import copy import operator import types as pytypes import operator import warnings from dataclasses import make_dataclass import llvmlite.ir import numpy as np import numba from numba.parfors import parfor from numba.core import types, ir, config, compiler, sigutils, cgutils from numba.core.ir_utils import ( add_offset_to_labels, replace_var_names, remove_dels, legalize_names, rename_labels, get_name_var_table, visit_vars_inner, get_definition, guard, get_call_table, is_pure, get_np_ufunc_typ, get_unused_var_name, is_const_call, fixup_var_define_in_scope, transfer_scope, find_max_label, get_global_func_typ, find_topo_order, ) from numba.core.typing import signature from numba.core import lowering from numba.parfors.parfor import ensure_parallel_support from numba.core.errors import ( NumbaParallelSafetyWarning, NotDefinedError, CompilerError, InternalError, ) from numba.parfors.parfor_lowering_utils import ParforLoweringBuilder class ParforLower(lowering.Lower): """This is a custom lowering class that extends standard lowering so as to accommodate parfor.Parfor nodes.""" # custom instruction lowering to handle parfor nodes def lower_inst(self, inst): if isinstance(inst, parfor.Parfor): _lower_parfor_parallel(self, inst) else: super().lower_inst(inst) @property def _disable_sroa_like_opt(self): """ Force disable this because Parfor use-defs is incompatible---it only considers use-defs in blocks that must be executing. See https://github.com/numba/numba/commit/017e2ff9db87fc34149b49dd5367ecbf0bb45268 """ return True def _lower_parfor_parallel(lowerer, parfor): if parfor.lowerer is None: return _lower_parfor_parallel_std(lowerer, parfor) else: return parfor.lowerer(lowerer, parfor) def _lower_parfor_parallel_std(lowerer, parfor): """Lowerer that handles LLVM code generation for parfor. This function lowers a parfor IR node to LLVM. The general approach is as follows: 1) The code from the parfor's init block is lowered normally in the context of the current function. 2) The body of the parfor is transformed into a gufunc function. 3) Code is inserted into the main function that calls do_scheduling to divide the iteration space for each thread, allocates reduction arrays, calls the gufunc function, and then invokes the reduction function across the reduction arrays to produce the final reduction values. """ from numba.np.ufunc.parallel import get_thread_count ensure_parallel_support() typingctx = lowerer.context.typing_context targetctx = lowerer.context builder = lowerer.builder # We copy the typemap here because for race condition variable we'll # update their type to array so they can be updated by the gufunc. orig_typemap = lowerer.fndesc.typemap # replace original typemap with copy and restore the original at the end. lowerer.fndesc.typemap = copy.copy(orig_typemap) if config.DEBUG_ARRAY_OPT: print("lowerer.fndesc", lowerer.fndesc, type(lowerer.fndesc)) typemap = lowerer.fndesc.typemap varmap = lowerer.varmap if config.DEBUG_ARRAY_OPT: print("_lower_parfor_parallel") parfor.dump() loc = parfor.init_block.loc scope = parfor.init_block.scope # produce instructions for init_block if config.DEBUG_ARRAY_OPT: print("init_block = ", parfor.init_block, " ", type(parfor.init_block)) for instr in parfor.init_block.body: if config.DEBUG_ARRAY_OPT: print("lower init_block instr = ", instr) lowerer.lower_inst(instr) for racevar in parfor.races: if racevar not in varmap: rvtyp = typemap[racevar] rv = ir.Var(scope, racevar, loc) lowerer._alloca_var(rv.name, rvtyp) alias_map = {} arg_aliases = {} numba.parfors.parfor.find_potential_aliases_parfor(parfor, parfor.params, typemap, lowerer.func_ir, alias_map, arg_aliases) if config.DEBUG_ARRAY_OPT: print("alias_map", alias_map) print("arg_aliases", arg_aliases) # run get_parfor_outputs() and get_parfor_reductions() before gufunc creation # since Jumps are modified so CFG of loop_body dict will become invalid assert parfor.params is not None parfor_output_arrays = numba.parfors.parfor.get_parfor_outputs( parfor, parfor.params) parfor_redvars, parfor_reddict = parfor.redvars, parfor.reddict if config.DEBUG_ARRAY_OPT: print("parfor_redvars:", parfor_redvars) print("parfor_reddict:", parfor_reddict) # init reduction array allocation here. nredvars = len(parfor_redvars) redarrs = {} to_cleanup = [] if nredvars > 0: # reduction arrays outer dimension equal to thread count scope = parfor.init_block.scope loc = parfor.init_block.loc pfbdr = ParforLoweringBuilder(lowerer=lowerer, scope=scope, loc=loc) # Get the Numba internal function to call to get the thread count. get_num_threads = pfbdr.bind_global_function( fobj=numba.np.ufunc.parallel._iget_num_threads, ftype=get_global_func_typ(numba.np.ufunc.parallel._iget_num_threads), args=() ) # Insert the call to assign the thread count to a variable. num_threads_var = pfbdr.assign( rhs=pfbdr.call(get_num_threads, args=[]), typ=types.intp, name="num_threads_var") # For each reduction variable... for i in range(nredvars): red_name = parfor_redvars[i] # Get the type of the reduction variable. redvar_typ = lowerer.fndesc.typemap[red_name] # Get the ir.Var for the reduction variable. redvar = ir.Var(scope, red_name, loc) # Get the type of the array that holds the per-thread # reduction variables. redarrvar_typ = redtyp_to_redarraytype(redvar_typ) reddtype = redarrvar_typ.dtype if config.DEBUG_ARRAY_OPT: print( "reduction_info", red_name, redvar_typ, redarrvar_typ, reddtype, types.DType(reddtype), num_threads_var, type(num_threads_var) ) # If this is reduction over an array, # the reduction array has just one added per-worker dimension. if isinstance(redvar_typ, types.npytypes.Array): redarrdim = redvar_typ.ndim + 1 else: redarrdim = 1 # Reduction array is created and initialized to the initial reduction value. # First create a var for the numpy empty ufunc. glbl_np_empty = pfbdr.bind_global_function( fobj=np.empty, ftype=get_np_ufunc_typ(np.empty), args=( types.UniTuple(types.intp, redarrdim), ), kws={'dtype': types.DType(reddtype)} ) size_var_list = [num_threads_var] # If this is a reduction over an array... if isinstance(redvar_typ, types.npytypes.Array): # Add code to get the shape of the array being reduced over. redshape_var = pfbdr.assign( rhs=ir.Expr.getattr(redvar, "shape", loc), typ=types.UniTuple(types.intp, redvar_typ.ndim), name="redarr_shape", ) # Add the dimension sizes of the array being reduced over to the tuple of sizes pass to empty. for j in range(redvar_typ.ndim): onedimvar = pfbdr.assign( rhs=ir.Expr.static_getitem(redshape_var, j, None, loc), typ=types.intp, name="redshapeonedim", ) size_var_list.append(onedimvar) # Empty call takes tuple of sizes. Create here and fill in outer dimension (num threads). size_var = pfbdr.make_tuple_variable( size_var_list, name='tuple_size_var', ) # Resolve dtype cval = pfbdr._typingctx.resolve_value_type(reddtype) dt = pfbdr.make_const_variable(cval=cval, typ=types.DType(reddtype)) # Add call to empty passing the size var tuple. empty_call = pfbdr.call(glbl_np_empty, args=[size_var, dt]) redarr_var = pfbdr.assign( rhs=empty_call, typ=redarrvar_typ, name="redarr", ) # Remember mapping of original reduction array to the newly created per-worker reduction array. redarrs[redvar.name] = redarr_var to_cleanup.append(redarr_var) init_val = parfor_reddict[red_name].init_val if init_val is not None: if isinstance(redvar_typ, types.npytypes.Array): # Create an array of identity values for the reduction. # First, create