COMPMID-440, COMPMID-441 - Port CLConvolutionLayer and CLFullyConnectedLayer to support 16 bit fixed point

Change-Id: I8d8ef2cb5ec453eb83fba8d8077550b96ed4bceb
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79837
Reviewed-by: Moritz Pflanzer <moritz.pflanzer@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>

17 files changed, 37 insertions, 31 deletions

diff --git a/arm_compute/core/CL/kernels/CLCol2ImKernel.h b/arm_compute/core/CL/kernels/CLCol2ImKernel.h
index 63b0b63f20..9866c44011 100644
--- a/arm_compute/core/CL/kernels/CLCol2ImKernel.h
+++ b/arm_compute/core/CL/kernels/CLCol2ImKernel.h
@@ -66,7 +66,7 @@ public:
 
     /** Set the input and output of the kernel.
      *
-     * @param[in]  input          The input tensor to convert. Data types supported: QS8/F16/F32
+     * @param[in]  input          The input tensor to convert. Data types supported: QS8/QS16/F16/F32
      * @param[out] output         The output tensor. 3 lower dimensions represent a single output [width, height, OFM],
      *                            while the rest represent batch of outputs. Data types supported: Same as @p input
      * @param[in]  convolved_dims Output convolved dimensions.
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h
index 74a7a0e4a6..167664f493 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h
@@ -46,7 +46,7 @@ public:
     CLGEMMMatrixAccumulateBiasesKernel &operator=(CLGEMMMatrixAccumulateBiasesKernel &&) = default;
     /** Set the accumulate buffer and the biases of the kernel.
      *
-     * @param[in, out] accum  The accumulate tensor to convert. Data types supported: QS8/F16/F32
+     * @param[in, out] accum  The accumulate tensor to convert. Data types supported: QS8/QS16/F16/F32
      * @param[in]      biases The shared biases tensor to append. It must be 1D tensor. Data types supported: Same as @p input
      */
     void configure(ICLTensor *accum, const ICLTensor *biases);
diff --git a/arm_compute/core/CL/kernels/CLIm2ColKernel.h b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
index e9f1a3f8e2..b9eeb2e088 100644
--- a/arm_compute/core/CL/kernels/CLIm2ColKernel.h
+++ b/arm_compute/core/CL/kernels/CLIm2ColKernel.h
@@ -69,7 +69,7 @@ public:
     /** Set the input and output of the kernel.
      *
      * @param[in]  input       The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM],
-     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QS8/F16/F32
+     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QS8/QS16/F16/F32
      * @param[out] output      The output tensor. First 2 lower dimensions represent a transform of each 3D input,
      *                         while every dimension above represents a batch. Data types supported: Same as @p input
      * @param[in]  kernel_dims The kernel dimensions (width and height).
diff --git a/arm_compute/core/CL/kernels/CLTransposeKernel.h b/arm_compute/core/CL/kernels/CLTransposeKernel.h
index 79596f34a1..98c69f4d67 100644
--- a/arm_compute/core/CL/kernels/CLTransposeKernel.h
+++ b/arm_compute/core/CL/kernels/CLTransposeKernel.h
@@ -40,7 +40,7 @@ class CLTransposeKernel : public ICLSimple2DKernel
 public:
     /** Initialise the kernel's input and output.
      *
-     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/F16/U32/S32/F32
+     * @param[in]  input  Input tensor. Data types supported: U8/S8/QS8/U16/S16/QS16/F16/U32/S32/F32
      * @param[out] output Output tensor. Data type supported: Same as @p input
      */
     void configure(const ICLTensor *input, ICLTensor *output);
diff --git a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
index 8732c6094b..099348fb15 100644
--- a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
+++ b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h
@@ -47,7 +47,7 @@ public:
     /** Set the input and output of the kernel.
      *
      * @param[in]  input  The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared,
-     *                    and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QS8/F16/F32
+     *                    and 5D tensor with dimensions [kernel_x, kernel_y, IFM, OFM,  num_patches] if unshared. Data types supported: QS8/QS16/F16/F32
      * @param[in]  biases The shared biases tensor to append.  Bias is 1D tensor with dimensions [OFM] if shared and 2D tensor with
      *                    dimensions [OFM, num_patches] if unshared. Data types supported: Same as @p input
      * @param[out] output The output tensor. Should be a 2D Tensor. Data types supported: Same as @p input
diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
index 50a7dc95eb..aba88bd856 100644
--- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h
@@ -53,7 +53,7 @@ public:
     CLConvolutionLayerReshapeWeights();
     /** Set the input and output tensors.
      *
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/F16/F32.
+     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: QS8/QS16/F16/F32.
      * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
      * @param[out] output       Destination tensor. Data types supported: Same as @p weights.
      * @param[in]  transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
@@ -88,7 +88,7 @@ public:
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
      *                          while every optional dimension from 4 and above represent a batch of inputs.
-     *                          Data types supported: QS8/F16/F32.
+     *                          Data types supported: QS8/QS16/F16/F32.
      * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input.
      * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input.
      * @param[out] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
diff --git a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
index 807ff693bc..64d56894d3 100644
--- a/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
+++ b/arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h
@@ -50,7 +50,7 @@ public:
     CLFullyConnectedLayerReshapeWeights();
     /** Set the input and output tensors.
      *
-     * @param[in]  input               Weights tensor. The weights must be 2 dimensional. Data types supported: QS8/F16/F32.
+     * @param[in]  input               Weights tensor. The weights must be 2 dimensional. Data types supported: QS8/QS16/F16/F32.
      * @param[out] output              Destination tensor. Data type supported: Same as @p input.
      * @param[in]  transpose_weights   True if the weights must be transposed. Data types supported: Same as @p weights.
      * @param[in]  is_batched_fc_layer True if it is a batched fully connected layer
diff --git a/src/core/CL/cl_kernels/fixed_point.h b/src/core/CL/cl_kernels/fixed_point.h
index b0dab0affc..bf126a2fdf 100644
--- a/src/core/CL/cl_kernels/fixed_point.h
+++ b/src/core/CL/cl_kernels/fixed_point.h
@@ -137,6 +137,11 @@ ADDQ_SAT_IMPL(qs8x2)
 ADDQ_SAT_IMPL(qs8x4)
 ADDQ_SAT_IMPL(qs8x8)
 ADDQ_SAT_IMPL(qs8x16)
+ADDQ_SAT_IMPL(qs16x1)
+ADDQ_SAT_IMPL(qs16x2)
+ADDQ_SAT_IMPL(qs16x4)
+ADDQ_SAT_IMPL(qs16x8)
+ADDQ_SAT_IMPL(qs16x16)
 
