COMPMID-2456: NEDeconvolutionLayer.cpp, NHWC is not supported

Support of NHWC for NEDeconvolutionLayer
Bugfix for QASYMM8 in CPPUpsample when offset is different than 0
QASYMM8 tests added in NEUpsample with offset different than 0

Change-Id: I8283fa5e5e323fd4d5777136359ddb33025674bb
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1517
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>

7 files changed, 225 insertions, 63 deletions

diff --git a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
index 62977a7647..360bb23f22 100644
--- a/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDeconvolutionLayer.h
@@ -118,8 +118,15 @@ private:
     NEConvolutionLayer   _conv_f;
     CPPUpsample          _upsample_f;
     CPPFlipWeightsKernel _flip_weights;
+    NEPermute            _permute_input;
+    NEPermute            _permute_weights;
+    NEPermute            _permute_output;
     Tensor               _scaled_output;
     Tensor               _weights_flipped;
+    Tensor               _permuted_input;
+    Tensor               _permuted_weights;
+    Tensor               _permuted_output;
+    bool                 _is_nchw;
     const ITensor       *_original_weights;
     ITensor             *_input;
     PadStrideInfo        _info;
diff --git a/src/core/CPP/kernels/CPPUpsampleKernel.cpp b/src/core/CPP/kernels/CPPUpsampleKernel.cpp
index 6620ce2aeb..ad2d54a0f8 100644
--- a/src/core/CPP/kernels/CPPUpsampleKernel.cpp
+++ b/src/core/CPP/kernels/CPPUpsampleKernel.cpp
@@ -34,8 +34,8 @@
 #include <cstddef>
 #include <cstdint>
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 CPPUpsampleKernel::CPPUpsampleKernel()
     : _input(nullptr), _output(nullptr), _info()
 {
@@ -82,7 +82,10 @@ void CPPUpsampleKernel::run(const Window &window, const ThreadInfo &info)
     const int    end_x         = width_scaled - _info.pad().first;
     const size_t element_size  = _input->info()->element_size();
 
-    std::fill_n(_output->buffer(), _output->info()->total_size(), 0);
+    //The fill value is normally 0, but for QASYMM8 the '0' corresponds to the offset
+    const uint8_t fill_value = _output->info()->data_type() == DataType::QASYMM8 ? utility::clamp<uint8_t>(_output->info()->quantization_info().uniform().offset) : 0;
+    //Filling a value different than 0 works only for QASYMM8 datatype since we are filling 1byte values in a buffer of uint8_ts
+    std::fill_n(_output->buffer(), _output->info()->total_size(), fill_value);
 
     // Create window
     Window window_out(window);
@@ -99,3 +102,4 @@ void CPPUpsampleKernel::run(const Window &window, const ThreadInfo &info)
     },
     in, out);
 }
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
index 7bbacb139c..581f257581 100644
--- a/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDeconvolutionLayer.cpp
@@ -38,8 +38,15 @@ NEDeconvolutionLayer::NEDeconvolutionLayer(std::shared_ptr<IMemoryManager> memor
       _conv_f(),
       _upsample_f(),
       _flip_weights(),
+      _permute_input(),
+      _permute_weights(),
+      _permute_output(),
       _scaled_output(),
       _weights_flipped(),
+      _permuted_input(),
+      _permuted_weights(),
+      _permuted_output(),
+      _is_nchw(false),
       _original_weights(nullptr),
       _input(nullptr),
       _info(),
@@ -52,18 +59,20 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, input);
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) != weights->dimension(1));
-    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(0) < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(weights, input);
+    const unsigned int width_idx  = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::WIDTH);
+    const unsigned int height_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::HEIGHT);
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) != weights->dimension(height_idx));
+    ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(width_idx) < 1);
     ARM_COMPUTE_RETURN_ERROR_ON(!info.padding_is_symmetric());
 
     const unsigned int stride_x = info.stride().first;
     const unsigned int stride_y = info.stride().second;
 
-    auto out_dims = deconvolution_output_dimensions(input->dimension(0), input->dimension(1), weights->dimension(0), weights->dimension(1),
+    auto out_dims = deconvolution_output_dimensions(input->dimension(width_idx), input->dimension(height_idx), weights->dimension(width_idx), weights->dimension(height_idx),
                                                     info.pad().first, info.pad().second, stride_x, stride_y);
 
