compmid-2573: Investigate FP16 Winograd reference implementations

Change-Id: I5a3e692c046a5ad28a676c03e3e51950c64cf503
Signed-off-by: giuros01 <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1845
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 60 insertions, 12 deletions

diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp
index 00165cd6c2..3646facab2 100644
--- a/utils/GraphUtils.cpp
+++ b/utils/GraphUtils.cpp
@@ -63,17 +63,33 @@ TFPreproccessor::TFPreproccessor(float min_range, float max_range)
 }
 void TFPreproccessor::preprocess(ITensor &tensor)
 {
+    if(tensor.info()->data_type() == DataType::F32)
+    {
+        preprocess_typed<float>(tensor);
+    }
+    else if(tensor.info()->data_type() == DataType::F16)
+    {
+        preprocess_typed<half>(tensor);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("NOT SUPPORTED!");
+    }
+}
+
+template <typename T>
+void TFPreproccessor::preprocess_typed(ITensor &tensor)
+{
     Window window;
     window.use_tensor_dimensions(tensor.info()->tensor_shape());
 
     const float range = _max_range - _min_range;
-
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const float value                                     = *reinterpret_cast<float *>(tensor.ptr_to_element(id));
-        float       res                                       = value / 255.f;            // Normalize to [0, 1]
-        res                                                   = res * range + _min_range; // Map to [min_range, max_range]
-        *reinterpret_cast<float *>(tensor.ptr_to_element(id)) = res;
+        const T value                                     = *reinterpret_cast<T *>(tensor.ptr_to_element(id));
+        float   res                                       = value / 255.f;            // Normalize to [0, 1]
+        res                                               = res * range + _min_range; // Map to [min_range, max_range]
+        *reinterpret_cast<T *>(tensor.ptr_to_element(id)) = res;
     });
 }
 
@@ -88,15 +104,31 @@ CaffePreproccessor::CaffePreproccessor(std::array<float, 3> mean, bool bgr, floa
 
 void CaffePreproccessor::preprocess(ITensor &tensor)
 {
+    if(tensor.info()->data_type() == DataType::F32)
+    {
+        preprocess_typed<float>(tensor);
+    }
+    else if(tensor.info()->data_type() == DataType::F16)
+    {
+        preprocess_typed<half>(tensor);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("NOT SUPPORTED!");
+    }
+}
+
+template <typename T>
+void CaffePreproccessor::preprocess_typed(ITensor &tensor)
+{
     Window window;
     window.use_tensor_dimensions(tensor.info()->tensor_shape());
-
     const int channel_idx = get_data_layout_dimension_index(tensor.info()->data_layout(), DataLayoutDimension::CHANNEL);
 
     execute_window_loop(window, [&](const Coordinates & id)
     {
-        const float value                                     = *reinterpret_cast<float *>(tensor.ptr_to_element(id)) - _mean[id[channel_idx]];
-        *reinterpret_cast<float *>(tensor.ptr_to_element(id)) = value * _scale;
+        const T value                                     = *reinterpret_cast<T *>(tensor.ptr_to_element(id)) - T(_mean[id[channel_idx]]);
+        *reinterpret_cast<T *>(tensor.ptr_to_element(id)) = value * T(_scale);
     });
 }
 
@@ -370,6 +402,9 @@ bool ValidationOutputAccessor::access_tensor(arm_compute::ITensor &tensor)
             case DataType::QASYMM8:
                 tensor_results = access_predictions_tensor<uint8_t>(tensor);
                 break;
+            case DataType::F16:
+                tensor_results = access_predictions_tensor<half>(tensor);
+                break;
             case DataType::F32:
                 tensor_results = access_predictions_tensor<float>(tensor);
                 break;
diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h
index 3417135f17..4c25dd2460 100644
--- a/utils/GraphUtils.h
+++ b/utils/GraphUtils.h
@@ -70,6 +70,9 @@ public:
     void preprocess(ITensor &tensor) override;
 
 private:
+    template <typename T>
+    void preprocess_typed(ITensor &tensor);
+
     std::array<float, 3> _mean;
     bool  _bgr;
     float _scale;
@@ -90,6 +93,9 @@ public:
     void preprocess(ITensor &tensor) override;
 
 private:
+    template <typename T>
+    void preprocess_typed(ITensor &tensor);
+
     float _min_range;
     float _max_range;
 };
diff --git a/utils/ImageLoader.h b/utils/ImageLoader.h
index 24fcbe179a..5d3a84c59a 100644
--- a/utils/ImageLoader.h
+++ b/utils/ImageLoader.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -249,14 +249,14 @@ public:
      *
      * @note If the image is a CLImage, the function maps and unmaps the image
      *
-     * @param[in,out] tensor Tensor with 3 planes to fill (Must be allocated, and of matching dimensions with the opened image). Data types supported: U8/F32
+     * @param[in,out] tensor Tensor with 3 planes to fill (Must be allocated, and of matching dimensions with the opened image). Data types supported: U8/F16/F32
      * @param[in]     bgr    (Optional) Fill the first plane with blue channel (default = false)
      */
     template <typename T>
     void fill_planar_tensor(T &tensor, bool bgr = false)
     {
         ARM_COMPUTE_ERROR_ON(!is_open());
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::U8, DataType::F32);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::U8, DataType::F32, DataType::F16);
 
         const DataLayout  data_layout  = tensor.info()->data_layout();
         const TensorShape tensor_shape = tensor.info()->tensor_shape();
@@ -324,6 +324,13 @@ public:
                         *reinterpret_cast<float *>(out.ptr() + 2 * stride_z) = static_cast<float>(bgr ? red : blue);
                         break;
                     }
+                    case DataType::F16:
+                    {
+                        *reinterpret_cast<half *>(out.ptr() + 0 * stride_z) = static_cast<half>(bgr ? blue : red);
+                        *reinterpret_cast<half *>(out.ptr() + 1 * stride_z) = static_cast<half>(green);
+                        *reinterpret_cast<half *>(out.ptr() + 2 * stride_z) = static_cast<half>(bgr ? red : blue);
+                        break;
+                    }
                     default:
                     {
                         ARM_COMPUTE_ERROR("Unsupported data type");
diff --git a/utils/Utils.h b/utils/Utils.h
index ec08896257..7fa74ab08b 100644
--- a/utils/Utils.h
+++ b/utils/Utils.h
@@ -401,7 +401,7 @@ public:
     void fill_tensor(T &tensor)
     {
         ARM_COMPUTE_ERROR_ON(!is_open());
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_NOT_IN(&tensor, arm_compute::DataType::QASYMM8, arm_compute::DataType::S32, arm_compute::DataType::F32);
+        ARM_COMPUTE_ERROR_ON_DATA_TYPE_NOT_IN(&tensor, arm_compute::DataType::QASYMM8, arm_compute::DataType::S32, arm_compute::DataType::F32, arm_compute::DataType::F16);
         try
         {
             // Map buffer if creating a CLTensor