def generate(impero_c, args, precision, scalar_type, kernel_name="loopy_kernel", index_names=[]):
"""Generates loopy code.
:arg impero_c: ImperoC tuple with Impero AST and other data
:arg args: list of loopy.GlobalArgs
:arg precision: floating-point precision for printing
:arg scalar_type: type of scalars as C typename string
:arg kernel_name: function name of the kernel
:arg index_names: pre-assigned index names
:returns: loopy kernel
"""
ctx = LoopyContext()
ctx.indices = impero_c.indices
ctx.index_names = defaultdict(lambda: "i", index_names)
ctx.precision = precision
ctx.scalar_type = scalar_type
ctx.epsilon = 10.0 ** (-precision)
# Create arguments
data = list(args)
for i, temp in enumerate(impero_c.temporaries):
name = "t%d" % i
if isinstance(temp, gem.Constant):
data.append(lp.TemporaryVariable(name, shape=temp.shape, dtype=temp.array.dtype, initializer=temp.array, address_space=lp.AddressSpace.LOCAL, read_only=True))
else:
shape = tuple([i.extent for i in ctx.indices[temp]]) + temp.shape
data.append(lp.TemporaryVariable(name, shape=shape, dtype=numpy.float64, initializer=None, address_space=lp.AddressSpace.LOCAL, read_only=False))
ctx.gem_to_pymbolic[temp] = p.Variable(name)
# Create instructions
instructions = statement(impero_c.tree, ctx)
# Create domains
domains = []
for idx, extent in ctx.index_extent.items():
inames = isl.make_zero_and_vars([idx])
domains.append(((inames[0].le_set(inames[idx])) & (inames[idx].lt_set(inames[0] + extent))))
if not domains:
domains = [isl.BasicSet("[] -> {[]}")]
# Create loopy kernel
knl = lp.make_function(domains, instructions, data, name=kernel_name, target=lp.CTarget(),
seq_dependencies=True, silenced_warnings=["summing_if_branches_ops"])
# Prevent loopy interchange by loopy
knl = lp.prioritize_loops(knl, ",".join(ctx.index_extent.keys()))
# Help loopy in scheduling by assigning priority to instructions
insn_new = []
for i, insn in enumerate(knl.instructions):
insn_new.append(insn.copy(priority=len(knl.instructions) - i))
knl = knl.copy(instructions=insn_new)
return knl
def test_c_execution_with_global_temporaries():
# ensure that the "host" code of a bare ExecutableCTarget with
# global constant temporaries is None
from loopy.target.c import ExecutableCTarget
from loopy.kernel.data import temp_var_scope as scopes
n = 10
knl = lp.make_kernel(
'{[i]: 0 <= i < n}',
"""
a[i] = b[i]
""", [
lp.GlobalArg('a', shape=(n, ), dtype=np.int32),
lp.TemporaryVariable('b',
shape=(n, ),
initializer=np.arange(n, dtype=np.int32),
dtype=np.int32,
read_only=True,
scope=scopes.GLOBAL)
],
target=ExecutableCTarget())
knl = lp.fix_parameters(knl, n=n)
assert ('int b[%d]' % n) not in lp.generate_code_v2(knl).host_code()
assert np.allclose(knl(a=np.zeros(10, dtype=np.int32))[1], np.arange(10))
def test_forced_iname_deps_and_reduction():
# See https://github.com/inducer/loopy/issues/24
# This is (purposefully) somewhat un-idiomatic, to replicate the conditions
# under which the above bug was found. If assignees were phi[i], then the
# iname propagation heuristic would not assume that dependent instructions
# need to run inside of 'i', and hence the forced_iname_* bits below would not
# be needed.
i1 = lp.CInstruction("i", "doSomethingToGetPhi();", assignees="phi")
from pymbolic.primitives import Subscript, Variable
i2 = lp.Assignment("a",
lp.Reduction("sum", "j",
Subscript(Variable("phi"), Variable("j"))),
forced_iname_deps=frozenset(),
forced_iname_deps_is_final=True)
k = lp.make_kernel(
"{[i,j] : 0<=i,j<n}",
[i1, i2],
[
lp.GlobalArg("a", dtype=np.float32, shape=()),
lp.ValueArg("n", dtype=np.int32),
lp.TemporaryVariable("phi", dtype=np.float32, shape=("n", )),
],
target=lp.CTarget(),
)
k = lp.preprocess_kernel(k)
assert 'i' not in k.insn_inames("insn_0_j_update")
print(k.stringify(with_dependencies=True))
def make_kernels(self, seq_dependencies):
result = []
for sub in self.kernels:
# {{{ figure out arguments
kernel_data = []
for arg_name in sub.arg_names:
dims = sub.dim_map.get(arg_name)
if dims is not None:
# default order is set to "F" in kernel creation below
kernel_data.append(
lp.GlobalArg(
arg_name,
dtype=sub.get_type(arg_name),
shape=sub.get_loopy_shape(arg_name),
))
else:
kernel_data.append(
lp.ValueArg(arg_name,
dtype=sub.get_type(arg_name)))
# }}}
# {{{ figure out temporary variables
for var_name in (
sub.known_names()
- set(sub.arg_names)
- sub.all_inames()):
dtype = sub.get_type(var_name, none_ok=True)
if sub.implicit_types is None and dtype is None:
continue
kernel_data.append(
lp.TemporaryVariable(
var_name, dtype=dtype,
shape=sub.get_loopy_shape(var_name)))
# }}}
knl = lp.make_kernel(
sub.index_sets,
sub.instructions,
kernel_data,
name=sub.subprogram_name,
default_order="F",
index_dtype=self.index_dtype,
target=self.target,
seq_dependencies=seq_dependencies,
)
from loopy.loop import fuse_loop_domains
knl = fuse_loop_domains(knl)
knl = lp.fold_constants(knl)
result.append(knl)
return result
def test_c_execution_with_global_temporaries():
# ensure that the "host" code of a bare ExecutableCTarget with
# global constant temporaries is None
from loopy.target.c import ExecutableCTarget
AS = lp.AddressSpace # noqa
n = 10
knl = lp.make_kernel(
"{[i]: 0 <= i < n}",
"""
a[i] = b[i]
""", [
lp.GlobalArg("a", shape=(n, ), dtype=np.int32),
lp.TemporaryVariable("b",
shape=(n, ),
initializer=np.arange(n, dtype=np.int32),
dtype=np.int32,
read_only=True,
address_space=AS.GLOBAL)
],
target=ExecutableCTarget())
knl = lp.fix_parameters(knl, n=n)
assert ("int b[%d]" % n) not in lp.generate_code_v2(knl).host_code()
assert np.allclose(knl(a=np.zeros(10, dtype=np.int32))[1], np.arange(10))
def test_barrier_counter_barriers():
knl = lp.make_kernel(
"[n,m,ell] -> {[i,k,j]: 0<=i<50 and 1<=k<98 and 0<=j<10}",
[
"""
c[i,j,k] = 2*a[i,j,k] {id=first}
e[i,j,k] = c[i,j,k+1]+c[i,j,k-1] {dep=first}
"""
], [
lp.TemporaryVariable("c", lp.auto, shape=(50, 10, 99)),
"..."
],
name="weird2",
)
knl = lp.add_and_infer_dtypes(knl, dict(a=np.int32))
knl = lp.split_iname(knl, "k", 128, inner_tag="l.0")
sync_map = lp.get_synchronization_map(knl)
print(sync_map)
n = 512
m = 256
ell = 128
params = {'n': n, 'm': m, 'ell': ell}
barrier_count = sync_map["barrier_local"].eval_with_dict(params)
assert barrier_count == 50*10*2
def test_to_batched_temp(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
''' { [i,j]: 0<=i,j<n } ''', ''' cnst = 2.0
out[i] = sum(j, cnst*a[i,j]*x[j])''', [
lp.TemporaryVariable("cnst",
dtype=np.float32,
shape=(),
scope=lp.temp_var_scope.PRIVATE), '...'
])
knl = lp.add_and_infer_dtypes(
knl, dict(out=np.float32, x=np.float32, a=np.float32))
ref_knl = lp.make_kernel(''' { [i,j]: 0<=i,j<n } ''',
'''out[i] = sum(j, 2.0*a[i,j]*x[j])''')
ref_knl = lp.add_and_infer_dtypes(
ref_knl, dict(out=np.float32, x=np.float32, a=np.float32))
bknl = lp.to_batched(knl, "nbatches", "out,x")
bref_knl = lp.to_batched(ref_knl, "nbatches", "out,x")
# checking that cnst is not being bathced
assert bknl.temporary_variables['cnst'].shape == ()
a = np.random.randn(5, 5)
x = np.random.randn(7, 5)
# Checking that the program compiles and the logic is correct
lp.auto_test_vs_ref(bref_knl,
ctx,
bknl,
parameters=dict(a=a, x=x, n=5, nbatches=7))
def test_to_batched_temp(ctx_factory):
ctx = ctx_factory()
knl = lp.make_kernel(
""" { [i,j]: 0<=i,j<n } """, """ cnst = 2.0
out[i] = sum(j, cnst*a[i,j]*x[j])""", [
lp.TemporaryVariable("cnst",
dtype=np.float32,
shape=(),
address_space=lp.AddressSpace.PRIVATE), "..."
