add autograd fallback handler

backends/interpreter.cpp

@@ -2,7 +2,6 @@
 // Distributed under the MIT license that can be found in the LICENSE file.
 
 #include "dispatch.h"
-#include "ops.h"
 #include "tensor.h"
 #include "trace.h"
 #include <ATen/core/List.h>

dispatch.h

@@ -5,3 +5,5 @@
 
 #define DISPATCHKEY_NO_NS PrivateUse1
 #define DISPATCHKEY DispatchKey::DISPATCHKEY_NO_NS
+
+#define AUTOGRADDISPATCHKEY_NO_NS AutogradPrivateUse1

tensor.cpp

@@ -7,7 +7,6 @@
 #undef NDEBUG
 #include "tensor.h"
 #include "dispatch.h"
-#include "ops.h"
 #include "trace.h"
 #include <ATen/RedispatchFunctions.h>
 #include <torch/library.h>
@@ -235,12 +234,8 @@ TORCH_LIBRARY_IMPL(aten, DISPATCHKEY_NO_NS, m) {
 #include "autogen/torch_library_table.h"
 }
 
-#if 0
-TORCH_LIBRARY_IMPL(aten, AutogradPrivateUse1, m) {
-  m.impl("isfinite", isfinite);
-  m.impl("reshape", reshape);
-  m.impl("to.device", to_device);
+TORCH_LIBRARY_IMPL(_, AUTOGRADDISPATCHKEY_NO_NS, m) {
+  m.fallback(torch::CppFunction::makeFallthrough());
 }
-#endif
 
 }

trace.cpp

@@ -50,6 +50,11 @@ public:
 };
 }
 
+void TensorOp::incref() {
+  assert(isObservable());
+  ++refs;
+}
+
 void TensorOp::decref(TensorOp *ops) {
   assert(refs > 0);
   --refs;
@@ -176,6 +181,18 @@ void Trace::incref(const List<optional<Tensor>> &l) {
   }
 }
 
+void Trace::set_unobservable(unsigned idx) {
+  auto &op = ops[idx];
+  assert(op.tensor);
+  op.tensor = nullptr;
+  op.decref(ops);
+
+  // reclaim slot if this was the last created tensor
+  if (op.refs == 0 && idx+1 == next_op) {
+    --next_op;
+  }
+}
+
 void Trace::flush() {
 #if 1
   cout << "Flush trace\n" << *this;

trace.h

@@ -3,6 +3,7 @@
 // Copyright (c) 2021-present The Torchy Authors.
 // Distributed under the MIT license that can be found in the LICENSE file.
 
+#include "ops.h"
 #include <ATen/Tensor.h>
 #include <c10/core/DispatchKeySet.h>
 #include <c10/util/variant.h>
@@ -51,18 +52,13 @@ struct TensorOp {
   TorchyTensor *tensor;
   std::vector<UnionInputTys> args;
   c10::DispatchKeySet dispatch_key;
-  unsigned id;
+  TorchOp id;
   unsigned refs;
 
-  void incref() {
-    assert(isObservable());
-    ++refs;
-  }
-
+  void incref();
   void decref(TensorOp *ops);
 
   bool isObservable() const {
-    assert(!tensor || refs > 0);
     return tensor;
   }
 
@@ -91,7 +87,7 @@ class Trace {
   std::vector<std::unique_ptr<unsigned char[]>> deep_copies;
 
   template<typename T>
-  at::ArrayRef<T> deep_copy(at::ArrayRef<T> arr) {
+  at::ArrayRef<T> deep_copy(const at::ArrayRef<T> &arr) {
     if (arr.empty())
       return arr;
     size_t size = arr.size() * sizeof(T);
@@ -139,7 +135,7 @@ public:
   TensorOp* getOps() { return ops; }
 
   template<typename... T>
-  unsigned register_tensor(TorchyTensor *tensor, unsigned op_id,
+  unsigned register_tensor(TorchyTensor *tensor, TorchOp op_id,
                            c10::DispatchKeySet ks, T&... args) {
     assert(!flushing);
     if (next_op == MAX_TRACE_LENGTH)
@@ -155,18 +151,7 @@ public:
     return next_op++;
   }
 
-  void set_unobservable(unsigned idx) {
-    auto &op = ops[idx];
-    assert(op.tensor);
-    op.tensor = nullptr;
-    op.decref(ops);
-
-    // reclaim slot if this was the last created tensor
-    if (op.refs == 0 && idx+1 == next_op) {
-      --next_op;
-    }
-  }
-
+  void set_unobservable(unsigned idx);
   void flush();
 
   friend std::ostream& operator<<(std::ostream &os, const Trace &t);