COMPMID-661: Convolution quantized (#32)

Change-Id: Id69df4ce98d1d89bdf9c9aa5c4d909659909b30f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110456
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

15 files changed, 263 insertions, 120 deletions

diff --git a/src/core/CL/cl_kernels/convolution_layer.cl b/src/core/CL/cl_kernels/convolution_layer.cl
index ce0849bf7a..77b9b64945 100644
--- a/src/core/CL/cl_kernels/convolution_layer.cl
+++ b/src/core/CL/cl_kernels/convolution_layer.cl
@@ -97,13 +97,14 @@ __kernel void reshape_to_columns(
     }
 }
 
-#if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
+#if defined(CONVOLVED_WIDTH) && defined(STRIDE_X) && defined(STRIDE_Y) && defined(PAD_LEFT) && defined(PAD_TOP) && defined(PAD_RIGHT) && defined(PAD_BOTTOM) && defined(KERNEL_WIDTH) && defined(KERNEL_HEIGHT) && defined(KERNEL_DEPTH) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(PAD_VALUE)
 /** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
+ * @note The value to use for the paddings must be passed at compile time using -DPAD_VALUE: e.g. -DPAD_VALUE=0
  * @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: QS8/QS16/F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/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)
@@ -149,14 +150,10 @@ __kernel void im2col_generic(
         {
 #if PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
             *output_ptr = *((__global DATA_TYPE *)(input_ptr + x * src_stride_x + y * src_stride_y));
-#else // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
+#else  // PAD_LEFT == 0 && PAD_TOP == 0 && PAD_RIGHT == 0 && PAD_BOTTOM == 0
             if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
             {
-#if defined(OFFSET)
-                *output_ptr = OFFSET;
-#else  /* OFFSET */
-                *output_ptr = 0;
-#endif /* OFFSET */
+                *output_ptr = PAD_VALUE;
             }
             else
             {
@@ -183,7 +180,7 @@ __kernel void im2col_generic(
  * @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: QS8/QS16/F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/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)
@@ -252,7 +249,7 @@ __kernel void im2col_kernel3x3_padx0_pady0(
  *
  * @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: QS8/QS16/F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/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)
@@ -291,7 +288,7 @@ __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: QS8/F16/F32
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QS8/QASYMM8/QS16/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)
diff --git a/src/core/CL/cl_kernels/gemmlowp.cl b/src/core/CL/cl_kernels/gemmlowp.cl
index a8e8e600fe..a92881320e 100644
--- a/src/core/CL/cl_kernels/gemmlowp.cl
+++ b/src/core/CL/cl_kernels/gemmlowp.cl
@@ -380,6 +380,7 @@ __kernel void gemmlowp_matrix_b_reduction(TENSOR3D_DECLARATION(src),
  * @attention The k_offset = a_offset * b_offset * k (where k is the number of matrix A columns) needs to be passed at compile time using -DK_OFFSET (i.e. -DK_OFFSET=1200)
  * @note In case the offset contribution due to a_offset is required, a_offset needs to be passed at compile time using -DA_OFFSET (i.e. -DA_OFFSET=1)
  * @note In case the offset contribution due to b_offset is required, b_offset needs to be passed at compile time using -DB_OFFSET (i.e. -DB_OFFSET=6)
+ * @note In case sum_col has batches, -DSUM_COL_HAS_BATCHES must be passed at compile time. Usually if gemmlowp is used to accelerate convolution layer, sum_col will not have batches
  *
  * The final result is:
  *
@@ -429,7 +430,12 @@ __kernel void gemmlowp_offset_contribution(TENSOR3D_DECLARATION(mm_result)
     Image sum_col = CONVERT_TO_IMAGE_STRUCT(sum_col);
 
     // Compute the offset contribution due to A_OFFSET
+#if defined(SUM_COL_HAS_BATCHES)
+    a_offset_s32 = vload16(0, (__global int *)(sum_col.ptr + get_global_id(2) * sum_col_stride_y));
+#else  // defined(MATRIX_B_HAS_BATCHES)
     a_offset_s32 = vload16(0, (__global int *)(sum_col.ptr));
+#endif // defined(MATRIX_B_HAS_BATCHES)
+
     a_offset_s32 *= (int16)A_OFFSET;
 #endif // defined(A_OFFSET)
 
