def compile_hsa(pyfunc, return_type, args, debug):
# First compilation will trigger the initialization of the HSA backend.
from .descriptor import HSATargetDesc
typingctx = HSATargetDesc.typingctx
targetctx = HSATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
flags.set('no_cpython_wrapper')
flags.unset('nrt')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
# Linking depending libraries
# targetctx.link_dependencies(cres.llvm_module, cres.target_context.linking)
library = cres.library
library.finalize()
return cres
def compile(self, sig, locals={}, **targetoptions):
locs = self.locals.copy()
locs.update(locals)
topt = self.targetoptions.copy()
topt.update(targetoptions)
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, topt)
flags.set("no_compile")
flags.set("no_cpython_wrapper")
flags.set("error_model", "numpy")
# Disable loop lifting
# The feature requires a real python function
flags.unset("enable_looplift")
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx, targetctx, self.py_func,
args=args, return_type=return_type,
flags=flags, locals=locals)
self.overloads[cres.signature] = cres
return cres
def compile(self, sig):
with self._compile_lock:
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
# (e.g. if another thread compiled it before we got the lock)
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
cres = compiler.compile_extra(
self.typingctx,
self.targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=self.locals,
)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def __init__(self, pyfunc, signature):
self.py_func = pyfunc
self.signature = signature
self.name = pyfunc.__name__
# recreate for each UDF, as linking is destructive to the
# precompiled module
impala_typing = impala_typing_context()
impala_targets = ImpalaTargetContext(impala_typing)
args, return_type = sigutils.normalize_signature(signature)
flags = Flags()
flags.set('no_compile')
self._cres = compile_extra(typingctx=impala_typing,
targetctx=impala_targets, func=pyfunc,
args=args, return_type=return_type,
flags=flags, locals={})
llvm_func = impala_targets.finalize(self._cres.llvm_func, return_type,
args)
self.llvm_func = llvm_func
# numba_module = llvm_func.module
self.llvm_module = llvm_func.module
# link in the precompiled module
# bc it's destructive, load a fresh version
precompiled = lc.Module.from_bitcode(
pkgutil.get_data("impala.udf", "precompiled/impyla.bc"))
self.llvm_module.link_in(precompiled)
def compile(self, sig, locals={}, **targetoptions):
with self._compile_lock():
locs = self.locals.copy()
locs.update(locals)
topt = self.targetoptions.copy()
topt.update(targetoptions)
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, topt)
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exist.
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing
cres = compiler.compile_extra(self.typingctx,
self.targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=locs)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def compile_cuda(pyfunc, return_type, args, debug, inline):
# First compilation will trigger the initialization of the CUDA backend.
from .descriptor import CUDATargetDesc
typingctx = CUDATargetDesc.typingctx
targetctx = CUDATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
flags.set('no_cpython_wrapper')
if debug:
flags.set('boundcheck')
if inline:
flags.set('forceinline')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
library = cres.library
library.finalize()
return cres
def compile_cuda(pyfunc, return_type, args, debug):
