Skip to content

optimize phi::CPUPlace in paddle/fluid/pybind/ #77207

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
luotao1 merged 1 commit into from
Jan 6, 2026
Merged

optimize phi::CPUPlace in paddle/fluid/pybind/ #77207

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 4 additions & 6 deletions paddle/fluid/pybind/args_mapper.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -133,15 +133,13 @@ void ArgMaxMinMapper(PyObject* args,
if (axis_obj == Py_None || axis_obj == nullptr) {
*flatten = true;
*axis = paddle::dialect::full(
std::vector<int64_t>{1}, 0, phi::DataType::INT64, phi::CPUPlace());
std::vector<int64_t>{1}, 0, phi::DataType::INT64, CPUPlace());
} else if (PyObject_CheckIRValue(axis_obj)) {
*axis = CastPyArg2Value(axis_obj, "argmax", 1);
} else {
int64_t axis_tmp = CastPyArg2Long(axis_obj, "argmax", 1);
*axis = paddle::dialect::full(std::vector<int64_t>{1},
axis_tmp,
phi::DataType::INT64,
phi::CPUPlace());
*axis = paddle::dialect::full(
std::vector<int64_t>{1}, axis_tmp, phi::DataType::INT64, CPUPlace());
}
*keepdims = CastPyArg2Boolean(keepdims_obj, "argmax", 2, false);

Expand Down Expand Up @@ -251,7 +249,7 @@ void ArgSumMapper(PyObject* args,
} else {
std::vector<int64_t> axis_tmp = CastPyArg2Longs(axis_obj, "sum", 1, {});
*axis = paddle::dialect::full_int_array(
axis_tmp, phi::DataType::INT64, phi::CPUPlace());
axis_tmp, phi::DataType::INT64, CPUPlace());
}

PyObject* py_obj_1 = GetItemFromArgsOrKWArgs(
Expand Down
20 changes: 9 additions & 11 deletions paddle/fluid/pybind/cuda_streams_py.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -67,7 +67,7 @@ PY_STREAM_TYPE get_legacy_default_stream(int device_id) {
if (device_id == -1) {
device_id = phi::backends::gpu::GetCurrentDeviceId();
}
phi::GPUPlace place(device_id);
GPUPlace place(device_id);

legacy_default_streams.try_emplace(
device_id, place, static_cast<gpuStream_t>(0));
Expand Down Expand Up @@ -128,8 +128,7 @@ void BindCudaStream(py::module *m_ptr) {
#endif
gpuStream_t raw = reinterpret_cast<gpuStream_t>(data_ptr);

return std::make_unique<phi::CUDAStream>(phi::GPUPlace(device_id),
raw);
return std::make_unique<phi::CUDAStream>(GPUPlace(device_id), raw);
#else
PADDLE_THROW(common::errors::Unavailable(
"Paddle is not compiled with CUDA/HIP, "
Expand Down Expand Up @@ -357,12 +356,11 @@ void BindCudaStream(py::module *m_ptr) {