])
knl = lp.add_and_infer_dtypes(
knl, dict(out=np.float32, x=np.float32, a=np.float32))
ref_knl = lp.make_kernel(""" { [i,j]: 0<=i,j<n } """,
"""out[i] = sum(j, 2.0*a[i,j]*x[j])""")
ref_knl = lp.add_and_infer_dtypes(
ref_knl, dict(out=np.float32, x=np.float32, a=np.float32))
bknl = lp.to_batched(knl, "nbatches", "out,x")
bref_knl = lp.to_batched(ref_knl, "nbatches", "out,x")
# checking that cnst is not being bathced
assert bknl["loopy_kernel"].temporary_variables["cnst"].shape == ()
a = np.random.randn(5, 5)
x = np.random.randn(7, 5)
# Checking that the program compiles and the logic is correct
lp.auto_test_vs_ref(bref_knl,
ctx,
bknl,
parameters=dict(a=a, x=x, n=5, nbatches=7))
def generate_wrapper_kernel_args(self, tensor2temp, templated_subkernels):
coords_extent = self.extent(self.expression.ufl_domain().coordinates)
args = [loopy.GlobalArg(self.coordinates_arg, shape=coords_extent,
dtype=self.tsfc_parameters["scalar_type"])]
for loopy_inner in templated_subkernels:
for arg in loopy_inner.args[1:]:
if arg.name == self.cell_orientations_arg or\
arg.name == self.cell_size_arg:
if arg not in args:
args.append(arg)
for coeff in self.bag.coefficients.values():
if isinstance(coeff, OrderedDict):
for (name, extent) in coeff.values():
arg = loopy.GlobalArg(name, shape=extent,
dtype=self.tsfc_parameters["scalar_type"])
args.append(arg)
else:
(name, extent) = coeff
arg = loopy.GlobalArg(name, shape=extent,
dtype=self.tsfc_parameters["scalar_type"])
args.append(arg)
if self.bag.needs_cell_facets:
# Arg for is exterior (==0)/interior (==1) facet or not
args.append(loopy.GlobalArg(self.cell_facets_arg, shape=(self.num_facets, 2),
dtype=np.int8))
args.append(
loopy.TemporaryVariable(self.local_facet_array_arg,
shape=(self.num_facets,),
dtype=np.uint32,
address_space=loopy.AddressSpace.LOCAL,
read_only=True,
initializer=np.arange(self.num_facets, dtype=np.uint32),))
if self.bag.needs_mesh_layers:
args.append(loopy.GlobalArg(self.layer_count, shape=(),
dtype=np.int32))
args.append(loopy.TemporaryVariable(self.layer_arg, shape=(),
dtype=np.int32, address_space=loopy.AddressSpace.GLOBAL))
for tensor_temp in tensor2temp.values():
args.append(tensor_temp)
return args
def initialise_terminals(self, var2terminal, coefficients):
""" Initilisation of the variables in which coefficients
and the Tensors coming from TSFC are saved.
:arg var2terminal: dictionary that maps Slate Tensors to gem Variables
"""
tensor2temp = OrderedDict()
inits = []
for gem_tensor, slate_tensor in var2terminal.items():
assert slate_tensor.terminal, "Only terminal tensors need to be initialised in Slate kernels."
(_, dtype), = assign_dtypes([gem_tensor],
self.tsfc_parameters["scalar_type"])
loopy_tensor = loopy.TemporaryVariable(
gem_tensor.name,
dtype=dtype,
shape=gem_tensor.shape,
address_space=loopy.AddressSpace.LOCAL)
tensor2temp[slate_tensor] = loopy_tensor
if not slate_tensor.assembled:
indices = self.bag.index_creator(self.shape(slate_tensor))
inames = {var.name for var in indices}
var = pym.Subscript(pym.Variable(loopy_tensor.name), indices)
inits.append(
loopy.Assignment(var,
"0.",
id="init%d" % len(inits),
within_inames=frozenset(inames)))
else:
f = slate_tensor.form if isinstance(
slate_tensor.form, tuple) else (slate_tensor.form, )
coeff = tuple(coefficients[c] for c in f)
offset = 0
ismixed = tuple(
(type(c.ufl_element()) == MixedElement) for c in f)
names = []
for (im, c) in zip(ismixed, coeff):
names += [name
for (name, ext) in c.values()] if im else [c[0]]
# Mixed coefficients come as seperate parameter (one per space)
for i, shp in enumerate(*slate_tensor.shapes.values()):
indices = self.bag.index_creator((shp, ))
inames = {var.name for var in indices}
offset_index = (pym.Sum((offset, indices[0])), )
name = names[i] if ismixed else names
var = pym.Subscript(pym.Variable(loopy_tensor.name),
offset_index)
c = pym.Subscript(pym.Variable(name), indices)
inits.append(
loopy.Assignment(var,
c,
id="init%d" % len(inits),
within_inames=frozenset(inames)))
offset += shp
return inits, tensor2temp
def expression_variable(expr, parameters):
name = expr.name
shape = expr.shape
dtype = expr.dtype
if name not in parameters.temporaries:
parameters.temporaries[name] = loopy.TemporaryVariable(
name, dtype=dtype, shape=shape, address_space=loopy.auto)
return pym.Variable(name)
def get_loopy_temporary(name: str, expr: Array) -> lp.TemporaryVariable:
is_shape_symbolic = not all(isinstance(dim, int) for dim in expr.shape)
# Only global variables can have symbolic shape.
address_space = lp.AddressSpace.GLOBAL if is_shape_symbolic else lp.auto
return lp.TemporaryVariable(name,
dtype=expr.dtype,
shape=expr.shape,
address_space=address_space)
def get_kernel(self, **kwargs):
extra_kernel_kwarg_types = ()
if "extra_kernel_kwarg_types" in kwargs:
extra_kernel_kwarg_types = kwargs["extra_kernel_kwarg_types"]
eval_inames = frozenset(["itgt"])
scalar_assignment = lp.Assignment(
id=None,
assignee="expr_val",
expression=self.get_normalised_expr(),
temp_var_type=None,
)
eval_insns = [
insn.copy(within_inames=insn.within_inames | eval_inames)
for insn in [scalar_assignment]
]
loopy_knl = lp.make_kernel( # NOQA
"{ [itgt]: 0<=itgt<n_targets }",
[
"""
for itgt
VAR_ASSIGNMENT
end
""".replace("VAR_ASSIGNMENT",
self.get_variable_assignment_code())
] + eval_insns + [
"""
for itgt
result[itgt] = expr_val
end
"""
],
[
lp.ValueArg("dim, n_targets", np.int32),
lp.GlobalArg("target_points", np.float64, "dim, n_targets"),
lp.TemporaryVariable("expr_val", None, ()),
] + list(extra_kernel_kwarg_types) + [
"...",
],
name="eval_expr",
lang_version=(2018, 2),
)
loopy_knl = lp.fix_parameters(loopy_knl, dim=self.dim)
loopy_knl = lp.set_options(loopy_knl, write_cl=False)
loopy_knl = lp.set_options(loopy_knl, return_dict=True)
if self.function_manglers is not None:
loopy_knl = lp.register_function_manglers(loopy_knl,
self.function_manglers)
if self.preamble_generators is not None:
loopy_knl = lp.register_preamble_generators(
loopy_knl, self.preamble_generators)
return loopy_knl
def test_pyopencl_target_with_global_temps_with_base_storage(ctx_factory):
from pyopencl.tools import ImmediateAllocator
class RecordingAllocator(ImmediateAllocator):
def __init__(self, queue):
super().__init__(queue)
self.allocated_nbytes = 0
def __call__(self, size):
self.allocated_nbytes += size
return super().__call__(size)
ctx = ctx_factory()
cq = cl.CommandQueue(ctx)
knl = lp.make_kernel(
"{[i, j]: 0<=i, j<10}",
"""
tmp1[i] = 2*i {id=w_tmp1}
y[i] = tmp1[i] {nosync=w_tmp1}
... gbarrier
tmp2[j] = 3*j {id=w_tmp2}
z[j] = tmp2[j] {nosync=w_tmp2}
""", [
lp.TemporaryVariable("tmp1",
base_storage="base",
address_space=lp.AddressSpace.GLOBAL),
lp.TemporaryVariable("tmp2",
base_storage="base",
address_space=lp.AddressSpace.GLOBAL), ...
],
seq_dependencies=True)
knl = lp.tag_inames(knl, {"i": "g.0", "j": "g.0"})
knl = lp.set_options(knl, "return_dict")
my_allocator = RecordingAllocator(cq)
_, out = knl(cq, allocator=my_allocator)
np.testing.assert_allclose(out["y"].get(), 2 * np.arange(10))
np.testing.assert_allclose(out["z"].get(), 3 * np.arange(10))
assert my_allocator.allocated_nbytes == (
40 # base
+ 40 # y
+ 40 # z
)
def expression_namedliteral(expr, parameters):
name = expr.name
val = loopy.TemporaryVariable(name,
dtype=expr.dtype,
shape=expr.shape,
address_space=loopy.AddressSpace.LOCAL,
read_only=True,
initializer=expr.value)
parameters.temporaries[name] = val
return pym.Variable(name)
def get_loopy_temporary(name: str, expr: Array, cgen_mapper: CodeGenMapper,
state: CodeGenState) -> lp.TemporaryVariable:
# always allocating to global address space to avoid stack overflow
address_space = lp.AddressSpace.GLOBAL
return lp.TemporaryVariable(name,
shape=shape_to_scalar_expression(
expr.shape, cgen_mapper, state),
dtype=expr.dtype,
address_space=address_space,
tags=_filter_tags_not_of_type(
expr,
cgen_mapper.array_tag_t_to_not_propagate))
def __call__(self, preamble_info):
from loopy.kernel.data import temp_var_scope as scopes
# find a function matching our name
func_match = next((x for x in preamble_info.seen_functions
if x.name == self.func_name), None)
desc = 'custom_funcs_indirect'
if func_match is not None:
from loopy.types import to_loopy_type
# check types
if tuple(to_loopy_type(x) for x in self.func_arg_dtypes) == \
func_match.arg_dtypes:
# if match, create our temporary
var = lp.TemporaryVariable('lookup',
initializer=self.arr,
dtype=self.arr.dtype,
shape=self.arr.shape,
scope=scopes.GLOBAL,
read_only=True)
# and code
code = """
int {name}(int start, int end, int match)
{{
int result = start;
for (int i = start + 1; i < end; ++i)
{{
if (lookup[i] == match)
result = i;
}}
return result;
}}
""".format(name=self.func_name)
# generate temporary variable code
from cgen import Initializer
from loopy.target.c import generate_array_literal
codegen_state = preamble_info.codegen_state.copy(
is_generating_device_code=True)
kernel = preamble_info.kernel
ast_builder = codegen_state.ast_builder
target = kernel.target
decl_info, = var.decl_info(target, index_dtype=kernel.index_dtype)
decl = ast_builder.wrap_global_constant(
ast_builder.get_temporary_decl(codegen_state, None, var,
decl_info))
if var.initializer is not None:
decl = Initializer(
decl,
generate_array_literal(codegen_state, var, var.initializer))
# return generated code
yield (desc, '\n'.join([str(decl), code]))
def get_vars(self, cname):
vars = []
comp = self.templates[cname]
# Parameters are constant arguments
for p in list(comp.parameters):
p_edge = comp.edge_info[p]
dtype = hu.get_attribute_value(p_edge.attributes['type'])
dims = self.get_edge_dims(p_edge)
param_var = lp.ConstantArg(p, dtype, shape=dims)
vars.append(param_var)
# Need to add read only option for inputs
# Inputs, outputs, and states represent
global_args = list(comp.input) + list(comp.output) + list(comp.state)
dim_vars = []
for g in global_args:
g_edge = comp.edge_info[g]
dtype = hu.get_attribute_value(g_edge.attributes['type'])
dims = self.get_edge_dims(g_edge)
for d in dims:
if d not in dim_vars:
dim_vars.append(d)
dim_var = lp.ValueArg(d, dtype=np.int32)
vars.append(dim_var)
if len(dims) == 0:
g_var = lp.ValueArg(g, dtype=dtype)
else:
g_var = lp.GlobalArg(g, dtype=dtype, shape=dims)
vars.append(g_var)
# Each flow declaration is a temporary variable declared in the comp scope
for s in comp.statements:
if s.op_type == 'declaration' and s.op_cat == 'declaration':
d_vars = list(s.output)
for d in d_vars:
edge_decl = comp.edge_info[d]
dims = self.get_edge_dims(edge_decl)
dtype = hu.get_attribute_value(
edge_decl.attributes['type'])
t_var = lp.TemporaryVariable(d, dtype=dtype, shape=dims)
vars.append(t_var)
return vars
def test_get_kernel_input_and_output():
# make a kernel
knl = lp.make_kernel('{[i]: 0 <= i < 2}',
'<> a = 1')
assert not len(find_inputs_and_outputs(knl))
knl = lp.make_kernel('{[i]: 0 <= i < 2}',
'<> a = b[i]',
[lp.GlobalArg('b', shape=(2,))])
assert find_inputs_and_outputs(knl) == set(['b'])
knl = lp.make_kernel('{[i]: 0 <= i < 2}',
'<> a = b[i] + c[i]',
[lp.GlobalArg('b', shape=(2,)),
lp.TemporaryVariable('c', shape=(2,), scope=scopes.GLOBAL)
],
silenced_warnings=['read_no_write(c)'])
assert find_inputs_and_outputs(knl) == set(['b', 'c'])
def initialise_terminals(self, var2terminal, coefficients):
""" Initilisation of the variables in which coefficients
and the Tensors coming from TSFC are saved.