# First compilation will trigger the initialization of the CUDA backend.
from .descriptor import CUDATargetDesc
typingctx = CUDATargetDesc.typingctx
targetctx = CUDATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
# Linking depending libraries
targetctx.link_dependencies(cres.llvm_module, cres.target_context.linking)
# Fix global naming
for gv in cres.llvm_module.global_variables:
if '.' in gv.name:
gv.name = gv.name.replace('.', '_')
return cres
def test_use_car_move(self):
tyctx = typing.Context()
tyctx.insert_class(Car, self.carattrs)
cgctx = CPUContext(tyctx)
cgctx.insert_class(Car, self.carattrs)
car_object = types.Object(Car)
argtys = (car_object, types.int32)
flags = compiler.Flags()
cr = compiler.compile_extra(tyctx, cgctx, use_car_move, args=argtys,
return_type=None, flags=flags, locals={})
func = cr.entry_point
if cr.typing_error:
raise cr.typing_error
car1 = Car(value=123)
car2 = Car(value=123)
self.assertEqual(use_car_move(car1, 321), func(car2, 321))
def bm_python():
use_car_move(car1, 321)
def bm_numba():
func(car2, 321)
python = utils.benchmark(bm_python, maxsec=.1)
numba = utils.benchmark(bm_numba, maxsec=.1)
print(python)
print(numba)
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
"""
self.exported_function_types = {}
self.function_environments = {}
codegen = self.context.codegen()
library = codegen.create_library(self.module_name)
# Generate IR for all exported functions
flags = Flags()
flags.set("no_compile")
if not self.export_python_wrap:
flags.set("no_cpython_wrapper")
if self.use_nrt:
flags.set("nrt")
# Compile NRT helpers
nrt_module, _ = atomicops.create_nrt_module(self.context)
library.add_ir_module(nrt_module)
for entry in self.export_entries:
cres = compile_extra(self.typing_context,
self.context,
entry.function,
entry.signature.args,
entry.signature.return_type,
flags,
locals={},
library=library)
func_name = cres.fndesc.llvm_func_name
llvm_func = cres.library.get_function(func_name)
if self.export_python_wrap:
llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = cres.fndesc.llvm_cpython_wrapper_name
wrapper = cres.library.get_function(wrappername)
wrapper.name = self._mangle_method_symbol(entry.symbol)
wrapper.linkage = lc.LINKAGE_EXTERNAL
fnty = cres.target_context.call_conv.get_function_type(
cres.fndesc.restype, cres.fndesc.argtypes)
self.exported_function_types[entry] = fnty
self.function_environments[entry] = cres.environment
else:
llvm_func.name = entry.symbol
self.dll_exports.append(entry.symbol)
if self.export_python_wrap:
wrapper_module = library.create_ir_module("wrapper")
self._emit_python_wrapper(wrapper_module)
library.add_ir_module(wrapper_module)
# Hide all functions in the DLL except those explicitly exported
library.finalize()
for fn in library.get_defined_functions():
if fn.name not in self.dll_exports:
fn.visibility = "hidden"
return library
def compile_cuda(pyfunc, return_type, args, debug, inline):
# First compilation will trigger the initialization of the CUDA backend.
from .descriptor import CUDATargetDesc
typingctx = CUDATargetDesc.typingctx
targetctx = CUDATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
flags.set('no_cpython_wrapper')
if debug:
flags.set('boundcheck')
if inline:
flags.set('forceinline')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
library = cres.library
library.finalize()
return cres
def compile_hsa(pyfunc, return_type, args, debug):
# First compilation will trigger the initialization of the HSA backend.
from .descriptor import HSATargetDesc
typingctx = HSATargetDesc.typingctx
targetctx = HSATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
flags.set('no_cpython_wrapper')
flags.unset('nrt')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
# Linking depending libraries
# targetctx.link_dependencies(cres.llvm_module, cres.target_context.linking)
library = cres.library
library.finalize()
return cres
def compile(self, sig):
with self._compile_lock:
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
# (e.g. if another thread compiled it before we got the lock)
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
cres = compiler.compile_extra(self.typingctx,
self.targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=self.locals)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def compile(self, sig, locals={}, **targetoptions):
locs = self.locals.copy()
locs.update(locals)
topt = self.targetoptions.copy()
topt.update(targetoptions)
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, topt)
flags.set("no_compile")
flags.set("no_cpython_wrapper")
# Disable loop lifting
# The feature requires a real python function
flags.unset("enable_looplift")
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx,
targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=locals)
self.overloads[cres.signature] = cres
return cres
def __init__(self, pyfunc, signature):
self.py_func = pyfunc
self.signature = signature
self.name = pyfunc.__name__
# recreate for each UDF, as linking is destructive to the
# precompiled module
impala_typing = impala_typing_context()
impala_targets = ImpalaTargetContext(impala_typing)
args, return_type = sigutils.normalize_signature(signature)
flags = Flags()
flags.set('no_compile')
self._cres = compile_extra(typingctx=impala_typing,
targetctx=impala_targets,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
llvm_func = impala_targets.finalize(self._cres.llvm_func, return_type,
args)
self.llvm_func = llvm_func
# numba_module = llvm_func.module
self.llvm_module = llvm_func.module
# link in the precompiled module
# bc it's destructive, load a fresh version
precompiled = lc.Module.from_bitcode(
pkgutil.get_data("impala.udf", "precompiled/impyla.bc"))
self.llvm_module.link_in(precompiled)
def compile_cuda(pyfunc, return_type, args, debug):
