COMPMID-420, COMPMID-414 - Port CLConvolutionLayer and CLFullyConnectedLayer to use 8 bit fixed point

Change-Id: I1cb1b4d7711ad7b569ee691e13a5df1b3430292b
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/79565
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

8 files changed, 78 insertions, 54 deletions

diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl
index a5cbe3d5c4..a875911140 100644
--- a/src/core/CL/cl_kernels/convolution_layer.cl
+++ b/src/core/CL/cl_kernels/convolution_layer.cl
@@ -21,6 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+#include "fixed_point.h"
 #include "helpers.h"
 
 /** This kernel reshapes the tensor's low three dimensions to single column
@@ -99,7 +100,7 @@ __kernel void reshape_to_columns(
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
  * @note In case biases will be added to the convolution -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/F16/F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
@@ -148,17 +149,21 @@ __kernel void im2col_generic(
         }
     }
 
-#if defined(HAS_BIAS)
-    *((__global DATA_TYPE *)output_ptr) = (DATA_TYPE)1;
-#endif /* HAS_BIAS */
+#ifdef HAS_BIAS
+#ifdef FIXED_POINT_POSITION
+    *((__global DATA_TYPE *)output_ptr) = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
+#else  // FIXED_POINT_POSITION
+    *((__global DATA_TYPE *)output_ptr) = 1.0f;
+#endif // FIXED_POINT_POSITION
+#endif // HAS_BIAS
 }
-#endif //(CONVOLVED_WIDTH && STRIDE_X && STRIDE_Y && PAD_X && PAD_Y && KERNEL_WIDTH && KERNEL_HEIGHT && KERNEL_DEPTH && SRC_WIDTH && SRC_HEIGHT)
+#endif //defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_X) && defined(PAD_Y) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
 
 /** This kernel performs a reshaping of the output of the convolution layer.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/F16/F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
@@ -192,7 +197,7 @@ __kernel void col2im(
  * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
  * @note In case biases will be added in late stage, -DHAS_BIAS has to be passed to append the final matrix with 1 in each row.
  *
- * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/F16/F32
  * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
  * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
@@ -225,7 +230,11 @@ __kernel void im2col_reduced(
     if(get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1))
     {
         tmp_out_ptr += dst_stride_x;
+#ifdef FIXED_POINT_POSITION
+        *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)(1 << FIXED_POINT_POSITION);
+#else  // FIXED_POINT_POSITION
         *((__global DATA_TYPE *)tmp_out_ptr) = (DATA_TYPE)1;
+#endif // FIXED_POINT_POSITION
     }
-#endif /* HAS_BIAS */
+#endif // HAS_BIAS
 }
diff --git a/src/core/CL/cl_kernels/gemm.cl b/src/core/CL/cl_kernels/gemm.cl
index 46f1645aa7..db15720ad0 100644
--- a/src/core/CL/cl_kernels/gemm.cl
+++ b/src/core/CL/cl_kernels/gemm.cl
@@ -21,9 +21,12 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "fixed_point.h"
 #include "helpers.h"
 
+#ifdef FIXED_POINT_POSITION
+#include "fixed_point.h"
+#endif // FIXED_POINT_POSITION
+
 /** This OpenCL kernel computes the "vector" 1x4 transposition of input matrix
  *
  * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: U32/S32/F32
@@ -274,7 +277,11 @@ __kernel void gemm_accumulate_biases(
     accum_value = vload16(0, (__global DATA_TYPE *)accum.ptr);
     VEC_DATA_TYPE(DATA_TYPE, 16)
     biases_value = vload16(0, (__global DATA_TYPE *)biases.ptr);
-    accum_value  = biases_value + accum_value;
+#ifdef FIXED_POINT_POSITION
+    accum_value = ADD_SAT_OP_EXPAND(biases_value, accum_value, DATA_TYPE, 16);
+#else  // FIXED_POINT_POSITION
+    accum_value = biases_value + accum_value;
+#endif // FIXED_POINT_POSITION
 
     // Store result in the accummulate buffer
     vstore16(accum_value, 0, (__global DATA_TYPE *)accum.ptr);
diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index 679943ba3e..6b2a18b261 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -43,9 +43,9 @@ 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::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
     _input          = input;
     _output         = output;
diff --git a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
index 75c1a6e629..a7ca6f2f01 100644
--- a/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.cpp
@@ -53,6 +53,10 @@ 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)
+    {
+        build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(accum->info()->fixed_point_position()));
+    }
 
     // Create kernel
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_accumulate_biases", build_opts));
diff --git a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
index 73c8429055..27b215f2c8 100644
--- a/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMTranspose1xWKernel.cpp
@@ -56,7 +56,7 @@ void CLGEMMTranspose1xWKernel::configure(const ICLTensor *input, ICLTensor *outp
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
-    const unsigned int num_elems_processed_per_iteration = max_cl_vector_width / data_size_from_type(input->info()->data_type());
+    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
     const float        scale_x                           = num_elems_processed_per_iteration;
     ARM_COMPUTE_ERROR_ON((0 == static_cast<int>(input->info()->dimension(0) * (1.f / scale_x))));
 
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index 092f495f92..51922e0925 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -46,8 +46,9 @@ 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::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
 
