From 351bd137e48c5276963274ac741b172483e98d21 Mon Sep 17 00:00:00 2001
From: giuros01 <giuseppe.rossini@arm.com>
Date: Fri, 23 Aug 2019 14:27:30 +0100
Subject: 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>
---
 utils/GraphUtils.cpp | 51 +++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 43 insertions(+), 8 deletions(-)

(limited to 'utils/GraphUtils.cpp')

diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp
index 00165cd6c2..3646facab2 100644
--- a/utils/GraphUtils.cpp
+++ b/utils/GraphUtils.cpp
@@ -62,18 +62,34 @@ 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;
     });
 }
 
@@ -87,16 +103,32 @@ 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;
-- 
cgit v1.2.1