a variable for np.full. full_func_node = pfbdr.bind_global_function( fobj=np.full, ftype=get_np_ufunc_typ(np.full), args=( types.UniTuple(types.intp, redvar_typ.ndim), reddtype, ), kws={'dtype': types.DType(reddtype)}, ) # Then create a var with the identify value. init_val_var = pfbdr.make_const_variable( cval=init_val, typ=reddtype, name="init_val", ) # Then, call np.full with the shape of the reduction array and the identity value. full_call = pfbdr.call( full_func_node, args=[redshape_var, init_val_var, dt], ) redtoset = pfbdr.assign( rhs=full_call, typ=redvar_typ, name="redtoset", ) # rettoset is an array from np.full() and must be released to_cleanup.append(redtoset) else: redtoset = pfbdr.make_const_variable( cval=init_val, typ=reddtype, name="redtoset", ) else: redtoset = redvar if config.DEBUG_ARRAY_OPT_RUNTIME: res_print_str = "res_print1 for redvar " + str(redvar) + ":" strconsttyp = types.StringLiteral(res_print_str) lhs = pfbdr.make_const_variable( cval=res_print_str, typ=strconsttyp, name="str_const", ) res_print = ir.Print(args=[lhs, redvar], vararg=None, loc=loc) lowerer.fndesc.calltypes[res_print] = signature(types.none, typemap[lhs.name], typemap[redvar.name]) print("res_print_redvar", res_print) lowerer.lower_inst(res_print) # For each thread, initialize the per-worker reduction array to # the current reduction array value. # Get the Numba type of the variable that holds the thread count. num_thread_type = typemap[num_threads_var.name] # Get the LLVM type of the thread count variable. ntllvm_type = targetctx.get_value_type(num_thread_type) # Create a LLVM variable to hold the loop index. alloc_loop_var = cgutils.alloca_once(builder, ntllvm_type) # Associate this LLVM variable to a Numba IR variable so that # we can use setitem IR builder. # Create a Numba IR variable. numba_ir_loop_index_var = scope.redefine("$loop_index", loc) # Give that variable the right type. typemap[numba_ir_loop_index_var.name] = num_thread_type # Associate this Numba variable to the LLVM variable in the # lowerer's varmap. lowerer.varmap[numba_ir_loop_index_var.name] = alloc_loop_var # Insert a loop into the outputed LLVM that goes from 0 to # the current thread count. with cgutils.for_range(builder, lowerer.loadvar(num_threads_var.name), intp=ntllvm_type) as loop: # Store the loop index into the alloca'd LLVM loop index variable. builder.store(loop.index, alloc_loop_var) # Initialize one element of the reduction array using the Numba # IR variable associated with this loop's index. pfbdr.setitem(obj=redarr_var, index=numba_ir_loop_index_var, val=redtoset) # compile parfor body as a separate function to be used with GUFuncWrapper flags = parfor.flags.copy() flags.error_model = "numpy" # Can't get here unless flags.auto_parallel == ParallelOptions(True) index_var_typ = typemap[parfor.loop_nests[0].index_variable.name] # index variables should have the same type, check rest of indices for l in parfor.loop_nests[1:]: assert typemap[l.index_variable.name] == index_var_typ numba.parfors.parfor.sequential_parfor_lowering = True try: (func, func_args, func_sig, func_arg_types, exp_name_to_tuple_var) = _create_gufunc_for_parfor_body( lowerer, parfor, typemap, typingctx, targetctx, flags, {}, bool(alias_map), index_var_typ, parfor.races) finally: numba.parfors.parfor.sequential_parfor_lowering = False # get the shape signature func_args = ['sched'] + func_args num_reductions = len(parfor_redvars) num_inputs = len(func_args) - len(parfor_output_arrays) - num_reductions if config.DEBUG_ARRAY_OPT: print("func_args = ", func_args) print("num_inputs = ", num_inputs) print("parfor_outputs = ", parfor_output_arrays) print("parfor_redvars = ", parfor_redvars) print("num_reductions = ", num_reductions) gu_signature = _create_shape_signature( parfor.get_shape_classes, num_inputs, num_reductions, func_args, func_sig, parfor.races, typemap) if config.DEBUG_ARRAY_OPT: print("gu_signature = ", gu_signature) # call the func in parallel by wrapping it with ParallelGUFuncBuilder loop_ranges = [(l.start, l.stop, l.step) for l in parfor.loop_nests] if config.DEBUG_ARRAY_OPT: print("loop_nests = ", parfor.loop_nests) print("loop_ranges = ", loop_ranges) call_parallel_gufunc( lowerer, func, gu_signature, func_sig, func_args, func_arg_types, loop_ranges, parfor_redvars, parfor_reddict, redarrs, parfor.init_block, index_var_typ, parfor.races, exp_name_to_tuple_var) if nredvars > 0: _parfor_lowering_finalize_reduction( parfor, redarrs, lowerer, parfor_reddict, num_threads_var, ) # Cleanup reduction variable for v in to_cleanup: lowerer.lower_inst(ir.Del(v.name, loc=loc)) # Restore the original typemap of the function that was replaced temporarily at the # Beginning of this function. lowerer.fndesc.typemap = orig_typemap if config.DEBUG_ARRAY_OPT: print("_lower_parfor_parallel done") _ReductionInfo = make_dataclass( "_ReductionInfo", [ "redvar_info", "redvar_name", "redvar_typ", "redarr_var", "redarr_typ", "init_val", ], frozen=True, ) def _parfor_lowering_finalize_reduction( parfor, redarrs, lowerer, parfor_reddict, thread_count_var, ): """Emit code to finalize the reduction from the intermediate values of each thread. """ # For each reduction variable for redvar_name, redarr_var in redarrs.items(): # Pseudo-code for this loop body: # tmp = redarr[0] # for i in range(1, thread_count): # tmp = reduce_op(redarr[i], tmp) # reduction_result = tmp redvar_typ = lowerer.fndesc.typemap[redvar_name] redarr_typ = lowerer.fndesc.typemap[redarr_var.name] init_val = lowerer.loadvar(redvar_name) reduce_info = _ReductionInfo( redvar_info = parfor_reddict[redvar_name], redvar_name=redvar_name, redvar_typ=redvar_typ, redarr_var=redarr_var, redarr_typ=redarr_typ, init_val=init_val, ) # generate code for combining reduction variable with thread output handler = (_lower_trivial_inplace_binops if reduce_info.redvar_info.redop is not None else _lower_non_trivial_reduce) handler(parfor, lowerer, thread_count_var, reduce_info) class ParforsUnexpectedReduceNodeError(InternalError): def __init__(self, inst): super().__init__(f"Unknown reduce instruction node: {inst}") def _lower_trivial_inplace_binops(parfor, lowerer, thread_count_var, reduce_info): """Lower trivial inplace-binop reduction. """ for inst in reduce_info.redvar_info.reduce_nodes: # Var assigns to Var? if _lower_var_to_var_assign(lowerer, inst): pass # Is inplace-binop for the reduction? elif _is_right_op_and_rhs_is_init(inst, reduce_info.redvar_name, "inplace_binop"): fn = inst.value.fn redvar_result = _emit_binop_reduce_call( fn, lowerer, thread_count_var, reduce_info, ) lowerer.storevar(redvar_result, name=inst.target.name) # Is binop for the reduction? elif _is_right_op_and_rhs_is_init(inst, reduce_info.redvar_name, "binop"): fn = inst.value.fn redvar_result = _emit_binop_reduce_call( fn, lowerer, thread_count_var, reduce_info, ) lowerer.storevar(redvar_result, name=inst.target.name) # Otherwise? else: raise ParforsUnexpectedReduceNodeError(inst) # XXX: This seems like a hack to stop the loop with this condition. if _fix_redvar_name_ssa_mismatch(parfor, lowerer, inst, reduce_info.redvar_name): break if config.DEBUG_ARRAY_OPT_RUNTIME: varname = reduce_info.redvar_name lowerer.print_variable( f"{parfor.loc}: parfor {fn.