MLCE-545 INT8 TFLite model execution abnormal

 * Add functionality to print output tensors to file in tempdir
 * UnitTests

Signed-off-by: Keith Davis <keith.davis@arm.com>
Change-Id: Idfb4c186544187db1fecdfca11c662540f645439

3 files changed, 74 insertions, 42 deletions

diff --git a/src/backends/reference/workloads/Debug.cpp b/src/backends/reference/workloads/Debug.cpp
index 24000d45e6..fdadfef590 100644
--- a/src/backends/reference/workloads/Debug.cpp
+++ b/src/backends/reference/workloads/Debug.cpp
@@ -5,22 +5,27 @@
 
 #include "Debug.hpp"
 #include <common/include/ProfilingGuid.hpp>
+#include <armnnUtils/Filesystem.hpp>
 
 #include <BFloat16.hpp>
 #include <Half.hpp>
 
 #include <algorithm>
 #include <iostream>
+#include <iosfwd>
+#include <fstream>
+#include <sys/stat.h>
 
 namespace armnn
 {
 
-template <typename T>
-void Debug(const TensorInfo& inputInfo,
-           const T* inputData,
-           LayerGuid guid,
-           const std::string& layerName,
-           unsigned int slotIndex)
+template<typename T>
+void PrintOutput(const TensorInfo& inputInfo,
+                 const T* inputData,
+                 LayerGuid guid,
+                 const std::string& layerName,
+                 unsigned int slotIndex,
+                 std::ostream& os)
 {
     const unsigned int numDims = inputInfo.GetNumDimensions();
     const unsigned int numElements = inputInfo.GetNumElements();
@@ -34,30 +39,30 @@ void Debug(const TensorInfo& inputInfo,
         strides[numDims - i] = strides[numDims - i + 1] * inputShape[numDims - i];
     }
 
-    std::cout << "{ ";
-    std::cout << "\"layerGuid\": " << guid << ", ";
-    std::cout << "\"layerName\": \"" << layerName << "\", ";
-    std::cout << "\"outputSlot\": " << slotIndex << ", ";
-    std::cout << "\"shape\": ";
+    os << "{ ";
+    os << "\"layerGuid\": " << guid << ", ";
+    os << "\"layerName\": \"" << layerName << "\", ";
+    os << "\"outputSlot\": " << slotIndex << ", ";
+    os << "\"shape\": ";
 
-    std::cout << "[";
+    os << "[";
     for (unsigned int i = 0; i < numDims; i++)
     {
-        std::cout << inputShape[i];
+        os << inputShape[i];
         if (i != numDims - 1)
         {
-            std::cout << ", ";
+            os << ", ";
         }
     }
-    std::cout << "], ";
+    os << "], ";
 
-    std::cout << "\"min\": "
-        << static_cast<float>(*std::min_element(inputData, inputData + numElements)) << ", ";
+    os << "\"min\": "
+              << static_cast<float>(*std::min_element(inputData, inputData + numElements)) << ", ";
 
-    std::cout << "\"max\": "
-        << static_cast<float>(*std::max_element(inputData, inputData + numElements)) << ", ";
+    os << "\"max\": "
+              << static_cast<float>(*std::max_element(inputData, inputData + numElements)) << ", ";
 
-    std::cout << "\"data\": ";
+    os << "\"data\": ";
 
     for (unsigned int i = 0; i < numElements; i++)
     {
@@ -65,69 +70,96 @@ void Debug(const TensorInfo& inputInfo,
         {
             if (i % strides[j] == 0)
             {
-                std::cout << "[" ;
+                os << "[";
             }
         }
 
-        std::cout << static_cast<float>(inputData[i]);
+        os << static_cast<float>(inputData[i]);
 
         for (unsigned int j = 0; j < numDims; j++)
         {
-            if ((i+1) % strides[j] == 0)
+            if ((i + 1) % strides[j] == 0)
             {
-                std::cout << "]" ;
+                os << "]";
             }
         }
 
         if (i != numElements - 1)
         {
-            std::cout << ", ";
+            os << ", ";
         }
     }
 