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-
     if(is_data_type_quantized_asymmetric(input->data_type()))
     {
         ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
@@ -90,10 +99,10 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
     TensorInfo          scale_out_info(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(scale_out_shape));
     const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
 
-    for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i)
-    {
-        ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != scale_out_info.dimension(i));
-    }
+    const unsigned int batches_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::BATCHES);
+    const unsigned int channel_idx = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::CHANNEL);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(batches_idx) != scale_out_info.dimension(batches_idx));
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(channel_idx) != scale_out_info.dimension(channel_idx));
 
     ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayer::validate(&scale_out_info, weights, bias, output, conv_info, WeightsInfo()));
 
@@ -103,21 +112,24 @@ Status NEDeconvolutionLayer::validate(const ITensorInfo *input, const ITensorInf
 void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *bias, ITensor *output, const PadStrideInfo &info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+    ARM_COMPUTE_ERROR_THROW_ON(NEDeconvolutionLayer::validate(input->info(), weights->info(), bias->info(), output->info(), info));
+
+    const DataLayout data_layout = input->info()->data_layout();
 
     _input            = input;
     _original_weights = weights;
     _info             = info;
     _is_prepared      = false;
+    _is_nchw          = data_layout == DataLayout::NCHW;
 
-    const DataLayout   data_layout = input->info()->data_layout();
-    const unsigned int stride_x    = info.stride().first;
-    const unsigned int stride_y    = info.stride().second;
+    const unsigned int stride_x = info.stride().first;
+    const unsigned int stride_y = info.stride().second;
 
-    _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
-    _flip_weights.configure(weights, &_weights_flipped);
-
-    auto out_dims = deconvolution_output_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights->info()->dimension(0), weights->info()->dimension(1),
-                                                    info.pad().first, info.pad().second, stride_x, stride_y);
+    const unsigned int width_idx  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
+    const unsigned int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
+    auto               out_dims   = deconvolution_output_dimensions(input->info()->dimension(width_idx), input->info()->dimension(height_idx), weights->info()->dimension(width_idx),
+                                                                    weights->info()->dimension(height_idx),
+                                                                    info.pad().first, info.pad().second, stride_x, stride_y);
 
     const TensorShape output_shape = compute_deconvolution_output_shape(out_dims, *input->info(), *weights->info());
     // Output auto initialization if not yet initialized
@@ -128,20 +140,73 @@ void NEDeconvolutionLayer::configure(ITensor *input, const ITensor *weights, con
 
     _memory_group.manage(&_scaled_output);
 