__name__} reduction {varname} =", varname, ) def _lower_non_trivial_reduce(parfor, lowerer, thread_count_var, reduce_info): """Lower non-trivial reduction such as call to `functools.reduce()`. """ init_name = f"{reduce_info.redvar_name}#init" # The init_name variable is not defined at this point. lowerer.fndesc.typemap.setdefault(init_name, reduce_info.redvar_typ) # Emit a sequence of the reduction operation for each intermediate result # of each thread. num_thread_llval = lowerer.loadvar(thread_count_var.name) with cgutils.for_range(lowerer.builder, num_thread_llval) as loop: tid = loop.index for inst in reduce_info.redvar_info.reduce_nodes: # Var assigns to Var? if _lower_var_to_var_assign(lowerer, inst): pass # The reduction operation? elif (isinstance(inst, ir.Assign) and any(var.name == init_name for var in inst.list_vars())): elem = _emit_getitem_call(tid, lowerer, reduce_info) lowerer.storevar(elem, init_name) lowerer.lower_inst(inst) # Otherwise? else: raise ParforsUnexpectedReduceNodeError(inst) # XXX: This seems like a hack to stop the loop with this condition. if _fix_redvar_name_ssa_mismatch(parfor, lowerer, inst, reduce_info.redvar_name): break if config.DEBUG_ARRAY_OPT_RUNTIME: varname = reduce_info.redvar_name lowerer.print_variable( f"{parfor.loc}: parfor non-trivial reduction {varname} =", varname, ) def _lower_var_to_var_assign(lowerer, inst): """Lower Var->Var assignment. Returns True if-and-only-if `inst` is a Var->Var assignment. """ if isinstance(inst, ir.Assign) and isinstance(inst.value, ir.Var): loaded = lowerer.loadvar(inst.value.name) lowerer.storevar(loaded, name=inst.target.name) return True return False def _emit_getitem_call(idx, lowerer, reduce_info): """Emit call to ``redarr_var[idx]`` """ def reducer_getitem(redarr, index): return redarr[index] builder = lowerer.builder ctx = lowerer.context redarr_typ = reduce_info.redarr_typ arg_arr = lowerer.loadvar(reduce_info.redarr_var.name) args = (arg_arr, idx) sig = signature(reduce_info.redvar_typ, redarr_typ, types.intp) elem = ctx.compile_internal(builder, reducer_getitem, sig, args) return elem def _emit_binop_reduce_call(binop, lowerer, thread_count_var, reduce_info): """Emit call to the ``binop`` for the reduction variable. """ def reduction_add(thread_count, redarr, init): c = init for i in range(thread_count): c += redarr[i] return c def reduction_mul(thread_count, redarr, init): c = init for i in range(thread_count): c *= redarr[i] return c kernel = { operator.iadd: reduction_add, operator.isub: reduction_add, operator.add: reduction_add, operator.sub: reduction_add, operator.imul: reduction_mul, operator.ifloordiv: reduction_mul, operator.itruediv: reduction_mul, operator.mul: reduction_mul, operator.floordiv: reduction_mul, operator.truediv: reduction_mul, }[binop] ctx = lowerer.context builder = lowerer.builder redarr_typ = reduce_info.redarr_typ arg_arr = lowerer.loadvar(reduce_info.redarr_var.name) if config.DEBUG_ARRAY_OPT_RUNTIME: init_var = reduce_info.redarr_var.scope.get(reduce_info.redvar_name) res_print = ir.Print( args=[reduce_info.redarr_var, init_var], vararg=None, loc=lowerer.loc, ) typemap = lowerer.fndesc.typemap lowerer.fndesc.calltypes[res_print] = signature( types.none, typemap[reduce_info.redarr_var.name], typemap[init_var.name], ) lowerer.lower_inst(res_print) arg_thread_count = lowerer.loadvar(thread_count_var.name) args = (arg_thread_count, arg_arr, reduce_info.init_val) sig = signature( reduce_info.redvar_typ, types.uintp, redarr_typ, reduce_info.redvar_typ, ) redvar_result = ctx.compile_internal(builder, kernel, sig, args) return redvar_result def _is_right_op_and_rhs_is_init(inst, redvar_name, op): """Is ``inst`` an inplace-binop and the RHS is the reduction init? """ if not isinstance(inst, ir.Assign): return False rhs = inst.value if not isinstance(rhs, ir.Expr): return False if rhs.op != op: return False if rhs.rhs.name != f"{redvar_name}#init": return False return True def _fix_redvar_name_ssa_mismatch(parfor, lowerer, inst, redvar_name): """Fix reduction variable name mismatch due to SSA. """ # Only process reduction statements post-gufunc execution # until we see an assignment with a left-hand side to the # reduction variable's name. This fixes problems with # cases where there are multiple assignments to the # reduction variable in the parfor. scope = parfor.init_block.scope if isinstance(inst, ir.Assign): try: reduction_var = scope.get_exact(redvar_name) except NotDefinedError: # Ideally, this shouldn't happen. The redvar name # missing from scope indicates an error from # other rewrite passes. is_same_source_var = redvar_name == inst.target.name else: # Because of SSA, the redvar and target var of # the current assignment would be different even # though they refer to the same source-level var. redvar_unver_name = reduction_var.unversioned_name target_unver_name = inst.target.unversioned_name is_same_source_var = redvar_unver_name == target_unver_name if is_same_source_var: # If redvar is different from target var, add an # assignment to put target var into redvar. if redvar_name != inst.target.name: val = lowerer.loadvar(inst.target.name) lowerer.storevar(val, name=redvar_name) return True return False def _create_shape_signature( get_shape_classes, num_inputs, num_reductions, args, func_sig, races, typemap): '''Create shape signature for GUFunc ''' if config.DEBUG_ARRAY_OPT: print("_create_shape_signature", num_inputs, num_reductions, args, races) for i in args[1:]: print("argument", i, type(i), get_shape_classes(i, typemap=typemap)) num_inouts = len(args) - num_reductions # maximum class number for array shapes classes = [get_shape_classes(var, typemap=typemap) if var not in races else (-1,) for var in args[1:]] class_set = set() for _class in classes: if _class: for i in _class: class_set.add(i) max_class = max(class_set) + 1 if class_set else 0 classes.insert(0, (max_class,)) # force set the class of 'sched' argument class_set.add(max_class) thread_num_class = max_class + 1 class_set.add(thread_num_class) class_map = {} # TODO: use prefix + class number instead of single char alphabet = ord('a') for n in class_set: if n >= 0: class_map[n] = chr(alphabet) alphabet += 1 threadcount_ordinal = chr(alphabet) alpha_dict = {'latest_alpha' : alphabet} def bump_alpha(c, class_map): if c >= 0: return class_map[c] else: alpha_dict['latest_alpha'] += 1 return chr(alpha_dict['latest_alpha']) gu_sin = [] gu_sout = [] count = 0 syms_sin = () if config.DEBUG_ARRAY_OPT: print("args", args) print("classes", classes) print("threadcount_ordinal", threadcount_ordinal) for cls, arg in zip(classes, args): count = count + 1 if cls: dim_syms = tuple(bump_alpha(c, class_map) for c in cls) else: dim_syms = () if (count > num_inouts): # Add the threadcount_ordinal to represent the thread count # to the start of the reduction array. gu_sin.append(tuple([threadcount_ordinal] + list(dim_syms[1:]))) else: gu_sin.append(dim_syms) syms_sin += dim_syms return (gu_sin, gu_sout) def _print_block(block): for i, inst in enumerate(block.body): print(" ", i, " ", inst) def _print_body(body_dict): '''Pretty-print a set of IR blocks. ''' topo_order = wrap_find_topo(body_dict) for label in topo_order: block = body_dict[label] print("label: ", label) _print_block(block) def wrap_loop_body(loop_body): blocks = loop_body.copy() # shallow copy is enough first_label = min(blocks.keys()) last_label = max(blocks.keys()) loc = blocks[last_label].loc blocks[last_label].body.append(ir.Jump(first_label, loc)) return blocks def unwrap_loop_body(loop_body): last_label = max(loop_body.keys()) loop_body[last_label].body = loop_body[last_label].body[:-1] def