:arg var2terminal: dictionary that maps Slate Tensors to gem Variables
"""
tensor2temp = OrderedDict()
inits = []
for gem_tensor, slate_tensor in var2terminal.items():
loopy_tensor = loopy.TemporaryVariable(gem_tensor.name,
shape=gem_tensor.shape,
address_space=loopy.AddressSpace.LOCAL)
tensor2temp[slate_tensor] = loopy_tensor
if isinstance(slate_tensor, slate.Tensor):
indices = self.bag.index_creator(self.shape(slate_tensor))
inames = {var.name for var in indices}
var = pym.Subscript(pym.Variable(loopy_tensor.name), indices)
inits.append(loopy.Assignment(var, "0.", id="init%d" % len(inits),
within_inames=frozenset(inames)))
elif isinstance(slate_tensor, slate.AssembledVector):
f = slate_tensor._function
coeff = coefficients[f]
offset = 0
ismixed = (type(f.ufl_element()) == MixedElement)
names = [name for (name, ext) in coeff.values()] if ismixed else coeff[0]
# Mixed coefficients come as seperate parameter (one per space)
for i, shp in enumerate(*slate_tensor.shapes.values()):
indices = self.bag.index_creator((shp,))
inames = {var.name for var in indices}
offset_index = (pym.Sum((offset, indices[0])),)
name = names[i] if ismixed else names
var = pym.Subscript(pym.Variable(loopy_tensor.name), offset_index)
c = pym.Subscript(pym.Variable(name), indices)
inits.append(loopy.Assignment(var, c, id="init%d" % len(inits),
within_inames=frozenset(inames)))
offset += shp
return inits, tensor2temp
def test_memory_tools_defn():
wrapper = __test_cases()
for opts in wrapper:
# create a dummy callgen
callgen = CallgenResult(order=opts.order, lang=opts.lang,
dev_mem_type=wrapper.state['dev_mem_type'],
type_map=type_map(opts.lang))
# create a memory manager
mem = get_memory(callgen, host_namer=HostNamer(), device_namer=DeviceNamer())
a1 = lp.GlobalArg('a1', shape=(arc.problem_size), dtype=np.int32)
a2 = lp.GlobalArg('a2', shape=(arc.problem_size, 10), dtype=np.int64)
d3 = lp.GlobalArg('d3', shape=(arc.problem_size, 10, 10), dtype=np.float64)
a4 = lp.ValueArg('a4', dtype=np.int64)
a5 = lp.ValueArg('a5', dtype=np.int32)
a6 = lp.TemporaryVariable('a6', initializer=np.array([0, 1, 2]),
read_only=True)
if opts.lang == 'opencl':
assert mem.define(True, a1) == 'cl_mem d_a1;'
assert mem.define(False, a2) == 'long int* h_a2;'
assert mem.define(True, d3) == 'cl_mem d_d3;'
assert mem.define(False, a4) == 'long int h_a4;'
assert mem.define(True, a5) == 'cl_uint d_a5;'
assert mem.define(True, a5) == 'cl_uint d_a5;'
with assert_raises(Exception):
mem.define(True, a6, host_constant=True)
assert mem.define(False, a6, host_constant=True) == \
'const long int h_a6[3] = {0, 1, 2};'
elif opts.lang == 'c':
assert mem.define(True, a1) == 'int* d_a1;'
assert mem.define(False, a2) == 'long int* h_a2;'
assert mem.define(True, d3) == 'double* d_d3;'
assert mem.define(False, a4) == 'long int h_a4;'
assert mem.define(True, a5) == 'int d_a5;'
with assert_raises(Exception):
mem.define(True, a6, host_constant=True)
assert mem.define(False, a6, host_constant=True) == \
'const long int h_a6[3] = {0, 1, 2};'
else:
raise NotImplementedError
def test_c_instruction(ctx_factory):
#logging.basicConfig(level=logging.DEBUG)
ctx = ctx_factory()
knl = lp.make_kernel("{[i,j]: 0<=i,j<n }", [
lp.CInstruction("i,j",
"""
x = sin((float) i*j);
""",
assignees="x"),
"a[i,j] = x",
], [
lp.GlobalArg("a", shape=lp.auto, dtype=np.float32),
lp.TemporaryVariable("x", np.float32),
"...",
],
assumptions="n>=1")
knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
print(knl)
print(lp.CompiledKernel(ctx, knl).get_highlighted_code())
def get_diff_var(self, var_name):
"""
:return: a string containing the name of a new variable
holding the derivative of *var_name* by the desired
*diff_context.by_name*, or *None* if no dependency exists.
"""
new_var_name = self.rule_mapping_context.make_unique_var_name(
var_name + "_d" + self.by_name)
writers = self.kernel.writer_map().get(var_name, [])
if not writers:
# FIXME: There should be hooks to supply earlier dvar_dby
# This would be the spot to think about them.
return None
if len(writers) > 1:
raise LoopyError("%s is written in more than one place" % var_name)
orig_writer_id, = writers
orig_writer_insn = self.kernel.id_to_insn[orig_writer_id]
diff_inames = self.add_diff_inames()
diff_iname_exprs = tuple(var(diname) for diname in diff_inames)
# {{{ write code
diff_mapper = LoopyDiffMapper(self.rule_mapping_context, self,
diff_inames)
diff_expr = diff_mapper(orig_writer_insn.expression, self.kernel,
orig_writer_insn)
if not diff_expr:
return None
assert isinstance(orig_writer_insn, lp.Assignment)
if isinstance(orig_writer_insn.assignee, p.Subscript):
lhs_ind = orig_writer_insn.assignee.index_tuple
elif isinstance(orig_writer_insn.assignee, p.Variable):
lhs_ind = ()
else:
raise LoopyError("Unrecognized LHS type in differentiation: %s" %
type(orig_writer_insn.assignee).__name__)
new_insn_id = self.generate_instruction_id()
insn = lp.Assignment(id=new_insn_id,
assignee=var(new_var_name)[lhs_ind +
diff_iname_exprs],
expression=diff_expr,
within_inames=(orig_writer_insn.within_inames
| frozenset(diff_inames)))
self.new_instructions.append(insn)
# }}}
# {{{ manage variable declaration
if var_name in self.kernel.arg_dict:
arg = self.kernel.arg_dict[var_name]
orig_shape = arg.shape
elif var_name in self.kernel.temporary_variables:
tv = self.kernel.temporary_variables[var_name]
orig_shape = tv.shape
else:
raise ValueError("%s: variable not found" % var_name)
shape = orig_shape + self.additional_shape
dim_tags = ("c", ) * len(shape)
if var_name in self.kernel.arg_dict:
self.new_args.append(
lp.GlobalArg(
new_var_name,
arg.dtype,
shape=shape,
dim_tags=dim_tags,
))
elif var_name in self.kernel.temporary_variables:
self.new_temporary_variables[new_var_name] = lp.TemporaryVariable(
new_var_name, tv.dtype, shape=shape, dim_tags=dim_tags)
# }}}
return new_var_name
def buffer_array(kernel, var_name, buffer_inames, init_expression=None,
store_expression=None, within=None, default_tag="l.auto",
temporary_scope=None, temporary_is_local=None,
fetch_bounding_box=False):
"""Replace accesses to *var_name* with ones to a temporary, which is
created and acts as a buffer. To perform this transformation, the access
footprint to *var_name* is determined and a temporary of a suitable
:class:`loopy.AddressSpace` and shape is created.
By default, the value of the buffered cells in *var_name* are read prior to
any (read/write) use, and the modified values are written out after use has
concluded, but for special use cases (e.g. additive accumulation), the
behavior can be modified using *init_expression* and *store_expression*.
:arg buffer_inames: The inames across which the buffer should be usable--i.e.
all possible values of these inames will be covered by the buffer footprint.
A tuple of inames or a comma-separated string.
:arg init_expression: Either *None* (indicating the prior value of the buffered
array should be read) or an expression optionally involving the
variable 'base' (which references the associated location in the array
being buffered).
:arg store_expression: Either *None*, *False*, or an expression involving
variables 'base' and 'buffer' (without array indices).
(*None* indicates that a default storage instruction should be used,
*False* indicates that no storing of the temporary should occur
at all.)
:arg within: If not None, limit the action of the transformation to
matching contexts. See :func:`loopy.match.parse_stack_match`
for syntax.
:arg temporary_scope: If given, override the choice of
:class:`AddressSpace` for the created temporary.
:arg default_tag: The default :ref:`iname-tags` to be assigned to the
inames used for fetching and storing
:arg fetch_bounding_box: If the access footprint is non-convex
(resulting in an error), setting this argument to *True* will force a
rectangular (and hence convex) superset of the footprint to be
fetched.