# First compilation will trigger the initialization of the CUDA backend.
from .descriptor import CUDATargetDesc
typingctx = CUDATargetDesc.typingctx
targetctx = CUDATargetDesc.targetctx
# TODO handle debug flag
flags = compiler.Flags()
# Do not compile (generate native code), just lower (to LLVM)
flags.set('no_compile')
# Run compilation pipeline
cres = compiler.compile_extra(typingctx=typingctx,
targetctx=targetctx,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
# Linking depending libraries
targetctx.link_dependencies(cres.llvm_module, cres.target_context.linking)
# Fix global naming
for gv in cres.llvm_module.global_variables:
if '.' in gv.name:
gv.name = gv.name.replace('.', '_')
return cres
def compile(self, sig, locals={}, **targetoptions):
with self._compile_lock():
locs = self.locals.copy()
locs.update(locals)
topt = self.targetoptions.copy()
topt.update(targetoptions)
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, topt)
glctx = self.targetdescr
typingctx = glctx.typing_context
targetctx = glctx.target_context
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exist.
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing.entry_point
cres = compiler.compile_extra(typingctx, targetctx, self.py_func,
args=args, return_type=return_type,
flags=flags, locals=locs)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
return cres.entry_point
def compile(self, sig, locals={}, **targetoptions):
locs = self.locals.copy()
locs.update(locals)
topt = self.targetoptions.copy()
topt.update(targetoptions)
if topt.get('nopython', True) == False:
raise TypeError("nopython option must be False")
topt['nopython'] = True
flags = compiler.Flags()
flags.set("no_compile")
self.targetdescr.options.parse_as_flags(flags, topt)
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx, targetctx, self.py_func,
args=args, return_type=return_type,
flags=flags, locals=locals)
self.overloads[cres.signature] = cres
return cres
def compile(self, sig):
with self._compile_lock:
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing
# Try to load from disk cache
cres = self._cache.load_overload(sig, self.targetctx)
if cres is not None:
# XXX fold this in add_overload()? (also see compiler.py)
if not cres.objectmode and not cres.interpmode:
self.targetctx.insert_user_function(cres.entry_point,
cres.fndesc, [cres.library])
self.add_overload(cres)
return cres.entry_point
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
cres = compiler.compile_extra(self.typingctx, self.targetctx,
self.py_func,
args=args, return_type=return_type,
flags=flags, locals=self.locals)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
self._cache.save_overload(sig, cres)
return cres.entry_point
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
"""
self.exported_function_types = {}
self.function_environments = {}
self.environment_gvs = {}
codegen = self.context.codegen()
library = codegen.create_library(self.module_name)
# Generate IR for all exported functions
flags = Flags()
flags.set("no_compile")
if not self.export_python_wrap:
flags.set("no_cpython_wrapper")
if self.use_nrt:
flags.set("nrt")
# Compile NRT helpers
nrt_module, _ = nrtdynmod.create_nrt_module(self.context)
library.add_ir_module(nrt_module)
for entry in self.export_entries:
cres = compile_extra(self.typing_context, self.context,
entry.function,
entry.signature.args,
entry.signature.return_type, flags,
locals={}, library=library)
func_name = cres.fndesc.llvm_func_name
llvm_func = cres.library.get_function(func_name)
if self.export_python_wrap:
llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = cres.fndesc.llvm_cpython_wrapper_name
wrapper = cres.library.get_function(wrappername)
wrapper.name = self._mangle_method_symbol(entry.symbol)
wrapper.linkage = lc.LINKAGE_EXTERNAL
fnty = cres.target_context.call_conv.get_function_type(
cres.fndesc.restype, cres.fndesc.argtypes)
self.exported_function_types[entry] = fnty
self.function_environments[entry] = cres.environment
self.environment_gvs[entry] = cres.fndesc.env_name
else:
llvm_func.name = entry.symbol
self.dll_exports.append(entry.symbol)
if self.export_python_wrap:
wrapper_module = library.create_ir_module("wrapper")
self._emit_python_wrapper(wrapper_module)
library.add_ir_module(wrapper_module)
# Hide all functions in the DLL except those explicitly exported
library.finalize()
for fn in library.get_defined_functions():
if fn.name not in self.dll_exports:
fn.visibility = "hidden"
return library
def test_scalar(self):
flags = Flags()
global cnd_jitted
cr1 = compile_isolated(cnd, (types.float64, ))
cnd_jitted = cr1.entry_point
tyctx = cr1.typing_context
ctx = cr1.target_context
ctx.dynamic_map_function(cnd_jitted)
tyctx.insert_user_function(cnd_jitted,
ctx.get_user_function(cnd_jitted))
array = types.Array(types.float64, 1, 'C')
argtys = (array, ) * 5 + (types.float64, types.float64)
cr2 = compile_extra(tyctx,
ctx,
blackscholes_scalar_jitted,
args=argtys,
return_type=None,
flags=flags,
locals={})
jitted_bs = cr2.entry_point
OPT_N = 400
iterations = 10
callResultGold = np.zeros(OPT_N)
putResultGold = np.zeros(OPT_N)
callResultNumba = np.zeros(OPT_N)
putResultNumba = np.zeros(OPT_N)
stockPrice = randfloat(np.random.random(OPT_N), 5.0, 30.0)
optionStrike = randfloat(np.random.random(OPT_N), 1.0, 100.0)
optionYears = randfloat(np.random.random(OPT_N), 0.25, 10.0)
args = stockPrice, optionStrike, optionYears, RISKFREE, VOLATILITY
ts = timer()
for i in range(iterations):
blackscholes_scalar(callResultGold, putResultGold, *args)
te = timer()
pytime = te - ts
ts = timer()
for i in range(iterations):
jitted_bs(callResultNumba, putResultNumba, *args)
te = timer()
jittime = te - ts
print("Python", pytime)
print("Numba", jittime)
print("Speedup: %s" % (pytime / jittime))
delta = np.abs(callResultGold - callResultNumba)
L1norm = delta.sum() / np.abs(callResultGold).sum()
print("L1 norm: %E" % L1norm)
print("Max absolute error: %E" % delta.max())
self.assertAlmostEqual(delta.max(), 0)
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
Resets the export environment afterwards.
"""
self.exported_signatures = export_registry
# Create new module containing everything
llvm_module = lc.Module.new(self.module_name)
# Compile all exported functions
typing_ctx = CPUTarget.typing_context
# TODO Use non JIT-ing target
target_ctx = CPUTarget.target_context
modules = []
flags = Flags()
if not self.export_python_wrap:
flags.set("no_compile")
for entry in self.exported_signatures:
cres = compile_extra(typing_ctx,
target_ctx,
entry.function,
entry.signature.args,
entry.signature.return_type,
flags,
locals={})
if self.export_python_wrap:
module = cres.llvm_func.module
cres.llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = "wrapper." + cres.llvm_func.name
wrapper = module.get_function_named(wrappername)
wrapper.name = entry.symbol
else:
cres.llvm_func.name = entry.symbol
modules.append(cres.llvm_module)
# Link all exported functions
for mod in modules:
llvm_module.link_in(mod, preserve=self.export_python_wrap)
# Optimize
tm = le.TargetMachine.new(opt=3)
pms = lp.build_pass_managers(tm=tm,
opt=3,
loop_vectorize=True,