add_to_def_once_sets(a_def, def_once, def_more): '''If the variable is already defined more than once, do nothing. Else if defined exactly once previously then transition this variable to the defined more than once set (remove it from def_once set and add to def_more set). Else this must be the first time we've seen this variable defined so add to def_once set. ''' if a_def in def_more: pass elif a_def in def_once: def_more.add(a_def) def_once.remove(a_def) else: def_once.add(a_def) def compute_def_once_block(block, def_once, def_more, getattr_taken, typemap, module_assigns): '''Effect changes to the set of variables defined once or more than once for a single block. block - the block to process def_once - set of variable names known to be defined exactly once def_more - set of variable names known to be defined more than once getattr_taken - dict mapping variable name to tuple of object and attribute taken module_assigns - dict mapping variable name to the Global that they came from ''' # The only "defs" occur in assignments, so find such instructions. assignments = block.find_insts(ir.Assign) # For each assignment... for one_assign in assignments: # Get the LHS/target of the assignment. a_def = one_assign.target.name # Add variable to def sets. add_to_def_once_sets(a_def, def_once, def_more) rhs = one_assign.value if isinstance(rhs, ir.Global): # Remember assignments of the form "a = Global(...)" # Is this a module? if isinstance(rhs.value, pytypes.ModuleType): module_assigns[a_def] = rhs.value.__name__ if isinstance(rhs, ir.Expr) and rhs.op == 'getattr' and rhs.value.name in def_once: # Remember assignments of the form "a = b.c" getattr_taken[a_def] = (rhs.value.name, rhs.attr) if isinstance(rhs, ir.Expr) and rhs.op == 'call' and rhs.func.name in getattr_taken: # If "a" is being called then lookup the getattr definition of "a" # as above, getting the module variable "b" (base_obj) # and the attribute "c" (base_attr). base_obj, base_attr = getattr_taken[rhs.func.name] if base_obj in module_assigns: # If we know the definition of the module variable then get the module # name from module_assigns. base_mod_name = module_assigns[base_obj] if not is_const_call(base_mod_name, base_attr): # Calling a method on an object could modify the object and is thus # like a def of that object. We call is_const_call to see if this module/attribute # combination is known to not modify the module state. If we don't know that # the combination is safe then we have to assume there could be a modification to # the module and thus add the module variable as defined more than once. add_to_def_once_sets(base_obj, def_once, def_more) else: # Assume the worst and say that base_obj could be modified by the call. add_to_def_once_sets(base_obj, def_once, def_more) if isinstance(rhs, ir.Expr) and rhs.op == 'call': # If a mutable object is passed to a function, then it may be changed and # therefore can't be hoisted. # For each argument to the function... for argvar in rhs.args: # Get the argument's type. if isinstance(argvar, ir.Var): argvar = argvar.name avtype = typemap[argvar] # If that type doesn't have a mutable attribute or it does and it's set to # not mutable then this usage is safe for hoisting. if getattr(avtype, 'mutable', False): # Here we have a mutable variable passed to a function so add this variable # to the def lists. add_to_def_once_sets(argvar, def_once, def_more) def wrap_find_topo(loop_body): blocks = wrap_loop_body(loop_body) topo_order = find_topo_order(blocks) unwrap_loop_body(loop_body) return topo_order def compute_def_once_internal(loop_body, def_once, def_more, getattr_taken, typemap, module_assigns): '''Compute the set of variables defined exactly once in the given set of blocks and use the given sets for storing which variables are defined once, more than once and which have had a getattr call on them. ''' # For each block in topological order... topo_order = wrap_find_topo(loop_body) for label in topo_order: block = loop_body[label] # Scan this block and effect changes to def_once, def_more, and getattr_taken # based on the instructions in that block. compute_def_once_block(block, def_once, def_more, getattr_taken, typemap, module_assigns) # Have to recursively process parfors manually here. for inst in block.body: if isinstance(inst, parfor.Parfor): # Recursively compute for the parfor's init block. compute_def_once_block(inst.init_block, def_once, def_more, getattr_taken, typemap, module_assigns) # Recursively compute for the parfor's loop body. compute_def_once_internal(inst.loop_body, def_once, def_more, getattr_taken, typemap, module_assigns) def compute_def_once(loop_body, typemap): '''Compute the set of variables defined exactly once in the given set of blocks. ''' def_once = set() # set to hold variables defined exactly once def_more = set() # set to hold variables defined more than once getattr_taken = {} module_assigns = {} compute_def_once_internal(loop_body, def_once, def_more, getattr_taken, typemap, module_assigns) return def_once, def_more def find_vars(var, varset): assert isinstance(var, ir.Var) varset.add(var.name) return var def _hoist_internal(inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, stored_arrays): if inst.target.name in stored_arrays: not_hoisted.append((inst, "stored array")) if config.DEBUG_ARRAY_OPT >= 1: print("Instruction", inst, "could not be hoisted because the created array is stored.") return False target_type = typemap[inst.target.name] uses = set() # Get vars used by this statement. visit_vars_inner(inst.value, find_vars, uses) # Filter out input parameters from the set of variable usages. unhoistable = {assgn.target.name for assgn, _ in not_hoisted} use_unhoist = uses & unhoistable diff = uses.difference(dep_on_param) diff |= use_unhoist if config.DEBUG_ARRAY_OPT >= 1: print("_hoist_internal:", inst, "uses:", uses, "diff:", diff) if len(diff) == 0 and is_pure(inst.value, None, call_table): if config.DEBUG_ARRAY_OPT >= 1: print("Will hoist instruction", inst, target_type) hoisted.append(inst) if not isinstance(target_type, types.npytypes.Array): dep_on_param += [inst.target.name] return True else: if len(diff) > 0: not_hoisted.append((inst, "dependency")) if config.DEBUG_ARRAY_OPT >= 1: print("Instruction", inst, "could not be hoisted because of a dependency.") else: not_hoisted.append((inst, "not pure")) if config.DEBUG_ARRAY_OPT >= 1: print("Instruction", inst, "could not be hoisted because it isn't pure.") return False def find_setitems_block(setitems, itemsset, block, typemap): for inst in block.body: if isinstance(inst, (ir.StaticSetItem, ir.SetItem)): setitems.add(inst.target.name) # If we store a non-mutable object into an array then that is safe to hoist. # If the stored object is mutable and you hoist then multiple entries in the # outer array could reference the same object and changing one index would then # change other indices. if getattr(typemap[inst.value.name], "mutable", False): itemsset.add(inst.value.name) elif isinstance(inst, parfor.Parfor): find_setitems_block(setitems, itemsset, inst.init_block, typemap) find_setitems_body(setitems, itemsset, inst.loop_body, typemap) elif isinstance(inst, ir.Assign): # If something of mutable type is given to a build_tuple or # used in a call then consider it unanalyzable and so # unavailable for hoisting. rhs = inst.value def add_to_itemset(item): assert isinstance(item, ir.Var), rhs if