-    // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
-    unsigned int      padx            = 0;
-    unsigned int      pady            = 0;
-    const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, padx, pady);
-
-    TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
-    _scaled_output.allocator()->init(scale_out_info);
-
-    const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
-    _upsample_f.configure(input, &_scaled_output, upsample_info);
-
-    // setup the function to convolve the upscaled output
-    const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
-    _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info);
+    if(!_is_nchw)
+    {
+        _memory_group.manage(&_permuted_input);
+        _memory_group.manage(&_permuted_weights);
+        _memory_group.manage(&_permuted_output);
+
+        // Configure the function to transform the input tensor from NHWC -> NCHW
+        _permuted_input.info()->set_quantization_info(input->info()->quantization_info());
+        _permute_input.configure(input, &_permuted_input, PermutationVector(1U, 2U, 0U));
+        _permuted_input.info()->set_data_layout(DataLayout::NCHW);
+
+        // Configure the function to transform the weights tensor from NHWC -> NCHW
+        _permuted_weights.info()->set_quantization_info(weights->info()->quantization_info());
+        _permute_weights.configure(weights, &_permuted_weights, PermutationVector(1U, 2U, 0U));
+        _permuted_weights.info()->set_data_layout(DataLayout::NCHW);
+
+        // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
+        unsigned int      padx            = 0;
+        unsigned int      pady            = 0;
+        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*_permuted_input.info(), *_permuted_weights.info(), stride_x, stride_y, out_dims, padx,
+                                                                                  pady);
+
+        TensorInfo scale_out_info(scale_out_shape, 1, _permuted_input.info()->data_type(), _permuted_input.info()->quantization_info());
+        scale_out_info.set_data_layout(DataLayout::NCHW);
+        _scaled_output.allocator()->init(scale_out_info);
+
+        const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+        _upsample_f.configure(&_permuted_input, &_scaled_output, upsample_info);
+
+        _weights_flipped.allocator()->init(*_permuted_weights.info()->clone());
+        _weights_flipped.info()->set_quantization_info(weights->info()->quantization_info());
+        _flip_weights.configure(&_permuted_weights, &_weights_flipped);
+
+        // setup the function to convolve the upscaled output
+        const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
+
+        _permuted_output.info()->set_quantization_info(output->info()->quantization_info());
+        _conv_f.configure(&_scaled_output, &_weights_flipped, bias, &_permuted_output, conv_info);
+        _permuted_output.info()->set_data_layout(DataLayout::NCHW);
+
+        // Configure the function to transform the convoluted output to NHWC
+        _permute_output.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U));
+
+        _permuted_input.allocator()->allocate();
+        _permuted_weights.allocator()->allocate();
+        _permuted_output.allocator()->allocate();
+    }
+    else
+    {
+        // Find the upsampled dimensions and the padding needed for the convolution with stride 1 in order to match output shape
+        unsigned int      padx            = 0;
+        unsigned int      pady            = 0;
+        const TensorShape scale_out_shape = compute_deconvolution_upsampled_shape(*input->info(), *weights->info(), stride_x, stride_y, out_dims, padx, pady);
+
+        TensorInfo scale_out_info(scale_out_shape, 1, input->info()->data_type(), input->info()->quantization_info());
+        scale_out_info.set_data_layout(data_layout);
+        _scaled_output.allocator()->init(scale_out_info);
+        const PadStrideInfo upsample_info(stride_x, stride_y, padx / 2, pady / 2);
+        _upsample_f.configure(input, &_scaled_output, upsample_info);
+
+        _weights_flipped.allocator()->init(weights->info()->clone()->set_data_layout(data_layout));
+        _flip_weights.configure(weights, &_weights_flipped);
+
+        // setup the function to convolve the upscaled output
+        const PadStrideInfo conv_info(1, 1, 0, 0, 0, 0, DimensionRoundingType::CEIL);
+        _conv_f.configure(&_scaled_output, &_weights_flipped, bias, output, conv_info);
+    }
     _scaled_output.allocator()->allocate();
 }
 
@@ -151,8 +216,20 @@ void NEDeconvolutionLayer::run()
 
     MemoryGroupResourceScope scope_mg(_memory_group);
 
+    // Permute input
+    if(!_is_nchw)
+    {
+        _permute_input.run();
+    }
+
     _upsample_f.run();
     _conv_f.run();
+
+    // Permute output
+    if(!_is_nchw)
+    {
+        _permute_output.run();
+    }
 }
 
 void NEDeconvolutionLayer::prepare()
@@ -163,6 +240,11 @@ void NEDeconvolutionLayer::prepare()
 
         // Run weights flipping and mark original weights tensor as unused
         _weights_flipped.allocator()->allocate();
+        // Permute weights
+        if(!_is_nchw)
+        {
+            _permute_weights.run();
+        }
         NEScheduler::get().schedule(&_flip_weights, Window::DimZ);
         _original_weights->mark_as_unused();
 
diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp
index 4d0ce63089..34a1bf5a45 100644
--- a/tests/validation/NEON/DeconvolutionLayer.cpp
+++ b/tests/validation/NEON/DeconvolutionLayer.cpp
@@ -62,7 +62,7 @@ const auto data3x3_precommit = datasets::SmallDeconvolutionShapes() * framework:
 const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1)
                      * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", { 3 });
 
-const auto data_layouts_dataset = framework::dataset::make("DataLayout", { DataLayout::NCHW });
+const auto data_layouts_dataset = framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC });
 } // namespace
 