 #define ADD_SAT_OP_EXPAND_STR(a, b, type, size) add_sat_##type##x##size((a), (b))
 #define ADD_SAT_OP_EXPAND(a, b, type, size) ADD_SAT_OP_EXPAND_STR(a, b, type, size)
diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index 6b2a18b261..cfbe7408e2 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -43,7 +43,7 @@ CLCol2ImKernel::CLCol2ImKernel()
 
 void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::pair<unsigned int, unsigned int> convolved_dims)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
diff --git a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
index a7ca6f2f01..b8fb488fed 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
@@ -43,7 +43,7 @@ CLGEMMMatrixAccumulateBiasesKernel::CLGEMMMatrixAccumulateBiasesKernel()
 
 void CLGEMMMatrixAccumulateBiasesKernel::configure(ICLTensor *accum, const ICLTensor *biases)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(accum, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(biases, accum);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(biases, accum);
     ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() != 1);
@@ -53,7 +53,7 @@ void CLGEMMMatrixAccumulateBiasesKernel::configure(ICLTensor *accum, const ICLTe
 
     std::set<std::string> build_opts;
     build_opts.insert(("-DDATA_TYPE=" + get_cl_type_from_data_type(accum->info()->data_type())));
-    if(accum->info()->data_type() == DataType::QS8)
+    if(is_data_type_fixed_point(accum->info()->data_type()))
     {
         build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(accum->info()->fixed_point_position()));
     }
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index 51922e0925..7d7732d5da 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -46,7 +46,7 @@ CLIm2ColKernel::CLIm2ColKernel()
 
 void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
@@ -58,7 +58,7 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
     build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
     build_opts.emplace((has_bias ? "-DHAS_BIAS" : ""));
 