@@ -615,4 +621,4 @@ __kernel void gemmlowp_output_stage_quantize_down_fixedpoint(TENSOR3D_DECLARATIO
     // Store the result
     vstore16(res, 0, dst.ptr);
 }
-#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
\ No newline at end of file
+#endif // defined(RESULT_OFFSET_AFTER_SHIFT) && defined(RESULT_FIXEDPOINT_MULTIPLIER) && defined(RESULT_SHIFT)
diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index f2886c569a..499e1e8fe0 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::QS16, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
@@ -52,7 +52,7 @@ void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::p
     output_shape.set(2, input->info()->tensor_shape()[0]);
 
     // Output auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position(), input->info()->quantization_info());
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
@@ -62,15 +62,15 @@ void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::p
     _output         = output;
     _convolved_dims = convolved_dims;
 
+    const DataType data_type = input->info()->data_type();
+
     // Create kernel
-    std::set<std::string> build_opts = { ("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())) };
-    build_opts.emplace("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.first));
-    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()));
-    }
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
+    build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.first));
+    build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
 
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts.options()));
 
     // Configure the local work size for Bifrost with a value obtained
     // via exhaustive autotuning over 30 representative tensor shapes.
diff --git a/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp
index d49aed3171..2877a74be8 100644
--- a/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.cpp
@@ -63,6 +63,7 @@ void CLGEMMLowpOffsetContributionKernel::configure(ICLTensor *mm_result, const I
         ARM_COMPUTE_ERROR_ON(vector_sum_col->info()->dimension(0) != mm_result->info()->dimension(0));
 
         build_opts.add_option("-DA_OFFSET=" + support::cpp11::to_string(a_offset));
+        build_opts.add_option_if(vector_sum_col->info()->tensor_shape().num_dimensions() > 1, "-DSUM_COL_HAS_BATCHES");
     }
 
     // If b_offset == 0, vector_sum_row can be a nullptr
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 37a430e8b0..62288cb771 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -41,8 +41,6 @@ namespace
 Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
     ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
 
@@ -53,6 +51,13 @@ Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, cons
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
     }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Error{};
 }
 
@@ -64,11 +69,17 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
 
     bool window_changed = update_window_and_padding(win,
-                                                    input_access,
-                                                    output_result_access);
+                                                    input_access);
+
+    if(output->total_size() != 0)
+    {
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
 
     if(bias != nullptr)
     {
@@ -76,8 +87,6 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
         window_changed = window_changed || update_window_and_padding(win, bias_access);
     }
 
-    output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
     Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
     return std::make_pair(err, win);
 }
@@ -93,6 +102,7 @@ CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::CLGEMMLowpQuantizeDow
 
 Error CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
                                                               (bias != nullptr) ? bias->clone().get() : nullptr,
diff --git a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
index 343c31c73d..5d4b25c142 100644
--- a/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -41,8 +41,6 @@ namespace
 Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
     ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
 
@@ -53,6 +51,13 @@ Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, cons
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
     }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Error{};
 }
 
@@ -64,11 +69,17 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
 
     bool window_changed = update_window_and_padding(win,
-                                                    input_access,
-                                                    output_result_access);
+                                                    input_access);
+
+    if(output->total_size() != 0)
+    {
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
 
     if(bias != nullptr)
     {
@@ -76,8 +87,6 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
         window_changed = window_changed || update_window_and_padding(win, bias_access);
     }
 
-    output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
     Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
     return std::make_pair(err, win);
 }
@@ -92,6 +101,7 @@ CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::CLGEMMLowpQuantizeDownInt32ToUint
 }
 Error CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
                                                               (bias != nullptr) ? bias->clone().get() : nullptr,
@@ -163,4 +173,4 @@ void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, cl
         enqueue(queue, *this, slice);
     }
     while(collapsed.slide_window_slice_3D(slice));
-}
\ No newline at end of file
+}
diff --git a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
index 6f410d3b14..bcf04b0982 100644
--- a/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMLowpReductionKernel.cpp
@@ -126,7 +126,7 @@ void CLGEMMLowpMatrixBReductionKernel::configure(const ICLTensor *mtx_b, ICLTens
     // Configure kernel window
     Window win = calculate_max_window(*vector_sum_col->info(), Steps(num_elems_processed_per_iteration));
 
