COMPMID-2317: Implement CLROIAlignLayer

Change-Id: Iaa61b7a3528d3f82339d2ff8a2d77e77a1c68603
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1821
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 177 insertions, 53 deletions

diff --git a/tests/validation/CL/ROIAlignLayer.cpp b/tests/validation/CL/ROIAlignLayer.cpp
index 566e1985b3..b213c6815f 100644
--- a/tests/validation/CL/ROIAlignLayer.cpp
+++ b/tests/validation/CL/ROIAlignLayer.cpp
@@ -41,11 +41,13 @@ namespace validation
 {
 namespace
 {
-RelativeTolerance<float> relative_tolerance_f32(0.01f);
-AbsoluteTolerance<float> absolute_tolerance_f32(0.001f);
+constexpr RelativeTolerance<float> relative_tolerance_f32(0.01f);
+constexpr AbsoluteTolerance<float> absolute_tolerance_f32(0.001f);
 
-RelativeTolerance<float> relative_tolerance_f16(0.01f);
-AbsoluteTolerance<float> absolute_tolerance_f16(0.001f);
+constexpr RelativeTolerance<float> relative_tolerance_f16(0.01f);
+constexpr AbsoluteTolerance<float> absolute_tolerance_f16(0.001f);
+
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);
 } // namespace
 
 TEST_SUITE(CL)
@@ -55,13 +57,14 @@ TEST_SUITE(RoiAlign)
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
                framework::dataset::make("InputInfo", { TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),
-                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32), // Mismatching data type input/rois
-                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32), // Mismatching data type input/output
-                                                       TensorInfo(TensorShape(250U, 128U, 2U), 1, DataType::F32), // Mismatching depth size input/output
-                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32), // Mismatching number of rois and output batch size
-                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32), // Invalid number of values per ROIS
-                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32), // Mismatching height and width input/output
-
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),                                         // Mismatching data type input/rois
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),                                         // Mismatching data type input/output
+                                                       TensorInfo(TensorShape(250U, 128U, 2U), 1, DataType::F32),                                         // Mismatching depth size input/output
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),                                         // Mismatching number of rois and output batch size
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),                                         // Invalid number of values per ROIS
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::F32),                                         // Mismatching height and width input/output
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::QASYMM8, QuantizationInfo(1.f / 255.f, 127)), // Invalid ROIS data type
+                                                       TensorInfo(TensorShape(250U, 128U, 3U), 1, DataType::QASYMM8, QuantizationInfo(1.f / 255.f, 127)), // Invalid ROIS Quantization Info
                                                      }),
                framework::dataset::make("RoisInfo", { TensorInfo(TensorShape(5, 4U), 1, DataType::F32),
                                                       TensorInfo(TensorShape(5, 4U), 1, DataType::F16),
@@ -70,6 +73,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
                                                       TensorInfo(TensorShape(5, 10U), 1, DataType::F32),
                                                       TensorInfo(TensorShape(4, 4U), 1, DataType::F32),
                                                       TensorInfo(TensorShape(5, 4U), 1, DataType::F32),
+                                                      TensorInfo(TensorShape(5, 4U), 1, DataType::F32),
+                                                      TensorInfo(TensorShape(5, 4U), 1, DataType::QASYMM16, QuantizationInfo(0.2f, 0)),
                                                     })),
                framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::F32),
                                                        TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::F32),
@@ -78,6 +83,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
                                                        TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::F32),
                                                        TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::F32),
                                                        TensorInfo(TensorShape(5U, 5U, 3U, 4U), 1, DataType::F32),
+                                                       TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::QASYMM8, QuantizationInfo(1.f / 255.f, 120)),
+                                                       TensorInfo(TensorShape(7U, 7U, 3U, 4U), 1, DataType::QASYMM8, QuantizationInfo(1.f / 255.f, 120)),
                                                      })),
                framework::dataset::make("PoolInfo", { ROIPoolingLayerInfo(7U, 7U, 1./8),
                                                       ROIPoolingLayerInfo(7U, 7U, 1./8),
@@ -86,8 +93,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
                                                       ROIPoolingLayerInfo(7U, 7U, 1./8),
                                                       ROIPoolingLayerInfo(7U, 7U, 1./8),
                                                       ROIPoolingLayerInfo(7U, 7U, 1./8),
+                                                      ROIPoolingLayerInfo(7U, 7U, 1./8),
                                                       })),
-               framework::dataset::make("Expected", { true, false, false, false, false, false, false })),
+               framework::dataset::make("Expected", { true, false, false, false, false, false, false, false, false })),
                input_info, rois_info, output_info, pool_info, expected)
 {
     ARM_COMPUTE_EXPECT(bool(CLROIAlignLayer::validate(&input_info.clone()->set_is_resizable(true), &rois_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), pool_info)) == expected, framework::LogLevel::ERRORS);
@@ -99,24 +107,46 @@ template <typename T>
 using CLROIAlignLayerFixture = ROIAlignLayerFixture<CLTensor, CLAccessor, CLROIAlignLayer, T>;
 
