Passing error to TF instead of exit

[njzjz]

This commit does three little things:
(1) create an exception called deepmd::deepmd_exception (based on std::runtime_error);
(2) throw this exception instead of exit or std::runtime_error;
(3) catch this exception in the op, and pass to TF using OP_REQUIRES_OK.
One more, the OOM error will raise ResourceExhausted, as the same as TF ops.

The benifit of doing so is that the TF side and Python side can processing other things, catch the error, and print the traceback.
This commit can also fix #802, where the Python didn't save the buffer to the file before exit.

The idea was inspired by TF: https://github.com/tensorflow/tensorflow/blob/5dcfc51118817f27fad5246812d83e5dccdc5f72/tensorflow/core/kernels/mkl/mkl_tfconv_op.h#L120-L126

[codecov-commenter]

Codecov Report

Merging #918 (01cad22) into devel (ee0ed99) will increase coverage by 7.87%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##            devel     #918      +/-   ##
==========================================
+ Coverage   75.42%   83.29%   +7.87%     
==========================================
  Files          85      118      +33     
  Lines        6730     9916    +3186     
==========================================
+ Hits         5076     8260    +3184     
- Misses       1654     1656       +2     

Impacted Files	Coverage Δ
source/lib/tests/test_pair_tab.cc	100.00% <0.00%> (ø)
source/lib/tests/test_prod_virial_grad_r.cc	100.00% <0.00%> (ø)
source/lib/tests/test_tabulate.cc	100.00% <0.00%> (ø)
source/lib/tests/test_env_mat_r.cc	100.00% <0.00%> (ø)
source/api_cc/tests/test_utils.h	97.22% <0.00%> (ø)
source/api_cc/tests/test_deeppot_model_devi.cc	100.00% <0.00%> (ø)
source/lib/tests/test_soft_min_switch_force.cc	100.00% <0.00%> (ø)
source/lib/tests/test_env_mat_a.cc	100.00% <0.00%> (ø)
source/api_cc/tests/test_ewald.cc	100.00% <0.00%> (ø)
source/lib/tests/test_soft_min_switch_virial.cc	100.00% <0.00%> (ø)
... and 23 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ee0ed99...01cad22. Read the comment docs.

[amcadmus]

We should avoid copying the try ... catch ... in all OP implementations.

Shall we define some function like

#define void ComputeFunction(OpKernelContext* context);

void save_compute(OpKernelContext* context, const ComputeFunction &ff) {
  try{
    ff(context);
  } catch (...) {
    ....
  } catch (...) {
    ....
  }
}

where ComputeFunction is the implementation of the OP.

[njzjz]

See

deepmd-kit/source/op/custom_op.cc

Lines 4 to 20 in 9c8a0da

	namespace deepmd {
	void save_compute(OpKernelContext* context, std::function<void(OpKernelContext*)> ff) {
	try{
	ff(context);
	} catch (deepmd::deepmd_exception_oom& e){
	OP_REQUIRES_OK(
	context,
	errors::ResourceExhausted("Operation received an exception: ", e.what(),
	", in file ",__FILE__, ":", __LINE__));
	} catch (deepmd::deepmd_exception& e) {
	OP_REQUIRES_OK(
	context,
	errors::Internal("Operation received an exception: ", e.what(),
	", in file ",__FILE__, ":", __LINE__));
	}
	}
	};

and

deepmd-kit/source/op/descrpt.cc

Lines 52 to 56 in 9c8a0da

	void Compute(OpKernelContext* context) override {
	deepmd::save_compute(context, [this](OpKernelContext* context) {this->_Compute(context);});
	}
	void _Compute(OpKernelContext* context) {

[amcadmus]

deepmd-kit/source/op/custom_op.cc

Line 5 in 9c8a0da

void save_compute(OpKernelContext* context, std::function<void(OpKernelContext*)> ff) {

The name should have been safe_compute, sry my mistake

[galeselee]

There is no problem on ROCm platform about UT and compiling