)DOC")
.def_property_readonly(
"place",
[](phi::CUDAStream &self) { return phi::GPUPlace(self.place()); })
"place", [](phi::CUDAStream &self) { return GPUPlace(self.place()); })
#endif
.def(
"__init__",
[](phi::CUDAStream &self, phi::GPUPlace *place, int priority) {
[](phi::CUDAStream &self, GPUPlace *place, int priority) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if (priority != 1 && priority != 2) {
PADDLE_THROW(common::errors::InvalidArgument(
Expand All @@ -372,7 +370,7 @@ void BindCudaStream(py::module *m_ptr) {
auto stream_flag = phi::CUDAStream::StreamFlag::kStreamNonBlocking;
if (place == nullptr) {
int curr_device_id = platform::GetCurrentDeviceId();
auto place_tmp = phi::GPUPlace(curr_device_id);
auto place_tmp = GPUPlace(curr_device_id);
new (&self) phi::CUDAStream(place_tmp, priority - 2, stream_flag);
} else {
// setting priority 1(high) and 2(normal) correspond to the actual
Expand Down Expand Up @@ -409,8 +407,8 @@ void BindCudaStream(py::module *m_ptr) {
auto stream_flag = phi::CUDAStream::StreamFlag::kStreamNonBlocking;
// setting priority 1(high) and 2(normal) correspond to the actual
// cuda stream priority -1 and 0.
new (&self) phi::CUDAStream(
phi::GPUPlace(device), priority - 2, stream_flag);
new (&self)
phi::CUDAStream(GPUPlace(device), priority - 2, stream_flag);
#else
PADDLE_THROW(common::errors::Unavailable(
"Class CUDAStream can only be initialized on the GPU platform."));
Expand All @@ -422,8 +420,8 @@ void BindCudaStream(py::module *m_ptr) {
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
int device_id = platform::GetCurrentDeviceId();
auto stream_flag = phi::CUDAStream::StreamFlag::kStreamNonBlocking;
new (&self) phi::CUDAStream(
phi::GPUPlace(device_id), /*priority=*/0, stream_flag);
new (&self)
phi::CUDAStream(GPUPlace(device_id), /*priority=*/0, stream_flag);
#else
PADDLE_THROW(common::errors::Unavailable(
"Class CUDAStream can only be initialized on the GPU platform."));
Expand Down
4 changes: 2 additions & 2 deletions paddle/fluid/pybind/cudart_py.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -169,7 +169,7 @@ void BindCudaRt(py::module* m) {
#endif

cudart.def("cudaMemGetInfo", [](int device) -> std::pair<size_t, size_t> {
const auto& place = phi::GPUPlace(device);
const auto& place = GPUPlace(device);
platform::CUDADeviceGuard cuda_guard(place);
size_t device_free = 0;
size_t device_total = 0;
Expand Down Expand Up @@ -399,7 +399,7 @@ void BindCudaRt(py::module* m) {
"cuda"
"MemGetInfo",
[](int device) -> std::pair<size_t, size_t> {
const auto& place = phi::GPUPlace(device);
const auto& place = GPUPlace(device);
platform::CUDADeviceGuard cuda_guard(place);
size_t device_free = 0;
size_t device_total = 0;
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/pybind/data_set_py.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,7 +102,7 @@ class IterableDatasetWrapper {
"Device number does not match reader number"));
for (size_t i = 0; i < places_.size(); ++i) {
data_feeds_[i]->AssignFeedVar(*scopes_[i]);
data_feeds_[i]->SetPlace(phi::CPUPlace());
data_feeds_[i]->SetPlace(CPUPlace());
PADDLE_ENFORCE_EQ(data_feeds_[i]->Start(),
true,
common::errors::Unavailable(
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/pybind/deep_ep_api.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,7 @@ void BindDeepEPApi(pybind11::module *m) {
cudaStream_t comm_stream = self.get_comm_stream();
auto s = phi::Stream(reinterpret_cast<phi::StreamId>(comm_stream));
#if defined(PADDLE_WITH_CUDA)
return phi::CUDAStream(phi::GPUPlace(device_id), s);
return phi::CUDAStream(GPUPlace(device_id), s);
#elif defined(PADDLE_WITH_XPU)
return phi::XPUCUDAStream(phi::XPUPlace(device_id), s);
#endif
Expand Down
6 changes: 2 additions & 4 deletions paddle/fluid/pybind/distributed_py.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,8 +90,7 @@ using GlooOptions = paddle::distributed::ProcessGroupGloo::GlooOptions;
using ProcessGroupFlagcx = paddle::distributed::ProcessGroupFlagcx;
#endif

static UNUSED void *use_ccl_comm_func =
phi::detail::GetCCLComm(phi::CPUPlace());
static UNUSED void *use_ccl_comm_func = phi::detail::GetCCLComm(CPUPlace());