getattr(typemap[item.name], "mutable", False): itemsset.add(item.name) if isinstance(rhs, ir.Expr): if rhs.op in ["build_tuple", "build_list", "build_set"]: for item in rhs.items: add_to_itemset(item) elif rhs.op == "build_map": for pair in rhs.items: for item in pair: add_to_itemset(item) elif rhs.op == "call": for item in list(rhs.args) + [x[1] for x in rhs.kws]: add_to_itemset(item) def find_setitems_body(setitems, itemsset, loop_body, typemap): """ Find the arrays that are written into (goes into setitems) and the mutable objects (mostly arrays) that are written into other arrays (goes into itemsset). """ for label, block in loop_body.items(): find_setitems_block(setitems, itemsset, block, typemap) def empty_container_allocator_hoist(inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, stored_arrays): if (isinstance(inst, ir.Assign) and isinstance(inst.value, ir.Expr) and inst.value.op == 'call' and inst.value.func.name in call_table): call_list = call_table[inst.value.func.name] if call_list == ['empty', np]: return _hoist_internal(inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, stored_arrays) return False def hoist(parfor_params, loop_body, typemap, wrapped_blocks): dep_on_param = copy.copy(parfor_params) hoisted = [] not_hoisted = [] # Compute the set of variable defined exactly once in the loop body. def_once, def_more = compute_def_once(loop_body, typemap) (call_table, reverse_call_table) = get_call_table(wrapped_blocks) setitems = set() itemsset = set() find_setitems_body(setitems, itemsset, loop_body, typemap) dep_on_param = list(set(dep_on_param).difference(setitems)) if config.DEBUG_ARRAY_OPT >= 1: print("hoist - def_once:", def_once, "setitems:", setitems, "itemsset:", itemsset, "dep_on_param:", dep_on_param, "parfor_params:", parfor_params) for si in setitems: add_to_def_once_sets(si, def_once, def_more) for label, block in loop_body.items(): new_block = [] for inst in block.body: if empty_container_allocator_hoist(inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, itemsset): continue elif isinstance(inst, ir.Assign) and inst.target.name in def_once: if _hoist_internal(inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, itemsset): # don't add this instruction to the block since it is # hoisted continue elif isinstance(inst, parfor.Parfor): new_init_block = [] if config.DEBUG_ARRAY_OPT >= 1: print("parfor") inst.dump() for ib_inst in inst.init_block.body: if empty_container_allocator_hoist(ib_inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, itemsset): continue elif (isinstance(ib_inst, ir.Assign) and ib_inst.target.name in def_once): if _hoist_internal(ib_inst, dep_on_param, call_table, hoisted, not_hoisted, typemap, itemsset): # don't add this instruction to the block since it is hoisted continue new_init_block.append(ib_inst) inst.init_block.body = new_init_block new_block.append(inst) block.body = new_block return hoisted, not_hoisted def redtyp_is_scalar(redtype): return not isinstance(redtype, types.npytypes.Array) def redtyp_to_redarraytype(redtyp): """Go from a reducation variable type to a reduction array type used to hold per-worker results. """ redarrdim = 1 # If the reduction type is an array then allocate reduction array with ndim+1 dimensions. if isinstance(redtyp, types.npytypes.Array): redarrdim += redtyp.ndim # We don't create array of array but multi-dimensional reduction array with same dtype. redtyp = redtyp.dtype return types.npytypes.Array(redtyp, redarrdim, "C") def redarraytype_to_sig(redarraytyp): """Given a reduction array type, find the type of the reduction argument to the gufunc. """ assert isinstance(redarraytyp, types.npytypes.Array) return types.npytypes.Array(redarraytyp.dtype, redarraytyp.ndim, redarraytyp.layout) def legalize_names_with_typemap(names, typemap): """ We use ir_utils.legalize_names to replace internal IR variable names containing illegal characters (e.g. period) with a legal character (underscore) so as to create legal variable names. The original variable names are in the typemap so we also need to add the legalized name to the typemap as well. """ outdict = legalize_names(names) # For each pair in the dict of legalized names... for x, y in outdict.items(): # If the name had some legalization change to it... if x != y: # Set the type of the new name the same as the type of the old name. typemap[y] = typemap[x] return outdict def to_scalar_from_0d(x): if isinstance(x, types.ArrayCompatible): if x.ndim == 0: return x.dtype return x 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. ''' if config.DEBUG_ARRAY_OPT >= 1: print("starting _create_gufunc_for_parfor_body") 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.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) # 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)) # ------------------------------------------------------------------------- # Convert tuples to individual parameters. tuple_expanded_parfor_inputs = [] tuple_var_to_expanded_names = {} expanded_name_to_tuple_var = {} next_expanded_tuple_var = 0 parfor_tuple_params = [] # For each input to the parfor. for pi in parfor_inputs: # Get the type of the input. pi_type = typemap[pi] # If it is a UniTuple or Tuple we will do the conversion. if isinstance(pi_type, types.UniTuple) or isinstance(pi_type, types.NamedUniTuple): # Get the size and dtype of the tuple. tuple_count = pi_type.count tuple_dtype = pi_type.dtype # Only do tuples up to config.PARFOR_MAX_TUPLE_SIZE length. assert(tuple_count <= config.PARFOR_MAX_TUPLE_SIZE) this_var_expansion = [] for i in range(tuple_count): # Generate a new name for the individual part of the tuple var. expanded_name = "expanded_tuple_var_" + str(next_expanded_tuple_var) # Add that name to the new list of inputs to the gufunc. tuple_expanded_parfor_inputs.append(expanded_name) this_var_expansion.append(expanded_name) # Remember a mapping from new param name to original tuple # var and the index within the tuple. expanded_name_to_tuple_var[expanded_name] = (pi, i) next_expanded_tuple_var += 1 # Set the type of the new parameter. typemap[expanded_name] = tuple_dtype # Remember a mapping from the original tuple var to the # individual parts. tuple_var_to_expanded_names[pi] = this_var_expansion parfor_tuple_params.append(pi) elif isinstance(pi_type, types.Tuple) or isinstance(pi_type, types.NamedTuple): # This is the same as above for UniTuple except that each part of # the tuple can have a different type and we fetch that type with # pi_type.types[offset]. tuple_count = pi_type.count tuple_types = pi_type.types # Only do tuples up to config.PARFOR_MAX_TUPLE_SIZE length. assert(tuple_count <= config.PARFOR_MAX_TUPLE_SIZE) this_var_expansion = [] for i in range(tuple_count): expanded_name = "expanded_tuple_var_" + str(next_expanded_tuple_var) tuple_expanded_parfor_inputs.append(expanded_name) this_var_expansion.append(expanded_name) expanded_name_to_tuple_var[expanded_name] = (pi, i) next_expanded_tuple_var += 1 typemap[expanded_name] = tuple_types[i] tuple_var_to_expanded_names[pi] = this_var_expansion parfor_tuple_params.append(pi) else: tuple_expanded_parfor_inputs.append(pi) parfor_inputs = tuple_expanded_parfor_inputs if config.DEBUG_ARRAY_OPT >= 1: print("parfor_inputs post tuple handling = ", parfor_inputs, " ", type(parfor_inputs)) # ------------------------------------------------------------------------- races = races.difference(set(parfor_redvars)) 