[cker/ruy] Introduce EXPERIMENTAL_RUY_FEATURE build option (Samsung#3446

)

Introduce EXPERIMENTAL_RUY_FEATURE build option. For now, deallocating
prepack cached const backend tensors has not been matured. (Samsung#3225)
Because of the internal issue, however, the feature is needed on some
target. To do it, enable it when EXPERIMENTAL_RUY_FEATURE is set.

ONE-DCO-1.0-Signed-off-by: Yongseop Kim <[email protected]>

compute/cker/CMakeLists.txt

@@ -8,6 +8,9 @@ target_link_libraries(nnfw_lib_cker INTERFACE gemmlowp)
 target_link_libraries(nnfw_lib_cker INTERFACE ruy)
 target_link_libraries(nnfw_lib_cker INTERFACE ruy_instrumentation)
 target_compile_definitions(nnfw_lib_cker INTERFACE USE_RUY_GEMV)
+if(EXPERIMENTAL_RUY_FEATURE)
+  target_compile_definitions(nnfw_lib_cker INTERFACE EXPERIMENTAL_RUY_FEATURE)
+endif(EXPERIMENTAL_RUY_FEATURE)
 if(PROFILE_RUY)
   target_link_libraries(nnfw_lib_cker INTERFACE ruy_profiler)
 endif(PROFILE_RUY)

runtime/onert/backend/cpu/ops/FullyConnectedLayer.cc

@@ -113,23 +113,40 @@ void FullyConnectedLayer::fullyConnectedHybrid()
       getTensorShape(_bias), reinterpret_cast<const float *>(_bias ? _bias->buffer() : nullptr),
       getTensorShape(_output), reinterpret_cast<float *>(_output->buffer()), temp_arena);
 
-// TODO Enable calling decrease_ref
-#if 0
+// TODO Remove this ifdef
+#ifdef EXPERIMENTAL_RUY_FEATURE
   if (_cached_weights == nullptr || _is_weights_freed)
     return;
 
+  // '_cached_weights is not nullptr and _is_weights_freed is false' means
+  // this weight shape is satisfied with the ruy kernel's prepack cache's condition.
+  // After entering here, it will not enter again except below the case - input is zero-vector
+
+  // if input's elements are filled with zero, it by-passes(does not enter ruy-kernel path)
+  // so that handle this case
+  const int input_size = getTensorShape(_input).FlatSize();
+  if (nnfw::cker::IsZeroVector(reinterpret_cast<float *>(_input->buffer()), input_size))
+    return;
+
+  // This weight tensor could be other ops' const tensor.
+  // Therefore, below reference should be checked like following
   auto weight_tensor = dynamic_cast<const Tensor *>(_weights);
   if (weight_tensor)
   {
     auto tensor = const_cast<Tensor *>(weight_tensor);
+    if (tensor->buffer() == nullptr) // ref is already 0?
+    {
+      _is_weights_freed = true;
+      return;
+    }
 
     tensor->decrease_ref();
     if (tensor->buffer() == nullptr) // ref == 0?
     {
       _is_weights_freed = true;
     }
   }
-#endif // if 0
+#endif
 #endif
 }
 
@@ -178,6 +195,7 @@ void FullyConnectedLayer::prepare()
   }
 
 #ifdef USE_RUY_GEMV
+#ifdef EXPERIMENTAL_RUY_FEATURE
   // TODO This is workaround
   // The only fc hybrid will use ruy kernel
   if (_input->data_type() != OperandType::FLOAT32 ||
@@ -209,6 +227,7 @@ void FullyConnectedLayer::prepare()
     }
   }
 #endif
+#endif
 }
 
 } // namespace ops

runtime/onert/backend/cpu/ops/FullyConnectedLayer.h

@@ -72,6 +72,7 @@ class FullyConnectedLayer : public ::onert::exec::IFunction
 
 #ifdef USE_RUY_GEMV
   uint8_t *_cached_weights = nullptr; // weights to be cached and a key
+  bool _is_weights_freed = false;     // is weights freed?
 #endif
 };