void BindDistributed(py::module *m) {
py::enum_<distributed::ReduceOp>(*m, "ReduceOp")
Expand Down Expand Up @@ -576,8 +575,7 @@ void BindDistributed(py::module *m) {
auto *dev_ctx =
self.GetDeviceContext(in_tensor.place(), use_calc_stream);
SplitTensor(*dev_ctx, *out_dense, &out_tensor_list);
if (!use_calc_stream &&
dev_ctx->GetPlace() != phi::CPUPlace()) {
if (!use_calc_stream && dev_ctx->GetPlace() != CPUPlace()) {
// calculate stream will wait comm stream
task->UpdateWaitChain(*dev_ctx);
}
Expand Down
8 changes: 4 additions & 4 deletions paddle/fluid/pybind/eager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -137,7 +137,7 @@ void EmptyStringTensorInitializer(TensorObject* self,
auto ddims = common::make_ddim(dims);
self->tensor.set_name(name);
// Note(zhoushunjie): Only support CPUPlace when create StringTensor
auto actual_place = phi::CPUPlace();
auto actual_place = CPUPlace();
// Allocate memory
paddle::experimental::DefaultAllocator string_allocator(actual_place);
std::shared_ptr<phi::StringTensor> string_tensor =
Expand Down Expand Up @@ -165,7 +165,7 @@ void InitTensorWithNumpyValue(TensorObject* self,
phi::DenseTensor* impl_ptr =
static_cast<phi::DenseTensor*>(self->tensor.impl().get());
if (phi::is_cpu_place(place)) {
SetTensorFromPyArray<phi::CPUPlace>(impl_ptr, array, place, zero_copy);
SetTensorFromPyArray<CPUPlace>(impl_ptr, array, place, zero_copy);
} else if (phi::is_xpu_place(place)) {
#if defined(PADDLE_WITH_XPU)
phi::backends::xpu::SetXPUDeviceId(place.device);
Expand All @@ -186,7 +186,7 @@ void InitTensorWithNumpyValue(TensorObject* self,
PADDLE_THROW(common::errors::PreconditionNotMet(
"PaddlePaddle should compile with GPU if use CUDAPlace."));
#endif
SetTensorFromPyArray<phi::GPUPlace>(impl_ptr, array, place, zero_copy);
SetTensorFromPyArray<GPUPlace>(impl_ptr, array, place, zero_copy);
} else if (phi::is_cuda_pinned_place(place)) {
SetTensorFromPyArray<phi::GPUPinnedPlace>(
impl_ptr, array, place, zero_copy);
Expand Down Expand Up @@ -227,7 +227,7 @@ void InitStringTensorWithNumpyValue(TensorObject* self, const py::object& obj) {
phi::Place place = impl_ptr->place();
auto array = obj.cast<py::array>();
if (phi::is_cpu_place(place)) {
SetStringTensorFromPyArray<phi::CPUPlace>(impl_ptr, array, place);
SetStringTensorFromPyArray<CPUPlace>(impl_ptr, array, place);
} else {
PADDLE_THROW(common::errors::InvalidArgument(
"StringTensor only support CPUPlace now, but receive %s",
Expand Down
2 changes: 1 addition & 1 deletion paddle/fluid/pybind/eager_functions.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,7 +102,7 @@ class EagerNumpyAllocation : public phi::Allocation {
: Allocation(
static_cast<void*>(pybind11::detail::array_proxy(numpy_data)->data),
phi::SizeOf(dtype) * PyArray_Size_(numpy_data),
phi::CPUPlace()),
CPUPlace()),
arr_(numpy_data) {
PADDLE_ENFORCE_NOT_NULL(
arr_,
Expand Down
4 changes: 2 additions & 2 deletions paddle/fluid/pybind/eager_math_op_patch.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,14 +88,14 @@ void InitTensorWithNumpyValue(const py::object& array,
phi::DenseTensor* impl_ptr =
static_cast<phi::DenseTensor*>(self->impl().get());
if (phi::is_cpu_place(place)) {
SetTensorFromPyArray<phi::CPUPlace>(impl_ptr, array, place, zero_copy);
SetTensorFromPyArray<CPUPlace>(impl_ptr, array, place, zero_copy);
} else if (phi::is_xpu_place(place)) {
SetTensorFromPyArray<phi::XPUPlace>(impl_ptr, array, place, zero_copy);
} else if (phi::is_xpu_pinned_place(place)) {
SetTensorFromPyArray<phi::XPUPinnedPlace>(
impl_ptr, array, place, zero_copy);
} else if (phi::is_gpu_place(place)) {
SetTensorFromPyArray<phi::GPUPlace>(impl_ptr, array, place, zero_copy);
SetTensorFromPyArray<GPUPlace>(impl_ptr, array, place, zero_copy);
} else if (phi::is_cuda_pinned_place(place)) {
SetTensorFromPyArray<phi::GPUPinnedPlace>(
impl_ptr, array, place, zero_copy);
Expand Down
13 changes: 5 additions & 8 deletions paddle/fluid/pybind/eager_method.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,12 +250,12 @@ static PyObject* tensor_method_numpy(TensorObject* self,
}