"""
# {{{ unify temporary_scope / temporary_is_local
from loopy.kernel.data import AddressSpace
if temporary_is_local is not None:
from warnings import warn
warn("temporary_is_local is deprecated. Use temporary_scope instead",
DeprecationWarning, stacklevel=2)
if temporary_scope is not None:
raise LoopyError("may not specify both temporary_is_local and "
"temporary_scope")
if temporary_is_local:
temporary_scope = AddressSpace.LOCAL
else:
temporary_scope = AddressSpace.PRIVATE
del temporary_is_local
# }}}
# {{{ process arguments
if isinstance(init_expression, str):
from loopy.symbolic import parse
init_expression = parse(init_expression)
if isinstance(store_expression, str):
from loopy.symbolic import parse
store_expression = parse(store_expression)
if isinstance(buffer_inames, str):
buffer_inames = [s.strip()
for s in buffer_inames.split(",") if s.strip()]
for iname in buffer_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname"
% iname)
buffer_inames = list(buffer_inames)
buffer_inames_set = frozenset(buffer_inames)
from loopy.match import parse_stack_match
within = parse_stack_match(within)
if var_name in kernel.arg_dict:
var_descr = kernel.arg_dict[var_name]
elif var_name in kernel.temporary_variables:
var_descr = kernel.temporary_variables[var_name]
else:
raise ValueError("variable '%s' not found" % var_name)
from loopy.kernel.data import ArrayBase
if isinstance(var_descr, ArrayBase):
var_shape = var_descr.shape
else:
var_shape = ()
if temporary_scope is None:
import loopy as lp
temporary_scope = lp.auto
# }}}
# {{{ caching
from loopy import CACHING_ENABLED
from loopy.preprocess import prepare_for_caching
key_kernel = prepare_for_caching(kernel)
cache_key = (key_kernel, var_name, tuple(buffer_inames),
PymbolicExpressionHashWrapper(init_expression),
PymbolicExpressionHashWrapper(store_expression), within,
default_tag, temporary_scope, fetch_bounding_box)
if CACHING_ENABLED:
try:
result = buffer_array_cache[cache_key]
logger.info("%s: buffer_array cache hit" % kernel.name)
return result
except KeyError:
pass
# }}}
var_name_gen = kernel.get_var_name_generator()
within_inames = set()
access_descriptors = []
for insn in kernel.instructions:
if not within(kernel, insn.id, ()):
continue
from pymbolic.primitives import Variable, Subscript
from loopy.symbolic import LinearSubscript
for assignee in insn.assignees:
if isinstance(assignee, Variable):
assignee_name = assignee.name
index = ()
elif isinstance(assignee, Subscript):
assignee_name = assignee.aggregate.name
index = assignee.index_tuple
elif isinstance(assignee, LinearSubscript):
if assignee.aggregate.name == var_name:
raise LoopyError("buffer_array may not be applied in the "
"presence of linear write indexing into '%s'" % var_name)
else:
raise LoopyError("invalid lvalue '%s'" % assignee)
if assignee_name == var_name:
within_inames.update(
(get_dependencies(index) & kernel.all_inames())
- buffer_inames_set)
access_descriptors.append(
AccessDescriptor(
identifier=insn.id,
storage_axis_exprs=index))
# {{{ find fetch/store inames
init_inames = []
store_inames = []
new_iname_to_tag = {}
for i in range(len(var_shape)):
dim_name = str(i)
if isinstance(var_descr, ArrayBase) and var_descr.dim_names is not None:
dim_name = var_descr.dim_names[i]
init_iname = var_name_gen(f"{var_name}_init_{dim_name}")
store_iname = var_name_gen(f"{var_name}_store_{dim_name}")
new_iname_to_tag[init_iname] = default_tag
new_iname_to_tag[store_iname] = default_tag
init_inames.append(init_iname)
store_inames.append(store_iname)
# }}}
# {{{ modify loop domain
non1_init_inames = []
non1_store_inames = []
if var_shape:
# {{{ find domain to be changed
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, buffer_inames_set | within_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in buffer_inames_set:
if kernel.get_home_domain_index(iname) != domch.leaf_domain_index:
raise RuntimeError("buffer iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, buffer_inames,
access_descriptors, len(var_shape))
for i in range(len(var_shape)):
if abm.non1_storage_axis_flags[i]:
non1_init_inames.append(init_inames[i])
non1_store_inames.append(store_inames[i])
else:
del new_iname_to_tag[init_inames[i]]
del new_iname_to_tag[store_inames[i]]
new_domain = domch.domain
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_init_inames,
boxify_sweep=fetch_bounding_box)
new_domain = abm.augment_domain_with_sweep(
new_domain, non1_store_inames,
boxify_sweep=fetch_bounding_box)
new_kernel_domains = domch.get_domains_with(new_domain)
del new_domain
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
abm = NoOpArrayToBufferMap()
# }}}
# {{{ set up temp variable
import loopy as lp
buf_var_name = var_name_gen(based_on=var_name+"_buf")
new_temporary_variables = kernel.temporary_variables.copy()
temp_var = lp.TemporaryVariable(
name=buf_var_name,
dtype=var_descr.dtype,
base_indices=(0,)*len(abm.non1_storage_shape),
shape=tuple(abm.non1_storage_shape),
address_space=temporary_scope)
new_temporary_variables[buf_var_name] = temp_var
# }}}
new_insns = []
buf_var = var(buf_var_name)
# {{{ generate init instruction
buf_var_init = buf_var
if non1_init_inames:
buf_var_init = buf_var_init.index(
tuple(var(iname) for iname in non1_init_inames))
init_base = var(var_name)
init_subscript = []
init_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_init_inames[init_iname_idx])
init_iname_idx += 1
init_subscript.append(ax_subscript)
if init_subscript:
init_base = init_base.index(tuple(init_subscript))
if init_expression is None:
init_expression = init_base
else:
init_expression = init_expression
init_expression = SubstitutionMapper(
make_subst_func({
"base": init_base,
}))(init_expression)
init_insn_id = kernel.make_unique_instruction_id(based_on="init_"+var_name)
from loopy.kernel.data import Assignment
init_instruction = Assignment(id=init_insn_id,
assignee=buf_var_init,
expression=init_expression,
within_inames=(
frozenset(within_inames)
| frozenset(non1_init_inames)),
depends_on=frozenset(),
depends_on_is_final=True)
# }}}
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
aar = ArrayAccessReplacer(rule_mapping_context, var_name,
within, abm, buf_var)
kernel = rule_mapping_context.finish_kernel(aar.map_kernel(kernel))
did_write = False
for insn_id in aar.modified_insn_ids:
insn = kernel.id_to_insn[insn_id]
if buf_var_name in insn.assignee_var_names():
did_write = True
# {{{ add init_insn_id to depends_on
new_insns = []
def none_to_empty_set(s):
if s is None:
return frozenset()
else:
return s
for insn in kernel.instructions:
if insn.id in aar.modified_insn_ids:
new_insns.append(
insn.copy(
depends_on=(
none_to_empty_set(insn.depends_on)
| frozenset([init_insn_id]))))
else:
new_insns.append(insn)
# }}}
# {{{ generate store instruction
buf_var_store = buf_var
if non1_store_inames:
buf_var_store = buf_var_store.index(
tuple(var(iname) for iname in non1_store_inames))
store_subscript = []
store_iname_idx = 0
if var_shape:
for i in range(len(var_shape)):
ax_subscript = abm.storage_base_indices[i]
if abm.non1_storage_axis_flags[i]:
ax_subscript += var(non1_store_inames[store_iname_idx])
store_iname_idx += 1
store_subscript.append(ax_subscript)
store_target = var(var_name)
if store_subscript:
store_target = store_target.index(tuple(store_subscript))
if store_expression is None:
store_expression = buf_var_store
else:
store_expression = SubstitutionMapper(
make_subst_func({
"base": store_target,
"buffer": buf_var_store,
}))(store_expression)
if store_expression is not False:
from loopy.kernel.data import Assignment
store_instruction = Assignment(
id=kernel.make_unique_instruction_id(based_on="store_"+var_name),
depends_on=frozenset(aar.modified_insn_ids),
no_sync_with=frozenset([(init_insn_id, "any")]),
assignee=store_target,
expression=store_expression,
within_inames=(
frozenset(within_inames)
| frozenset(non1_store_inames)))
else:
did_write = False
# }}}
new_insns.append(init_instruction)
if did_write:
new_insns.append(store_instruction)
else:
for iname in store_inames:
del new_iname_to_tag[iname]
kernel = kernel.copy(
domains=new_kernel_domains,
instructions=new_insns,
temporary_variables=new_temporary_variables)
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
if CACHING_ENABLED:
from loopy.preprocess import prepare_for_caching
buffer_array_cache.store_if_not_present(
cache_key, prepare_for_caching(kernel))
return kernel
def generate(impero_c,
args,
scalar_type,
kernel_name="loopy_kernel",
index_names=[],
return_increments=True):
"""Generates loopy code.
:arg impero_c: ImperoC tuple with Impero AST and other data
:arg args: list of loopy.GlobalArgs
:arg scalar_type: type of scalars as C typename string
:arg kernel_name: function name of the kernel
:arg index_names: pre-assigned index names
:arg return_increments: Does codegen for Return nodes increment the lvalue, or assign?
:returns: loopy kernel
"""
ctx = LoopyContext()
ctx.indices = impero_c.indices
ctx.index_names = defaultdict(lambda: "i", index_names)
ctx.epsilon = numpy.finfo(scalar_type).resolution
ctx.scalar_type = scalar_type
ctx.return_increments = return_increments
# Create arguments
data = list(args)
for i, (temp, dtype) in enumerate(
assign_dtypes(impero_c.temporaries, scalar_type)):
name = "t%d" % i
if isinstance(temp, gem.Constant):
data.append(
lp.TemporaryVariable(name,
shape=temp.shape,
dtype=dtype,
initializer=temp.array,
address_space=lp.AddressSpace.LOCAL,
read_only=True))
else:
shape = tuple([i.extent for i in ctx.indices[temp]]) + temp.shape
data.append(
lp.TemporaryVariable(name,
shape=shape,
dtype=dtype,
initializer=None,
address_space=lp.AddressSpace.LOCAL,
read_only=False))
ctx.gem_to_pymbolic[temp] = p.Variable(name)
# Create instructions
instructions = statement(impero_c.tree, ctx)
# Create domains
domains = create_domains(ctx.index_extent.items())
# Create loopy kernel
knl = lp.make_function(domains,
instructions,
data,
name=kernel_name,
target=lp.CTarget(),
seq_dependencies=True,
silenced_warnings=["summing_if_branches_ops"],
lang_version=(2018, 2))
# Prevent loopy interchange by loopy
knl = lp.prioritize_loops(knl, ",".join(ctx.index_extent.keys()))
return knl
def __init__(self, fft, effective_k, dk, dx):
self.fft = fft
if not callable(effective_k):
if effective_k != 0:
from pystella.derivs import FirstCenteredDifference
h = effective_k
effective_k = FirstCenteredDifference(h).get_eigenvalues
else:
def effective_k(k, dx): # pylint: disable=function-redefined
return k
queue = self.fft.sub_k["momenta_x"].queue
sub_k = list(x.get().astype("int") for x in self.fft.sub_k.values())
eff_mom_names = ("eff_mom_x", "eff_mom_y", "eff_mom_z")
self.eff_mom = {}
for mu, (name, kk) in enumerate(zip(eff_mom_names, sub_k)):
eff_k = effective_k(dk[mu] * kk.astype(fft.rdtype), dx[mu])
eff_k[abs(sub_k[mu]) == fft.grid_shape[mu] // 2] = 0.