     _input  = input;
     _output = output;
@@ -57,6 +58,11 @@ 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)
+    {
+        build_opts.emplace("-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
+    }
+
     int pad_x    = 0;
     int pad_y    = 0;
     int stride_x = 0;
diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index b29bf8f136..96d04dc143 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::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, 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);
@@ -63,8 +63,9 @@ void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const
         const unsigned int mat_weights_cols = weights->info()->dimension(3);
         const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
         TensorShape        shape_wr(mat_weights_cols, mat_weights_rows);
-        const DataType     dt = weights->info()->data_type();
-        TensorInfo         info_wr(shape_wr, 1, dt);
+        const DataType     dt                   = weights->info()->data_type();
+        const int          fixed_point_position = weights->info()->fixed_point_position();
+        TensorInfo         info_wr(shape_wr, 1, dt, fixed_point_position);
 
         _weights_reshaped.allocator()->init(info_wr);
         _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
@@ -95,23 +96,27 @@ 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::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, 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));
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4);
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F16, DataType::F32);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
         ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3));
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
 
+    const DataType dt                   = input->info()->data_type();
+    const int      fixed_point_position = input->info()->fixed_point_position();
+
     _has_bias             = (biases != nullptr);
     _are_weights_reshaped = weights_info.are_reshaped();
 
-    // Get parameters for conv_info
+    // Get parameters from conv_info
     unsigned int stride_x = 0;
     unsigned int stride_y = 0;
     unsigned int pad_x    = 0;
@@ -123,8 +128,8 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
 
-    const unsigned int kernel_width  = _are_weights_reshaped ? weights_info.kernel_size().first : weights->info()->dimension(0);
-    const unsigned int kernel_height = _are_weights_reshaped ? weights_info.kernel_size().second : weights->info()->dimension(1);
+    const unsigned int kernel_width  = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0);
+    const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1);
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height,
                                                  conv_info);
     ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one");
@@ -132,9 +137,10 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     // Check if its a "fully connected" convolution
     _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
 
-    // Create tensor to store the reshaped weights
-    size_t mat_weights_cols = weights->info()->dimension(3);
-    size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0);
+    unsigned int mat_weights_cols = weights->info()->dimension(3);
+    unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0);
+
+    // Reshape weights if needed
     if(_are_weights_reshaped)
     {
         mat_weights_cols                         = output->info()->dimension(2);
@@ -147,49 +153,48 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
         {
             // Create tensor to store the reshaped weights
             TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
-            TensorInfo  info_wr(shape_wr, 1, weights->info()->data_type());
+            TensorInfo  info_wr(shape_wr, 1, dt, fixed_point_position);
             _weights_reshaped.allocator()->init(info_wr);
-            _reshape_weights.configure(weights, biases, &_weights_reshaped, false);
-            weights = &_weights_reshaped;
+            _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
         }
         else
         {
             // Create tensor to store transposed weights
-            TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f)));
-            TensorInfo  info_wt(shape_wt, 1, weights->info()->data_type());
-            _weights_transposed.allocator()->init(info_wt);
-            _reshape_weights.configure(weights, biases, &_weights_transposed, true);
-            weights = &_weights_transposed;
+            const float transpose_width = 16.0f / input->info()->element_size();
+            TensorShape shape_wt(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
+            TensorInfo  info_wt(shape_wt, 1, dt, fixed_point_position);
+            _weights_reshaped.allocator()->init(info_wt);
+            _reshape_weights.configure(weights, biases, &_weights_reshaped, true /* 1xW transpose */);
         }
+        weights = &_weights_reshaped;
     }
+
     // Create tensor to store im2col reshaped inputs
-    const size_t mat_input_cols = mat_weights_rows;
-    const size_t mat_input_rows = conv_w * conv_h;
-    TensorShape  shape_im2col   = input->info()->tensor_shape();
+    const unsigned int mat_input_cols = mat_weights_rows;
+    const unsigned int mat_input_rows = conv_w * conv_h;
+    TensorShape        shape_im2col   = input->info()->tensor_shape();
     shape_im2col.set(0, mat_input_cols);
     shape_im2col.set(1, mat_input_rows);
     shape_im2col.set(2, 1);
-    _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type()));
+    _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, dt, fixed_point_position));
 
     // Create tensor (interleave) to prepare input tensor for GEMM
     if(!_is_fully_connected_convolution)
     {
         TensorShape shape_interleaved = shape_im2col;
         shape_interleaved.set(0, shape_interleaved.x() * 4);
-        shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4.f));
-        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type()));
+        shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
+        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position));
     }
 
     // Create GEMM output tensor
     TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape();
     shape_gemm.set(0, mat_weights_cols);
     shape_gemm.set(1, mat_input_rows);
-    _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type()));
+    _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, dt, fixed_point_position));
 
     // Configure kernels
     _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
-    _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
-
     if(_is_fully_connected_convolution)
     {
         _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f);
@@ -199,19 +204,13 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
         _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
         _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f);
     }
+    _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
 
+    // Allocate intermediate tensor
     if(!_are_weights_reshaped)
     {
-        if(!_is_fully_connected_convolution)
-        {
-            _weights_transposed.allocator()->allocate();
-        }
-        else
-        {
-            _weights_reshaped.allocator()->allocate();
-        }
+        _weights_reshaped.allocator()->allocate();
     }
-
     _input_im2col_reshaped.allocator()->allocate();
     if(!_is_fully_connected_convolution)
     {
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index b51e709927..11e670c98e 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::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, 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));
@@ -196,8 +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::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output);
     ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() != 2);