-    AccessWindowStatic     input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1));
+    AccessWindowStatic     input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), num_elems_processed_per_iteration), _input->info()->dimension(1));
     AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration);
 
     update_window_and_padding(win,
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index f7cf9a3cb4..6514d6cf91 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -61,7 +61,6 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
     build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)));
     build_opts.add_option_if(has_bias, "-DHAS_BIAS");
     build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
-    build_opts.add_option_if(is_data_type_quantized_asymmetric(data_type), "-DOFFSET=" + support::cpp11::to_string(input->info()->quantization_info().offset));
 
     int stride_x = 0;
     int stride_y = 0;
@@ -95,6 +94,7 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
         build_opts.add_option("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
         build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
         build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
+        build_opts.add_option_if_else(is_data_type_quantized(data_type), "-DPAD_VALUE=" + support::cpp11::to_string(input->info()->quantization_info().offset), "-DPAD_VALUE=0");
 
         if(kernel_dims.width == 3 && kernel_dims.height == 3 && !conv_info.has_padding())
         {
diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
index be633b2418..3a9a32e58f 100644
--- a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
@@ -41,12 +41,12 @@ 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::QS16, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
-    const DataType dt                   = input->info()->data_type();
-    const int      fixed_point_position = input->info()->fixed_point_position();
+    const DataType data_type = input->info()->data_type();
 
+    // Calculate output shape
     TensorShape output_shape{ input->info()->tensor_shape() };
     output_shape.collapse(3);
     const size_t tmp_dim = output_shape[0];
@@ -54,7 +54,7 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *
     output_shape.set(1, tmp_dim + (biases != nullptr ? 1 : 0));
 
     // Output tensor auto inizialitation if not yet initialized
-    auto_init_if_empty(*output->info(), output_shape, 1, dt, fixed_point_position, input->info()->quantization_info());
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
 
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
@@ -62,6 +62,7 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *
 
     if(biases != nullptr)
     {
+        ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(data_type));
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
         ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
         ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (biases->info()->num_dimensions() != 1));
@@ -75,16 +76,13 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *
     _input  = input;
 
     // Create build options
-    std::set<std::string> build_opts;
-    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())));
-    build_opts.emplace(((biases != nullptr) ? "-DHAS_BIAS" : ""));
-    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()));
-    }
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
+    build_opts.add_option_if(biases != nullptr, "-DHAS_BIAS");
+    build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reshape_to_columns", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reshape_to_columns", build_opts.options()));
 
     // Set static arguments
     unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 102d08c7ba..c6f7ca4fb2 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -44,8 +44,6 @@ namespace
 Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
     ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
 
@@ -56,6 +54,13 @@ Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, cons
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
     }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Error{};
 }
 
@@ -67,11 +72,17 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
 
     bool window_changed = update_window_and_padding(win,
-                                                    input_access,
-                                                    output_result_access);
+                                                    input_access);
+
+    if(output->total_size() != 0)
+    {
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
 
     if(bias != nullptr)
     {
@@ -79,8 +90,6 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
         window_changed = window_changed || update_window_and_padding(win, bias_access);
     }
 
-    output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
     Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
     return std::make_pair(err, win);
 }
@@ -255,6 +264,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const
 
 Error NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
                                                               (bias != nullptr) ? bias->clone().get() : nullptr,
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
index edd6a9fd80..68b81d8a79 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.cpp
@@ -43,8 +43,6 @@ namespace
 Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON(max > 255);
     ARM_COMPUTE_RETURN_ERROR_ON(min < 0 || min > max);
 
@@ -55,6 +53,13 @@ Error validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, cons
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
     }
+
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+    }
+
     return Error{};
 }
 
@@ -66,11 +71,17 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
     Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
     AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
-    AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
 
     bool window_changed = update_window_and_padding(win,
-                                                    input_access,
-                                                    output_result_access);
+                                                    input_access);
+
+    if(output->total_size() != 0)
+    {
+        AccessWindowHorizontal output_result_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = window_changed || update_window_and_padding(win, output_result_access);
+
+        output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
 
     if(bias != nullptr)
     {
@@ -78,8 +89,6 @@ std::pair<Error, Window> validate_and_configure_window(ITensorInfo *input, ITens
         window_changed = window_changed || update_window_and_padding(win, bias_access);
     }
 
-    output_result_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
-
     Error err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Error{};
     return std::make_pair(err, win);
 }
@@ -262,6 +271,7 @@ void NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ITensor *inp
 
