void enableProfiler(ProfilerState new_state) {
TORCH_ASSERT(new_state != ProfilerState::Disabled);
#ifndef WITH_CUDA
if (new_state == ProfilerState::NVTX)
throw std::runtime_error("Can't use NVTX profiler - PyTorch was compiled without CUDA");
#endif
if (state != ProfilerState::Disabled && new_state != state) {
throw std::runtime_error("can't change kind of profiling (e.g. NVTX to CPU) while profiler is running");
}
state = new_state;
#ifdef WITH_CUDA
if(state == ProfilerState::CUDA) {
// event recording appears to have some startup overhead, so we need to
// to generate some dummy events first before recording syncrhonization events
for(int i = 0; i < 5; i++) {
onEachDevice([](int d) {
mark("__cuda_startup");
cudaDeviceSynchronize();
});
}
// cuda events must be on the same device, so we need a start event recorded
// for each gpu. we then use this event to synchronize time on the GPU
// with the CPU clock.
onEachDevice([](int d) {
mark("__cuda_start_event");
});
}
#endif
mark("__start_profile", false);
}
PyObject * THPStorage_(New)(THStorage *ptr)
{
TORCH_ASSERT(ptr);
PyTypeObject *type = (PyTypeObject *)THPStorageClass;
PyObject *obj = type->tp_alloc(type, 0);
if (obj) {
((THPStorage *)obj)->cdata = ptr;
} else {
THStorage_(free)(LIBRARY_STATE ptr);
}
return obj;
}
void InputBuffer::add(size_t pos, Variable var) {
TORCH_ASSERT(pos < buffer.size());
if (!var.defined()) {
return;
}
auto& old_var = buffer[pos];
if (!old_var.defined()) {
buffer[pos] = std::move(var);
} else {
AutoGPU auto_gpu(var);
// ATen doesn't route sparse additions correctly...
if (old_var.type().is_sparse()) {
buffer[pos] = var + old_var;
} else {
buffer[pos] = old_var + var;
}
}
}
at::Type& get_default_tensor_type() {
TORCH_ASSERT(default_tensor_type);
return *default_tensor_type;
}
