def __init__(self):
# TODO: come up with a less naive way of handling the symtab and types
self.classes = None
self.symtab = {}
self.typetab = {}
self.is_break = False
self.current_class = None
self.loop_end_blocks = []
self.loop_cond_blocks = []
context = ir.Context()
self.module = Module(name='opal-lang', context=context)
self.blocks = []
self.scope = {}
self._add_builtins()
func_ty = ir.FunctionType(ir.VoidType(), [])
func = Function(self.module, func_ty, 'main')
self.current_function = func
entry_block = self.add_block('entry')
exit_block = self.add_block('exit')
self.function_stack = [func]
self.builder = Builder(entry_block)
self.exit_blocks = [exit_block]
self.block_stack = [entry_block]
def _get_equal(context, module, datamodel, container_element_type):
assert datamodel.contains_nrt_meminfo()
fe_type = datamodel.fe_type
data_ptr_ty = datamodel.get_data_type().as_pointer()
wrapfnty = context.call_conv.get_function_type(types.int32,
[fe_type, fe_type])
argtypes = [fe_type, fe_type]
def build_wrapper(fn):
builder = Builder(fn.append_basic_block())
args = context.call_conv.decode_arguments(builder, argtypes, fn)
sig = typing.signature(types.boolean, fe_type, fe_type)
op = operator.eq
fnop = context.typing_context.resolve_value_type(op)
fnop.get_call_type(context.typing_context, sig.args, {})
eqfn = context.get_function(fnop, sig)
res = eqfn(builder, args)
intres = context.cast(builder, res, types.boolean, types.int32)
context.call_conv.return_value(builder, intres)
wrapfn = cgutils.get_or_insert_function(
module,
wrapfnty,
name='.numba_{}.{}_equal.wrap'.format(context.fndesc.mangled_name,
container_element_type))
build_wrapper(wrapfn)
equal_fnty = ir.FunctionType(ir.IntType(32), [data_ptr_ty, data_ptr_ty])
equal_fn = cgutils.get_or_insert_function(
module,
equal_fnty,
name='.numba_{}.{}_equal'.format(context.fndesc.mangled_name,
container_element_type),
)
builder = Builder(equal_fn.append_basic_block())
lhs = datamodel.load_from_data_pointer(builder, equal_fn.args[0])
rhs = datamodel.load_from_data_pointer(builder, equal_fn.args[1])
status, retval = context.call_conv.call_function(
builder,
wrapfn,
types.boolean,
argtypes,
[lhs, rhs],
)
with builder.if_then(status.is_ok, likely=True):
with builder.if_then(status.is_none):
builder.ret(context.get_constant(types.int32, 0))
retval = context.cast(builder, retval, types.boolean, types.int32)
builder.ret(retval)
# Error out
builder.ret(context.get_constant(types.int32, -1))
return equal_fn
def build_wrapper(fn):
builder = Builder(fn.append_basic_block())
args = context.call_conv.decode_arguments(builder, argtypes, fn)
sig = typing.signature(types.boolean, fe_type, fe_type)
op = operator.eq
fnop = context.typing_context.resolve_value_type(op)
fnop.get_call_type(context.typing_context, sig.args, {})
eqfn = context.get_function(fnop, sig)
res = eqfn(builder, args)
intres = context.cast(builder, res, types.boolean, types.int32)
context.call_conv.return_value(builder, intres)
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(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("test_nvvm_from_llvm")
fty = Type.function(Type.void(), [Type.int()])
kernel = m.add_function(fty, name='mycudakernel')
bldr = Builder(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 code(self, codegen):
klass = codegen.current_class
method_name = f'{klass.name}::{self.name}'
func = list(
filter(lambda f: f.name == method_name,
codegen.module.functions))[0]
codegen.function_stack.append(func)
old_func = codegen.current_function
old_builder = codegen.builder
codegen.current_function = func
entry_block = codegen.add_block('entry')
exit_block = codegen.add_block('exit')
codegen.exit_blocks.append(exit_block)
codegen.builder = Builder(entry_block)
if self.is_constructor:
this = codegen.gep(func.args[0], INDICES)
codegen.builder.store(
codegen.module.get_global(f'{klass.name}_vtable'), this)
body = self.body
if body:
ret = codegen.visit(body)
else:
ret = None
codegen.branch(exit_block)
if not ret:
codegen.position_at_end(exit_block)
codegen.builder.ret_void()
codegen.current_function = old_func
codegen.builder = old_builder
codegen.exit_blocks.pop()
codegen.function_stack.pop()
return func
def test_inline_rsqrt(self):
mod = Module(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, 'cu_rsqrt')
bldr = Builder(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 = self.fndesc.llvm_cpython_wrapper_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(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_wrapper(self, library, name):
"""
The LLVM IRBuilder code to create the gufunc wrapper.
The *library* arg is the CodeLibrary for which the wrapper should
be added to. The *name* arg is the name of the wrapper function being
created.
"""
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 = library.create_ir_module('')
func_type = self.call_conv.get_function_type(self.fndesc.restype,
self.fndesc.argtypes)
fname = self.fndesc.llvm_func_name
func = wrapper_module.add_function(func_type, name=fname)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty, 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(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()
# Link
library.add_ir_module(wrapper_module)
library.add_linking_library(self.library)
def build_ufunc_wrapper(library, context, fname, signature, objmode, envptr,
env):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
assert isinstance(fname, str)
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])
wrapperlib = context.codegen().create_library('ufunc_wrapper')
wrapper_module = wrapperlib.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)
func = wrapper_module.add_function(func_type, name=fname)
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(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
# Link and finalize
wrapperlib.add_ir_module(wrapper_module)
wrapperlib.add_linking_library(library)
return wrapperlib.get_pointer_to_function(wrapper.name)
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(self.entry_block)
self.call_helper = self.call_conv.init_call_helper(self.builder)
def generate_kernel_wrapper(self, library, fname, argtypes, debug):
"""
Generate the kernel wrapper in the given ``library``.
The function being wrapped have the name ``fname`` and argument types
``argtypes``. The wrapper function is returned.
"""
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=fname)
prefixed = itanium_mangler.prepend_namespace(func.name, ns='cudapy')
wrapfn = wrapper_module.add_function(wrapfnty, name=prefixed)
builder = Builder(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)
if debug:
# 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()
nvvm.set_cuda_kernel(wrapfn)
library.add_ir_module(wrapper_module)
library.finalize()
wrapfn = library.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(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