 Error NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
     ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(),
                                                               (bias != nullptr) ? bias->clone().get() : nullptr,
diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp
index 8d45416b30..66548d19b2 100644
--- a/src/runtime/CL/functions/CLConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 
 #include <cmath>
@@ -42,19 +43,22 @@ CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights(std::shared_p
 
 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::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, 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);
 
     if(biases != nullptr)
     {
+        ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->info()->data_type()));
         ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
         ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3));
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
 
-    const bool _has_bias = (biases != nullptr);
+    const bool       append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
+    const unsigned   bias_element  = (append_biases) ? 1 : 0;
+    const ICLTensor *biases_to_use = (append_biases) ? biases : nullptr;
 
     _transpose1xW = transpose1xW;
 
@@ -62,7 +66,7 @@ void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const
     {
         // Create tensor to store the reshaped weights
         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);
+        const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
         TensorShape        shape_wr(mat_weights_cols, mat_weights_rows);
         const DataType     dt                   = weights->info()->data_type();
         const int          fixed_point_position = weights->info()->fixed_point_position();
@@ -70,13 +74,13 @@ void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const
 
         _weights_reshaped.allocator()->init(info_wr);
         _memory_group.manage(&_weights_reshaped);
-        _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
+        _weights_reshape_kernel.configure(weights, biases_to_use, &_weights_reshaped);
         _weights_transposed_kernel.configure(&_weights_reshaped, output);
         _weights_reshaped.allocator()->allocate();
     }
     else
     {
-        _weights_reshape_kernel.configure(weights, biases, output);
+        _weights_reshape_kernel.configure(weights, biases_to_use, output);
     }
 }
 
@@ -95,36 +99,73 @@ void CLConvolutionLayerReshapeWeights::run()
 }
 
 CLConvolutionLayer::CLConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(),
-      _input_interleaved_reshaped(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false)
+    : _memory_group(memory_manager), _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _output_col2im_kernel(),
+      _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _append_bias(false), _is_fully_connected_convolution(false),
+      _are_weights_reshaped(false), _is_quantized(false)
 {
 }
 
+void CLConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed)
+{
+    if(_is_quantized)
+    {
+        // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
+        // Extract and negate input and weights offset
+        const QuantizationInfo input_quantization_info   = input->info()->quantization_info();
+        const QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
+
+        input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
+        weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
+
+        _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/));
+
+        // Revert back QuantizatioInfo as input and weights could be used in other convolution layers
+        input->info()->set_quantization_info(input_quantization_info);
+        weights->info()->set_quantization_info(weights_quantization_info);
+    }
+    else
+    {
+        _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed);
+    }
+}
+
 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::QS16, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights);
     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);
+    ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->info()->data_type()));
+
+    _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        if(_is_quantized)
+        {
+            ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            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();
+    const DataType dt = input->info()->data_type();
 
     // Set the GPU target for matrix multiply
     _mm_kernel.set_target(CLScheduler::get().target());
 
-    _has_bias             = (biases != nullptr);
+    _append_bias          = (biases != nullptr) && (!_is_quantized);
     _are_weights_reshaped = weights_info.are_reshaped();
 
+    const unsigned   bias_element  = (_append_bias) ? 1 : 0;
+    const ICLTensor *biases_to_use = (_append_bias) ? biases : nullptr;
+
     // Get parameters from conv_info
     unsigned int stride_x = 0;
     unsigned int stride_y = 0;
@@ -141,36 +182,36 @@ 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));
+    const bool run_interleaved      = (!_is_fully_connected_convolution && !_is_quantized);
 
     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);
+    unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element;
 
     // Reshape weights if needed
     if(_are_weights_reshaped)
     {
         mat_weights_cols                         = weights_info.num_kernels();
         const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4;
-        mat_weights_rows                         = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols);
+        mat_weights_rows                         = quarter_reshaped_cols + bias_element;
     }
     else
     {
-        if(_is_fully_connected_convolution)
+        if(_is_fully_connected_convolution || _is_quantized)
         {
             // Create tensor to store the reshaped weights
             TensorShape shape_wr(mat_weights_cols, mat_weights_rows);
-            TensorInfo  info_wr(shape_wr, 1, dt, fixed_point_position);
-            _weights_reshaped.allocator()->init(info_wr);
-            _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */);
+            _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wr));
+            _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, false /* 1xW transpose */);
         }
         else
         {
             // Create tensor to store transposed weights
             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_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_wt));
+            _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, true /* 1xW transpose */);
         }
+        _weights_reshaped.info()->set_quantization_info(weights->info()->quantization_info());
         weights = &_weights_reshaped;
     }
 