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) # Reduction variables are represented as arrays, so they go under # different names. parfor_redarrs = [] parfor_red_arg_types = [] for var in parfor_redvars: arr = var + "_arr" parfor_redarrs.append(arr) redarraytype = redtyp_to_redarraytype(typemap[var]) parfor_red_arg_types.append(redarraytype) redarrsig = redarraytype_to_sig(redarraytype) if arr in typemap: assert(typemap[arr] == redarrsig) else: typemap[arr] = redarrsig # 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_with_typemap(parfor_params + parfor_redvars + parfor_tuple_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] # Calculate types of args passed to gufunc. func_arg_types = [typemap[v] for v in (parfor_inputs + parfor_outputs)] + parfor_red_arg_types if config.DEBUG_ARRAY_OPT >= 1: print("new param_types:", param_types) print("new func_arg_types:", 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. # 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" globls = {"np": np, "numba": numba} # First thing in the gufunc, we reconstruct tuples from their # individual parts, e.g., orig_tup_name = (part1, part2,). # The rest of the code of the function will use the original tuple name. for tup_var, exp_names in tuple_var_to_expanded_names.items(): tup_type = typemap[tup_var] gufunc_txt += " " + param_dict[tup_var] # Determine if the tuple is a named tuple. if (isinstance(tup_type, types.NamedTuple) or isinstance(tup_type, types.NamedUniTuple)): named_tup = True else: named_tup = False if named_tup: # It is a named tuple so try to find the global that defines the # named tuple. func_def = guard(get_definition, lowerer.func_ir, tup_var) named_tuple_def = None if config.DEBUG_ARRAY_OPT: print("func_def:", func_def, type(func_def)) if func_def is not None: if (isinstance(func_def, ir.Expr) and func_def.op == 'call'): named_tuple_def = guard(get_definition, lowerer.func_ir, func_def.func) if config.DEBUG_ARRAY_OPT: print("named_tuple_def:", named_tuple_def, type(named_tuple_def)) elif isinstance(func_def, ir.Arg): named_tuple_def = typemap[func_def.name] if config.DEBUG_ARRAY_OPT: print("named_tuple_def:", named_tuple_def, type(named_tuple_def), named_tuple_def.name) if named_tuple_def is not None: if (isinstance(named_tuple_def, ir.Global) or isinstance(named_tuple_def, ir.FreeVar)): gval = named_tuple_def.value if config.DEBUG_ARRAY_OPT: print("gval:", gval, type(gval)) globls[named_tuple_def.name] = gval elif isinstance(named_tuple_def, types.containers.BaseNamedTuple): named_tuple_name = named_tuple_def.name.split('(')[0] if config.DEBUG_ARRAY_OPT: print("name:", named_tuple_name, named_tuple_def.instance_class, type(named_tuple_def.instance_class)) globls[named_tuple_name] = named_tuple_def.instance_class else: if config.DEBUG_ARRAY_OPT: print("Didn't find definition of namedtuple for globls.") raise CompilerError("Could not find definition of " + str(tup_var), tup_var.loc) gufunc_txt += " = " + tup_type.instance_class.__name__ + "(" for name, field_name in zip(exp_names, tup_type.fields): gufunc_txt += field_name + "=" + param_dict[name] + "," else: # Just a regular tuple so use (part0, part1, ...) gufunc_txt += " = (" + ", ".join([param_dict[x] for x in exp_names]) if len(exp_names) == 1: # Add comma for tuples with singular values. We can't unilaterally # add a comma always because (,) isn't valid. gufunc_txt += "," gufunc_txt += ")\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") gufunc_thread_id_var = "ParallelAcceleratorGufuncThreadId" if len(parfor_redarrs) > 0: gufunc_txt += " " + gufunc_thread_id_var + " = " gufunc_txt += "numba.np.ufunc.parallel._iget_thread_id()\n" # Add initialization of reduction variables for arr, var in zip(parfor_redarrs, parfor_redvars): gufunc_txt += " " + param_dict[var] + \ "=" + param_dict[arr] + "[" + gufunc_thread_id_var + "]\n" if config.DEBUG_ARRAY_OPT_RUNTIME: gufunc_txt += " print(\"thread id =\", ParallelAcceleratorGufuncThreadId)\n" gufunc_txt += " print(\"initial reduction value\",ParallelAcceleratorGufuncThreadId," + param_dict[var] + "," + param_dict[var] + ".shape)\n" gufunc_txt += " print(\"reduction array\",ParallelAcceleratorGufuncThreadId," + param_dict[arr] + "," + param_dict[arr] + ".shape)\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): if config.DEBUG_ARRAY_OPT_RUNTIME: gufunc_txt += " print(\"final reduction value\",ParallelAcceleratorGufuncThreadId," + param_dict[var] + ")\n" gufunc_txt += " print(\"final reduction array\",ParallelAcceleratorGufuncThreadId," + param_dict[arr] + ")\n" # After the gufunc loops, copy the accumulated temp value back to reduction array. gufunc_txt += " " + param_dict[arr] + \ "[" + gufunc_thread_id_var + "] = " + param_dict[var] + "\n" gufunc_txt += " return None\n" if config.DEBUG_ARRAY_OPT: print("gufunc_txt = ", type(gufunc_txt), "\n", gufunc_txt) print("globls:", globls, type(globls)) # Force gufunc outline into existence. 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] = parfor.init_block.scope.redefine(name, loc).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 = [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 = find_max_label(gufunc_ir.blocks) + 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 = find_max_label(loop_body) + 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.StringLiteral(strval) lhs = scope.redefine("str_const", loc) # 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.core.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, not_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) # store hoisted into diagnostics diagnostics = lowerer.metadata['parfor_diagnostics'] diagnostics.hoist_info[parfor.id] = {'hoisted': hoisted, 'not_hoisted': not_hoisted} 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] = transfer_scope(b, scope) 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 fixup_var_define_in_scope(gufunc_ir.blocks) class ParforGufuncCompiler(compiler.CompilerBase): def define_pipelines(self): from numba.core.compiler_machinery import PassManager dpb = compiler.DefaultPassBuilder pm = PassManager("full_parfor_gufunc") parfor_gufunc_passes = dpb.define_parfor_gufunc_pipeline(self.state) pm.passes.extend(parfor_gufunc_passes.passes) lowering_passes = dpb.define_parfor_gufunc_nopython_lowering_pipeline(self.state) pm.passes.extend(lowering_passes.passes) pm.finalize() return [pm] kernel_func = compiler.compile_ir( typingctx, targetctx, gufunc_ir, gufunc_param_types, types.none, flags, locals, pipeline_class=ParforGufuncCompiler) flags.noalias = old_alias kernel_sig = signature(types.none, *gufunc_param_types) if config.DEBUG_ARRAY_OPT: print("finished create_gufunc_for_parfor_body. kernel_sig = ", kernel_sig) return kernel_func, parfor_args, kernel_sig, func_arg_types, expanded_name_to_tuple_var def replace_var_with_array_in_block(vars, block, typemap, calltypes): new_block = [] for inst in block.body: if isinstance(inst, ir.Assign) and inst.target.name in vars: loc = inst.loc scope = inst.target.scope const_node = ir.Const(0, loc) const_var = scope.redefine("$const_ind_0", loc) typemap[const_var.name] = types.uintp const_assign = ir.Assign(const_node, const_var, loc) new_block.append(const_assign) val_var = scope.redefine("$val", loc) typemap[val_var.name] = typemap[inst.target.name] new_block.append(ir.Assign(inst.value, val_var, loc)) setitem_node = ir.SetItem(inst.target, const_var, val_var, loc) calltypes[setitem_node] = signature( types.none, types.npytypes.Array(typemap[inst.target.name], 