forked from M-Labs/artiq
1
0
Fork 0

compiler: Implement binary NumPy math functions (arctan2, …)

The bulk of the diff is just factoring out the implementation
for binary arithmetic implementations, to be reused for binary
function calls.
This commit is contained in:
David Nadlinger 2020-11-10 22:24:04 +01:00
parent fcf4763ae7
commit d95e619567
7 changed files with 211 additions and 135 deletions

View File

@ -53,23 +53,36 @@ unary_fp_runtime_calls = [
("cbrt", "cbrt"),
]
#: (float, float) -> float numpy.* math functions lowered to runtime calls.
binary_fp_runtime_calls = [
("arctan2", "atan2"),
("copysign", "copysign"),
("fmax", "fmax"),
("fmin", "fmin"),
# ("ldexp", "ldexp"), # One argument is an int; would need a bit more plumbing.
("hypot", "hypot"),
("nextafter", "nextafter"),
]
#: Array handling builtins (special treatment due to allocations).
numpy_builtins = ["transpose"]
def unary_fp_type(name):
return types.TExternalFunction(OrderedDict([("arg", builtins.TFloat())]),
def fp_runtime_type(name, arity):
args = [("arg{}".format(i), builtins.TFloat()) for i in range(arity)]
return types.TExternalFunction(OrderedDict(args),
builtins.TFloat(),
name,
# errno isn't observable from ARTIQ Python.
flags={"nounwind", "nowrite"},
broadcast_across_arrays=True)
numpy_map = {
getattr(numpy, symbol): unary_fp_type(mangle)
getattr(numpy, symbol): fp_runtime_type(mangle, arity=1)
for symbol, mangle in (unary_fp_intrinsics + unary_fp_runtime_calls)
}
for symbol, mangle in binary_fp_runtime_calls:
numpy_map[getattr(numpy, symbol)] = fp_runtime_type(mangle, arity=2)
for name in numpy_builtins:
numpy_map[getattr(numpy, name)] = types.TBuiltinFunction("numpy." + name)