-    if(input->info()->data_type() == DataType::QS8)
+    if(is_data_type_fixed_point(input->info()->data_type()))
     {
         build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
     }
diff --git a/src/core/CL/kernels/CLTransposeKernel.cpp b/src/core/CL/kernels/CLTransposeKernel.cpp
index c1a75f7dff..75d31d5f16 100644
--- a/src/core/CL/kernels/CLTransposeKernel.cpp
+++ b/src/core/CL/kernels/CLTransposeKernel.cpp
@@ -40,7 +40,8 @@ using namespace arm_compute;
 
 void CLTransposeKernel::configure(const ICLTensor *input, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::U32, DataType::S32, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::U16, DataType::S16, DataType::QS16, DataType::U32, DataType::S32, DataType::F16,
+                                                  DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape  output_shape{ input->info()->tensor_shape() };
diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
index 82634164de..b802c862fc 100644
--- a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
@@ -41,7 +41,7 @@ CLWeightsReshapeKernel::CLWeightsReshapeKernel()
 
 void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     const DataType dt                   = input->info()->data_type();
diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index 96d04dc143..4362cfd437 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -41,7 +41,7 @@ CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights()
 
 void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
@@ -96,7 +96,7 @@ CLConvolutionLayer::CLConvolutionLayer()
 
 void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights, output);
     ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2));
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 11e670c98e..5d3d104f92 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -39,7 +39,7 @@ CLFullyConnectedLayerReshapeWeights::CLFullyConnectedLayerReshapeWeights()
 
 void CLFullyConnectedLayerReshapeWeights::configure(const ICLTensor *input, ICLTensor *output, bool transpose_weights, bool is_batched_fc_layer)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON(output == nullptr);
     ARM_COMPUTE_ERROR_ON(input->info()->num_dimensions() != 2);
     ARM_COMPUTE_ERROR_ON((transpose_weights == false) && (is_batched_fc_layer == false));
@@ -196,7 +196,7 @@ void CLFullyConnectedLayer::configure_fc_fc_nb(const ICLTensor *input, const ICL
 
 void CLFullyConnectedLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose_weights, bool are_weights_reshaped)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2);
 
diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp
index f613f77e5e..7c5a1eb254 100644
--- a/tests/validation/CL/ConvolutionLayer.cpp
+++ b/tests/validation/CL/ConvolutionLayer.cpp
@@ -47,7 +47,7 @@ using namespace arm_compute::test::validation;
 namespace
 {
 const float tolerance_f32 = 1e-03f; /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-const float tolerance_qs8 = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS8 */
+const float tolerance_q   = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */
 
