def __init__(self, context, library, fndesc, interp):
self.context = context
self.library = library
self.fndesc = fndesc
self.blocks = utils.SortedMap(utils.iteritems(interp.blocks))
# Initialize LLVM
self.module = self.library.create_ir_module(self.fndesc.unique_name)
# Python execution environment (will be available to the compiled
# function).
self.env = _dynfunc.Environment(
globals=self.fndesc.lookup_module().__dict__)
# Mapping of error codes to exception classes or instances
self.exceptions = {}
# Setup function
self.function = context.declare_function(self.module, fndesc)
self.entry_block = self.function.append_basic_block('entry')
self.builder = Builder.new(self.entry_block)
# Internal states
self.blkmap = {}
self.varmap = {}
self.firstblk = min(self.blocks.keys())
self.loc = -1
# Subclass initialization
self.init()
def __init__(self, context, library, fndesc, interp):
self.context = context
self.library = library
self.fndesc = fndesc
self.blocks = utils.SortedMap(utils.iteritems(interp.blocks))
self.interp = interp
self.call_conv = context.call_conv
# Initialize LLVM
self.module = self.library.create_ir_module(self.fndesc.unique_name)
# Python execution environment (will be available to the compiled
# function).
self.env = _dynfunc.Environment(
globals=self.fndesc.lookup_module().__dict__)
# Setup function
self.function = context.declare_function(self.module, fndesc)
self.entry_block = self.function.append_basic_block('entry')
self.builder = Builder.new(self.entry_block)
self.call_helper = self.call_conv.init_call_helper(self.builder)
# Internal states
self.blkmap = {}
self.varmap = {}
self.firstblk = min(self.blocks.keys())
self.loc = -1
# Subclass initialization
self.init()
def lower_finalize_func(self, lower):
"""
Lower the generator's finalizer.
"""
fnty = Type.function(Type.void(), [self.context.get_value_type(self.gentype)])
function = lower.module.get_or_insert_function(fnty, name=self.gendesc.llvm_finalizer_name)
entry_block = function.append_basic_block("entry")
builder = Builder.new(entry_block)
genptrty = self.context.get_value_type(self.gentype)
genptr = builder.bitcast(function.args[0], genptrty)
self.lower_finalize_func_body(builder, genptr)
def lower_finalize_func(self, lower):
"""
Lower the generator's finalizer.
"""
fnty = Type.function(Type.void(),
[self.context.get_value_type(self.gentype)])
function = lower.module.get_or_insert_function(
fnty, name=self.gendesc.llvm_finalizer_name)
entry_block = function.append_basic_block('entry')
builder = Builder.new(entry_block)
genptrty = self.context.get_value_type(self.gentype)
genptr = builder.bitcast(function.args[0], genptrty)
self.lower_finalize_func_body(builder, genptr)
def test_nvvm_from_llvm(self):
m = Module.new("test_nvvm_from_llvm")
fty = Type.function(Type.void(), [Type.int()])
kernel = m.add_function(fty, name='mycudakernel')
bldr = Builder.new(kernel.append_basic_block('entry'))
bldr.ret_void()
set_cuda_kernel(kernel)
fix_data_layout(m)
ptx = llvm_to_ptx(str(m)).decode('utf8')
self.assertTrue('mycudakernel' in ptx)
if is64bit:
self.assertTrue('.address_size 64' in ptx)
else:
self.assertTrue('.address_size 32' in ptx)
def test_nvvm_from_llvm(self):
m = Module.new("test_nvvm_from_llvm")
fty = Type.function(Type.void(), [Type.int()])
kernel = m.add_function(fty, name='mycudakernel')
bldr = Builder.new(kernel.append_basic_block('entry'))
bldr.ret_void()
set_cuda_kernel(kernel)
fix_data_layout(m)
ptx = llvm_to_ptx(str(m)).decode('utf8')
self.assertTrue('mycudakernel' in ptx)
if is64bit:
self.assertTrue('.address_size 64' in ptx)
else:
self.assertTrue('.address_size 32' in ptx)
def test_inline_rsqrt(self):
mod = Module.new(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, "cu_rsqrt")
bldr = Builder.new(fn.append_basic_block("entry"))
rsqrt_approx_fnty = Type.function(Type.float(), [Type.float()])
inlineasm = InlineAsm.get(rsqrt_approx_fnty, "rsqrt.approx.f32 $0, $1;", "=f,f", side_effect=True)
val = bldr.load(fn.args[0])
res = bldr.call(inlineasm, [val])
bldr.store(res, fn.args[0])
bldr.ret_void()
# generate ptx
nvvm.fix_data_layout(mod)