View File

@ -1502,6 +1502,48 @@ class ARTIQIRGenerator(algorithm.Visitor):
self.final_branch = old_final_branch
self.unwind_target = old_unwind
def _make_array_elementwise_binop(self, name, result_type, lhs_type,
rhs_type, make_op):
def body_gen(result, lhs, rhs):
# At this point, shapes are assumed to match; could just pass buffer
# pointer for two of the three arrays as well.
result_buffer = self.append(ir.GetAttr(result, "buffer"))
shape = self.append(ir.GetAttr(result, "shape"))
num_total_elts = self._get_total_array_len(shape)
if builtins.is_array(lhs.type):
lhs_buffer = self.append(ir.GetAttr(lhs, "buffer"))
def get_left(index):
return self.append(ir.GetElem(lhs_buffer, index))
else:
def get_left(index):
return lhs
if builtins.is_array(rhs.type):
rhs_buffer = self.append(ir.GetAttr(rhs, "buffer"))
def get_right(index):
return self.append(ir.GetElem(rhs_buffer, index))
else:
def get_right(index):
return rhs
def loop_gen(index):
l = get_left(index)
r = get_right(index)
result = make_op(l, r)
self.append(ir.SetElem(result_buffer, index, result))
return self.append(
ir.Arith(ast.Add(loc=None), index,
ir.Constant(1, self._size_type)))
self._make_loop(
ir.Constant(0, self._size_type), lambda index: self.append(
ir.Compare(ast.Lt(loc=None), index, num_total_elts)),
loop_gen)
return self._make_array_binop(name, result_type, lhs_type, rhs_type,
body_gen)
def _mangle_arrayop_types(self, types):
def name_error(typ):
assert False, "Internal compiler error: No RPC tag for {}".format(typ)
@ -1517,53 +1559,16 @@ class ARTIQIRGenerator(algorithm.Visitor):
return "_".join(mangle_name(t) for t in types)
def _get_array_binop(self, op, result_type, lhs_type, rhs_type):
def _get_array_elementwise_binop(self, name, make_op, result_type, lhs_type, rhs_type):
# Currently, we always have any type coercions resolved explicitly in the AST.
# In the future, this might no longer be true and the three types might all
# differ.
name = "_array_{}_{}".format(
type(op).__name__,
name,
self._mangle_arrayop_types([result_type, lhs_type, rhs_type]))
if name not in self.array_op_funcs:
def body_gen(result, lhs, rhs):
# At this point, shapes are assumed to match; could just pass buffer
# pointer for two of the three arrays as well.
result_buffer = self.append(ir.GetAttr(result, "buffer"))
shape = self.append(ir.GetAttr(result, "shape"))
num_total_elts = self._get_total_array_len(shape)
if builtins.is_array(lhs.type):
lhs_buffer = self.append(ir.GetAttr(lhs, "buffer"))
def get_left(index):
return self.append(ir.GetElem(lhs_buffer, index))
else:
def get_left(index):
return lhs
if builtins.is_array(rhs.type):
rhs_buffer = self.append(ir.GetAttr(rhs, "buffer"))
def get_right(index):
return self.append(ir.GetElem(rhs_buffer, index))
else:
def get_right(index):
return rhs
def loop_gen(index):
l = get_left(index)
r = get_right(index)
result = self.append(ir.Arith(op, l, r))
self.append(ir.SetElem(result_buffer, index, result))
return self.append(
ir.Arith(ast.Add(loc=None), index,
ir.Constant(1, self._size_type)))
self._make_loop(
ir.Constant(0, self._size_type), lambda index: self.append(
ir.Compare(ast.Lt(loc=None), index, num_total_elts)), loop_gen)
self.array_op_funcs[name] = self._make_array_binop(
name, result_type, lhs_type, rhs_type, body_gen)
self.array_op_funcs[name] = self._make_array_elementwise_binop(
name, result_type, lhs_type, rhs_type, make_op)
return self.array_op_funcs[name]
def _invoke_arrayop(self, func, params):
@ -1721,6 +1726,34 @@ class ARTIQIRGenerator(algorithm.Visitor):
return self.append(ir.Alloc([result_buffer, shape], node.type))
return self.append(ir.GetElem(result_buffer, ir.Constant(0, self._size_type)))