phi::DenseTensor cpu_tensor;
phi::CPUPlace cpu_place;
CPUPlace cpu_place;

if (self->tensor.is_cpu() || self->tensor.is_gpu_pinned() ||
self->tensor.is_xpu_pinned()) {
eager_gil_scoped_release guard;
phi::CPUPlace place;
CPUPlace place;
if (self->tensor.is_selected_rows()) {
VLOG(6) << "Getting SelectedRows's numpy value";
auto* selected_rows =
Expand Down Expand Up @@ -382,7 +382,7 @@ static PyObject* tensor_method_numpy(TensorObject* self,
#endif
#if defined(PADDLE_WITH_XPU)
} else if (self->tensor.is_xpu()) {
phi::CPUPlace place;
CPUPlace place;
if (self->tensor.is_selected_rows()) {
VLOG(6) << "Getting SelectedRows's numpy value";
auto* selected_rows =
Expand Down Expand Up @@ -3165,11 +3165,8 @@ static PyObject* tensor_method__share_memory(TensorObject* self,
const std::string& ipc_name = shared_writer_holder->ipc_name();
memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);
// 4. copy data & reset holder
memory::Copy(phi::CPUPlace(),
shared_writer_holder->ptr(),
phi::CPUPlace(),
data_ptr,
data_size);
memory::Copy(
CPUPlace(), shared_writer_holder->ptr(), CPUPlace(), data_ptr, data_size);
t->ResetHolder(shared_writer_holder);
return ToPyObject(t);
#else
Expand Down
6 changes: 3 additions & 3 deletions paddle/fluid/pybind/eager_utils.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -735,9 +735,9 @@ phi::Place CastPyArg2Place(PyObject* obj, ssize_t arg_pos) {
if (PyObject_TypeCheck(obj, g_place_pytype)) { // NOLINT
place = ::pybind11::handle(obj).cast<phi::Place>();
} else if (PyObject_TypeCheck(obj, g_cudaplace_pytype)) {
place = ::pybind11::handle(obj).cast<phi::GPUPlace>();
place = ::pybind11::handle(obj).cast<GPUPlace>();
} else if (PyObject_TypeCheck(obj, g_cpuplace_pytype)) {
place = ::pybind11::handle(obj).cast<phi::CPUPlace>();
place = ::pybind11::handle(obj).cast<CPUPlace>();
} else if (PyObject_TypeCheck(obj, g_xpuplace_pytype)) {
place = ::pybind11::handle(obj).cast<phi::XPUPlace>();
} else if (PyObject_TypeCheck(obj, g_cudapinnedplace_pytype)) {
Expand Down Expand Up @@ -2464,7 +2464,7 @@ std::vector<pir::Value> CastPyArg2VectorOfValueOrLong(
} else if (PyObject_CheckLong(item)) {
int64_t k_tmp = CastPyArg2Long(item, op_type, arg_pos);
value_list.emplace_back(
paddle::dialect::full(shape, k_tmp, dtype, phi::CPUPlace()));
paddle::dialect::full(shape, k_tmp, dtype, CPUPlace()));
} else if (item == Py_None) {
continue; // skip
} else {
Expand Down
42 changes: 20 additions & 22 deletions paddle/fluid/pybind/imperative.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -120,10 +120,10 @@ class PyVariableWrapperHook : public imperative::VariableWrapperHook {
};