nvvm.set_cuda_kernel(fn)
nvvmir = str(mod)
ptx = nvvm.llvm_to_ptx(nvvmir)
self.assertTrue("rsqrt.approx.f32" in str(ptx))
def test_inline_rsqrt(self):
mod = Module.new(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, 'cu_rsqrt')
bldr = Builder.new(fn.append_basic_block('entry'))
rsqrt_approx_fnty = Type.function(Type.float(), [Type.float()])
inlineasm = InlineAsm.get(rsqrt_approx_fnty,
'rsqrt.approx.f32 $0, $1;',
'=f,f', side_effect=True)
val = bldr.load(fn.args[0])
res = bldr.call(inlineasm, [val])
bldr.store(res, fn.args[0])
bldr.ret_void()
# generate ptx
nvvm.fix_data_layout(mod)
nvvm.set_cuda_kernel(fn)
nvvmir = str(mod)
ptx = nvvm.llvm_to_ptx(nvvmir)
self.assertTrue('rsqrt.approx.f32' in str(ptx))
def build(self):
wrapname = "wrapper.%s" % self.func.name
# This is the signature of PyCFunctionWithKeywords
# (see CPython's methodobject.h)
pyobj = self.context.get_argument_type(types.pyobject)
wrapty = Type.function(pyobj, [pyobj, pyobj, pyobj])
wrapper = self.module.add_function(wrapty, name=wrapname)
builder = Builder.new(wrapper.append_basic_block('entry'))
# - `closure` will receive the `self` pointer stored in the
# PyCFunction object (see _dynfunc.c)
# - `args` and `kws` will receive the tuple and dict objects
# of positional and keyword arguments, respectively.
closure, args, kws = wrapper.args
closure.name = 'py_closure'
args.name = 'py_args'
kws.name = 'py_kws'
api = self.context.get_python_api(builder)
self.build_wrapper(api, builder, closure, args, kws)
return wrapper, api
def build_ufunc_wrapper(library, context, func, signature, objmode, envptr, env):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
intp_ptr_t, byte_ptr_t])
wrapper_module = library.create_ir_module('')
if objmode:
func_type = context.call_conv.get_function_type(
types.pyobject, [types.pyobject] * len(signature.args))
else:
func_type = context.call_conv.get_function_type(
signature.return_type, signature.args)
oldfunc = func
func = wrapper_module.add_function(func_type,
name=func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(UArrayArg(context, builder, arg_args, arg_steps, i, typ))
# Prepare output
out = UArrayArg(context, builder, arg_args, arg_steps, len(arrays),
signature.return_type)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
pyapi = context.get_python_api(builder)
if objmode:
# General loop
gil = pyapi.gil_ensure()
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_obj_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
pyapi, envptr, env)
pyapi.gil_release(gil)
builder.ret_void()
else:
with builder.if_else(unit_strided) as (is_unit_strided, is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as loop:
fastloop = build_fast_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
loop.index, pyapi)
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
pyapi)
builder.ret_void()
del builder
# Run optimizer
library.add_ir_module(wrapper_module)
wrapper = library.get_function(wrapper.name)
return wrapper
def build_ufunc_wrapper(library, context, func, signature, objmode, env):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(),
[byte_ptr_ptr_t, intp_ptr_t, intp_ptr_t, byte_ptr_t])
wrapper_module = library.create_ir_module('')
if objmode:
func_type = context.call_conv.get_function_type(
types.pyobject, [types.pyobject] * len(signature.args))
else:
func_type = context.call_conv.get_function_type(
signature.return_type, signature.args)
oldfunc = func
func = wrapper_module.add_function(func_type, name=func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
actual_args = context.call_conv.get_arguments(func)
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(
UArrayArg(context, builder, arg_args, arg_steps, i,
context.get_argument_type(typ)))
# Prepare output
valty = context.get_data_type(signature.return_type)
out = UArrayArg(context, builder, arg_args, arg_steps, len(actual_args),
valty)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
if objmode:
# General loop
pyapi = context.get_python_api(builder)
gil = pyapi.gil_ensure()
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_obj_loop_body(context, func, builder, arrays, out,