def _broadcast_binop(self, name, make_op, result_type, lhs, rhs):
# Broadcast scalars (broadcasting higher dimensions is not yet allowed in the
# language).
broadcast = False
array_arg = lhs
if not builtins.is_array(lhs.type):
broadcast = True
array_arg = rhs
elif not builtins.is_array(rhs.type):
broadcast = True
shape = self.append(ir.GetAttr(array_arg, "shape"))
if not broadcast:
rhs_shape = self.append(ir.GetAttr(rhs, "shape"))
self._make_check(
self.append(ir.Compare(ast.Eq(loc=None), shape, rhs_shape)),
lambda: self.alloc_exn(
builtins.TException("ValueError"),
ir.Constant("operands could not be broadcast together",
builtins.TStr())))
elt = result_type.find()["elt"]
result, _ = self._allocate_new_array(elt, shape)
func = self._get_array_elementwise_binop(name, make_op, result_type, lhs.type,
rhs.type)
self._invoke_arrayop(func, [result, lhs, rhs])
return result
def visit_BinOpT(self, node):
if isinstance(node.op, ast.MatMult):
@ -1728,31 +1761,10 @@ class ARTIQIRGenerator(algorithm.Visitor):
elif builtins.is_array(node.type):
lhs = self.visit(node.left)
rhs = self.visit(node.right)
# Broadcast scalars.
broadcast = False
array_arg = lhs
if not builtins.is_array(lhs.type):
broadcast = True
array_arg = rhs
elif not builtins.is_array(rhs.type):
broadcast = True
shape = self.append(ir.GetAttr(array_arg, "shape"))
if not broadcast:
rhs_shape = self.append(ir.GetAttr(rhs, "shape"))
self._make_check(
self.append(ir.Compare(ast.Eq(loc=None), shape, rhs_shape)),
lambda: self.alloc_exn(
builtins.TException("ValueError"),
ir.Constant("operands could not be broadcast together",
builtins.TStr())))
result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
func = self._get_array_binop(node.op, node.type, lhs.type, rhs.type)
self._invoke_arrayop(func, [result, lhs, rhs])
return result
name = type(node.op).__name__
def make_op(l, r):
return self.append(ir.Arith(node.op, l, r))
return self._broadcast_binop(name, make_op, node.type, lhs, rhs)
elif builtins.is_numeric(node.type):
lhs = self.visit(node.left)
rhs = self.visit(node.right)
@ -2427,25 +2439,27 @@ class ARTIQIRGenerator(algorithm.Visitor):
if types.is_builtin(node.func.type):
insn = self.visit_builtin_call(node)
elif (types.is_broadcast_across_arrays(node.func.type) and len(args) >= 1
and builtins.is_array(args[0].type)):
and any(builtins.is_array(arg.type) for arg in args)):
# The iodelay machinery set up in the surrounding code was
# deprecated/a relic from the past when array broadcasting support
# was added, so no attempt to keep the delay tracking intact is
# made.
assert len(args) == 1, "Broadcasting for multiple arguments not implemented"
def make_call(val):
return self._user_call(ir.Constant(None, callee.type), [val], {},
def make_call(*args):
return self._user_call(ir.Constant(None, callee.type), args, {},
node.arg_exprs)
shape = self.append(ir.GetAttr(args[0], "shape"))
result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
# TODO: Generate more generically if non-externals are allowed.
name = node.func.type.find().name
func = self._get_array_unaryop(name, make_call, node.type, args[0].type)
self._invoke_arrayop(func, [result, args[0]])
insn = result
if len(args) == 1:
shape = self.append(ir.GetAttr(args[0], "shape"))
result, _ = self._allocate_new_array(node.type.find()["elt"], shape)
func = self._get_array_unaryop(name, make_call, node.type, args[0].type)
self._invoke_arrayop(func, [result, args[0]])
insn = result
elif len(args) == 2:
insn = self._broadcast_binop(name, make_call, node.type, *args)
else:
assert False, "Broadcasting for {} arguments not implemented".format(len)
else:
insn = self._user_call(callee, args, keywords, node.arg_exprs)
if isinstance(node.func, asttyped.AttributeT):

