[PyTorch] Use plain old function pointer for RecordFunctionCallback (…

…reapply) (pytorch#49408)

Summary:
Pull Request resolved: pytorch#49408

Nearly every non-test callsite doesn't need to capture any variables anyway, and this saves 48 bytes per callback.
ghstack-source-id: 118665808

Test Plan:
Wait for GitHub CI since we had C++14-specific issues with
this one in previous PR pytorch#48629

Reviewed By: malfet

Differential Revision: D25563207

fbshipit-source-id: 6a2831205917d465f8248ca37429ba2428d5626d

aten/src/ATen/record_function.cpp

@@ -277,10 +277,12 @@ class CallbackManager {
       bool is_start) {
     try {
       if (is_start) {
-        ctx = rfcb.start()(rf);
+        ctx = rfcb.start() ? rfcb.start()(rf) : nullptr;
       }
       else {
-        rfcb.end()(rf, ctx.get());
+        if (rfcb.end()) {
+          rfcb.end()(rf, ctx.get());
+        }
       }
       return true;
     } catch (const std::exception &e) {

aten/src/ATen/record_function.h

@@ -305,14 +305,16 @@ struct TORCH_API RecordFunction {
  */
 class TORCH_API RecordFunctionCallback {
  public:
+  using StartCallback = std::unique_ptr<ObserverContext>(*)(const RecordFunction&);
+  using EndCallback = void (*)(const RecordFunction&, ObserverContext*);
+
   // This interface supports observers that require passing an ObserverContext
   // between start and end callbacks.
   explicit RecordFunctionCallback(
-      std::function<std::unique_ptr<ObserverContext>(const RecordFunction&)> start,
-      std::function<void(const RecordFunction&, ObserverContext*)> end =
-        [](const RecordFunction&, ObserverContext*) {}):
-      start_(std::move(start)),
-      end_(std::move(end)) {
+      StartCallback start,
+      EndCallback end = nullptr) :
+      start_(start),
+      end_(end) {
     scopes_.fill(true);
   }
 
@@ -368,18 +370,18 @@ class TORCH_API RecordFunctionCallback {
     return scopes_[(size_t)sc];
   }
 
-  inline const std::function<std::unique_ptr<ObserverContext>(const RecordFunction&)>& start() const {
+  inline StartCallback start() const {
     return start_;
   }
 
-  inline const std::function<void(const RecordFunction&, ObserverContext*)>& end() const {
+  inline EndCallback end() const {
     return end_;
   }
 
  private:
   friend class CallbackManager;
-  std::function<std::unique_ptr<ObserverContext>(const RecordFunction&)> start_;
-  std::function<void(const RecordFunction&, ObserverContext*)> end_;
+  StartCallback start_;
+  EndCallback end_;
   bool(*should_run_)(const RecordFunctionCallback&) = nullptr;
   double sampling_prob_ = 1.0;
   std::array<bool, static_cast<size_t>(RecordScope::NUM_SCOPES)> scopes_ = {};

binaries/record_function_benchmark.cc

@@ -19,10 +19,10 @@ const float kLowSamplingProb = 0.0001;
 
 void addTestCallback(
     double sampling_prob = 1.0,
-    std::function<std::unique_ptr<at::ObserverContext>(const at::RecordFunction&)> fn =
-        [](const at::RecordFunction&) { return nullptr; }) {
+    at::RecordFunctionCallback::StartCallback fn =
+        [](const at::RecordFunction&) -> std::unique_ptr<at::ObserverContext> { return nullptr; }) {
   auto cb = at::RecordFunctionCallback(
-      std::move(fn),
+      fn,
       [](const at::RecordFunction&, at::ObserverContext*) {})
     .needsInputs(false);
   if (sampling_prob < 1.0) {
@@ -106,10 +106,10 @@ int main(int argc, char** argv) {
   at::clearCallbacks();
 
   std::cout << "Checking number of sampled observer invocations" << std::endl;
-  int cb_count = 0;
+  static int cb_count = 0;
   addTestCallback(
       kLowSamplingProb,
-      [&](const at::RecordFunction& fn) {
+      [](const at::RecordFunction&) -> std::unique_ptr<at::ObserverContext> {
         ++cb_count;
         return nullptr;
       }