eff_k[sub_k[mu] == 0] = 0.
import pyopencl.array as cla
self.eff_mom[name] = cla.to_device(queue, eff_k)
from pymbolic import var, parse
from pymbolic.primitives import If, Comparison, LogicalAnd
from pystella import Field
indices = parse("i, j, k")
eff_k = tuple(
var(array)[mu] for array, mu in zip(eff_mom_names, indices))
fabs, sqrt, conj = parse("fabs, sqrt, conj")
kmag = sqrt(sum(kk**2 for kk in eff_k))
from pystella import ElementWiseMap
vector = Field("vector", shape=(3, ))
vector_T = Field("vector_T", shape=(3, ))
kvec_zero = LogicalAnd(
tuple(Comparison(fabs(eff_k[mu]), "<", 1e-14) for mu in range(3)))
# note: write all output via private temporaries to allow for in-place
div = var("div")
div_insn = [(div, sum(eff_k[mu] * vector[mu] for mu in range(3)))]
self.transversify_knl = ElementWiseMap(
{
vector_T[mu]: If(kvec_zero, 0,
vector[mu] - eff_k[mu] / kmag**2 * div)
for mu in range(3)
},
tmp_instructions=div_insn,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
import loopy as lp
def assign(asignee, expr, **kwargs):
default = dict(within_inames=frozenset(("i", "j", "k")),
no_sync_with=[("*", "any")])
default.update(kwargs)
return lp.Assignment(asignee, expr, **default)
kmag, Kappa = parse("kmag, Kappa")
eps_insns = [
assign(kmag, sqrt(sum(kk**2 for kk in eff_k))),
assign(Kappa, sqrt(sum(kk**2 for kk in eff_k[:2])))
]
zero = fft.cdtype.type(0)
kx_ky_zero = LogicalAnd(
tuple(Comparison(fabs(eff_k[mu]), "<", 1e-10) for mu in range(2)))
kz_nonzero = Comparison(fabs(eff_k[2]), ">", 1e-10)
eps = var("eps")
eps_insns.extend([
assign(
eps[0],
If(kx_ky_zero, If(kz_nonzero, fft.cdtype.type(1 / 2**.5),
zero),
(eff_k[0] * eff_k[2] / kmag - 1j * eff_k[1]) / Kappa /
2**.5)),
assign(
eps[1],
If(kx_ky_zero,
If(kz_nonzero, fft.cdtype.type(1j / 2**(1 / 2)),
zero), (eff_k[1] * eff_k[2] / kmag + 1j * eff_k[0]) /
Kappa / 2**.5)),
assign(eps[2], If(kx_ky_zero, zero, -Kappa / kmag / 2**.5))
])
plus, minus, lng = Field("plus"), Field("minus"), Field("lng")
plus_tmp, minus_tmp = parse("plus_tmp, minus_tmp")
pol_isns = [(plus_tmp,
sum(vector[mu] * conj(eps[mu]) for mu in range(3))),
(minus_tmp, sum(vector[mu] * eps[mu] for mu in range(3)))]
args = [
lp.TemporaryVariable("kmag"),
lp.TemporaryVariable("Kappa"),
lp.TemporaryVariable("eps", shape=(3, )), ...
]
self.vec_to_pol_knl = ElementWiseMap(
{
plus: plus_tmp,
minus: minus_tmp
},
tmp_instructions=eps_insns + pol_isns,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
vector_tmp = var("vector_tmp")
vec_insns = [(vector_tmp[mu], plus * eps[mu] + minus * conj(eps[mu]))
for mu in range(3)]
self.pol_to_vec_knl = ElementWiseMap(
{vector[mu]: vector_tmp[mu]
for mu in range(3)},
tmp_instructions=eps_insns + vec_insns,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
ksq = sum(kk**2 for kk in eff_k)
lng_rhs = If(kvec_zero, 0, -div / ksq * 1j)
self.vec_decomp_knl = ElementWiseMap(
{
plus: plus_tmp,
minus: minus_tmp,
lng: lng_rhs
},
tmp_instructions=eps_insns + pol_isns + div_insn,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
lng_rhs = If(kvec_zero, 0, -div / ksq**.5 * 1j)
self.vec_decomp_knl_times_abs_k = ElementWiseMap(
{
plus: plus_tmp,
minus: minus_tmp,
lng: lng_rhs
},
tmp_instructions=eps_insns + pol_isns + div_insn,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
from pystella.sectors import tensor_index as tid
eff_k_hat = tuple(kk / sqrt(sum(kk**2 for kk in eff_k))
for kk in eff_k)
hij = Field("hij", shape=(6, ))
hij_TT = Field("hij_TT", shape=(6, ))
Pab = var("P")
Pab_insns = [(Pab[tid(a, b)], (If(Comparison(a, "==", b), 1, 0) -
eff_k_hat[a - 1] * eff_k_hat[b - 1]))
for a in range(1, 4) for b in range(a, 4)]
hij_TT_tmp = var("hij_TT_tmp")
TT_insns = [(hij_TT_tmp[tid(a, b)],
sum((Pab[tid(a, c)] * Pab[tid(d, b)] -
Pab[tid(a, b)] * Pab[tid(c, d)] / 2) * hij[tid(c, d)]
for c in range(1, 4) for d in range(1, 4)))
for a in range(1, 4) for b in range(a, 4)]
# note: where conditionals (branch divergence) go can matter:
# this kernel is twice as fast when putting the branching in the global
# write, rather than when setting hij_TT_tmp
write_insns = [(hij_TT[tid(a,
b)], If(kvec_zero, 0, hij_TT_tmp[tid(a,
b)]))
for a in range(1, 4) for b in range(a, 4)]
self.tt_knl = ElementWiseMap(
write_insns,
tmp_instructions=Pab_insns + TT_insns,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
tensor_to_pol_insns = {
plus:
sum(hij[tid(c, d)] * conj(eps[c - 1]) * conj(eps[d - 1])
for c in range(1, 4) for d in range(1, 4)),
minus:
sum(hij[tid(c, d)] * eps[c - 1] * eps[d - 1] for c in range(1, 4)
for d in range(1, 4))
}
self.tensor_to_pol_knl = ElementWiseMap(
tensor_to_pol_insns,
tmp_instructions=eps_insns,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
pol_to_tensor_insns = {
hij[tid(a, b)]: (plus * eps[a - 1] * eps[b - 1] +
minus * conj(eps[a - 1]) * conj(eps[b - 1]))
for a in range(1, 4) for b in range(a, 4)
}
self.pol_to_tensor_knl = ElementWiseMap(
pol_to_tensor_insns,
tmp_instructions=eps_insns,
args=args,
lsize=(32, 1, 1),
rank_shape=fft.shape(True),
)
def __init__(self, decomp, histograms, num_bins, dtype, **kwargs):
self.decomp = decomp
self.histograms = histograms
self.num_bins = num_bins
num_hists = len(histograms)
from pymbolic import var
_bin = var("bin")
b = var("b")
bb = var("bb")
hist = var("hist")
temp = var("temp")
weight_val = var("weight")
args = kwargs.pop("args", [])
args += [
lp.TemporaryVariable("temp",
dtype,
shape=(
num_hists,
self.num_bins,
),
for_atomic=True,
address_space=lp.AddressSpace.LOCAL),
lp.TemporaryVariable("bin", "int", shape=(num_hists, )),
lp.TemporaryVariable("weight", dtype, shape=(num_hists, )),
lp.GlobalArg("hist",
dtype,
shape=(
num_hists,
self.num_bins,
),
for_atomic=True),
]
fixed_pars = kwargs.pop("fixed_parameters", dict())
fixed_pars.update(dict(num_bins=num_bins, num_hists=num_hists))
silenced_warnings = kwargs.pop("silenced_warnings", [])
silenced_warnings += ["write_race(tmp*)", "write_race(glb*)"]
domains = """
[Nx, Ny, Nz, num_bins] ->
{[i, j, k, b, bb]: 0<=i<Nx and 0<=j<Ny and 0<=k<Nz and 0<=b<num_bins
and 0<=bb<num_bins}
"""
insns = [
lp.Assignment(hist[j, bb],
0,
id=f"zero_hist_{j}",
within_inames=frozenset("bb"),
atomicity=(lp.AtomicInit(str(hist)), ))
for j in range(num_hists)
]
insns.append(
lp.BarrierInstruction("post_zero_barrier",
synchronization_kind="global"))
insns.extend([
lp.Assignment(temp[j, bb],
0,
id=f"zero_temp_{j}",
within_inames=frozenset(("j", "bb")),
atomicity=(lp.AtomicInit(str(temp)), ))
for j in range(num_hists)
])
for j, (bin_expr, weight_expr) in enumerate(histograms.values()):
insns.extend([
lp.Assignment(_bin[j],
var("floor")(bin_expr),
id=f"set_bin_{j}",
within_inames=frozenset(("i", "j", "k"))),
lp.Assignment(weight_val[j],
weight_expr,
id=f"set_weight_{j}",
within_inames=frozenset(("i", "j", "k"))),
lp.Assignment(temp[j, _bin[j]],
temp[j, _bin[j]] + weight_val[j],
id=f"tmp_{j}",
within_inames=frozenset(("i", "j", "k")),
atomicity=(lp.AtomicUpdate(str(temp)), ))
])
insns.extend([
lp.Assignment(hist[j, b],
hist[j, b] + temp[j, b],
id=f"glb_{j}",
within_inames=frozenset(("j", "b")),
atomicity=(lp.AtomicUpdate(str(hist)), ))
for j in range(num_hists)
])
lsize = [min(256, self.num_bins)]
super().__init__(insns,
args=args,
lsize=lsize,
fixed_parameters=fixed_pars,
domains=domains,
silenced_warnings=silenced_warnings,
**kwargs)
def precompute(
kernel,
subst_use,
sweep_inames=[],
within=None,
storage_axes=None,
temporary_name=None,
precompute_inames=None,
precompute_outer_inames=None,
storage_axis_to_tag={},
# "None" is a valid value here, distinct from the default.
default_tag=_not_provided,
dtype=None,
fetch_bounding_box=False,
temporary_address_space=None,
compute_insn_id=None,
**kwargs):
"""Precompute the expression described in the substitution rule determined by
*subst_use* and store it in a temporary array. A precomputation needs two
things to operate, a list of *sweep_inames* (order irrelevant) and an
ordered list of *storage_axes* (whose order will describe the axis ordering
of the temporary array).