fpm=False)
pms.pm.run(llvm_module)
if self.export_python_wrap:
self._emit_python_wrapper(llvm_module)
del self.exported_signatures[:]
print(llvm_module)
return llvm_module
def test_scalar(self):
flags = Flags()
# Compile the inner function
global cnd_jitted
cr1 = compile_isolated(cnd, (types.float64,))
cnd_jitted = cr1.entry_point
# Manually type the compiled function for calling into
tyctx = cr1.typing_context
ctx = cr1.target_context
signature = typing.make_concrete_template("cnd_jitted", cnd_jitted,
[cr1.signature])
tyctx.insert_user_function(cnd_jitted, signature)
# Compile the outer function
array = types.Array(types.float64, 1, 'C')
argtys = (array,) * 5 + (types.float64, types.float64)
cr2 = compile_extra(tyctx, ctx, blackscholes_scalar_jitted,
args=argtys, return_type=None, flags=flags,
locals={})
jitted_bs = cr2.entry_point
OPT_N = 400
iterations = 10
callResultGold = np.zeros(OPT_N)
putResultGold = np.zeros(OPT_N)
callResultNumba = np.zeros(OPT_N)
putResultNumba = np.zeros(OPT_N)
stockPrice = randfloat(self.random.random_sample(OPT_N), 5.0, 30.0)
optionStrike = randfloat(self.random.random_sample(OPT_N), 1.0, 100.0)
optionYears = randfloat(self.random.random_sample(OPT_N), 0.25, 10.0)
args = stockPrice, optionStrike, optionYears, RISKFREE, VOLATILITY
ts = timer()
for i in range(iterations):
blackscholes_scalar(callResultGold, putResultGold, *args)
te = timer()
pytime = te - ts
ts = timer()
for i in range(iterations):
jitted_bs(callResultNumba, putResultNumba, *args)
te = timer()
jittime = te - ts
print("Python", pytime)
print("Numba", jittime)
print("Speedup: %s" % (pytime / jittime))
delta = np.abs(callResultGold - callResultNumba)
L1norm = delta.sum() / np.abs(callResultGold).sum()
print("L1 norm: %E" % L1norm)
print("Max absolute error: %E" % delta.max())
self.assertAlmostEqual(delta.max(), 0)
def compile(self, py_func, sig, library=None):
if library is None:
library = self.library
args, return_type = sigutils.normalize_signature(sig)
return compiler.compile_extra(self.typingctx, self.targetctx,
py_func, args=args,
return_type=return_type,
flags=self.flags, locals=self.locals,
library=library)
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
Resets the export environment afterwards.
"""
self.exported_signatures = export_registry
# Create new module containing everything
llvm_module = lc.Module.new(self.module_name)
# Compile all exported functions
typing_ctx = CPUTarget.typing_context
# TODO Use non JIT-ing target
target_ctx = CPUTarget.target_context
modules = []
flags = Flags()
if not self.export_python_wrap:
flags.set("no_compile")
for entry in self.exported_signatures:
cres = compile_extra(typing_ctx, target_ctx, entry.function,
entry.signature.args,
entry.signature.return_type, flags,
locals={})
if self.export_python_wrap:
module = cres.llvm_func.module
cres.llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = "wrapper." + cres.llvm_func.name
wrapper = module.get_function_named(wrappername)
wrapper.name = entry.symbol
else:
cres.llvm_func.name = entry.symbol
modules.append(cres.llvm_module)
# Link all exported functions
for mod in modules:
llvm_module.link_in(mod, preserve=self.export_python_wrap)
# Optimize
tm = le.TargetMachine.new(opt=3)
pms = lp.build_pass_managers(tm=tm, opt=3, loop_vectorize=True,
fpm=False)
pms.pm.run(llvm_module)
if self.export_python_wrap:
self._emit_python_wrapper(llvm_module)
#del self.exported_signatures[:]
print(llvm_module)
return llvm_module
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
Resets the export environment afterwards.
