Receive output tensor buffers for the inference

This patch receives tensor buffers allocated by upper layer
and makes them to be used for the inference.

As for this, it implements GetOutputTensorBuffers callback.

Change-Id: Ic8a3978205b38de0b1b93cd61b6fec4e98c62f3a
Signed-off-by: Inki Dae <inki.dae@samsung.com>

diff --git a/src/inference_engine_armnn.cpp b/src/inference_engine_armnn.cpp
index 30eeff115d2e610f5e414243443d3274f513b401..c29384dc7f9b7206b26127933451b5812d9f6137 100644 (file)
--- a/src/inference_engine_armnn.cpp
+++ b/src/inference_engine_armnn.cpp
@@ -221,13 +221,25 @@ int InferenceARMNN::GetInputTensorBuffers(std::vector<inference_engine_tensor_bu
 {
     LOGI("ENTER");
 
-    // Upper layer will allocate input tensor buffer/buffers. 
+    // Upper layer will allocate input tensor buffer/buffers.
 
     LOGI("LEAVE");
 
     return INFERENCE_ENGINE_ERROR_NONE;
 }
 
+int InferenceARMNN::GetOutputTensorBuffers(std::vector<inference_engine_tensor_buffer> &buffers)
+{
+    LOGI("ENTER");
+
+    // Upper layer will allocate output tensor buffer/buffers.
+
+    LOGI("LEAVE");
+
+    return INFERENCE_ENGINE_ERROR_NONE;
+}
+
+
 int InferenceARMNN::GetInputLayerProperty(inference_engine_layer_property &property)
 {
     LOGI("ENTER");
@@ -303,7 +315,8 @@ int InferenceARMNN::GetBackendCapacity(inference_engine_capacity *capacity)
     return INFERENCE_ENGINE_ERROR_NONE;
 }
 
-int InferenceARMNN::Run(std::vector<inference_engine_tensor_buffer> &input_buffers)
+int InferenceARMNN::Run(std::vector<inference_engine_tensor_buffer> &input_buffers,
+                        std::vector<inference_engine_tensor_buffer> &output_buffers)
 {
     LOGI("ENTER");
 
@@ -324,18 +337,11 @@ int InferenceARMNN::Run(std::vector<inference_engine_tensor_buffer> &input_buffe
 
     // TODO. consider mutiple input and output.
 
-    void *outputData = AllocateTensorBuffer((armnn::DataType)outputTensorInfo.GetDataType(),
-        tensor_size);
-    if (outputData == nullptr) {
-        LOGE("Fail to allocate tensor buffer.");
-        return INFERENCE_ENGINE_ERROR_OUT_OF_MEMORY;
-    }
-
-    armnn::Tensor output_tensor(outputTensorInfo, outputData);
+    armnn::Tensor output_tensor(outputTensorInfo, output_buffers.front().buffer);
     armnn::Tensor input_tensor(inputTensorInfo, input_buffers.front().buffer);
 
-    mOutputTensor.clear();
-    mOutputTensor.push_back(output_tensor);
+    std::vector<armnn::Tensor> OutputTensors;
+    OutputTensors.push_back(output_tensor);
 
     std::vector<armnn::Tensor> InputTensors;
     InputTensors.push_back(input_tensor);
@@ -344,7 +350,7 @@ int InferenceARMNN::Run(std::vector<inference_engine_tensor_buffer> &input_buffe
     armnn::OutputTensors output_tensors;
 
     input_tensors.push_back({mInputBindingInfo[0].first, InputTensors.front()});
-    output_tensors.push_back({mOutputBindingInfo[0].first, mOutputTensor.front()});
+    output_tensors.push_back({mOutputBindingInfo[0].first, OutputTensors.front()});
 
     armnn::Status ret = mRuntime->EnqueueWorkload(mNetworkIdentifier,
                                                   input_tensors, output_tensors);
@@ -360,19 +366,6 @@ int InferenceARMNN::GetInferenceResult(tensor_t& results)
 {
     LOGI("ENTER");
 
-    armnn::Tensor output_tensor = mOutputTensor.front();
-    armnn::TensorShape shape = output_tensor.GetShape();
-
-    std::vector<int> tmpDimInfo;
-    for (int i = 0; i < (int)output_tensor.GetNumDimensions(); i++) {
-        tmpDimInfo.push_back(shape[i]);
-    }
-
-    results.dimInfo.push_back(tmpDimInfo);
-    results.data.push_back((void *)output_tensor.GetMemoryArea());
-
-    // TODO. when should output_tensor buffer be released?
-
     LOGI("LEAVE");
 
     return INFERENCE_ENGINE_ERROR_NONE;

diff --git a/src/inference_engine_armnn_private.h b/src/inference_engine_armnn_private.h
index ece87bdc34ac68e4577b7323622d6a81b2998073..645b538b6d560780ef743d28a385c64a2f511a49 100644 (file)
--- a/src/inference_engine_armnn_private.h
+++ b/src/inference_engine_armnn_private.h
@@ -53,13 +53,16 @@ public:
 
     int GetInputTensorBuffers(std::vector<inference_engine_tensor_buffer> &buffers) override;
 
+    int GetOutputTensorBuffers(std::vector<inference_engine_tensor_buffer> &buffers) override;
+
     int GetInputLayerProperty(inference_engine_layer_property &property) override;
 
     int GetOutputLayerProperty(inference_engine_layer_property &property) override;
 
     int GetBackendCapacity(inference_engine_capacity *capacity) override;
 
-    int Run(std::vector<inference_engine_tensor_buffer> &input_buffers) override;
+    int Run(std::vector<inference_engine_tensor_buffer> &input_buffers,
+            std::vector<inference_engine_tensor_buffer> &output_buffers) override;
 
     int GetInferenceResult(tensor_t& results) override;
 
@@ -79,8 +82,6 @@ private:
 
     std::vector<armnn::BindingPointInfo> mInputBindingInfo;
     std::vector<armnn::BindingPointInfo> mOutputBindingInfo;
-
-    std::vector<armnn::Tensor> mOutputTensor;
 };
 
 } /* InferenceEngineImpl */