 CLTensor compute_convolution_layer(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, DataType dt,
                                    const PadStrideInfo &conv_info, int fixed_point_position)
@@ -102,7 +102,7 @@ BOOST_AUTO_TEST_SUITE(GEMM)
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(Configuration,
-                     AlexNetConvolutionLayerDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8 }),
+                     AlexNetConvolutionLayerDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8, DataType::QS16 }),
                      conv_set, dt)
 {
     // Set fixed point position data type allowed
@@ -189,7 +189,7 @@ BOOST_AUTO_TEST_SUITE_END()
 BOOST_AUTO_TEST_SUITE(Quantized)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
 BOOST_DATA_TEST_CASE(SmallConvolutionLayer,
-                     SmallConvolutionLayerDataset() * boost::unit_test::data::make(DataType::QS8) * boost::unit_test::data::xrange(4, 7),
+                     SmallConvolutionLayerDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      conv_set, dt, fixed_point_position)
 {
     // Compute function
@@ -199,12 +199,12 @@ BOOST_DATA_TEST_CASE(SmallConvolutionLayer,
     RawTensor ref_dst = Reference::compute_reference_convolution_layer(conv_set.src_shape, conv_set.weights_shape, conv_set.bias_shape, conv_set.dst_shape, dt, conv_set.info, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(LargeConvolutionLayer,
-                     AlexNetConvolutionLayerDataset() * boost::unit_test::data::make(DataType::QS8) * boost::unit_test::data::xrange(4, 7),
+                     AlexNetConvolutionLayerDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      conv_set, dt, fixed_point_position)
 {
     // Compute function
@@ -214,7 +214,7 @@ BOOST_DATA_TEST_CASE(LargeConvolutionLayer,
     RawTensor ref_dst = Reference::compute_reference_convolution_layer(conv_set.src_shape, conv_set.weights_shape, conv_set.bias_shape, conv_set.dst_shape, dt, conv_set.info, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 BOOST_AUTO_TEST_SUITE_END()
 
diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp
index 4d00c30d16..974a0bd1d5 100644
--- a/tests/validation/CL/FullyConnectedLayer.cpp
+++ b/tests/validation/CL/FullyConnectedLayer.cpp
@@ -44,7 +44,7 @@ using namespace arm_compute::test::validation;
 namespace
 {
 const float tolerance_f32 = 1e-03f; /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-const float tolerance_qs8 = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS8 */
+const float tolerance_q   = 1.0f;   /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */
 
 CLTensor compute_fully_connected_layer(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, DataType dt,
                                        bool transpose_weights, int fixed_point_position)
@@ -109,7 +109,7 @@ BOOST_AUTO_TEST_SUITE(FullyConnectedLayer)
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit") * boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(Configuration,
-                     SmallFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8 }),
+                     SmallFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::F32, DataType::QS8, DataType::QS16 }),
                      fc_set, dt)
 {
     // Set fixed point position data type allowed
@@ -188,7 +188,7 @@ BOOST_AUTO_TEST_SUITE_END()
 BOOST_AUTO_TEST_SUITE(Quantized)
 BOOST_TEST_DECORATOR(*boost::unit_test::label("precommit"))
 BOOST_DATA_TEST_CASE(RunSmall,
-                     SmallFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::QS8 }) * boost::unit_test::data::xrange(4, 7),
+                     SmallFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      fc_set, dt, fixed_point_position)
 {
     // Compute function
@@ -198,12 +198,12 @@ BOOST_DATA_TEST_CASE(RunSmall,
     RawTensor ref_dst = Reference::compute_reference_fully_connected_layer(fc_set.src_shape, fc_set.weights_shape, fc_set.bias_shape, fc_set.dst_shape, dt, fc_set.transpose_weights, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 
 BOOST_TEST_DECORATOR(*boost::unit_test::label("nightly"))
 BOOST_DATA_TEST_CASE(RunLarge,
-                     LargeFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::QS8 }) * boost::unit_test::data::xrange(4, 7),
+                     LargeFullyConnectedLayerDataset() * boost::unit_test::data::make({ DataType::QS8, DataType::QS16 }) * boost::unit_test::data::xrange(4, 7),
                      fc_set, dt, fixed_point_position)
 {
     // Compute function
@@ -213,7 +213,7 @@ BOOST_DATA_TEST_CASE(RunLarge,
     RawTensor ref_dst = Reference::compute_reference_fully_connected_layer(fc_set.src_shape, fc_set.weights_shape, fc_set.bias_shape, fc_set.dst_shape, dt, fc_set.transpose_weights, fixed_point_position);
 
     // Validate output
-    validate(CLAccessor(dst), ref_dst, tolerance_qs8);
+    validate(CLAccessor(dst), ref_dst, tolerance_q);
 }
 BOOST_AUTO_TEST_SUITE_END()