View File

@ -405,6 +405,47 @@ class Inferencer(algorithm.Visitor):
else:
assert False
def _coerce_binary_broadcast_op(self, left, right, map_return_elt, op_loc):
def num_dims(typ):
if builtins.is_array(typ):
# TODO: If number of dimensions is ever made a non-fixed parameter,
# need to acutally unify num_dims in _coerce_binop/….
return typ.find()["num_dims"].value
return 0
left_dims = num_dims(left.type)
right_dims = num_dims(right.type)
if left_dims != right_dims and left_dims != 0 and right_dims != 0:
# Mismatch (only scalar broadcast supported for now).
note1 = diagnostic.Diagnostic("note", "operand of dimension {num_dims}",
{"num_dims": left_dims}, left.loc)
note2 = diagnostic.Diagnostic("note", "operand of dimension {num_dims}",
{"num_dims": right_dims}, right.loc)
diag = diagnostic.Diagnostic(
"error", "dimensions of '{op}' array operands must match",
{"op": op_loc.source()}, op_loc, [left.loc, right.loc], [note1, note2])
self.engine.process(diag)
return
def map_node_type(typ):
if not builtins.is_array(typ):
# This is a single value broadcast across the array.
return typ
return typ.find()["elt"]
# Figure out result type, handling broadcasts.
result_dims = left_dims if left_dims else right_dims
def map_return(typ):
elt = map_return_elt(typ)
result = builtins.TArray(elt=elt, num_dims=result_dims)
left = builtins.TArray(elt=elt, num_dims=left_dims) if left_dims else elt
right = builtins.TArray(elt=elt, num_dims=right_dims) if right_dims else elt
return (result, left, right)
return self._coerce_numeric((left, right),
map_return=map_return,
map_node_type=map_node_type)
def _coerce_binop(self, op, left, right):
if isinstance(op, ast.MatMult):
if types.is_var(left.type) or types.is_var(right.type):
@ -477,45 +518,11 @@ class Inferencer(algorithm.Visitor):
self.engine.process(diag)
return
def num_dims(typ):
if builtins.is_array(typ):
# TODO: If number of dimensions is ever made a non-fixed parameter,
# need to acutally unify num_dims in _coerce_binop.
return typ.find()["num_dims"].value
return 0
left_dims = num_dims(left.type)
right_dims = num_dims(right.type)
if left_dims != right_dims and left_dims != 0 and right_dims != 0:
# Mismatch (only scalar broadcast supported for now).
note1 = diagnostic.Diagnostic("note", "operand of dimension {num_dims}",
{"num_dims": left_dims}, left.loc)
note2 = diagnostic.Diagnostic("note", "operand of dimension {num_dims}",
{"num_dims": right_dims}, right.loc)
diag = diagnostic.Diagnostic(
"error", "dimensions of '{op}' array operands must match",
{"op": op.loc.source()}, op.loc, [left.loc, right.loc], [note1, note2])
self.engine.process(diag)
return
def map_node_type(typ):
if not builtins.is_array(typ):
# This is a single value broadcast across the array.
return typ
return typ.find()["elt"]
# Figure out result type, handling broadcasts.
result_dims = left_dims if left_dims else right_dims
def map_return(typ):
elt = builtins.TFloat() if isinstance(op, ast.Div) else typ
result = builtins.TArray(elt=elt, num_dims=result_dims)
left = builtins.TArray(elt=elt, num_dims=left_dims) if left_dims else elt
right = builtins.TArray(elt=elt, num_dims=right_dims) if right_dims else elt
return (result, left, right)
return self._coerce_numeric((left, right),
map_return=map_return,
map_node_type=map_node_type)
def map_result(typ):
if isinstance(op, ast.Div):
return builtins.TFloat()
return typ
return self._coerce_binary_broadcast_op(left, right, map_result, op.loc)
elif isinstance(op, (ast.BitAnd, ast.BitOr, ast.BitXor,
ast.LShift, ast.RShift)):
# bitwise operators require integers
@ -1290,15 +1297,26 @@ class Inferencer(algorithm.Visitor):
self.engine.process(diag)
return
# Array broadcasting for unary functions explicitly marked as such.
if len(node.args) == typ_arity == 1 and types.is_broadcast_across_arrays(typ):
arg_type = node.args[0].type.find()
if builtins.is_array(arg_type):
typ_arg, = typ_args.values()
self._unify(typ_arg, arg_type["elt"], node.args[0].loc, None)
self._unify(node.type, builtins.TArray(typ_ret, arg_type["num_dims"]),
node.loc, None)
return
# Array broadcasting for functions explicitly marked as such.
if len(node.args) == typ_arity and types.is_broadcast_across_arrays(typ):
if typ_arity == 1:
arg_type = node.args[0].type.find()
if builtins.is_array(arg_type):
typ_arg, = typ_args.values()
self._unify(typ_arg, arg_type["elt"], node.args[0].loc, None)
self._unify(node.type, builtins.TArray(typ_ret, arg_type["num_dims"]),
node.loc, None)
return
elif typ_arity == 2:
if any(builtins.is_array(arg.type) for arg in node.args):
ret, arg0, arg1 = self._coerce_binary_broadcast_op(
node.args[0], node.args[1], lambda t: typ_ret, node.loc)
node.args[0] = self._coerce_one(arg0, node.args[0],
other_node=node.args[1])
node.args[1] = self._coerce_one(arg1, node.args[1],
other_node=node.args[0])
self._unify(node.type, ret, node.loc, None)
return
for actualarg, (formalname, formaltyp) in \
zip(node.args, list(typ_args.items()) + list(typ_optargs.items())):