"""
self.exported_signatures = export_registry
self.exported_function_types = {}
typing_ctx = CPUTarget.typing_context
target_ctx = CPUTarget.target_context.subtarget(aot_mode=True)
codegen = target_ctx.aot_codegen(self.module_name)
library = codegen.create_library(self.module_name)
# Generate IR for all exported functions
flags = Flags()
flags.set("no_compile")
for entry in self.exported_signatures:
cres = compile_extra(typing_ctx,
target_ctx,
entry.function,
entry.signature.args,
entry.signature.return_type,
flags,
locals={},
library=library)
func_name = cres.fndesc.llvm_func_name
llvm_func = cres.library.get_function(func_name)
if self.export_python_wrap:
# XXX: unsupported (necessary?)
llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = cres.fndesc.llvm_cpython_wrapper_name
wrapper = cres.library.get_function(wrappername)
wrapper.name = entry.symbol
wrapper.linkage = lc.LINKAGE_EXTERNAL
fnty = cres.target_context.call_conv.get_function_type(
cres.fndesc.restype, cres.fndesc.argtypes)
self.exported_function_types[entry] = fnty
else:
llvm_func.linkage = lc.LINKAGE_EXTERNAL
llvm_func.name = entry.symbol
if self.export_python_wrap:
wrapper_module = library.create_ir_module("wrapper")
self._emit_python_wrapper(wrapper_module)
library.add_ir_module(wrapper_module)
return library
def compile(self, func, args, return_type=None, flags=DEFAULT_FLAGS):
"""
Compile the function or retrieve an already compiled result
from the cache.
"""
cache_key = (func, args, return_type, flags)
try:
cr = self.cr_cache[cache_key]
except KeyError:
cr = compile_extra(self.typingctx, self.targetctx, func,
args, return_type, flags, locals={})
self.cr_cache[cache_key] = cr
return cr
def compile(self, func, args, return_type=None, flags=DEFAULT_FLAGS):
"""
Compile the function or retrieve an already compiled result
from the cache.
"""
cache_key = (func, args, return_type, flags)
try:
cr = self.cr_cache[cache_key]
except KeyError:
cr = compile_extra(self.typingctx, self.targetctx, func,
args, return_type, flags, locals={})
self.cr_cache[cache_key] = cr
return cr
def compile(self, args, return_type):
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
impl = self._get_implementation(args, {})
cres = compiler.compile_extra(self.targetdescr.typing_context,
self.targetdescr.target_context,
impl,
args=args, return_type=return_type,
flags=flags, locals=self.locals)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
return cres
def compile(self, args, return_type):
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
impl = self._get_implementation(args, {})
cres = compiler.compile_extra(self.targetdescr.typing_context,
self.targetdescr.target_context,
impl,
args=args, return_type=return_type,
flags=flags, locals=self.locals)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
return cres
def test_scalar(self):
flags = Flags()
# Compile the inner function
global cnd_jitted
cr1 = compile_isolated(cnd, (types.float64, ))
cnd_jitted = cr1.entry_point
# Manually type the compiled function for calling into
tyctx = cr1.typing_context
ctx = cr1.target_context
signature = typing.make_concrete_template("cnd_jitted", cnd_jitted,
[cr1.signature])
tyctx.insert_user_function(cnd_jitted, signature)
# Compile the outer function
array = types.Array(types.float64, 1, 'C')
argtys = (array, ) * 5 + (types.float64, types.float64)
cr2 = compile_extra(tyctx,
ctx,
blackscholes_scalar_jitted,
args=argtys,
return_type=None,
flags=flags,
locals={})
jitted_bs = cr2.entry_point
OPT_N = 400
iterations = 10
callResultGold = np.zeros(OPT_N)
putResultGold = np.zeros(OPT_N)
callResultNumba = np.zeros(OPT_N)
putResultNumba = np.zeros(OPT_N)
stockPrice = randfloat(self.random.random_sample(OPT_N), 5.0, 30.0)
optionStrike = randfloat(self.random.random_sample(OPT_N), 1.0, 100.0)
optionYears = randfloat(self.random.random_sample(OPT_N), 0.25, 10.0)
args = stockPrice, optionStrike, optionYears, RISKFREE, VOLATILITY
blackscholes_scalar(callResultGold, putResultGold, *args)
jitted_bs(callResultNumba, putResultNumba, *args)
delta = np.abs(callResultGold - callResultNumba)
L1norm = delta.sum() / np.abs(callResultGold).sum()
print("L1 norm: %E" % L1norm)
print("Max absolute error: %E" % delta.max())
self.assertAlmostEqual(delta.max(), 0)
def _cull_exports(self):
"""Read all the exported functions/modules in the translator
environment, and join them into a single LLVM module.