 TEST_SUITE(Float)
-FIXTURE_DATA_TEST_CASE(SmallROIAlignLayerFloat, CLROIAlignLayerFixture<float>, framework::DatasetMode::ALL,
-                       framework::dataset::combine(framework::dataset::combine(datasets::SmallROIDataset(),
-                                                                               framework::dataset::make("DataType", { DataType::F32 })),
-                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(Small, CLROIAlignLayerFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(datasets::SmallROIDataset(),
+                                       framework::dataset::make("DataType", { DataType::F32 })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, relative_tolerance_f32, .02f, absolute_tolerance_f32);
 }
-FIXTURE_DATA_TEST_CASE(SmallROIAlignLayerHalf, CLROIAlignLayerFixture<half>, framework::DatasetMode::ALL,
-                       framework::dataset::combine(framework::dataset::combine(datasets::SmallROIDataset(),
-                                                                               framework::dataset::make("DataType", { DataType::F16 })),
-                                                   framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+TEST_SUITE_END() // FP32
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(Small, CLROIAlignLayerFixture<half>, framework::DatasetMode::ALL,
+                       combine(combine(datasets::SmallROIDataset(),
+                                       framework::dataset::make("DataType", { DataType::F16 })),
+                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, relative_tolerance_f16, .02f, absolute_tolerance_f16);
 }
+TEST_SUITE_END() // FP16
 TEST_SUITE_END() // Float
 
+template <typename T>
+using CLROIAlignLayerQuantizedFixture = ROIAlignLayerQuantizedFixture<CLTensor, CLAccessor, CLROIAlignLayer, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(Small, CLROIAlignLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL,
+                       combine(combine(combine(combine(datasets::SmallROIDataset(),
+                                                       framework::dataset::make("DataType", { DataType::QASYMM8 })),
+                                               framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+                                       framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 127) })),
+                               framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(2.f / 255.f, 120) })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
+
 TEST_SUITE_END() // RoiAlign
 TEST_SUITE_END() // CL
 } // namespace validation
diff --git a/tests/validation/Helpers.cpp b/tests/validation/Helpers.cpp
index 360859e487..a811cabf56 100644
--- a/tests/validation/Helpers.cpp
+++ b/tests/validation/Helpers.cpp
@@ -120,6 +120,18 @@ SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint8_t> &src)
     return dst;
 }
 
+SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint16_t> &src)
+{
+    const UniformQuantizationInfo &quantization_info = src.quantization_info().uniform();
+    SimpleTensor<float>            dst{ src.shape(), DataType::F32, 1, QuantizationInfo(), src.data_layout() };
+
+    for(int i = 0; i < src.num_elements(); ++i)
+    {
+        dst[i] = dequantize_qasymm16(src[i], quantization_info);
+    }
+    return dst;
+}
+
 SimpleTensor<uint8_t> convert_to_asymmetric(const SimpleTensor<float> &src, const QuantizationInfo &quantization_info)
 {
     SimpleTensor<uint8_t>          dst{ src.shape(), DataType::QASYMM8, 1, quantization_info };
diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h
index 44dd7a9b81..0d6515b5c5 100644
--- a/tests/validation/Helpers.h
+++ b/tests/validation/Helpers.h
@@ -177,7 +177,7 @@ void fill_lookuptable(T &&table)
     }
 }
 
