def timedelta_mod_timedelta(context, builder, sig, args):
# inspired by https://github.com/numpy/numpy/blob/fe8072a12d65e43bd2e0b0f9ad67ab0108cc54b3/numpy/core/src/umath/loops.c.src#L1424
# alg is basically as `a % b`:
# if a or b is NaT return NaT
# elseif b is 0 return NaT
# else pretend a and b are int and do pythonic int modulus
[va, vb] = args
[ta, tb] = sig.args
not_nan = are_not_nat(builder, [va, vb])
ll_ret_type = context.get_value_type(sig.return_type)
ret = alloc_timedelta_result(builder)
builder.store(NAT, ret)
zero = Constant.int(ll_ret_type, 0)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vapos = builder.icmp_signed('>', va, zero)
vbpos = builder.icmp_signed('>', vb, zero)
rem = builder.srem(va, vb)
cond = builder.or_(builder.and_(vapos, vbpos),
builder.icmp_signed('==', rem, zero))
with builder.if_else(cond) as (then, otherwise):
with then:
builder.store(rem, ret)
with otherwise:
builder.store(builder.add(rem, vb), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def timedelta_floor_div_timedelta(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ll_ret_type = context.get_value_type(sig.return_type)
not_nan = are_not_nat(builder, [va, vb])
ret = cgutils.alloca_once(builder, ll_ret_type, name='ret')
zero = Constant.int(ll_ret_type, 0)
one = Constant.int(ll_ret_type, 1)
builder.store(zero, ret)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vaneg = builder.icmp_signed('<', va, zero)
neg = builder.or_(vaneg, builder.icmp_signed('<', vb, zero))
with builder.if_else(neg) as (then, otherwise):
with then: # one or more value negative
with builder.if_else(vaneg) as (negthen, negotherwise):
with negthen:
top = builder.sub(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with negotherwise:
top = builder.add(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with otherwise:
div = builder.sdiv(va, vb)
builder.store(div, ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def unbox_buffer(typ, obj, c):
"""
Convert a Py_buffer-providing object to a native array structure.
"""
buf = c.pyapi.alloca_buffer()
res = c.pyapi.get_buffer(obj, buf)
is_error = cgutils.is_not_null(c.builder, res)
nativearycls = c.context.make_array(typ)
nativeary = nativearycls(c.context, c.builder)
aryptr = nativeary._getpointer()
with cgutils.if_likely(c.builder, c.builder.not_(is_error)):
ptr = c.builder.bitcast(aryptr, c.pyapi.voidptr)
if c.context.enable_nrt:
c.pyapi.nrt_adapt_buffer_from_python(buf, ptr)
else:
c.pyapi.numba_buffer_adaptor(buf, ptr)
def cleanup():
c.pyapi.release_buffer(buf)
return NativeValue(c.builder.load(aryptr),
is_error=is_error,
cleanup=cleanup)
def unbox_unicodecharseq(typ, obj, c):
lty = c.context.get_value_type(typ)
ok, buffer, size, kind, is_ascii, hashv = \
c.pyapi.string_as_string_size_and_kind(obj)
# If conversion is ok, copy the buffer to the output storage.
with cgutils.if_likely(c.builder, ok):
# Check if the returned string size fits in the charseq
storage_size = ir.Constant(size.type, typ.count)
size_fits = c.builder.icmp_unsigned("<=", size, storage_size)
# Allow truncation of string
size = c.builder.select(size_fits, size, storage_size)
# Initialize output to zero bytes
null_string = ir.Constant(lty, None)
outspace = cgutils.alloca_once_value(c.builder, null_string)
# We don't need to set the NULL-terminator because the storage
# is already zero-filled.
cgutils.memcpy(c.builder, c.builder.bitcast(outspace, buffer.type),
buffer, size)
ret = c.builder.load(outspace)
return NativeValue(ret, is_error=c.builder.not_(ok))
def raise_error(self, builder, api, status):
"""
Given a non-ok *status*, raise the corresponding Python exception.
"""
bbend = builder.function.append_basic_block()
with builder.if_then(status.is_user_exc):
