyp
new_block.append(assign_lhs)
# Make print node
print_node = ir.Print(args=[lhs, inst.target], vararg=None, loc=loc)
new_block.append(print_node)
sig = numba.typing.signature(types.none,
typemap[lhs.name],
typemap[inst.target.name])
lowerer.fndesc.calltypes[print_node] = sig
loop_body[label] = new_block
if config.DEBUG_ARRAY_OPT:
print("parfor loop body")
_print_body(loop_body)
wrapped_blocks = wrap_loop_body(loop_body)
hoisted = hoist(parfor_params, loop_body, typemap, wrapped_blocks)
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)
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:
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
kernel_func = compiler.compile_ir(
typingctx,
targetctx,
gufunc_ir,
gufunc_param_types,
types.none,
flags,
locals)
flags.noalias = old_alias
kernel_sig = signature(types.none, *gufunc_param_types)
if config.DEBUG_ARRAY_OPT:
print("kernel_sig = ", kernel_sig)
return kernel_func, parfor_args, kernel_sig