offsets, store_offset, signature,
pyapi, env)
pyapi.gil_release(gil)
builder.ret_void()
else:
with cgutils.ifelse(builder,
unit_strided) as (is_unit_strided, is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
fastloop = build_fast_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
ind)
builder.ret_void()
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature)
builder.ret_void()
builder.ret_void()
del builder
# Run optimizer
library.add_ir_module(wrapper_module)
wrapper = library.get_function(wrapper.name)
oldfunc.linkage = LINKAGE_INTERNAL
return wrapper
def build(self):
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = self.context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(
Type.void(), [byte_ptr_ptr_t, intp_ptr_t, intp_ptr_t, byte_ptr_t])
wrapper_module = self.library.create_ir_module('')
func_type = self.call_conv.get_function_type(self.fndesc.restype,
self.fndesc.argtypes)
func = wrapper_module.add_function(func_type, name=self.func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty,
"__gufunc__." + self.func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
# Unpack shapes
unique_syms = set()
for grp in (self.sin, self.sout):
for syms in grp:
unique_syms |= set(syms)
sym_map = {}
for syms in self.sin:
for s in syms:
if s not in sym_map:
sym_map[s] = len(sym_map)
sym_dim = {}
for s, i in sym_map.items():
sym_dim[s] = builder.load(
builder.gep(arg_dims,
[self.context.get_constant(types.intp, i + 1)]))
# Prepare inputs
arrays = []
step_offset = len(self.sin) + len(self.sout)
for i, (typ, sym) in enumerate(
zip(self.signature.args, self.sin + self.sout)):
ary = GUArrayArg(self.context, builder, arg_args, arg_dims,
arg_steps, i, step_offset, typ, sym, sym_dim)
if not ary.as_scalar:
step_offset += ary.ndim
arrays.append(ary)
bbreturn = cgutils.get_function(builder).append_basic_block('.return')
# Prologue
self.gen_prologue(builder)
# Loop
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
args = [a.array_value for a in arrays]
innercall, error = self.gen_loop_body(builder, func, args)
# If error, escape
cgutils.cbranch_or_continue(builder, error, bbreturn)
for a in arrays:
a.next(ind)
builder.branch(bbreturn)
builder.position_at_end(bbreturn)
# Epilogue
self.gen_epilogue(builder)
builder.ret_void()
self.library.add_ir_module(wrapper_module)
wrapper = self.library.get_function(wrapper.name)
# Set core function to internal so that it is not generated
self.func.linkage = LINKAGE_INTERNAL
return wrapper, self.env
def generate_kernel_wrapper(self, func, argtypes):
module = func.module
arginfo = self.get_arg_packer(argtypes)
argtys = list(arginfo.argument_types)
wrapfnty = Type.function(Type.void(), argtys)
wrapper_module = self.create_module("cuda.kernel.wrapper")
fnty = Type.function(Type.int(),
[self.call_conv.get_return_type(types.pyobject)] +
argtys)
func = wrapper_module.add_function(fnty, name=func.name)
wrapfn = wrapper_module.add_function(wrapfnty,
name="cudaPy_" + func.name)
builder = Builder.new(wrapfn.append_basic_block(''))
# Define error handling variables
def define_error_gv(postfix):
gv = wrapper_module.add_global_variable(Type.int(),
name=wrapfn.name + postfix)
gv.initializer = Constant.null(gv.type.pointee)
return gv
gv_exc = define_error_gv("__errcode__")
gv_tid = []
gv_ctaid = []
for i in 'xyz':
gv_tid.append(define_error_gv("__tid%s__" % i))
gv_ctaid.append(define_error_gv("__ctaid%s__" % i))
callargs = arginfo.from_arguments(builder, wrapfn.args)
status, _ = self.call_conv.call_function(builder, func, types.void,
argtypes, callargs)
# Check error status
with cgutils.if_likely(builder, status.is_ok):
builder.ret_void()
with builder.if_then(builder.not_(status.is_python_exc)):
# User exception raised
old = Constant.null(gv_exc.type.pointee)
# Use atomic cmpxchg to prevent rewriting the error status
# Only the first error is recorded
casfnty = lc.Type.function(old.type,
[gv_exc.type, old.type, old.type])
casfn = wrapper_module.add_function(casfnty,
name="___numba_cas_hack")
xchg = builder.call(casfn, [gv_exc, old, status.code])
changed = builder.icmp(ICMP_EQ, xchg, old)