# Unserialize user exception.
# Make sure another error may not interfere.
api.err_clear()
exc = api.unserialize(status.excinfoptr)
with cgutils.if_likely(builder,
cgutils.is_not_null(builder, exc)):
api.raise_object(exc) # steals ref
builder.branch(bbend)
with builder.if_then(status.is_stop_iteration):
api.err_set_none("PyExc_StopIteration")
builder.branch(bbend)
with builder.if_then(status.is_python_exc):
# Error already raised => nothing to do
builder.branch(bbend)
api.err_set_string("PyExc_SystemError",
"unknown error when calling native function")
builder.branch(bbend)
builder.position_at_end(bbend)
def unbox_funcptr(typ, obj, c):
if typ.get_pointer is None:
raise NotImplementedError(typ)
# Call get_pointer() on the object to get the raw pointer value
ptrty = c.context.get_function_pointer_type(typ)
ret = cgutils.alloca_once_value(c.builder, ir.Constant(ptrty, None), name="fnptr")
ser = c.pyapi.serialize_object(typ.get_pointer)
get_pointer = c.pyapi.unserialize(ser)
with cgutils.if_likely(c.builder, cgutils.is_not_null(c.builder, get_pointer)):
intobj = c.pyapi.call_function_objargs(get_pointer, (obj,))
c.pyapi.decref(get_pointer)
with cgutils.if_likely(c.builder, cgutils.is_not_null(c.builder, intobj)):
ptr = c.pyapi.long_as_voidptr(intobj)
c.pyapi.decref(intobj)
c.builder.store(c.builder.bitcast(ptr, ptrty), ret)
return NativeValue(c.builder.load(ret), is_error=c.pyapi.c_api_error())
def timedelta_sub_impl(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ret = alloc_timedelta_result(builder)
with cgutils.if_likely(builder, are_not_nat(builder, [va, vb])):
va = scale_timedelta(context, builder, va, ta, sig.return_type)
vb = scale_timedelta(context, builder, vb, tb, sig.return_type)
builder.store(builder.sub(va, vb), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def impl(context, builder, dt_arg, dt_unit, td_arg, td_unit, ret_unit):
ret = alloc_timedelta_result(builder)
with cgutils.if_likely(builder, are_not_nat(builder, [dt_arg, td_arg])):
dt_arg = convert_datetime_for_arith(builder, dt_arg, dt_unit, ret_unit)
td_factor = npdatetime_helpers.get_timedelta_conversion_factor(
td_unit, ret_unit
)
td_arg = scale_by_constant(builder, td_arg, td_factor)
ret_val = getattr(builder, ll_op_name)(dt_arg, td_arg)
builder.store(ret_val, ret)
return builder.load(ret)
def datetime_minus_datetime(context, builder, sig, args):
va, vb = args
ta, tb = sig.args
unit_a = ta.unit
unit_b = tb.unit
ret_unit = sig.return_type.unit
ret = alloc_timedelta_result(builder)
with cgutils.if_likely(builder, are_not_nat(builder, [va, vb])):
va = convert_datetime_for_arith(builder, va, unit_a, ret_unit)
vb = convert_datetime_for_arith(builder, vb, unit_b, ret_unit)
ret_val = builder.sub(va, vb)
builder.store(ret_val, ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def timedelta_over_timedelta(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
not_nan = are_not_nat(builder, [va, vb])
ll_ret_type = context.get_value_type(sig.return_type)
ret = cgutils.alloca_once(builder, ll_ret_type, name='ret')
builder.store(Constant(ll_ret_type, float('nan')), ret)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
va = builder.sitofp(va, ll_ret_type)
vb = builder.sitofp(vb, ll_ret_type)
builder.store(builder.fdiv(va, vb), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def _timedelta_times_number(context, builder, td_arg, td_type, number_arg,
number_type, return_type):
ret = alloc_timedelta_result(builder)
with cgutils.if_likely(builder, is_not_nat(builder, td_arg)):
if isinstance(number_type, types.Float):
val = builder.sitofp(td_arg, number_arg.type)
val = builder.fmul(val, number_arg)
val = _cast_to_timedelta(context, builder, val)
else:
val = builder.mul(td_arg, number_arg)
# The scaling is required for ufunc np.multiply() with an explicit
# output in a different unit.
val = scale_timedelta(context, builder, val, td_type, return_type)
builder.store(val, ret)
return builder.load(ret)
def iternext_zip(context, builder, sig, args, result):
genty, = sig.args
gen, = args
impl = context.get_generator_impl(genty)
status, retval = impl(context, builder, sig, args)
context.add_linking_libs(getattr(impl, 'libs', ()))
with cgutils.if_likely(builder, status.is_ok):
result.set_valid(True)
result.yield_(retval)
with cgutils.if_unlikely(builder, status.is_stop_iteration):
result.set_exhausted()
with cgutils.if_unlikely(builder,
builder.and_(status.is_error,
builder.not_(status.is_stop_iteration))):
context.call_conv.return_status_propagate(builder, status)
def timedelta_over_number(context, builder, sig, args):
td_arg, number_arg = args
number_type = sig.args[1]
ret = alloc_timedelta_result(builder)
ok = builder.and_(is_not_nat(builder, td_arg),
builder.not_(cgutils.is_scalar_zero_or_nan(builder, number_arg)))
with cgutils.if_likely(builder, ok):
# Denominator is non-zero, non-NaN
if isinstance(number_type, types.Float):
val = builder.sitofp(td_arg, number_arg.type)
val = builder.fdiv(val, number_arg)
val = _cast_to_timedelta(context, builder, val)
else:
val = builder.sdiv(td_arg, number_arg)
# The scaling is required for ufuncs np.*divide() with an explicit
# output in a different unit.
val = scale_timedelta(context, builder, val,
sig.args[0], sig.return_type)
builder.store(val, ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
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 = ir.FunctionType(ir.VoidType(), argtys)
wrapper_module = self.create_module("cuda.kernel.wrapper")
fnty = ir.FunctionType(
ir.IntType(32),
[self.call_conv.get_return_type(types.pyobject)] + argtys)
func = ir.Function(wrapper_module, fnty, fname)
prefixed = itanium_mangler.prepend_namespace(func.name, ns='cudapy')
wrapfn = ir.Function(wrapper_module, wrapfnty, prefixed)
builder = ir.IRBuilder(wrapfn.append_basic_block(''))
# Define error handling variables
def define_error_gv(postfix):
name = wrapfn.name + postfix
gv = cgutils.add_global_variable(wrapper_module, ir.IntType(32),
name)
gv.initializer = ir.Constant(gv.type.pointee, None)
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 = ir.Constant(gv_exc.type.pointee, None)
# Use atomic cmpxchg to prevent rewriting the error status
# Only the first error is recorded
if nvvm.NVVM().is_nvvm70:
xchg = builder.cmpxchg(gv_exc, old, status.code,
'monotonic', 'monotonic')
changed = builder.extract_value(xchg, 1)
else:
casfnty = ir.FunctionType(
old.type, [gv_exc.type, old.type, old.type])
cas_hack = "___numba_atomic_i32_cas_hack"
casfn = ir.Function(wrapper_module, casfnty, name=cas_hack)
xchg = builder.call(casfn, [gv_exc, old, status.code])
changed = builder.icmp_unsigned('==', 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