-/** Convert quantized simple tensor into float using tensor quantization information.
+/** Convert 8-bit asymmetric quantized simple tensor into float using tensor quantization information.
  *
  * @param[in] src Quantized tensor.
  *
@@ -185,6 +185,14 @@ void fill_lookuptable(T &&table)
  */
 SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint8_t> &src);
 
+/** Convert 16-bit asymmetric quantized simple tensor into float using tensor quantization information.
+ *
+ * @param[in] src Quantized tensor.
+ *
+ * @return Float tensor.
+ */
+SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<uint16_t> &src);
+
 /** Convert float simple tensor into quantized using specified quantization information.
  *
  * @param[in] src               Float tensor.
diff --git a/tests/validation/fixtures/ROIAlignLayerFixture.h b/tests/validation/fixtures/ROIAlignLayerFixture.h
index dfbb478a41..b9b85d3073 100644
--- a/tests/validation/fixtures/ROIAlignLayerFixture.h
+++ b/tests/validation/fixtures/ROIAlignLayerFixture.h
@@ -26,7 +26,7 @@
 
 #include "arm_compute/core/TensorShape.h"
 #include "arm_compute/core/Types.h"
-#include "arm_compute/runtime/CL/functions/CLROIAlignLayer.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "tests/AssetsLibrary.h"
 #include "tests/Globals.h"
 #include "tests/IAccessor.h"
@@ -42,14 +42,17 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ROIAlignLayerFixture : public framework::Fixture
+class ROIAlignLayerGenericFixture : public framework::Fixture
 {
 public:
+    using TRois = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value, uint16_t, T>::type;
+
     template <typename...>
-    void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout)
+    void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo)
     {
-        _target    = compute_target(input_shape, data_type, data_layout, pool_info, rois_shape);
-        _reference = compute_reference(input_shape, data_type, pool_info, rois_shape);
+        _rois_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::QASYMM16 : data_type;
+        _target         = compute_target(input_shape, data_type, data_layout, pool_info, rois_shape, qinfo, output_qinfo);
+        _reference      = compute_reference(input_shape, data_type, pool_info, rois_shape, qinfo, output_qinfo);
     }
 
 protected:
@@ -66,17 +69,17 @@ protected:
         const size_t num_rois       = rois_shape.y();
 
         std::mt19937 gen(library->seed());
-        T           *rois_ptr = static_cast<T *>(rois.data());
+        TRois       *rois_ptr = static_cast<TRois *>(rois.data());
 
         const float pool_width  = pool_info.pooled_width();
         const float pool_height = pool_info.pooled_height();
         const float roi_scale   = pool_info.spatial_scale();
 
         // Calculate distribution bounds
-        const auto scaled_width  = static_cast<T>((shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)] / roi_scale) / pool_width);
-        const auto scaled_height = static_cast<T>((shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)] / roi_scale) / pool_height);
-        const auto min_width     = static_cast<T>(pool_width / roi_scale);
-        const auto min_height    = static_cast<T>(pool_height / roi_scale);
+        const auto scaled_width  = static_cast<float>((shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)] / roi_scale) / pool_width);
+        const auto scaled_height = static_cast<float>((shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)] / roi_scale) / pool_height);
+        const auto min_width     = static_cast<float>(pool_width / roi_scale);
+        const auto min_height    = static_cast<float>(pool_height / roi_scale);
 
         // Create distributions
         std::uniform_int_distribution<int> dist_batch(0, shape[3] - 1);
@@ -93,11 +96,21 @@ protected:
             const auto x2        = x1 + dist_w(gen);
             const auto y2        = y1 + dist_h(gen);
 