:arg subst_use: Describes what to prefetch.
The following objects may be given for *subst_use*:
* The name of the substitution rule.
* The tagged name ("name$tag") of the substitution rule.
* A list of invocations of the substitution rule.
This list of invocations, when swept across *sweep_inames*, then serves
to define the footprint of the precomputation.
Invocations may be tagged ("name$tag") to filter out a subset of the
usage sites of the substitution rule. (Namely those usage sites that
use the same tagged name.)
Invocations may be given as a string or as a
:class:`pymbolic.primitives.Expression` object.
If only one invocation is to be given, then the only entry of the list
may be given directly.
If the list of invocations generating the footprint is not given,
all (tag-matching, if desired) usage sites of the substitution rule
are used to determine the footprint.
The following cases can arise for each sweep axis:
* The axis is an iname that occurs within arguments specified at
usage sites of the substitution rule. This case is assumed covered
by the storage axes provided for the argument.
* The axis is an iname that occurs within the *value* of the rule, but not
within its arguments. A new, dedicated storage axis is allocated for
such an axis.
:arg sweep_inames: A :class:`list` of inames to be swept.
May also equivalently be a comma-separated string.
:arg within: a stack match as understood by
:func:`loopy.match.parse_stack_match`.
:arg storage_axes: A :class:`list` of inames and/or rule argument
names/indices to be used as storage axes.
May also equivalently be a comma-separated string.
:arg temporary_name:
The temporary variable name to use for storing the precomputed data.
If it does not exist, it will be created. If it does exist, its properties
(such as size, type) are checked (and updated, if possible) to match
its use.
:arg precompute_inames:
A tuple of inames to be used to carry out the precomputation.
If the specified inames do not already exist, they will be
created. If they do already exist, their loop domain is verified
against the one required for this precomputation. This tuple may
be shorter than the (provided or automatically found) *storage_axes*
tuple, in which case names will be automatically created.
May also equivalently be a comma-separated string.
:arg precompute_outer_inames: A :class:`frozenset` of inames within which
the compute instruction is nested. If *None*, make an educated guess.
May also be specified as a comma-separated string.
:arg default_tag: The :ref:`iname tag <iname-tags>` to be applied to the
inames created to perform the precomputation. The current default will
make them local axes and automatically split them to fit the work
group size, but this default will disappear in favor of simply leaving them
untagged in 2019. For 2018, a warning will be issued if no *default_tag* is
specified.
:arg compute_insn_id: The ID of the instruction generated to perform the
precomputation.
If `storage_axes` is not specified, it defaults to the arrangement
`<direct sweep axes><arguments>` with the direct sweep axes being the
slower-varying indices.
Trivial storage axes (i.e. axes of length 1 with respect to the sweep) are
eliminated.
"""
# {{{ unify temporary_address_space / temporary_scope
temporary_scope = kwargs.pop("temporary_scope", None)
from loopy.kernel.data import AddressSpace
if temporary_scope is not None:
from warnings import warn
warn(
"temporary_scope is deprecated. Use temporary_address_space instead",
DeprecationWarning,
stacklevel=2)
if temporary_address_space is not None:
raise LoopyError(
"may not specify both temporary_address_space and "
"temporary_scope")
temporary_address_space = temporary_scope
del temporary_scope
# }}}
if kwargs:
raise TypeError("unrecognized keyword arguments: %s" %
", ".join(kwargs.keys()))
# {{{ check, standardize arguments
if isinstance(sweep_inames, str):
sweep_inames = [iname.strip() for iname in sweep_inames.split(",")]
for iname in sweep_inames:
if iname not in kernel.all_inames():
raise RuntimeError("sweep iname '%s' is not a known iname" % iname)
sweep_inames = list(sweep_inames)
sweep_inames_set = frozenset(sweep_inames)
if isinstance(storage_axes, str):
storage_axes = [ax.strip() for ax in storage_axes.split(",")]
if isinstance(precompute_inames, str):
precompute_inames = [
iname.strip() for iname in precompute_inames.split(",")
]
if isinstance(precompute_outer_inames, str):
precompute_outer_inames = frozenset(
iname.strip() for iname in precompute_outer_inames.split(","))
if isinstance(subst_use, str):
subst_use = [subst_use]
footprint_generators = None
subst_name = None
subst_tag = None
from pymbolic.primitives import Variable, Call
from loopy.symbolic import parse, TaggedVariable
for use in subst_use:
if isinstance(use, str):
use = parse(use)
if isinstance(use, Call):
if footprint_generators is None:
footprint_generators = []
footprint_generators.append(use)
subst_name_as_expr = use.function
else:
subst_name_as_expr = use
if isinstance(subst_name_as_expr, TaggedVariable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = subst_name_as_expr.tag
elif isinstance(subst_name_as_expr, Variable):
new_subst_name = subst_name_as_expr.name
new_subst_tag = None
else:
raise ValueError("unexpected type of subst_name")
if (subst_name, subst_tag) == (None, None):
subst_name, subst_tag = new_subst_name, new_subst_tag
else:
if (subst_name, subst_tag) != (new_subst_name, new_subst_tag):
raise ValueError("not all uses in subst_use agree "
"on rule name and tag")
from loopy.match import parse_stack_match
within = parse_stack_match(within)
try:
subst = kernel.substitutions[subst_name]
except KeyError:
raise LoopyError("substitution rule '%s' not found" % subst_name)
c_subst_name = subst_name.replace(".", "_")
# {{{ handle default_tag
from loopy.transform.data import _not_provided \
as transform_data_not_provided
if default_tag is _not_provided or default_tag is transform_data_not_provided:
# no need to warn for scalar precomputes
if sweep_inames:
from warnings import warn
warn(
"Not specifying default_tag is deprecated, and default_tag "
"will become mandatory in 2019.x. "
"Pass 'default_tag=\"l.auto\" to match the current default, "
"or Pass 'default_tag=None to leave the loops untagged, which "
"is the recommended behavior.",
DeprecationWarning,
stacklevel=(
# In this case, we came here through add_prefetch. Increase
# the stacklevel.
3 if default_tag is transform_data_not_provided else 2))
default_tag = "l.auto"
from loopy.kernel.data import parse_tag
default_tag = parse_tag(default_tag)
# }}}
# }}}
# {{{ process invocations in footprint generators, start access_descriptors
if footprint_generators:
from pymbolic.primitives import Variable, Call
access_descriptors = []
for fpg in footprint_generators:
if isinstance(fpg, Variable):
args = ()
elif isinstance(fpg, Call):
args = fpg.parameters
else:
raise ValueError("footprint generator must "
"be substitution rule invocation")
access_descriptors.append(
RuleAccessDescriptor(identifier=access_descriptor_id(
args, None),
args=args))
# }}}
# {{{ gather up invocations in kernel code, finish access_descriptors
if not footprint_generators:
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
invg = RuleInvocationGatherer(rule_mapping_context, kernel, subst_name,
subst_tag, within)
del rule_mapping_context
import loopy as lp
for insn in kernel.instructions:
if isinstance(insn, lp.MultiAssignmentBase):
for assignee in insn.assignees:
invg(assignee, kernel, insn)
invg(insn.expression, kernel, insn)
access_descriptors = invg.access_descriptors
if not access_descriptors:
raise RuntimeError("no invocations of '%s' found" % subst_name)
# }}}
# {{{ find inames used in arguments
expanding_usage_arg_deps = set()
for accdesc in access_descriptors:
for arg in accdesc.args:
expanding_usage_arg_deps.update(
get_dependencies(arg) & kernel.all_inames())
# }}}
var_name_gen = kernel.get_var_name_generator()
# {{{ use given / find new storage_axes
# extra axes made necessary because they don't occur in the arguments
extra_storage_axes = set(sweep_inames_set - expanding_usage_arg_deps)
from loopy.symbolic import SubstitutionRuleExpander
submap = SubstitutionRuleExpander(kernel.substitutions)
value_inames = (get_dependencies(submap(subst.expression)) -
frozenset(subst.arguments)) & kernel.all_inames()
if value_inames - expanding_usage_arg_deps < extra_storage_axes:
raise RuntimeError("unreferenced sweep inames specified: " +
", ".join(extra_storage_axes - value_inames -
expanding_usage_arg_deps))
new_iname_to_tag = {}
if storage_axes is None:
storage_axes = []
# Add sweep_inames (in given--rather than arbitrary--order) to
# storage_axes *if* they are part of extra_storage_axes.
for iname in sweep_inames:
if iname in extra_storage_axes:
extra_storage_axes.remove(iname)
storage_axes.append(iname)
if extra_storage_axes:
if (precompute_inames is not None
and len(storage_axes) < len(precompute_inames)):
raise LoopyError(
"must specify a sufficient number of "
"storage_axes to uniquely determine the meaning "
"of the given precompute_inames. (%d storage_axes "
"needed)" % len(precompute_inames))
storage_axes.extend(sorted(extra_storage_axes))
storage_axes.extend(range(len(subst.arguments)))
del extra_storage_axes
prior_storage_axis_name_dict = {}
storage_axis_names = []
storage_axis_sources = [] # number for arg#, or iname
# {{{ check for pre-existing precompute_inames
if precompute_inames is not None:
preexisting_precompute_inames = (set(precompute_inames)
& kernel.all_inames())
else:
preexisting_precompute_inames = set()
# }}}
for i, saxis in enumerate(storage_axes):
tag_lookup_saxis = saxis
if saxis in subst.arguments:
saxis = subst.arguments.index(saxis)
storage_axis_sources.append(saxis)
if isinstance(saxis, int):
# argument index
name = old_name = subst.arguments[saxis]
else:
old_name = saxis
name = "%s_%s" % (c_subst_name, old_name)
if (precompute_inames is not None and i < len(precompute_inames)
and precompute_inames[i]):
name = precompute_inames[i]
tag_lookup_saxis = name
if (name not in preexisting_precompute_inames
and var_name_gen.is_name_conflicting(name)):
raise RuntimeError("new storage axis name '%s' "
"conflicts with existing name" % name)
else:
name = var_name_gen(name)
storage_axis_names.append(name)
if name not in preexisting_precompute_inames:
new_iname_to_tag[name] = storage_axis_to_tag.get(
tag_lookup_saxis, default_tag)
prior_storage_axis_name_dict[name] = old_name
del storage_axis_to_tag
del storage_axes
del precompute_inames
# }}}
# {{{ fill out access_descriptors[...].storage_axis_exprs
access_descriptors = [
accdesc.copy(storage_axis_exprs=storage_axis_exprs(
storage_axis_sources, accdesc.args))
for accdesc in access_descriptors
]
# }}}
expanding_inames = sweep_inames_set | frozenset(expanding_usage_arg_deps)
assert expanding_inames <= kernel.all_inames()
if storage_axis_names:
# {{{ find domain to be changed
change_inames = expanding_inames | preexisting_precompute_inames
from loopy.kernel.tools import DomainChanger
domch = DomainChanger(kernel, change_inames)
if domch.leaf_domain_index is not None:
# If the sweep inames are at home in parent domains, then we'll add
# fetches with loops over copies of these parent inames that will end
# up being scheduled *within* loops over these parents.
for iname in sweep_inames_set:
if kernel.get_home_domain_index(
iname) != domch.leaf_domain_index:
raise RuntimeError(
"sweep iname '%s' is not 'at home' in the "
"sweep's leaf domain" % iname)
# }}}
abm = ArrayToBufferMap(kernel, domch.domain, sweep_inames,
access_descriptors, len(storage_axis_names))
non1_storage_axis_names = []
for i, saxis in enumerate(storage_axis_names):
if abm.non1_storage_axis_flags[i]:
non1_storage_axis_names.append(saxis)
else:
del new_iname_to_tag[saxis]
if saxis in preexisting_precompute_inames:
raise LoopyError(
"precompute axis %d (1-based) was "
"eliminated as "
"having length 1 but also mapped to existing "
"iname '%s'" % (i + 1, saxis))
mod_domain = domch.domain
# {{{ modify the domain, taking into account preexisting inames
# inames may already exist in mod_domain, add them primed to start
primed_non1_saxis_names = [
iname + "'" for iname in non1_storage_axis_names
]
mod_domain = abm.augment_domain_with_sweep(
domch.domain,
primed_non1_saxis_names,
boxify_sweep=fetch_bounding_box)
check_domain = mod_domain
for i, saxis in enumerate(non1_storage_axis_names):
var_dict = mod_domain.get_var_dict(isl.dim_type.set)
if saxis in preexisting_precompute_inames:
# add equality constraint between existing and new variable
dt, dim_idx = var_dict[saxis]
saxis_aff = isl.Aff.var_on_domain(mod_domain.space, dt,
dim_idx)
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
new_var_aff = isl.Aff.var_on_domain(mod_domain.space, dt,
dim_idx)
mod_domain = mod_domain.add_constraint(
isl.Constraint.equality_from_aff(new_var_aff - saxis_aff))
# project out the new one
mod_domain = mod_domain.project_out(dt, dim_idx, 1)
else:
# remove the prime from the new variable
dt, dim_idx = var_dict[primed_non1_saxis_names[i]]
mod_domain = mod_domain.set_dim_name(dt, dim_idx, saxis)
def add_assumptions(d):
assumption_non_param = isl.BasicSet.from_params(kernel.assumptions)
assumptions, domain = isl.align_two(assumption_non_param, d)
return assumptions & domain
# {{{ check that we got the desired domain
check_domain = add_assumptions(
check_domain.project_out_except(primed_non1_saxis_names,
[isl.dim_type.set]))
mod_check_domain = add_assumptions(mod_domain)
# re-add the prime from the new variable
var_dict = mod_check_domain.get_var_dict(isl.dim_type.set)
for saxis in non1_storage_axis_names:
dt, dim_idx = var_dict[saxis]
mod_check_domain = mod_check_domain.set_dim_name(
dt, dim_idx, saxis + "'")
mod_check_domain = mod_check_domain.project_out_except(
primed_non1_saxis_names, [isl.dim_type.set])
mod_check_domain, check_domain = isl.align_two(mod_check_domain,
check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError("domain of preexisting inames does not match "
"domain needed for precompute")
# }}}
# {{{ check that we didn't shrink the original domain
# project out the new names from the modified domain
orig_domain_inames = list(domch.domain.get_var_dict(isl.dim_type.set))
mod_check_domain = add_assumptions(
mod_domain.project_out_except(orig_domain_inames,
[isl.dim_type.set]))
check_domain = add_assumptions(domch.domain)
mod_check_domain, check_domain = isl.align_two(mod_check_domain,
check_domain)
# The modified domain can't get bigger by adding constraints
assert mod_check_domain <= check_domain
if not check_domain <= mod_check_domain:
print(check_domain)
print(mod_check_domain)
raise LoopyError(
"original domain got shrunk by applying the precompute")
# }}}
# }}}
new_kernel_domains = domch.get_domains_with(mod_domain)
else:
# leave kernel domains unchanged
new_kernel_domains = kernel.domains
non1_storage_axis_names = []
abm = NoOpArrayToBufferMap()
kernel = kernel.copy(domains=new_kernel_domains)
# {{{ set up compute insn
if temporary_name is None:
temporary_name = var_name_gen(based_on=c_subst_name)
assignee = var(temporary_name)
if non1_storage_axis_names:
assignee = assignee[tuple(
var(iname) for iname in non1_storage_axis_names)]
# {{{ process substitutions on compute instruction
storage_axis_subst_dict = {}
for arg_name, bi in zip(storage_axis_names, abm.storage_base_indices):
if arg_name in non1_storage_axis_names:
arg = var(arg_name)
else:
arg = 0
storage_axis_subst_dict[prior_storage_axis_name_dict.get(
arg_name, arg_name)] = arg + bi
rule_mapping_context = SubstitutionRuleMappingContext(
kernel.substitutions, kernel.get_var_name_generator())
from loopy.match import parse_stack_match
expr_subst_map = RuleAwareSubstitutionMapper(
rule_mapping_context,
make_subst_func(storage_axis_subst_dict),
within=parse_stack_match(None))
compute_expression = expr_subst_map(subst.expression, kernel, None)
# }}}
from loopy.kernel.data import Assignment
if compute_insn_id is None:
compute_insn_id = kernel.make_unique_instruction_id(
based_on=c_subst_name)
compute_insn = Assignment(
id=compute_insn_id,
assignee=assignee,
expression=compute_expression,
# within_inames determined below
)
compute_dep_id = compute_insn_id
added_compute_insns = [compute_insn]
if temporary_address_space == AddressSpace.GLOBAL:
barrier_insn_id = kernel.make_unique_instruction_id(
based_on=c_subst_name + "_barrier")
from loopy.kernel.instruction import BarrierInstruction
barrier_insn = BarrierInstruction(id=barrier_insn_id,
depends_on=frozenset(
[compute_insn_id]),
synchronization_kind="global",
mem_kind="global")
compute_dep_id = barrier_insn_id
added_compute_insns.append(barrier_insn)
# }}}
# {{{ substitute rule into expressions in kernel (if within footprint)
from loopy.symbolic import SubstitutionRuleExpander
expander = SubstitutionRuleExpander(kernel.substitutions)
invr = RuleInvocationReplacer(rule_mapping_context,
subst_name,
subst_tag,
within,
access_descriptors,
abm,
storage_axis_names,
storage_axis_sources,
non1_storage_axis_names,
temporary_name,
compute_insn_id,
compute_dep_id,
compute_read_variables=get_dependencies(
expander(compute_expression)))
kernel = invr.map_kernel(kernel)
kernel = kernel.copy(instructions=added_compute_insns +
kernel.instructions)
kernel = rule_mapping_context.finish_kernel(kernel)
# }}}
# {{{ add dependencies to compute insn
kernel = kernel.copy(instructions=[
insn.copy(depends_on=frozenset(invr.compute_insn_depends_on)) if insn.
