experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA] Rename Executable::SizeInBytes to Executable::SizeOfGeneratedCodeInBytes

Browse Source 
On the GPU backend a lot of library code is used: cuDNN, cuBLAS, etc.
We do not know the size of the used kernels, so we can not report the total
amount of loaded kernels.

Renaming to SizeOfGeneratedCodeInBytes makes it clear that the returned number
does not include library code.

PiperOrigin-RevId: 261026023
This commit is contained in:
George Karpenkov 2019-07-31 16:52:35 -07:00 committed by TensorFlower Gardener
parent ff1087ac16
commit a2ade2ecb4
7 changed files with 13 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/compiler/xla/python/local_client.h
View File

@ -212,7 +212,9 @@ class PyLocalExecutable {
return executable_->build_options().num_replicas(); return executable_->build_options().num_replicas();
} }
int64 SizeInBytes() const { return executable_->executable()->SizeInBytes(); } int64 SizeOfGeneratedCodeInBytes() const {
return executable_->executable()->SizeOfGeneratedCodeInBytes();
}
// Returns the device ordinals to which each replica is assigned. // Returns the device ordinals to which each replica is assigned.
std::vector<int> DeviceOrdinals() const; std::vector<int> DeviceOrdinals() const;

3
tensorflow/compiler/xla/python/xla.cc
View File

@ -407,7 +407,8 @@ PYBIND11_MODULE(xla_extension, m) {
.def_static("Compile", &PyLocalExecutable::Compile, .def_static("Compile", &PyLocalExecutable::Compile,
py::call_guard<py::gil_scoped_release>()) py::call_guard<py::gil_scoped_release>())
.def("DeviceOrdinals", &PyLocalExecutable::DeviceOrdinals) .def("DeviceOrdinals", &PyLocalExecutable::DeviceOrdinals)
.def("SizeInBytes", &PyLocalExecutable::SizeInBytes) .def("SizeOfGeneratedCodeInBytes",
&PyLocalExecutable::SizeOfGeneratedCodeInBytes)
.def("Delete", &PyLocalExecutable::Delete) .def("Delete", &PyLocalExecutable::Delete)
.def("Execute", &PyLocalExecutable::Execute, .def("Execute", &PyLocalExecutable::Execute,
py::call_guard<py::gil_scoped_release>(), py::arg("arguments")) py::call_guard<py::gil_scoped_release>(), py::arg("arguments"))

2
tensorflow/compiler/xla/python/xla_client.py
View File

@ -545,7 +545,7 @@ class Computation(object):
# def Execute(self, arguments : [Buffer]) -> Buffer: # def Execute(self, arguments : [Buffer]) -> Buffer:
# """Execute on one replica with Buffer arguments and return value.""" # """Execute on one replica with Buffer arguments and return value."""
# #
# def SizeInBytes(self) -> int: # def SizeOfGeneratedCodeInBytes(self) -> int:
# """Return generated binary size, or -1 if not known.""" # """Return generated binary size, or -1 if not known."""
# #
# def ExecutePerReplica(self, arguments: [[Buffer]]) -> [Buffer]: # def ExecutePerReplica(self, arguments: [[Buffer]]) -> [Buffer]:

4
tensorflow/compiler/xla/service/executable.cc
View File

@ -128,7 +128,7 @@ StatusOr<ScopedShapedBuffer> Executable::ExecuteOnStreamWrapper(
profile->set_compute_time_ns(profile->compute_and_transfer_time_ns()); profile->set_compute_time_ns(profile->compute_and_transfer_time_ns());
} }
const int64 executable_size_in_bytes = SizeInBytes(); const int64 executable_size_in_bytes = SizeOfGeneratedCodeInBytes();
if (executable_size_in_bytes != 0) { if (executable_size_in_bytes != 0) {
profile->set_executable_size_in_bytes(executable_size_in_bytes); profile->set_executable_size_in_bytes(executable_size_in_bytes);
} }
@ -143,6 +143,6 @@ StatusOr<ScopedShapedBuffer> Executable::ExecuteOnStreamWrapper(
return return_value; return return_value;
} }
int64 Executable::SizeInBytes() { return -1; } int64 Executable::SizeOfGeneratedCodeInBytes() { return -1; }
} // namespace xla } // namespace xla

6
tensorflow/compiler/xla/service/executable.h
View File

@ -219,11 +219,11 @@ class Executable {
return hlo_module_->config().entry_computation_layout().result_shape(); return hlo_module_->config().entry_computation_layout().result_shape();
} }
// Returns the size of the executable in bytes. Returns -1 by default if the // Returns the size of the executable in bytes. Returns -1 if this query is
// method is not overridden to support this kind of query. // not supported by the executable.
// //
// Does not include the size of used libraries (e.g. cuDNN, Eigen, etc.). // Does not include the size of used libraries (e.g. cuDNN, Eigen, etc.).
virtual int64 SizeInBytes(); virtual int64 SizeOfGeneratedCodeInBytes();
// Dumping helpers. // Dumping helpers.
void set_hlo_proto(std::unique_ptr<xla::HloProto> hlo_proto) { void set_hlo_proto(std::unique_ptr<xla::HloProto> hlo_proto) {

2
tensorflow/compiler/xla/service/gpu/gpu_executable.cc
View File

@ -428,7 +428,7 @@ const InstructionValueSet& GpuExecutable::GetRootValueSet() const {
module().entry_computation()->root_instruction()); module().entry_computation()->root_instruction());
} }
int64 GpuExecutable::SizeInBytes() { int64 GpuExecutable::SizeOfGeneratedCodeInBytes() {
// Non-empty PTX but empty cubin: compilation must have failed, return // Non-empty PTX but empty cubin: compilation must have failed, return
// "unknown". // "unknown".
if (binary().empty() && !text_.empty()) { if (binary().empty() && !text_.empty()) {

2
tensorflow/compiler/xla/service/gpu/gpu_executable.h
View File

@ -61,7 +61,7 @@ class GpuExecutable : public Executable {
std::unique_ptr<HloProfileIndexMap> hlo_profile_index_map); std::unique_ptr<HloProfileIndexMap> hlo_profile_index_map);
~GpuExecutable() override; ~GpuExecutable() override;
int64 SizeInBytes() override; int64 SizeOfGeneratedCodeInBytes() override;
// This should be called after set_ir_module_string. // This should be called after set_ir_module_string.
const string& ir_module_string() const { return ir_module_string_; } const string& ir_module_string() const { return ir_module_string_; }