-            rois_ptr[values_per_roi * pw]     = batch_idx;
-            rois_ptr[values_per_roi * pw + 1] = x1;
-            rois_ptr[values_per_roi * pw + 2] = y1;
-            rois_ptr[values_per_roi * pw + 3] = x2;
-            rois_ptr[values_per_roi * pw + 4] = y2;
+            rois_ptr[values_per_roi * pw] = batch_idx;
+            if(rois.data_type() == DataType::QASYMM16)
+            {
+                rois_ptr[values_per_roi * pw + 1] = quantize_qasymm16(static_cast<float>(x1), rois.quantization_info());
+                rois_ptr[values_per_roi * pw + 2] = quantize_qasymm16(static_cast<float>(y1), rois.quantization_info());
+                rois_ptr[values_per_roi * pw + 3] = quantize_qasymm16(static_cast<float>(x2), rois.quantization_info());
+                rois_ptr[values_per_roi * pw + 4] = quantize_qasymm16(static_cast<float>(y2), rois.quantization_info());
+            }
+            else
+            {
+                rois_ptr[values_per_roi * pw + 1] = static_cast<TRois>(x1);
+                rois_ptr[values_per_roi * pw + 2] = static_cast<TRois>(y1);
+                rois_ptr[values_per_roi * pw + 3] = static_cast<TRois>(x2);
+                rois_ptr[values_per_roi * pw + 4] = static_cast<TRois>(y2);
+            }
         }
     }
 
@@ -105,17 +118,23 @@ protected:
                               DataType                   data_type,
                               DataLayout                 data_layout,
                               const ROIPoolingLayerInfo &pool_info,
-                              const TensorShape          rois_shape)
+                              const TensorShape          rois_shape,
+                              const QuantizationInfo    &qinfo,
+                              const QuantizationInfo    &output_qinfo)
     {
         if(data_layout == DataLayout::NHWC)
         {
             permute(input_shape, PermutationVector(2U, 0U, 1U));
         }
 
+        const QuantizationInfo rois_qinfo = is_data_type_quantized(data_type) ? QuantizationInfo(0.125f, 0) : QuantizationInfo();
+
         // Create tensors
-        TensorType src         = create_tensor<TensorType>(input_shape, data_type, 1, QuantizationInfo(), data_layout);
-        TensorType rois_tensor = create_tensor<TensorType>(rois_shape, data_type);
-        TensorType dst;
+        TensorType src         = create_tensor<TensorType>(input_shape, data_type, 1, qinfo, data_layout);
+        TensorType rois_tensor = create_tensor<TensorType>(rois_shape, _rois_data_type, 1, rois_qinfo);
+
+        const TensorShape dst_shape = misc::shape_calculator::compute_roi_align_shape(*(src.info()), *(rois_tensor.info()), pool_info);
+        TensorType        dst       = create_tensor<TensorType>(dst_shape, data_type, 1, output_qinfo, data_layout);
 
         // Create and configure function
         FunctionType roi_align_layer;
@@ -147,23 +166,51 @@ protected:
     SimpleTensor<T> compute_reference(const TensorShape         &input_shape,
                                       DataType                   data_type,
                                       const ROIPoolingLayerInfo &pool_info,
-                                      const TensorShape          rois_shape)
+                                      const TensorShape          rois_shape,
+                                      const QuantizationInfo    &qinfo,
+                                      const QuantizationInfo    &output_qinfo)
     {
         // Create reference tensor
-        SimpleTensor<T> src{ input_shape, data_type };
-        SimpleTensor<T> rois_tensor{ rois_shape, data_type };
+        SimpleTensor<T>        src{ input_shape, data_type, 1, qinfo };
+        const QuantizationInfo rois_qinfo = is_data_type_quantized(data_type) ? QuantizationInfo(0.125f, 0) : QuantizationInfo();
+        SimpleTensor<TRois>    rois_tensor{ rois_shape, _rois_data_type, 1, rois_qinfo };
 
         // Fill reference tensor
         fill(src);
         generate_rois(rois_tensor, input_shape, pool_info, rois_shape);
 
-        return reference::roi_align_layer(src, rois_tensor, pool_info);
+        return reference::roi_align_layer(src, rois_tensor, pool_info, output_qinfo);
     }
 
     TensorType      _target{};
     SimpleTensor<T> _reference{};
+    DataType        _rois_data_type{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ROIAlignLayerFixture : public ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout)
+    {
+        ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, pool_info, rois_shape, data_type, data_layout,
+                                                                                      QuantizationInfo(), QuantizationInfo());
+    }
 };
 