1, "C"), types.intp, typemap[inst.target.name]) new_block.append(setitem_node) continue elif isinstance(inst, parfor.Parfor): replace_var_with_array_internal(vars, {0: inst.init_block}, typemap, calltypes) replace_var_with_array_internal(vars, inst.loop_body, typemap, calltypes) new_block.append(inst) return new_block def replace_var_with_array_internal(vars, loop_body, typemap, calltypes): for label, block in loop_body.items(): block.body = replace_var_with_array_in_block(vars, block, typemap, calltypes) def replace_var_with_array(vars, loop_body, typemap, calltypes): replace_var_with_array_internal(vars, loop_body, typemap, calltypes) for v in vars: el_typ = typemap[v] typemap.pop(v, None) typemap[v] = types.npytypes.Array(el_typ, 1, "C") def call_parallel_gufunc(lowerer, cres, gu_signature, outer_sig, expr_args, expr_arg_types, loop_ranges, redvars, reddict, redarrdict, init_block, index_var_typ, races, exp_name_to_tuple_var): ''' Adds the call to the gufunc function from the main function. ''' context = lowerer.context builder = lowerer.builder from numba.np.ufunc.parallel import (build_gufunc_wrapper, _launch_threads) if config.DEBUG_ARRAY_OPT: print("make_parallel_loop") print("outer_sig = ", outer_sig.args, outer_sig.return_type, outer_sig.recvr, outer_sig.pysig) print("loop_ranges = ", loop_ranges) print("expr_args", expr_args) print("expr_arg_types", expr_arg_types) print("gu_signature", gu_signature) # Build the wrapper for GUFunc args, return_type = sigutils.normalize_signature(outer_sig) llvm_func = cres.library.get_function(cres.fndesc.llvm_func_name) sin, sout = gu_signature # These are necessary for build_gufunc_wrapper to find external symbols _launch_threads() info = build_gufunc_wrapper(llvm_func, cres, sin, sout, cache=False, is_parfors=True) wrapper_name = info.name cres.library._ensure_finalized() if config.DEBUG_ARRAY_OPT: print("parallel function = ", wrapper_name, cres) # loadvars for loop_ranges def load_range(v): if isinstance(v, ir.Var): return lowerer.loadvar(v.name) else: return context.get_constant(types.uintp, v) num_dim = len(loop_ranges) for i in range(num_dim): start, stop, step = loop_ranges[i] start = load_range(start) stop = load_range(stop) assert(step == 1) # We do not support loop steps other than 1 step = load_range(step) loop_ranges[i] = (start, stop, step) if config.DEBUG_ARRAY_OPT: print("call_parallel_gufunc loop_ranges[{}] = ".format(i), start, stop, step) cgutils.printf(builder, "loop range[{}]: %d %d (%d)\n".format(i), start, stop, step) # Commonly used LLVM types and constants byte_t = llvmlite.ir.IntType(8) byte_ptr_t = llvmlite.ir.PointerType(byte_t) byte_ptr_ptr_t = llvmlite.ir.PointerType(byte_ptr_t) intp_t = context.get_value_type(types.intp) uintp_t = context.get_value_type(types.uintp) intp_ptr_t = llvmlite.ir.PointerType(intp_t) intp_ptr_ptr_t = llvmlite.ir.PointerType(intp_ptr_t) uintp_ptr_t = llvmlite.ir.PointerType(uintp_t) uintp_ptr_ptr_t = llvmlite.ir.PointerType(uintp_ptr_t) zero = context.get_constant(types.uintp, 0) one = context.get_constant(types.uintp, 1) one_type = one.type sizeof_intp = context.get_abi_sizeof(intp_t) # Prepare sched, first pop it out of expr_args, outer_sig, and gu_signature expr_args.pop(0) sched_sig = sin.pop(0) if config.DEBUG_ARRAY_OPT: print("Parfor has potentially negative start", index_var_typ.signed) if index_var_typ.signed: sched_type = intp_t sched_ptr_type = intp_ptr_t sched_ptr_ptr_type = intp_ptr_ptr_t else: sched_type = uintp_t sched_ptr_type = uintp_ptr_t sched_ptr_ptr_type = uintp_ptr_ptr_t # Call do_scheduling with appropriate arguments dim_starts = cgutils.alloca_once( builder, sched_type, size=context.get_constant( types.uintp, num_dim), name="dim_starts") dim_stops = cgutils.alloca_once( builder, sched_type, size=context.get_constant( types.uintp, num_dim), name="dim_stops") for i in range(num_dim): start, stop, step = loop_ranges[i] if start.type != one_type: start = builder.sext(start, one_type) if stop.type != one_type: stop = builder.sext(stop, one_type) if step.type != one_type: step = builder.sext(step, one_type) # substract 1 because do-scheduling takes inclusive ranges stop = builder.sub(stop, one) builder.store( start, builder.gep( dim_starts, [ context.get_constant( types.uintp, i)])) builder.store(stop, builder.gep(dim_stops, [context.get_constant(types.uintp, i)])) # Prepare to call get/set parallel_chunksize and get the number of threads. get_chunksize = cgutils.get_or_insert_function( builder.module, llvmlite.ir.FunctionType(uintp_t, []), name="get_parallel_chunksize") set_chunksize = cgutils.get_or_insert_function( builder.module, llvmlite.ir.FunctionType(llvmlite.ir.VoidType(), [uintp_t]), name="set_parallel_chunksize") get_num_threads = cgutils.get_or_insert_function( builder.module, llvmlite.ir.FunctionType(llvmlite.ir.IntType(types.intp.bitwidth), []), "get_num_threads") # Get the current number of threads. num_threads = builder.call(get_num_threads, []) # Get the current chunksize so we can use it and restore the value later. current_chunksize = builder.call(get_chunksize, []) with cgutils.if_unlikely(builder, builder.icmp_signed('<=', num_threads, num_threads.type(0))): cgutils.printf(builder, "num_threads: %d\n", num_threads) context.call_conv.return_user_exc(builder, RuntimeError, ("Invalid number of threads. " "This likely indicates a bug in Numba.",)) # Call get_sched_size from gufunc_scheduler.cpp that incorporates the size of the work, # the number of threads and the selected chunk size. This will tell us how many entries # in the schedule we will need. get_sched_size_fnty = llvmlite.ir.FunctionType(uintp_t, [uintp_t, uintp_t, intp_ptr_t, intp_ptr_t]) get_sched_size = cgutils.get_or_insert_function( builder.module, get_sched_size_fnty, name="get_sched_size") num_divisions = builder.call(get_sched_size, [num_threads, context.get_constant(types.uintp, num_dim), dim_starts, dim_stops]) # Set the chunksize to zero so that any nested calls get the default chunk size behavior. builder.call(set_chunksize, [zero]) # Each entry in the schedule is 2 times the number of dimensions long. multiplier = context.get_constant(types.uintp, num_dim * 2) # Compute the total number of entries in the schedule. sched_size = builder.mul(num_divisions, multiplier) # Prepare to dynamically allocate memory to hold the schedule. alloc_sched_fnty = llvmlite.ir.FunctionType(sched_ptr_type, [uintp_t]) alloc_sched_func = cgutils.get_or_insert_function( builder.module, alloc_sched_fnty, name="allocate_sched") # Call gufunc_scheduler.cpp to allocate the schedule. # This may or may not do pooling. alloc_space = builder.call(alloc_sched_func, [sched_size]) # Allocate a slot in the entry block to store the schedule pointer. sched = cgutils.alloca_once(builder, sched_ptr_type) # Store the schedule pointer into that slot. builder.store(alloc_space, sched) debug_flag = 1 if config.DEBUG_ARRAY_OPT else 0 scheduling_fnty = llvmlite.ir.FunctionType( intp_ptr_t, [uintp_t, intp_ptr_t, intp_ptr_t, uintp_t, sched_ptr_type, intp_t]) if index_var_typ.signed: do_scheduling = cgutils.get_or_insert_function(builder.module, scheduling_fnty, name="do_scheduling_signed") else: do_scheduling = cgutils.get_or_insert_function(builder.module, scheduling_fnty, name="do_scheduling_unsigned") # Call the scheduling routine that decides how to break up the work. builder.call( do_scheduling, [ context.get_constant( types.uintp, num_dim), dim_starts, dim_stops, num_divisions, builder.load(sched), context.get_constant( types.intp, debug_flag)]) # Get the LLVM vars for the Numba IR reduction array vars. redarrs = [lowerer.loadvar(redarrdict[x].name) for x in redvars] nredvars = len(redvars) ninouts = len(expr_args) - nredvars def load_potential_tuple_var(x): """Given a variable name, if that variable is not a new name introduced as the extracted part of a tuple then just return the variable loaded from its name. However, if the variable does represent part of a tuple, as recognized by the name of the variable being present in the exp_name_to_tuple_var dict, then we load the original tuple var instead that we get from the dict and then extract the corresponding element of the tuple, also stored and returned to use in the dict (i.e., offset). """ if x in exp_name_to_tuple_var: orig_tup, offset = exp_name_to_tuple_var[x] tup_var = lowerer.loadvar(orig_tup) res = builder.extract_value(tup_var, offset) return res else: return lowerer.loadvar(x) # ---------------------------------------------------------------------------- # Prepare arguments: args, shapes, steps, data all_args = [load_potential_tuple_var(x) for x in expr_args[:ninouts]] + redarrs num_args = len(all_args) num_inps = len(sin) + 1 args = cgutils.alloca_once( builder, byte_ptr_t, size=context.get_constant( types.intp, 1 + num_args), name="pargs") array_strides = [] # sched goes first builder.store(builder.bitcast(builder.load(sched), byte_ptr_t), args) array_strides.append(context.get_constant(types.intp, sizeof_intp)) rv_to_arg_dict = {} # followed by other arguments for i in range(num_args): arg = all_args[i] var = expr_args[i] aty = expr_arg_types[i] dst = builder.gep(args, [context.get_constant(types.intp, i + 1)]) if i >= ninouts: # reduction variables ary = context.make_array(aty)(context, builder, arg) strides = cgutils.unpack_tuple(builder, ary.strides, aty.ndim) # Start from 1 because we skip the first dimension of length num_threads just like sched. for j in range(len(strides)): array_strides.append(strides[j]) builder.store(builder.bitcast(ary.data, byte_ptr_t), dst) elif isinstance(aty, types.ArrayCompatible): if var in races: typ = (context.get_data_type(aty.dtype) if aty.dtype != types.boolean else llvmlite.ir.IntType(1)) rv_arg = cgutils.alloca_once(builder, typ) builder.store(arg, rv_arg) builder.store(builder.bitcast(rv_arg, byte_ptr_t), dst) rv_to_arg_dict[var] = (arg, rv_arg) array_strides.append(context.get_constant(types.intp, context.get_abi_sizeof(typ))) else: ary = context.make_array(aty)(context, builder, arg) strides = cgutils.unpack_tuple(builder, ary.strides, aty.ndim) for j in range(len(strides)): array_strides.append(strides[j]) builder.store(builder.bitcast(ary.data, byte_ptr_t), dst) else: if i < num_inps: # Scalar input, need to store the value in an array of size 1 if isinstance(aty, types.Optional): # Unpack optional type unpacked_aty = aty.type arg = context.cast(builder, arg, aty, unpacked_aty) else: unpacked_aty = aty typ = (context.get_data_type(unpacked_aty) if not isinstance(unpacked_aty, types.Boolean) else llvmlite.ir.IntType(1)) ptr = cgutils.alloca_once(builder, typ) builder.store(arg, ptr) else: # Scalar output, must allocate typ = (context.get_data_type(aty) if not isinstance(aty, types.Boolean) else llvmlite.ir.IntType(1)) ptr = cgutils.alloca_once(builder, typ) builder.store(builder.bitcast(ptr, byte_ptr_t), dst) # ---------------------------------------------------------------------------- # Next, we prepare the individual dimension info recorded in gu_signature sig_dim_dict = {} occurrences = [] occurrences = [sched_sig[0]] sig_dim_dict[sched_sig[0]] = context.get_constant(types.intp, 2 * num_dim) assert len(expr_args) == len(all_args) assert len(expr_args) == len(expr_arg_types) assert len(expr_args) == len(sin + sout) assert len(expr_args) == len(outer_sig.args[1:]) for var, arg, aty, gu_sig in zip(expr_args, all_args, expr_arg_types, sin + sout): if isinstance(aty, types.npytypes.Array): i = aty.ndim - len(gu_sig) else: i = 0 if config.DEBUG_ARRAY_OPT: print("var =", var, "gu_sig =", gu_sig, "type =", aty, "i =", i) for dim_sym in gu_sig: if config.DEBUG_ARRAY_OPT: print("var = ", var, " type = ", aty) if var in races: sig_dim_dict[dim_sym] = context.get_constant(types.intp, 1) else: ary = context.make_array(aty)(context, builder, arg) shapes = cgutils.unpack_tuple(builder, ary.shape, aty.ndim) sig_dim_dict[dim_sym] = shapes[i] if not (dim_sym in occurrences): if config.DEBUG_ARRAY_OPT: print("dim_sym = ", dim_sym, ", i = ", i) cgutils.printf(builder, dim_sym + " = %d\n", sig_dim_dict[dim_sym]) occurrences.append(dim_sym) i = i + 1 # ---------------------------------------------------------------------------- # Prepare shapes, which is a single number (outer loop size), followed by # the size of individual shape variables. nshapes = len(sig_dim_dict) + 1 shapes = cgutils.alloca_once(builder, intp_t, size=nshapes, name="pshape") # For now, outer loop size is the same as number of threads builder.store(num_divisions, shapes) # Individual shape variables go next i = 1 for dim_sym in occurrences: if config.DEBUG_ARRAY_OPT: cgutils.printf(builder, dim_sym + " = %d\n", sig_dim_dict[dim_sym]) builder.store( sig_dim_dict[dim_sym], builder.gep( shapes, [ context.get_constant( types.intp, i)])) i = i + 1 # ---------------------------------------------------------------------------- # Prepare steps for each argument. Note that all steps are counted in # bytes. num_steps = num_args + 1 + len(array_strides) steps = cgutils.alloca_once( builder, intp_t, size=context.get_constant( types.intp, num_steps), name="psteps") # First goes the step size for sched, which is 2 * num_dim builder.store(context.get_constant(types.intp, 2 * num_dim * sizeof_intp), steps) # The steps for all others are 0, except for reduction results. for i in range(num_args): # steps are strides from one thread to the next stepsize = zero dst = builder.gep(steps, [context.get_constant(types.intp, 1 + i)]) builder.store(stepsize, dst) for j in range(len(array_strides)): dst = builder.gep( steps, [ context.get_constant( types.intp, 1 + num_args + j)]) builder.store(array_strides[j], dst) # ---------------------------------------------------------------------------- # prepare data data = cgutils.get_null_value(byte_ptr_t) fnty = llvmlite.ir.FunctionType(llvmlite.ir.VoidType(), [byte_ptr_ptr_t, intp_ptr_t, intp_ptr_t, byte_ptr_t]) fn = cgutils.get_or_insert_function(builder.module, fnty, wrapper_name) context.active_code_library.add_linking_library(info.library) if config.DEBUG_ARRAY_OPT: cgutils.printf(builder, "before calling kernel %p\n", fn) builder.call(fn, [args, shapes, steps, data]) if config.DEBUG_ARRAY_OPT: cgutils.printf(builder, "after calling kernel %p\n", fn) builder.call(set_chunksize, [current_chunksize]) # Deallocate the schedule's memory. dealloc_sched_fnty = llvmlite.ir.FunctionType(llvmlite.ir.VoidType(), [sched_ptr_type]) dealloc_sched_func = cgutils.get_or_insert_function( builder.module, dealloc_sched_fnty, name="deallocate_sched") builder.call(dealloc_sched_func, [builder.load(sched)]) for k, v in rv_to_arg_dict.items(): arg, rv_arg = v only_elem_ptr = builder.gep(rv_arg, [context.get_constant(types.intp, 0)]) builder.store(builder.load(only_elem_ptr), lowerer.getvar(k)) context.active_code_library.add_linking_library(cres.library)