 TEST_SUITE(NEON)
@@ -250,7 +250,7 @@ TEST_SUITE(W4x4)
 FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture4x4<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data4x4, framework::dataset::make("DataType",
                                                                                                                        DataType::QASYMM8)),
                                                                                                                        data_layouts_dataset),
-                                                                                                                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 0))))
+                                                                                                                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 10))))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8, tolerance_num);
@@ -261,7 +261,7 @@ TEST_SUITE(W3x3)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data3x3_precommit, framework::dataset::make("DataType",
                        DataType::QASYMM8)),
                        data_layouts_dataset),
-                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 0))))
+                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 10))))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8, tolerance_num);
@@ -269,7 +269,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t
 FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType",
                        DataType::QASYMM8)),
                        data_layouts_dataset),
-                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 0))))
+                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 10))))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8, tolerance_num);
@@ -280,7 +280,7 @@ TEST_SUITE(W1x1)
 FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture1x1<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data1x1, framework::dataset::make("DataType",
                                                                                                                        DataType::QASYMM8)),
                                                                                                                        data_layouts_dataset),
-                                                                                                                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 0))))
+                                                                                                                       framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 10))))
 {
     // Validate output
     validate(Accessor(_target), _reference, tolerance_qasymm8, tolerance_num);
diff --git a/tests/validation/NEON/Upsample.cpp b/tests/validation/NEON/Upsample.cpp
index 39b69ee1e3..9ddfbe0f8e 100644
--- a/tests/validation/NEON/Upsample.cpp
+++ b/tests/validation/NEON/Upsample.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -105,6 +105,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
 template <typename T>
 using NEUpsampleLayerFixture = UpsampleLayerFixture<Tensor, Accessor, NEUpsampleLayer, T>;
 
+template <typename T>
+using NEUpsampleLayerQuantizedFixture = UpsampleLayerQuantizedFixture<Tensor, Accessor, NEUpsampleLayer, T>;
+
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(),
@@ -136,11 +139,12 @@ TEST_SUITE_END() // Float
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(),
-                                                                                                                     framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                                     framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
-                                                                                                                     framework::dataset::make("PadInfo", { Size2D(2, 2) })),
-                                                                                                             framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEUpsampleLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(),
+                                                                                                                      framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                                      framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+                                                                                                                      framework::dataset::make("PadInfo", { Size2D(2, 2) })),
+                                                                                                                      framework::dataset::make("UpsamplingPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR })),
+                                                                                                                      framework::dataset::make("QuantizationInfo", QuantizationInfo(2.f / 255.f, 10))))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/fixtures/UpsampleLayerFixture.h b/tests/validation/fixtures/UpsampleLayerFixture.h
index 40229e2866..0a72e44dad 100644
--- a/tests/validation/fixtures/UpsampleLayerFixture.h
+++ b/tests/validation/fixtures/UpsampleLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,37 +40,51 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class UpsampleLayerFixture : public framework::Fixture
+class UpsampleLayerFixtureBase : public framework::Fixture
 {
 public:
     template <typename...>
     void setup(TensorShape input_shape, DataType data_type, DataLayout data_layout,
-               Size2D info, const InterpolationPolicy &policy)
+               Size2D info, const InterpolationPolicy &policy, QuantizationInfo quantization_info)
     {
         _data_type = data_type;
 
-        _target    = compute_target(input_shape, info, policy, data_type, data_layout);
-        _reference = compute_reference(input_shape, info, policy, data_type);
+        _target    = compute_target(input_shape, info, policy, data_type, data_layout, quantization_info);
+        _reference = compute_reference(input_shape, info, policy, data_type, quantization_info);
     }
 
 protected:
     template <typename U>
     void fill(U &&tensor, int i)
     {
-        library->fill_tensor_uniform(tensor, i);
+        if(_data_type == DataType::QASYMM8)
+        {
+            const auto                             bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f);
+            std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second);
+            library->fill(tensor, distribution, i);
+        }
+        else
+        {
+            library->fill_tensor_uniform(tensor, i);
+        }
     }
 
-    TensorType compute_target(TensorShape   input_shape,
-                              const Size2D &info, const InterpolationPolicy &policy, DataType data_type, DataLayout data_layout)
+    TensorType compute_target(TensorShape input_shape, const Size2D &info, const InterpolationPolicy &policy,
+                              DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info)
     {
+        TensorShape output_shape(input_shape);
+        output_shape.set(0, info.x() * input_shape[0]);
+        output_shape.set(1, info.y() * input_shape[1]);
+
         if(data_layout == DataLayout::NHWC)
         {
             permute(input_shape, PermutationVector(2U, 0U, 1U));
+            permute(output_shape, PermutationVector(2U, 0U, 1U));
         }
 
