pytorch/torch/csrc/jit/init.cpp
2017-09-05 17:48:55 -04:00

105 lines
3.5 KiB
C++

#include <Python.h>
#include "THP.h"
#include "torch/csrc/jit/ir.h"
#include "torch/csrc/jit/graph_fuser.h"
#include "torch/csrc/jit/init_pass.h"
#include "torch/csrc/jit/dead_code_elimination.h"
#include "torch/csrc/jit/python_tracer.h"
#include "torch/csrc/utils/python_strings.h"
#include "torch/csrc/DynamicTypes.h"
#ifdef WITH_TOFFEE
#include "torch/csrc/toffee/export.h"
#endif
PyObject * THPJIT_initExtension(PyObject *_unused)
{
// Leaving this code here, because it will likely be useful at some point
//PyObject *jit_module = PyImport_ImportModule("torch.jit");
//THPUtils_assert(jit_module, "class loader couldn't access "
//"torch.jit module");
//PyObject *jit_dict = PyModule_GetDict(jit_module);
Py_RETURN_TRUE;
}
// stub to run all C++ only tests for the JIT
// the stuff in test_jit.cpp is kept separate from the rest of PyTorch
// so we can build and iterate on it faster.
// from test_jit.cpp
namespace torch { namespace jit { extern void runJITCPPTests(); } };
namespace {
using namespace torch::jit;
using pass_type = void (std::shared_ptr<Graph>&);
template<pass_type pass>
PyObject * wrap_pass(PyObject *_unused, PyObject *py_state) {
HANDLE_TH_ERRORS
THPUtils_assert(THPTracingState_Check(py_state), "expected a TracingState instance");
THPTracingState *state = (THPTracingState*)py_state;
pass(state->cdata->graph);
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
#ifdef WITH_TOFFEE
PyObject * export_graph(PyObject *_unused, PyObject *args) {
HANDLE_TH_ERRORS
PyObject* py_state;
PyObject* initializers = nullptr;
if (!PyArg_ParseTuple(args, "O|O", &py_state, &initializers)) {
return NULL;
}
std::vector<at::Tensor> initializers_;
if(initializers) {
if (!PySequence_Check(initializers))
throw std::runtime_error("expected initializers to be a sequence");
auto N = PySequence_Length(initializers);
for(Py_ssize_t i = 0; i < N; ++i) {
auto tensor = PySequence_GetItem(initializers,i);
if(!THPModule_isTensor(tensor))
throw std::runtime_error("expected a tensor value in initializer dict");
initializers_.push_back(torch::createTensor(tensor));
}
}
THPUtils_assert(THPTracingState_Check(py_state), "expected a TracingState instance");
THPTracingState *state = (THPTracingState*)py_state;
return THPUtils_packString(ExportGraph(state->cdata->graph, state->cdata->buffer_map, initializers_));
END_HANDLE_TH_ERRORS
}
#endif // WITH_TOFFEE
PyObject * run_cpp_tests(PyObject *_unused, PyObject *_unused2) {
HANDLE_TH_ERRORS
runJITCPPTests();
Py_RETURN_NONE;
END_HANDLE_TH_ERRORS
}
struct PyMethodDef _THPJIT_methods[] = {
{"_jit_init", (PyCFunction)THPJIT_initExtension, METH_NOARGS, NULL},
{"_tracer_enter", (PyCFunction)THPTracer_enter, METH_VARARGS, NULL},
{"_tracer_exit", (PyCFunction)THPTracer_exit, METH_VARARGS, NULL},
{"_jit_pass_init", (PyCFunction)wrap_pass<MatchJITOps>, METH_O, "init"},
{"_jit_pass_fuse", (PyCFunction)wrap_pass<FuseGraph>, METH_O, "fuse"},
{"_jit_pass_dco", (PyCFunction)wrap_pass<EliminateDeadCode>, METH_O, "dco"},
{"_jit_pass_lint", (PyCFunction)wrap_pass<LintGraph>, METH_O, "lint"},
#ifdef WITH_TOFFEE
{"_jit_pass_export", (PyCFunction)export_graph, METH_VARARGS, "export"},
#endif // WITH_TOFFEE
{"_jit_run_cpp_tests",(PyCFunction)run_cpp_tests, METH_NOARGS, NULL},
{NULL}
};
} // anonymous namespace
PyMethodDef* THPJIT_methods() {
return _THPJIT_methods;
}