static const phi::Place PyObjectToPlace(const py::object &place_obj) {
if (py::isinstance<phi::CPUPlace>(place_obj)) {
return place_obj.cast<phi::CPUPlace>();
} else if (py::isinstance<phi::GPUPlace>(place_obj)) {
return place_obj.cast<phi::GPUPlace>();
if (py::isinstance<CPUPlace>(place_obj)) {
return place_obj.cast<CPUPlace>();
} else if (py::isinstance<GPUPlace>(place_obj)) {
return place_obj.cast<GPUPlace>();
} else if (py::isinstance<phi::XPUPlace>(place_obj)) {
return place_obj.cast<phi::XPUPlace>();
} else if (py::isinstance<phi::GPUPinnedPlace>(place_obj)) {
Expand Down Expand Up @@ -177,11 +177,11 @@ static void InitVarBaseAndTensor(imperative::VarBase *self,
auto *tensor = self->MutableVar()->GetMutable<phi::DenseTensor>();
VLOG(4) << "zero_copy: " << zero_copy;
if (phi::is_cpu_place(place)) {
SetTensorFromPyArray<phi::CPUPlace>(tensor, array, place, zero_copy);
SetTensorFromPyArray<CPUPlace>(tensor, array, place, zero_copy);
} else if (phi::is_xpu_place(place)) {
SetTensorFromPyArray<phi::XPUPlace>(tensor, array, place, zero_copy);
} else if (phi::is_gpu_place(place)) {
SetTensorFromPyArray<phi::GPUPlace>(tensor, array, place, zero_copy);
SetTensorFromPyArray<GPUPlace>(tensor, array, place, zero_copy);
} else if (phi::is_cuda_pinned_place(place)) {
SetTensorFromPyArray<phi::GPUPinnedPlace>(tensor, array, place, zero_copy);
} else if (phi::is_xpu_pinned_place(place)) {
Expand Down Expand Up @@ -528,7 +528,7 @@ void BindImperative(py::module *m_ptr) {
"_generator' to locate the data causes this issue."));
// 2. construct DenseTensor
phi::DenseTensor t;
SetTensorFromPyArray<phi::CPUPlace>(&t, array, phi::CPUPlace(), true);
SetTensorFromPyArray<CPUPlace>(&t, array, CPUPlace(), true);
// 3. allocate shared memory
void *data_ptr = t.data();
size_t data_size = t.numel() * phi::SizeOf(t.dtype());
Expand All @@ -538,9 +538,9 @@ void BindImperative(py::module *m_ptr) {
const std::string &ipc_name = shared_writer_holder->ipc_name();
memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);
// 5. copy data & reset holder
memory::Copy(phi::CPUPlace(),
memory::Copy(CPUPlace(),
shared_writer_holder->ptr(),
phi::CPUPlace(),
CPUPlace(),
data_ptr,
data_size);
t.ResetHolder(shared_writer_holder);
Expand All @@ -567,7 +567,7 @@ void BindImperative(py::module *m_ptr) {
"_generator' to locate the data causes this issue."));
// 2. construct DenseTensor
phi::DenseTensor t;
SetTensorFromPyArray<phi::CPUPlace>(&t, array, phi::CPUPlace(), true);
SetTensorFromPyArray<CPUPlace>(&t, array, CPUPlace(), true);
// 3. allocate shared memory
void *data_ptr = t.data();
size_t data_size = t.numel() * phi::SizeOf(t.dtype());
Expand All @@ -577,9 +577,9 @@ void BindImperative(py::module *m_ptr) {
const std::string &ipc_name = shared_writer_holder->ipc_name();
memory::allocation::MemoryMapFdSet::Instance().Insert(ipc_name);
// 5. copy data & reset holder
memory::Copy(phi::CPUPlace(),
memory::Copy(CPUPlace(),
shared_writer_holder->ptr(),
phi::CPUPlace(),
CPUPlace(),
data_ptr,
data_size);
t.ResetHolder(shared_writer_holder);
Expand Down Expand Up @@ -692,8 +692,8 @@ void BindImperative(py::module *m_ptr) {
return py::cast(self.ExpectedPlace());
},
[](imperative::Tracer &self, const py::object &obj) {
if (py::isinstance<phi::GPUPlace>(obj)) {
auto p = obj.cast<phi::GPUPlace *>();
if (py::isinstance<GPUPlace>(obj)) {
auto p = obj.cast<GPUPlace *>();
self.SetExpectedPlace(*p);
// TODO(jiabin): Support eager here when we need to make all
// dygraph in eager mode
Expand All @@ -704,8 +704,8 @@ void BindImperative(py::module *m_ptr) {
self.SetExpectedPlace(*p);
VLOG(4) << "Tracer(" << &self << ")"
<< " set expected place " << *p;
} else if (py::isinstance<phi::CPUPlace>(obj)) {
auto p = obj.cast<phi::CPUPlace *>();
} else if (py::isinstance<CPUPlace>(obj)) {
auto p = obj.cast<CPUPlace *>();
self.SetExpectedPlace(*p);
VLOG(4) << "Tracer(" << &self << ")"
<< " set expected place " << *p;
Expand Down Expand Up @@ -844,8 +844,8 @@ void BindImperative(py::module *m_ptr) {
});