View File

@ -76,6 +76,7 @@ static mut API: &'static [(&'static str, *const ())] = &[
api!(atanh),
api!(cbrt),
api!(ceil),
api!(copysign),
api!(cos),
api!(cosh),
api!(erf),
@ -86,6 +87,8 @@ static mut API: &'static [(&'static str, *const ())] = &[
api!(expm1),
api!(fabs),
api!(floor),
// api!(fmax),
// api!(fmin),
//api!(fma),
api!(fmod),
api!(hypot),
@ -96,6 +99,7 @@ static mut API: &'static [(&'static str, *const ())] = &[
api!(log),
//api!(log2),
api!(log10),
api!(nextafter),
api!(pow),
api!(round),
api!(sin),

View File

@ -27,7 +27,7 @@ class CompareHostDeviceTest(ExperimentCase):
def _test_binop(self, op, a, b):
exp = self.create(_RunOnDevice)
exp.run = kernel_from_string(["a", "b", "callback", "numpy"],
"callback(a " + op + "b)",
"callback(" + op + ")",
decorator=portable)
checked = False
@ -74,16 +74,17 @@ class CompareHostDeviceTest(ExperimentCase):
for typ in [numpy.int32, numpy.int64, numpy.float]]
for op in ELEM_WISE_BINOPS:
for arg in args:
self._test_binop(op, *arg)
self._test_binop("a" + op + "b", *arg)
def test_scalar_matrix_binops(self):
for typ in [numpy.int32, numpy.int64, numpy.float]:
scalar = typ(3)
matrix = numpy.array([[4, 5, 6], [7, 8, 9]], dtype=typ)
for op in ELEM_WISE_BINOPS:
self._test_binop(op, scalar, matrix)
self._test_binop(op, matrix, scalar)
self._test_binop(op, matrix, matrix)
code = "a" + op + "b"
self._test_binop(code, scalar, matrix)
self._test_binop(code, matrix, scalar)
self._test_binop(code, matrix, matrix)
def test_unary_math_fns(self):
names = [
@ -99,3 +100,17 @@ class CompareHostDeviceTest(ExperimentCase):
# Avoid 0.5, as numpy.rint's rounding mode currently doesn't match.
self._test_unaryop(op, 0.51)
self._test_unaryop(op, numpy.array([[0.3, 0.4], [0.51, 0.6]]))
def test_binary_math_fns(self):
names = [name for name, _ in math_fns.binary_fp_runtime_calls]
exp = self.create(_RunOnDevice)
if exp.core.target_cls != CortexA9Target:
names.remove("fmax")
names.remove("fmin")
for name in names:
code = "numpy.{}(a, b)".format(name)
# Avoid 0.5, as numpy.rint's rounding mode currently doesn't match.
self._test_binop(code, 1.0, 2.0)
self._test_binop(code, 1.0, numpy.array([2.0, 3.0]))
self._test_binop(code, numpy.array([1.0, 2.0]), 3.0)
self._test_binop(code, numpy.array([1.0, 2.0]), numpy.array([3.0, 4.0]))

View File

@ -6,10 +6,21 @@ import numpy as np
@kernel
def entrypoint():
# Just make sure everything compiles.
# LLVM intrinsic:
a = np.array([1.0, 2.0, 3.0])
b = np.sin(a)
assert b.shape == a.shape
# libm:
c = np.array([1.0, 2.0, 3.0])
d = np.arctan(c)
assert d.shape == c.shape
# libm, binary:
e = np.array([1.0, 2.0, 3.0])
f = np.array([4.0, 5.0, 6.0])
g = np.arctan2(e, f)
# g = np.arctan2(e, 0.0)
# g = np.arctan2(0.0, f)
assert g.shape == e.shape

View File

@ -28,3 +28,4 @@ def entrypoint():
assert numpy.arcsin(0.0) == 0.0
assert numpy.arccos(1.0) == 0.0
assert numpy.arctan(0.0) == 0.0
assert numpy.arctan2(0.0, 1.0) == 0.0