Resets the export environment afterwards.
"""
self.exported_signatures = export_registry
self.exported_function_types = {}
typing_ctx = CPUTarget.typing_context
target_ctx = CPUTarget.target_context
codegen = target_ctx.aot_codegen(self.module_name)
library = codegen.create_library(self.module_name)
# Generate IR for all exported functions
flags = Flags()
flags.set("no_compile")
for entry in self.exported_signatures:
cres = compile_extra(typing_ctx, target_ctx, entry.function,
entry.signature.args,
entry.signature.return_type, flags,
locals={}, library=library)
func_name = cres.fndesc.llvm_func_name
llvm_func = cres.library.get_function(func_name)
if self.export_python_wrap:
# XXX: unsupported (necessary?)
llvm_func.linkage = lc.LINKAGE_INTERNAL
wrappername = cres.fndesc.llvm_cpython_wrapper_name
wrapper = cres.library.get_function(wrappername)
wrapper.name = entry.symbol
wrapper.linkage = lc.LINKAGE_EXTERNAL
fnty = cres.target_context.call_conv.get_function_type(
cres.fndesc.restype, cres.fndesc.argtypes)
self.exported_function_types[entry] = fnty
else:
llvm_func.linkage = lc.LINKAGE_EXTERNAL
llvm_func.name = entry.symbol
if self.export_python_wrap:
wrapper_module = library.create_ir_module("wrapper")
self._emit_python_wrapper(wrapper_module)
library.add_ir_module(wrapper_module)
return library
def __init__(self, pyfunc, signature):
self.py_func = pyfunc
self.signature = signature
self.name = pyfunc.__name__
args, return_type = sigutils.normalize_signature(signature)
flags = Flags()
flags.set('no_compile')
self._cres = compile_extra(typingctx=impala_typing,
targetctx=impala_targets, func=pyfunc,
args=args, return_type=return_type,
flags=flags, locals={})
llvm_func = impala_targets.finalize(self._cres.llvm_func, return_type,
args)
self.llvm_func = llvm_func
self.llvm_module = llvm_func.module
def _compile_core(self, sig, flags, locals):
"""
Trigger the compiler on the core function or load a previously
compiled version from the cache. Returns the CompileResult.
"""
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
cres = self.cache.load_overload(sig, targetctx)
if cres is not None:
# Use cached version
return cres
# Compile
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx, targetctx, self.py_func,
args=args, return_type=return_type,
flags=flags, locals=locals)
self.cache.save_overload(sig, cres)
return cres
def compile(self, func, args, return_type=None, flags=DEFAULT_FLAGS):
"""
Compile the function or retrieve an already compiled result
from the cache.
"""
from numba.targets.registry import cpu_target
cache_key = (func, args, return_type, flags)
try:
cr = self.cr_cache[cache_key]
except KeyError:
# Register the contexts in case for nested @jit or @overload calls
# (same as compile_isolated())
with cpu_target.nested_context(self.typingctx, self.targetctx):
cr = compile_extra(self.typingctx, self.targetctx, func,
args, return_type, flags, locals={})
self.cr_cache[cache_key] = cr
return cr
def compile(self, func, args, return_type=None, flags=DEFAULT_FLAGS):
"""
Compile the function or retrieve an already compiled result
from the cache.