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ROIAlignLayerQuantizedFixture : public ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type,
+               DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo)
+    {
+        ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, pool_info, rois_shape,
+                                                                                      data_type, data_layout, qinfo, output_qinfo);
+    }
+};
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/ROIAlignLayer.cpp b/tests/validation/reference/ROIAlignLayer.cpp
index 8a76983d44..8ad78ff915 100644
--- a/tests/validation/reference/ROIAlignLayer.cpp
+++ b/tests/validation/reference/ROIAlignLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -112,15 +112,31 @@ T clamp(T value, T lower, T upper)
 {
     return std::max(lower, std::min(value, upper));
 }
+
+SimpleTensor<float> convert_rois_from_asymmetric(SimpleTensor<uint16_t> rois)
+{
+    const UniformQuantizationInfo &quantization_info = rois.quantization_info().uniform();
+    SimpleTensor<float>            dst{ rois.shape(), DataType::F32, 1, QuantizationInfo(), rois.data_layout() };
+
+    for(int i = 0; i < rois.num_elements(); i += 5)
+    {
+        dst[i]     = static_cast<float>(rois[i]); // batch idx
+        dst[i + 1] = dequantize_qasymm16(rois[i + 1], quantization_info);
+        dst[i + 2] = dequantize_qasymm16(rois[i + 2], quantization_info);
+        dst[i + 3] = dequantize_qasymm16(rois[i + 3], quantization_info);
+        dst[i + 4] = dequantize_qasymm16(rois[i + 4], quantization_info);
+    }
+    return dst;
+}
 } // namespace
-template <typename T>
-SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &rois, const ROIPoolingLayerInfo &pool_info)
+template <typename T, typename TRois>
+SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const SimpleTensor<TRois> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo)
 {
     const size_t values_per_roi = rois.shape()[0];
     const size_t num_rois       = rois.shape()[1];
     DataType     dst_data_type  = src.data_type();
 
-    const auto *rois_ptr = static_cast<const T *>(rois.data());
+    const auto *rois_ptr = static_cast<const TRois *>(rois.data());
 
     TensorShape     input_shape = src.shape();
     TensorShape     output_shape(pool_info.pooled_width(), pool_info.pooled_height(), src.shape()[2], num_rois);
@@ -183,8 +199,19 @@ SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const SimpleTensor<T
     }
     return dst;
 }
-template SimpleTensor<float> roi_align_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &rois, const ROIPoolingLayerInfo &pool_info);
-template SimpleTensor<half> roi_align_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &rois, const ROIPoolingLayerInfo &pool_info);
+
+template SimpleTensor<float> roi_align_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo);
+template SimpleTensor<half> roi_align_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo);
+
+template <>
+SimpleTensor<uint8_t> roi_align_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint16_t> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo)
+{
+    SimpleTensor<float>   src_tmp  = convert_from_asymmetric(src);
+    SimpleTensor<float>   rois_tmp = convert_rois_from_asymmetric(rois);
+    SimpleTensor<float>   dst_tmp  = roi_align_layer<float, float>(src_tmp, rois_tmp, pool_info, output_qinfo);
+    SimpleTensor<uint8_t> dst      = convert_to_asymmetric(dst_tmp, output_qinfo);
+    return dst;
+}
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/ROIAlignLayer.h b/tests/validation/reference/ROIAlignLayer.h
index b67ff42166..e1568133e7 100644
--- a/tests/validation/reference/ROIAlignLayer.h
+++ b/tests/validation/reference/ROIAlignLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -36,8 +36,8 @@ namespace validation
 {
 namespace reference
 {
-template <typename T>
-SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &rois, const ROIPoolingLayerInfo &pool_info);
+template <typename T, typename TRois>
+SimpleTensor<T> roi_align_layer(const SimpleTensor<T> &src, const SimpleTensor<TRois> &rois, const ROIPoolingLayerInfo &pool_info, const QuantizationInfo &output_qinfo);
 } // namespace reference
 } // namespace validation
 } // namespace test