m.def("varbase_copy", &VarBaseCopy<phi::Place>);
m.def("varbase_copy", &VarBaseCopy<phi::CPUPlace>);
m.def("varbase_copy", &VarBaseCopy<phi::GPUPlace>);
m.def("varbase_copy", &VarBaseCopy<CPUPlace>);
m.def("varbase_copy", &VarBaseCopy<GPUPlace>);
m.def("varbase_copy", &VarBaseCopy<phi::XPUPlace>);
m.def("varbase_copy", &VarBaseCopy<phi::GPUPinnedPlace>);
m.def("varbase_copy", &VarBaseCopy<phi::XPUPinnedPlace>);
Expand Down Expand Up @@ -925,8 +925,7 @@ void BindImperative(py::module *m_ptr) {
imperative::ParallelContext,
std::shared_ptr<imperative::NCCLParallelContext>>(
m, "NCCLParallelContext")
.def(py::init<const imperative::ParallelStrategy &,
const phi::GPUPlace &>())
.def(py::init<const imperative::ParallelStrategy &, const GPUPlace &>())
.def("init", [](imperative::NCCLParallelContext &self) { self.Init(); })
.def("init_with_ring_id",
&imperative::NCCLParallelContext::InitWithRingID,
Expand Down Expand Up @@ -965,8 +964,7 @@ void BindImperative(py::module *m_ptr) {
imperative::ParallelContext,
std::shared_ptr<imperative::GLOOParallelContext>>(
m, "GLOOParallelContext")
.def(py::init<const imperative::ParallelStrategy &,
const phi::CPUPlace &>())
.def(py::init<const imperative::ParallelStrategy &, const CPUPlace &>())
.def("init", [](imperative::GLOOParallelContext &self) { self.Init(); })
.def("init_with_ring_id",
&imperative::GLOOParallelContext::InitWithRingID,
Expand Down
Loading
Loading