"""
from numba.targets.registry import cpu_target
cache_key = (func, args, return_type, flags)
try:
cr = self.cr_cache[cache_key]
except KeyError:
# Register the contexts in case for nested @jit or @overload calls
# (same as compile_isolated())
with cpu_target.nested_context(self.typingctx, self.targetctx):
cr = compile_extra(self.typingctx, self.targetctx, func,
args, return_type, flags, locals={})
self.cr_cache[cache_key] = cr
return cr
def __init__(self, pyfunc, signature):
self.py_func = pyfunc
self.signature = signature
self.name = pyfunc.__name__
args, return_type = sigutils.normalize_signature(signature)
flags = Flags()
flags.set('no_compile')
self._cres = compile_extra(typingctx=impala_typing,
targetctx=impala_targets,
func=pyfunc,
args=args,
return_type=return_type,
flags=flags,
locals={})
llvm_func = impala_targets.finalize(self._cres.llvm_func, return_type,
args)
self.llvm_func = llvm_func
self.llvm_module = llvm_func.module
def _compile_core(self, sig, flags, locals):
"""
Trigger the compiler on the core function or load a previously
compiled version from the cache. Returns the CompileResult.
"""
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
@contextmanager
def store_overloads_on_success():
# use to ensure overloads are stored on success
try:
yield
except:
raise
else:
exists = self.overloads.get(cres.signature)
if exists is None:
self.overloads[cres.signature] = cres
# Use cache and compiler in a critical section
with compiler.lock_compiler:
with store_overloads_on_success():
# attempt look up of existing
cres = self.cache.load_overload(sig, targetctx)
if cres is not None:
return cres
# Compile
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx,
targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=locals)
# cache lookup failed before so safe to save
self.cache.save_overload(sig, cres)
return cres
def _compile_core(self, sig, flags, locals):
"""
Trigger the compiler on the core function or load a previously
compiled version from the cache. Returns the CompileResult.
"""
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
# Use cache and compiler in a critical section
with compiler.lock_compiler:
cres = self.cache.load_overload(sig, targetctx)
if cres is not None:
# Use cached version
return cres
# Compile
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx, targetctx, self.py_func,
args=args, return_type=return_type,
flags=flags, locals=locals)
self.cache.save_overload(sig, cres)
return cres
def _compile_core(self, sig, flags, locals):
"""
Trigger the compiler on the core function or load a previously
compiled version from the cache. Returns the CompileResult.
"""
typingctx = self.targetdescr.typing_context
targetctx = self.targetdescr.target_context
@contextmanager
def store_overloads_on_success():
# use to ensure overloads are stored on success
try:
yield
except:
raise
else:
exists = self.overloads.get(cres.signature)
if exists is None:
self.overloads[cres.signature] = cres
# Use cache and compiler in a critical section
with global_compiler_lock:
with store_overloads_on_success():
# attempt look up of existing
cres = self.cache.load_overload(sig, targetctx)
if cres is not None:
return cres
# Compile
args, return_type = sigutils.normalize_signature(sig)
cres = compiler.compile_extra(typingctx, targetctx,
self.py_func, args=args,
return_type=return_type,
flags=flags, locals=locals)
# cache lookup failed before so safe to save
self.cache.save_overload(sig, cres)
return cres
def compile(self, sig):
with self._compile_lock:
args, return_type = sigutils.normalize_signature(sig)
# Don't recompile if signature already exists
existing = self.overloads.get(tuple(args))
if existing is not None:
return existing
# Try to load from disk cache
cres = self._cache.load_overload(sig, self.targetctx)
if cres is not None:
# XXX fold this in add_overload()? (also see compiler.py)
if not cres.objectmode and not cres.interpmode:
self.targetctx.insert_user_function(
cres.entry_point, cres.fndesc, [cres.library])
self.add_overload(cres)
return cres.entry_point
flags = compiler.Flags()
self.targetdescr.options.parse_as_flags(flags, self.targetoptions)
cres = compiler.compile_extra(self.typingctx,
self.targetctx,
self.py_func,
args=args,
return_type=return_type,
flags=flags,
locals=self.locals)
# Check typing error if object mode is used
if cres.typing_error is not None and not flags.enable_pyobject:
raise cres.typing_error
self.add_overload(cres)
self._cache.save_overload(sig, cres)
return cres.entry_point