         // Create tensors
-        TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, QuantizationInfo(), data_layout);
-        TensorType dst;
+        TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, quantization_info, data_layout);
+        TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, quantization_info, data_layout);
 
         // Create and configure function
         FunctionType upsample;
@@ -95,11 +109,11 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &input_shape,
-                                      const Size2D &info, const InterpolationPolicy &policy, DataType data_type)
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const Size2D &info, const InterpolationPolicy &policy,
+                                      DataType data_type, QuantizationInfo quantization_info)
     {
         // Create reference
-        SimpleTensor<T> src{ input_shape, data_type };
+        SimpleTensor<T> src{ input_shape, data_type, 1, quantization_info };
 
         // Fill reference
         fill(src, 0);
@@ -111,6 +125,33 @@ protected:
     SimpleTensor<T> _reference{};
     DataType        _data_type{};
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class UpsampleLayerFixture : public UpsampleLayerFixtureBase<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, DataType data_type, DataLayout data_layout,
+               Size2D info, const InterpolationPolicy &policy)
+    {
+        UpsampleLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, data_type, data_layout,
+                                                                                   info, policy, QuantizationInfo());
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class UpsampleLayerQuantizedFixture : public UpsampleLayerFixtureBase<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, DataType data_type, DataLayout data_layout,
+               Size2D info, const InterpolationPolicy &policy, QuantizationInfo quantization_info)
+    {
+        UpsampleLayerFixtureBase<TensorType, AccessorType, FunctionType, T>::setup(input_shape, data_type, data_layout,
+                                                                                   info, policy, quantization_info);
+    }
+};
+
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/reference/UpsampleLayer.cpp b/tests/validation/reference/UpsampleLayer.cpp
index 876f6d794a..8e36ee857e 100644
--- a/tests/validation/reference/UpsampleLayer.cpp
+++ b/tests/validation/reference/UpsampleLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -33,9 +33,10 @@ namespace validation
 {
 namespace reference
 {
+namespace
+{
 template <typename T>
-SimpleTensor<T> upsample_layer(const SimpleTensor<T> &src,
-                               const Size2D &info, const InterpolationPolicy policy)
+SimpleTensor<T> upsample_function(const SimpleTensor<T> &src, const Size2D &info, const InterpolationPolicy policy)
 {
     ARM_COMPUTE_ERROR_ON(policy != InterpolationPolicy::NEAREST_NEIGHBOR);
     ARM_COMPUTE_UNUSED(policy);
@@ -72,16 +73,39 @@ SimpleTensor<T> upsample_layer(const SimpleTensor<T> &src,
             }
         }
     }
-
     return out;
 }
 
+} // namespace
+
+template <typename T>
+SimpleTensor<T> upsample_layer(const SimpleTensor<T> &src, const Size2D &info, const InterpolationPolicy policy)
+{
+    return upsample_function<T>(src, info, policy);
+}
+
+template <>
+SimpleTensor<uint8_t> upsample_layer(const SimpleTensor<uint8_t> &src, const Size2D &info, const InterpolationPolicy policy)
+{
+    SimpleTensor<uint8_t> dst(src.shape(), src.data_type(), 1, src.quantization_info());
+
+    if(is_data_type_quantized_asymmetric(src.data_type()))
+    {
+        SimpleTensor<float> src_tmp = convert_from_asymmetric(src);
+        SimpleTensor<float> dst_tmp = upsample_function<float>(src_tmp, info, policy);
+        dst                         = convert_to_asymmetric(dst_tmp, src.quantization_info());
+    }
+    else
+    {
+        dst = upsample_function<uint8_t>(src, info, policy);
+    }
+    return dst;
+}
+
 template SimpleTensor<float> upsample_layer(const SimpleTensor<float> &src,
                                             const Size2D &info, const InterpolationPolicy policy);
 template SimpleTensor<half> upsample_layer(const SimpleTensor<half> &src,
                                            const Size2D &info, const InterpolationPolicy policy);
-template SimpleTensor<uint8_t> upsample_layer(const SimpleTensor<uint8_t> &src,
-                                              const Size2D &info, const InterpolationPolicy policy);
 } // namespace reference
 } // namespace validation
 } // namespace test