id == compute_insn_id else insn for insn in kernel.instructions
])
# }}}
# {{{ propagate storage iname subst to dependencies of compute instructions
from loopy.kernel.tools import find_recursive_dependencies
compute_deps = find_recursive_dependencies(kernel,
frozenset([compute_insn_id]))
# FIXME: Need to verify that there are no outside dependencies
# on compute_deps
prior_storage_axis_names = frozenset(storage_axis_subst_dict)
new_insns = []
for insn in kernel.instructions:
if (insn.id in compute_deps
and insn.within_inames & prior_storage_axis_names):
insn = (insn.with_transformed_expressions(
lambda expr: expr_subst_map(expr, kernel, insn)).copy(
within_inames=frozenset(
storage_axis_subst_dict.get(iname, var(iname)).name
for iname in insn.within_inames)))
new_insns.append(insn)
else:
new_insns.append(insn)
kernel = kernel.copy(instructions=new_insns)
# }}}
# {{{ determine inames for compute insn
if precompute_outer_inames is None:
from loopy.kernel.tools import guess_iname_deps_based_on_var_use
precompute_outer_inames = (
frozenset(non1_storage_axis_names)
| frozenset((expanding_usage_arg_deps | value_inames) -
sweep_inames_set)
| guess_iname_deps_based_on_var_use(kernel, compute_insn))
else:
if not isinstance(precompute_outer_inames, frozenset):
raise TypeError("precompute_outer_inames must be a frozenset")
precompute_outer_inames = precompute_outer_inames \
| frozenset(non1_storage_axis_names)
kernel = kernel.copy(instructions=[
insn.copy(within_inames=precompute_outer_inames) if insn.id ==
compute_insn_id else insn for insn in kernel.instructions
])
# }}}
# {{{ set up temp variable
import loopy as lp
if dtype is not None:
dtype = np.dtype(dtype)
if temporary_address_space is None:
temporary_address_space = lp.auto
new_temp_shape = tuple(abm.non1_storage_shape)
new_temporary_variables = kernel.temporary_variables.copy()
if temporary_name not in new_temporary_variables:
temp_var = lp.TemporaryVariable(
name=temporary_name,
dtype=dtype,
base_indices=(0, ) * len(new_temp_shape),
shape=tuple(abm.non1_storage_shape),
address_space=temporary_address_space,
dim_names=tuple(non1_storage_axis_names))
else:
temp_var = new_temporary_variables[temporary_name]
# {{{ check and adapt existing temporary
if temp_var.dtype is lp.auto:
pass
elif temp_var.dtype is not lp.auto and dtype is lp.auto:
dtype = temp_var.dtype
elif temp_var.dtype is not lp.auto and dtype is not lp.auto:
if temp_var.dtype != dtype:
raise LoopyError("Existing and new dtype of temporary '%s' "
"do not match (existing: %s, new: %s)" %
(temporary_name, temp_var.dtype, dtype))
temp_var = temp_var.copy(dtype=dtype)
if len(temp_var.shape) != len(new_temp_shape):
raise LoopyError(
"Existing and new temporary '%s' do not "
"have matching number of dimensions ('%d' vs. '%d') " %
(temporary_name, len(temp_var.shape), len(new_temp_shape)))
if temp_var.base_indices != (0, ) * len(new_temp_shape):
raise LoopyError(
"Existing and new temporary '%s' do not "
"have matching number of dimensions ('%d' vs. '%d') " %
(temporary_name, len(temp_var.shape), len(new_temp_shape)))
new_temp_shape = tuple(
max(i, ex_i) for i, ex_i in zip(new_temp_shape, temp_var.shape))
temp_var = temp_var.copy(shape=new_temp_shape)
if temporary_address_space == temp_var.address_space:
pass
elif temporary_address_space is lp.auto:
temporary_address_space = temp_var.address_space
elif temp_var.address_space is lp.auto:
pass
else:
raise LoopyError("Existing and new temporary '%s' do not "
"have matching scopes (existing: %s, new: %s)" %
(temporary_name,
AddressSpace.stringify(temp_var.address_space),
AddressSpace.stringify(temporary_address_space)))
temp_var = temp_var.copy(address_space=temporary_address_space)
# }}}
new_temporary_variables[temporary_name] = temp_var
kernel = kernel.copy(temporary_variables=new_temporary_variables)
# }}}
from loopy import tag_inames
kernel = tag_inames(kernel, new_iname_to_tag)
from loopy.kernel.data import AutoFitLocalIndexTag, filter_iname_tags_by_type
if filter_iname_tags_by_type(new_iname_to_tag.values(),
AutoFitLocalIndexTag):
from loopy.kernel.tools import assign_automatic_axes
kernel = assign_automatic_axes(kernel)
return kernel
def test_fuzz_expression_code_gen(ctx_factory, expr_type, random_seed):
from pymbolic import evaluate
def get_numpy_type(x):
if expr_type in ["real", "complex"]:
if isinstance(x, (complex, np.complexfloating)):
return np.complex128
else:
return np.float64
elif expr_type in ["int", "int_nonneg"]:
return np.int64
else:
raise ValueError("unknown expr_type: %s" % expr_type)
from random import seed
ctx = ctx_factory()
queue = cl.CommandQueue(ctx)
seed(random_seed)
data = []
instructions = []
ref_values = {}
if expr_type in ["real", "complex"]:
result_type = np.complex128
elif expr_type in ["int", "int_nonneg"]:
result_type = np.int64
else:
assert False
var_names = []
fuzz_iter = iter(generate_random_fuzz_examples(expr_type))
count = 0
while True:
if count == 10:
break
i, expr, var_values = next(fuzz_iter)
var_name = "expr%d" % i
print(expr)
#assert_parse_roundtrip(expr)
if expr_type in ["int", "int_nonneg"]:
result_type_iinfo = np.iinfo(np.int32)
bceval_mapper = BoundsCheckingEvaluationMapper(
var_values,
lbound=result_type_iinfo.min,
ubound=result_type_iinfo.max)
print(expr)
try:
ref_values[var_name] = bceval_mapper(expr)
except BoundsCheckError:
print(expr)
print("BOUNDS CHECK FAILED")
continue
else:
try:
ref_values[var_name] = evaluate(expr, var_values)
except ZeroDivisionError:
continue
count += 1
data.append(lp.GlobalArg(var_name, result_type, shape=()))
data.extend([
lp.TemporaryVariable(name, get_numpy_type(val))
for name, val in var_values.items()
])
instructions.extend([
lp.Assignment(name,
get_numpy_type(val)(val))
for name, val in var_values.items()
])
instructions.append(lp.Assignment(var_name, expr))
if expr_type == "int_nonneg":
var_names.extend(var_values)
knl = lp.make_kernel("{ : }", instructions, data, seq_dependencies=True)
import islpy as isl
knl = lp.assume(
knl,
isl.BasicSet(
"[%s] -> { : %s}" %
(", ".join(var_names), " and ".join("%s >= 0" % name
for name in var_names))))
knl = lp.set_options(knl, return_dict=True)
print(knl)
evt, lp_values = knl(queue, out_host=True)
for name, ref_value in ref_values.items():
lp_value = lp_values[name]
if expr_type in ["real", "complex"]:
err = abs(ref_value - lp_value) / abs(ref_value)
elif expr_type in ["int", "int_nonneg"]:
err = abs(ref_value - lp_value)
else:
assert False
if abs(err) > 1e-10:
print(80 * "-")
print(knl)
print(80 * "-")
print(lp.generate_code_v2(knl).device_code())
print(80 * "-")
print(f"WRONG: {name} rel error={err:g}")
print("reference=%r" % ref_value)
print("loopy=%r" % lp_value)
print(80 * "-")
1 / 0
print(lp.generate_code_v2(knl).device_code())
def get_kernel(self):
if self.integral_kernel.is_complex_valued:
potential_dtype = np.complex128
else:
potential_dtype = np.float64
lpknl = loopy.make_kernel( # NOQA
[
"{ [ tbox ] : 0 <= tbox < n_tgt_boxes }",
"{ [ tid, sbox ] : 0 <= tid < n_box_targets and \
sbox_begin <= sbox < sbox_end }",
"{ [ sid ] : 0 <= sid < n_box_sources }",
],
"""
for tbox
<> target_box_id = target_boxes[tbox]
<> box_target_beg = box_target_starts[target_box_id]
<> n_box_targets = box_target_counts_cumul[target_box_id]
<> sbox_begin = neighbor_source_boxes_starts[tbox]
<> sbox_end = neighbor_source_boxes_starts[tbox+1]
<> tbox_level = box_levels[target_box_id]
<> tbox_extent = root_extent * (1.0 / (2**tbox_level))
for tid
<> target_id = box_target_beg + tid
end
for tid, sbox
<> source_box_id = source_boxes[sbox]
<> n_box_sources = box_source_counts_cuml[source_box_id]
<> box_source_beg = box_source_starts[source_box_id]
<> sbox_level = box_levels[source_box_id]
<> sbox_extent = root_extent * (1.0 / (2**sbox_level))
table_lev_tmp = GET_TABLE_LEVEL {id=tab_lev_tmp}
table_lev = round(table_lev_tmp) {id=tab_lev,dep=tab_lev_tmp}
vec_id_tmp = COMPUTE_VEC_ID {id=vec_id_tmp}
vec_id = round(vec_id_tmp) {id=vec_id,dep=vec_id_tmp}
<> case_id = case_indices[vec_id] {dep=vec_id}
<> scaling = COMPUTE_SCALING
for sid
<> tgt_scaling = COMPUTE_TGT_SCALING
<> tgt_displacement = COMPUTE_TGT_DISPLACEMENT
tgt_table_lev_tmp = GET_TGT_TABLE_LEVEL {id=tgttab_lev_tmp}
tgt_table_lev = round(tgt_table_lev_tmp) \
{id=tgttab_lev,dep=tgttab_lev_tmp}
<> ext_nmlz = exterior_mode_nmlz[tgt_table_lev, tid] \
* tgt_scaling + tgt_displacement \
{id=extnmlz,dep=tgttab_lev}
<> source_id = box_source_beg + sid
<> pair_id = sid * n_box_targets + tid
<> entry_id = case_id * \
(n_box_targets * n_box_sources) \
+ pair_id
<> displacement = COMPUTE_DISPLACEMENT
<> integ = table_data[table_lev, entry_id] * scaling \
+ displacement {id=integ,dep=tab_lev}
# <> source_id_tree = user_source_ids[source_id]
<> coef = source_coefs[source_id] {id=coef}
# <> target_id_user = sorted_target_ids[target_id]
#db_table_lev[target_id] = table_lev_tmp {dep=tab_lev}
#db_case_id[target_id] = case_id
#db_vec_id[target_id] = vec_id
#db_n_box_targets[target_id] = n_box_targets
#db_n_box_sources[target_id] = n_box_sources
#db_entry_id[target_id] = entry_id
end
end
for tid
result[target_id] = sum((sbox, sid),
coef * integ) + EXTERIOR_PART {dep=integ:coef:extnmlz}
# Try inspecting case_id if something goes wrong
# (like segmentation fault) and look for -1's
# result[target_id] = min((sbox, sid), case_id)
# result[target_id] = vec_id_tmp
end
end
""".replace("COMPUTE_VEC_ID", self.codegen_vec_id()).replace(
"COMPUTE_SCALING", self.codegen_compute_scaling()).replace(
"COMPUTE_DISPLACEMENT",
self.codegen_compute_displacement()).replace(
"COMPUTE_TGT_SCALING",
self.codegen_compute_scaling('tbox')).replace(
"COMPUTE_TGT_DISPLACEMENT",
self.codegen_compute_displacement('tbox')).replace(
"GET_TABLE_LEVEL",
self.codegen_get_table_level()).replace(
"GET_TGT_TABLE_LEVEL",
self.codegen_get_table_level(
'tbox')).replace(
"EXTERIOR_PART",
self.codegen_exterior_part()),
[
loopy.TemporaryVariable("vec_id", np.int32),
loopy.TemporaryVariable("vec_id_tmp", np.float64),
loopy.TemporaryVariable("table_lev", np.int32),
loopy.TemporaryVariable("table_lev_tmp", np.float64),
loopy.TemporaryVariable("tgt_table_lev", np.int32),
loopy.TemporaryVariable("tgt_table_lev_tmp", np.float64),
loopy.ValueArg("encoding_base", np.int32),
loopy.GlobalArg("mode_nmlz", potential_dtype,
"n_tables, n_q_points"),
loopy.GlobalArg("exterior_mode_nmlz", potential_dtype,
"n_tables, n_q_points"),
loopy.GlobalArg("table_data", potential_dtype,
"n_tables, n_table_entries"),
loopy.GlobalArg("source_boxes", np.int32, "n_source_boxes"),
loopy.GlobalArg("box_centers", None, "dim, aligned_nboxes"),
loopy.ValueArg("aligned_nboxes", np.int32),
loopy.ValueArg("table_root_extent", np.float64),
loopy.ValueArg(
"dim, n_source_boxes, n_tables, "
"n_q_points, n_table_entries",
np.int32,
),
"...",
],
name="near_field",
lang_version=(2018, 2))
# lpknl = loopy.set_options(lpknl, write_code=True)
lpknl = loopy.set_options(lpknl, return_dict=True)
return lpknl