@@ -181,16 +222,16 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     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, dt, fixed_point_position));
+    _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
     _memory_group.manage(&_input_im2col_reshaped);
 
     // Create tensor (interleave) to prepare input tensor for GEMM
-    if(!_is_fully_connected_convolution)
+    if(run_interleaved)
     {
         TensorShape shape_interleaved = shape_im2col;
         shape_interleaved.set(0, shape_interleaved.x() * 4);
         shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f));
-        _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, dt, fixed_point_position));
+        _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved));
         _memory_group.manage(&_input_interleaved_reshaped);
     }
 
@@ -198,30 +239,51 @@ void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weig
     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, dt, fixed_point_position));
+    const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt;
+    // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
+    TensorInfo info_gemm(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_gemm).set_data_type(gemm_data_type).set_quantization_info(
+                             output->info()->quantization_info()));
+    _gemm_output.allocator()->init(info_gemm);
     _memory_group.manage(&_gemm_output);
 
     // Configure kernels
-
     _input_im2col_kernel.set_target(CLScheduler::get().target());
-    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
+    _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias);
 
     // Configure matrix multiply
-    if(_is_fully_connected_convolution)
+    if(run_interleaved)
     {
-        // The matrix A and Matrix B have not been reshaped
-        _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f, false);
+        _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
+        configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output);
+        _input_interleaved_reshaped.allocator()->allocate();
     }
     else
     {
-        _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
-        _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f);
-        _input_interleaved_reshaped.allocator()->allocate();
+        configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, false);
     }
     _input_im2col_reshaped.allocator()->allocate();
+
+    // Configure output stage for quantized case
+    if(_is_quantized)
+    {
+        float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale;
+        int   output_multiplier, output_shift;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+        _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+        _gemm_output.allocator()->allocate();
+    }
+
+    // Configure Col2Im
     _output_col2im_kernel.set_target(CLScheduler::get().target());
-    _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
-    _gemm_output.allocator()->allocate();
+    _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, std::make_pair(conv_w, conv_h));
+    if(_is_quantized)
+    {
+        _tmp_output.allocator()->allocate();
+    }
+    else
+    {
+        _gemm_output.allocator()->allocate();
+    }
 
     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");
 
@@ -243,15 +305,30 @@ void CLConvolutionLayer::run()
 
     _memory_group.acquire();
 
-    // Run input reshaping
+    // Run im2col
     CLScheduler::get().enqueue(_input_im2col_kernel);
-    if(!_is_fully_connected_convolution)
+
+    if(!_is_fully_connected_convolution && !_is_quantized)
     {
+        // Run interleave4x4
         CLScheduler::get().enqueue(_input_interleave_kernel);
     }
 
     // Runs matrix multiply on reshaped matrices
-    CLScheduler::get().enqueue(_mm_kernel);
+    if(_is_quantized)
+    {
+        _mm_gemmlowp.run();
+    }
+    else
+    {
+        CLScheduler::get().enqueue(_mm_kernel);
+    }
+
+    // Run output stage for quantized case
+    if(_is_quantized)
+    {
+        _gemmlowp_output_stage.run();
+    }
 
     // Reshape output matrix
     CLScheduler::get().enqueue(_output_col2im_kernel, false);
diff --git a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
index 6cc2f4bdb7..7fd81cdb94 100644
--- a/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLFullyConnectedLayer.cpp
@@ -50,13 +50,20 @@ void CLFullyConnectedLayer::configure_mm(const ICLTensor *input, const ICLTensor
 {
     if(_is_quantized)
     {
+        // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
         // Extract and negate input and weights offset
-        QuantizationInfo input_quantization_info   = input->info()->quantization_info();
-        QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
+        const QuantizationInfo input_quantization_info   = input->info()->quantization_info();
+        const QuantizationInfo weights_quantization_info = weights->info()->quantization_info();
+
         input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
         weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
+
         // Configure gemmlowp function
         _mm_gemmlowp.configure(input, weights, output);
+
+        // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
+        input->info()->set_quantization_info(input_quantization_info);
+        weights->info()->set_quantization_info(weights_quantization_info);
     }
     else
     {
diff --git a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
index 5d2d13e243..5c6f5b4ed0 100644
--- a/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.cpp
@@ -35,11 +35,11 @@ using namespace arm_compute;
 
 CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _mm_kernel(), _mtx_a_reshape_kernel(), _mtx_b_reshape_kernel(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _offset_contribution_kernel(),
-      _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _a_offset(0), _b_offset(0), _is_interleaved_transposed(true)
+      _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _a_offset(0), _b_offset(0), _is_interleaved_transposed(true), _is_first_run(true), _reshape_b_only_on_first_run(false)
 {
 }
 
-void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output)
+void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8);
     ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
@@ -47,9 +47,12 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B");
     ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A");
     ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The output matrix must have the same number of columns as the matrix B");
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
 
-    _a_offset = a->info()->quantization_info().offset;
-    _b_offset = b->info()->quantization_info().offset;
+    _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run();
+    _a_offset                    = a->info()->quantization_info().offset;
+    _b_offset                    = b->info()->quantization_info().offset;
 
     // If the input tensor has less than 16 rows, we run a special version of GEMMLowp without reshaping the input tensors
     _is_interleaved_transposed = a->info()->dimension(1) > 16;
@@ -93,7 +96,8 @@ void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor
     if(_a_offset != 0)
     {
         TensorShape shape_vector_sum_col = b->info()->tensor_shape();
-        if(b->info()->num_dimensions() > 1)
+
+        if(shape_vector_sum_col.num_dimensions() > 1)
         {
             shape_vector_sum_col.remove_dimension(1);
         }
@@ -152,8 +156,21 @@ void CLGEMMLowpMatrixMultiplyCore::run()
         // Run reshape matrix A
         CLScheduler::get().enqueue(_mtx_a_reshape_kernel, false);
 
-        // Run reshape matrix B
-        CLScheduler::get().enqueue(_mtx_b_reshape_kernel, false);
+        if(_is_first_run || !_reshape_b_only_on_first_run)
+        {
+            // Run reshape matrix B
+            CLScheduler::get().enqueue(_mtx_b_reshape_kernel, false);
+        }
+    }
+
+    // Note: if _reshape_b_only_on_first_run = true, the reduction kernel can be executed only once
+    if(_is_first_run || !_reshape_b_only_on_first_run)
+    {
+        // Run matrix B reduction kernel only if _a_offset is not equal to 0
+        if(_a_offset != 0)
+        {
+            CLScheduler::get().enqueue(_mtx_b_reduction_kernel, false);
+        }
     }
 
     // Run matrix multiply
@@ -165,14 +182,10 @@ void CLGEMMLowpMatrixMultiplyCore::run()
         CLScheduler::get().enqueue(_mtx_a_reduction_kernel, false);
     }
 
-    // Run matrix B reduction kernel only if _a_offset is not equal to 0
-    if(_a_offset != 0)
-    {
-        CLScheduler::get().enqueue(_mtx_b_reduction_kernel, false);
-    }
-
     // Run offset contribution kernel
     CLScheduler::get().enqueue(_offset_contribution_kernel, true);
 
     _memory_group.release();
+
+    _is_first_run = false;
 }
diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
index 2c6515c1df..a18f48d9a7 100644
--- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
+++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp
@@ -52,10 +52,11 @@ NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemo
 {
 }
 
-void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output)
+void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output);
-    ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info()));
+    ARM_COMPUTE_UNUSED(gemm_info);
+    ARM_COMPUTE_ERROR_THROW_ON(NEGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), output->info(), gemm_info));
 
     _a_offset                         = a->info()->quantization_info().offset;
     _b_offset                         = b->info()->quantization_info().offset;
@@ -198,7 +199,7 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b,
     }
 }
 
-Error NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output)
+Error NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, const GEMMInfo &gemm_info)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
@@ -209,6 +210,9 @@ Error NEGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensor
                                     "The output matrix must have the same number of rows as the matrix A");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG((b)->dimension(0) != (output)->dimension(0),
                                     "The output matrix must have the same number of columns as the matrix B");
+    ARM_COMPUTE_UNUSED(gemm_info);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported");
 
     int32_t a_offset                         = a->quantization_info().offset;
     int32_t b_offset                         = b->quantization_info().offset;