# If the xchange is successful, save the thread ID.
sreg = nvvmutils.SRegBuilder(builder)
with builder.if_then(changed):
for dim, ptr, in zip("xyz", gv_tid):
val = sreg.tid(dim)
builder.store(val, ptr)
for dim, ptr, in zip("xyz", gv_ctaid):
val = sreg.ctaid(dim)
builder.store(val, ptr)
builder.ret_void()
# force inline
# inline_function(status.code)
nvvm.set_cuda_kernel(wrapfn)
module.link_in(ll.parse_assembly(str(wrapper_module)))
module.verify()
wrapfn = module.get_function(wrapfn.name)
return wrapfn
def setup_function(self, fndesc):
# Setup function
self.function = self.context.declare_function(self.module, fndesc)
self.entry_block = self.function.append_basic_block('entry')
self.builder = Builder.new(self.entry_block)
self.call_helper = self.call_conv.init_call_helper(self.builder)
def setup_function(self, fndesc):
# Setup function
self.function = self.context.declare_function(self.module, fndesc)
self.entry_block = self.function.append_basic_block('entry')
self.builder = Builder.new(self.entry_block)
self.call_helper = self.call_conv.init_call_helper(self.builder)
def generate_kernel_wrapper(self, func, argtypes):
module = func.module
arginfo = self.get_arg_packer(argtypes)
argtys = list(arginfo.argument_types)
wrapfnty = Type.function(Type.void(), argtys)
wrapper_module = self.create_module("cuda.kernel.wrapper")
fnty = Type.function(Type.int(),
[self.call_conv.get_return_type(types.pyobject)] + argtys)
func = wrapper_module.add_function(fnty, name=func.name)
wrapfn = wrapper_module.add_function(wrapfnty, name="cudaPy_" + func.name)
builder = Builder.new(wrapfn.append_basic_block(''))
# Define error handling variables
def define_error_gv(postfix):
gv = wrapper_module.add_global_variable(Type.int(),
name=wrapfn.name + postfix)
gv.initializer = Constant.null(gv.type.pointee)
return gv
gv_exc = define_error_gv("__errcode__")
gv_tid = []
gv_ctaid = []
for i in 'xyz':
gv_tid.append(define_error_gv("__tid%s__" % i))
gv_ctaid.append(define_error_gv("__ctaid%s__" % i))
callargs = arginfo.from_arguments(builder, wrapfn.args)
status, _ = self.call_conv.call_function(
builder, func, types.void, argtypes, callargs)
# Check error status
with cgutils.if_likely(builder, status.is_ok):
builder.ret_void()
with builder.if_then(builder.not_(status.is_python_exc)):
# User exception raised
old = Constant.null(gv_exc.type.pointee)
# Use atomic cmpxchg to prevent rewriting the error status
# Only the first error is recorded
casfnty = lc.Type.function(old.type, [gv_exc.type, old.type,
old.type])
casfn = wrapper_module.add_function(casfnty,
name="___numba_cas_hack")
xchg = builder.call(casfn, [gv_exc, old, status.code])
changed = builder.icmp(ICMP_EQ, xchg, old)
# If the xchange is successful, save the thread ID.
sreg = nvvmutils.SRegBuilder(builder)
with builder.if_then(changed):
for dim, ptr, in zip("xyz", gv_tid):
val = sreg.tid(dim)
builder.store(val, ptr)
for dim, ptr, in zip("xyz", gv_ctaid):
val = sreg.ctaid(dim)
builder.store(val, ptr)
builder.ret_void()
# force inline
# inline_function(status.code)
nvvm.set_cuda_kernel(wrapfn)
module.link_in(ll.parse_assembly(str(wrapper_module)))
module.verify()
wrapfn = module.get_function(wrapfn.name)
return wrapfn
def build(self):
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = self.context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
intp_ptr_t, byte_ptr_t])
wrapper_module = self.library.create_ir_module('')
func_type = self.call_conv.get_function_type(self.fndesc.restype,
self.fndesc.argtypes)
func = wrapper_module.add_function(func_type, name=self.func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty,
"__gufunc__." + self.func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
pyapi = self.context.get_python_api(builder)
# Unpack shapes
unique_syms = set()
for grp in (self.sin, self.sout):
for syms in grp:
unique_syms |= set(syms)
sym_map = {}
for syms in self.sin:
for s in syms:
if s not in sym_map:
sym_map[s] = len(sym_map)
sym_dim = {}
for s, i in sym_map.items():
sym_dim[s] = builder.load(builder.gep(arg_dims,
[self.context.get_constant(
types.intp,
i + 1)]))
# Prepare inputs
arrays = []
step_offset = len(self.sin) + len(self.sout)
for i, (typ, sym) in enumerate(zip(self.signature.args,
self.sin + self.sout)):
ary = GUArrayArg(self.context, builder, arg_args,
arg_steps, i, step_offset, typ, sym, sym_dim)
step_offset += len(sym)
arrays.append(ary)
bbreturn = builder.append_basic_block('.return')
# Prologue
self.gen_prologue(builder, pyapi)
# Loop
with cgutils.for_range(builder, loopcount, intp=intp_t) as loop:
args = [a.get_array_at_offset(loop.index) for a in arrays]
innercall, error = self.gen_loop_body(builder, pyapi, func, args)
# If error, escape
cgutils.cbranch_or_continue(builder, error, bbreturn)
builder.branch(bbreturn)
builder.position_at_end(bbreturn)
# Epilogue
self.gen_epilogue(builder, pyapi)
builder.ret_void()
self.library.add_ir_module(wrapper_module)
wrapper = self.library.get_function(wrapper.name)
# Set core function to internal so that it is not generated
self.func.linkage = LINKAGE_INTERNAL
return wrapper, self.env