-    std::cout << " }" << std::endl;
+    os << " }" << std::endl;
+}
+
+template<typename T>
+void Debug(const TensorInfo& inputInfo,
+           const T* inputData,
+           LayerGuid guid,
+           const std::string& layerName,
+           unsigned int slotIndex,
+           bool outputsToFile)
+{
+    if (outputsToFile)
+    {
+        auto rootPathToFile = armnnUtils::Filesystem::CreateDirectory("/ArmNNIntermediateLayerOutputs");
+        std::ofstream out(rootPathToFile + layerName + ".numpy");
+        PrintOutput<T>(inputInfo, inputData, guid, layerName, slotIndex, out);
+    }
+    else
+    {
+        PrintOutput<T>(inputInfo, inputData, guid, layerName, slotIndex, std::cout);
+    }
 }
 
 template void Debug<BFloat16>(const TensorInfo& inputInfo,
-                          const BFloat16* inputData,
-                          LayerGuid guid,
-                          const std::string& layerName,
-                          unsigned int slotIndex);
+                              const BFloat16* inputData,
+                              LayerGuid guid,
+                              const std::string& layerName,
+                              unsigned int slotIndex,
+                              bool outputsToFile);
 
 template void Debug<Half>(const TensorInfo& inputInfo,
                           const Half* inputData,
                           LayerGuid guid,
                           const std::string& layerName,
-                          unsigned int slotIndex);
+                          unsigned int slotIndex,
+                          bool outputsToFile);
 
 template void Debug<float>(const TensorInfo& inputInfo,
                            const float* inputData,
                            LayerGuid guid,
                            const std::string& layerName,
-                           unsigned int slotIndex);
+                           unsigned int slotIndex,
+                           bool outputsToFile);
 
 template void Debug<uint8_t>(const TensorInfo& inputInfo,
                              const uint8_t* inputData,
                              LayerGuid guid,
                              const std::string& layerName,
-                             unsigned int slotIndex);
+                             unsigned int slotIndex,
+                             bool outputsToFile);
 
 template void Debug<int8_t>(const TensorInfo& inputInfo,
-                             const int8_t* inputData,
-                             LayerGuid guid,
-                             const std::string& layerName,
-                             unsigned int slotIndex);
+                            const int8_t* inputData,
+                            LayerGuid guid,
+                            const std::string& layerName,
+                            unsigned int slotIndex,
+                            bool outputsToFile);
 
 template void Debug<int16_t>(const TensorInfo& inputInfo,
                              const int16_t* inputData,
                              LayerGuid guid,
                              const std::string& layerName,
-                             unsigned int slotIndex);
+                             unsigned int slotIndex,
+                             bool outputsToFile);
 
 template void Debug<int32_t>(const TensorInfo& inputInfo,
                              const int32_t* inputData,
                              LayerGuid guid,
                              const std::string& layerName,
-                             unsigned int slotIndex);
+                             unsigned int slotIndex,
+                             bool outputsToFile);
 
 } // namespace armnn
diff --git a/src/backends/reference/workloads/Debug.hpp b/src/backends/reference/workloads/Debug.hpp
index 3f9920c543..a8802d1524 100644
--- a/src/backends/reference/workloads/Debug.hpp
+++ b/src/backends/reference/workloads/Debug.hpp
@@ -8,12 +8,12 @@
 
 namespace armnn
 {
-
 template <typename T>
 void Debug(const TensorInfo& inputInfo,
            const T* inputData,
            LayerGuid guid,
            const std::string& layerName,
-           unsigned int slotIndex);
+           unsigned int slotIndex,
+           bool outputsToFile);
 
 } //namespace armnn
diff --git a/src/backends/reference/workloads/RefDebugWorkload.cpp b/src/backends/reference/workloads/RefDebugWorkload.cpp
index 48b519f809..db67b3a782 100644
--- a/src/backends/reference/workloads/RefDebugWorkload.cpp
+++ b/src/backends/reference/workloads/RefDebugWorkload.cpp
@@ -45,7 +45,7 @@ void RefDebugWorkload<DataType>::Execute(std::vector<ITensorHandle*> inputs) con
     }
     else
     {
-        Debug(inputInfo, inputData, m_Data.m_Guid, m_Data.m_LayerName, m_Data.m_SlotIndex);
+        Debug(inputInfo, inputData, m_Data.m_Guid, m_Data.m_LayerName, m_Data.m_SlotIndex, m_Data.m_LayerOutputToFile);
     }
 
     std::memcpy(outputData, inputData, inputInfo.GetNumElements()*sizeof(T));