COMPMID-856: CL Depthwise Convolution QASYMM8 support

Change-Id: Ic6097e7cf160e8b829fb521b7b99d9a57d9799d3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118774
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

20 files changed, 601 insertions, 84 deletions

diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h
index e9a1fde71a..e7fa74d00e 100644
--- a/arm_compute/core/CL/CLKernels.h
+++ b/arm_compute/core/CL/CLKernels.h
@@ -53,6 +53,7 @@
 #include "arm_compute/core/CL/kernels/CLDerivativeKernel.h"
 #include "arm_compute/core/CL/kernels/CLDilateKernel.h"
 #include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerKernel.h"
+#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLErodeKernel.h"
 #include "arm_compute/core/CL/kernels/CLFastCornersKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
index 7e786e8df8..86b8b90e94 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -51,7 +51,7 @@ 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: F32
+     *                         while every optional dimension from 4 and above represent a batch of inputs. Data types supported: QASYMM8/F32
      * @param[out] output      The output tensor. First 3 lower dimensions represent a transform of each 3D input,
      *                         while every dimension above 3 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/CLDepthwiseVectorToTensorKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h
index 1dae9b2b5f..36d7cadbec 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,7 +52,7 @@ public:
     CLDepthwiseVectorToTensorKernel &operator=(CLDepthwiseVectorToTensorKernel &&) = default;
     /** Set the input and output of the kernel.
      *
-     * @param[in]  input  The input vector to convert. Data type supported: F32.
+     * @param[in]  input  The input vector to convert. Data type supported: QASYMM8/S32/F16/F32.
      * @param[out] output The output tensor. 3 lower dimensions represent a single input [width, height, IFM]. Data type supported: same as @p input.
      * @param[in]  conv_w The converted tensor's width.
      * @param[in]  conv_h The converted tensor's height.
diff --git a/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h
index 7989257d34..c8d1e2f1ad 100644
--- a/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h
+++ b/arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,7 +50,8 @@ public:
     CLDepthwiseWeightsReshapeKernel &operator=(CLDepthwiseWeightsReshapeKernel &&) = default;
     /** 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]. Data type supported: F32.
+     * @param[in]  input  The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM].
+     *                    Data type supported: QASYMM8/F32.
      * @param[out] output The output tensor. Data type supported: same as @p input.
      * @param[in]  biases (Optional) The input biases to add. Shape [IFM]. Data type supported: same as @p input.
      */
diff --git a/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h
new file mode 100644
index 0000000000..9340e9a8d8
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef __ARM_COMPUTE_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__
+#define __ARM_COMPUTE_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ITensor;
+/** OpenCL kernel to accumulate the biases, if provided, or downscale in case of quantized input.
+ *
+ * @note We assume bias to be shared
+ */
+class CLDirectConvolutionLayerOutputStageKernel : public ICLKernel
+{
+public:
+    /** Default constructor */
+    CLDirectConvolutionLayerOutputStageKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDirectConvolutionLayerOutputStageKernel(const CLDirectConvolutionLayerOutputStageKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    CLDirectConvolutionLayerOutputStageKernel &operator=(const CLDirectConvolutionLayerOutputStageKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    CLDirectConvolutionLayerOutputStageKernel(CLDirectConvolutionLayerOutputStageKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    CLDirectConvolutionLayerOutputStageKernel &operator=(CLDirectConvolutionLayerOutputStageKernel &&) = default;
+    /** Default destructor */
+    ~CLDirectConvolutionLayerOutputStageKernel() = default;
+    /** Set the accumulate buffer and the biases of the kernel.
+     *
+     * @param[in, out] input                        Input to add the bias to. If @p output is not specified then accumulation is done in-place.
+     *                                              Data type supported: S32/QS16/QS32/F16/F32
+     * @param[in]      bias                         (Optional) The shared bias tensor to add. It must be 1D Tensor. Data type supported: Same as @p input
+     * @param[out]     output                       (Optional) If the output tensor is specified the accumulation is done out-of-place. (Defaults to nullptr)
+     *                                              Required parameter if output is of QASYMM8 type.
+     *                                              Data types supported: QS8/QASYMM8/QS16/F16/F32
+     * @param[in]      result_fixedpoint_multiplier (Optional)Fixed point value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]      result_shift                 (Optional)Integer value used to round to nearest division by a power-of-two the result after the fixed point multiplication
+     * @param[in]      result_offset_after_shift    (Optional)Offset to be applied to result before converting it back to QASYMM8
+     */
+    void configure(ICLTensor *input, const ICLTensor *bias = nullptr, ICLTensor *output = nullptr,
+                   int result_fixedpoint_multiplier = 0, int result_shift = 0, int result_offset_after_shift = 0);
+    /** Static function to check if given info will lead to a valid configuration of @ref CLDirectConvolutionLayerOutputStageKernel
+     *
+     * @param[in] input  Input to add the bias to. If @p output is not specified then accumulation is done in-place.
+     *                   Data type supported: QS16/QS32/F16/F32
+     * @param[in] bias   (Optional) The shared bias tensor to add. It must be 1D Tensor. Data type supported: Same as @p input
+     * @param[in] output (Optional) If the output tensor is specified the accumulation is done out-of-place. (Defaults to nullptr)
+     *                         Data type supported: QS8/QS16/F16/F32
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *bias = nullptr, const ITensorInfo *output = nullptr);
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+
+private:
+    ICLTensor       *_input;
+    const ICLTensor *_bias;
+    ICLTensor       *_output;
+    int              _result_fixedpoint_multiplier;
+    int              _result_shift;
+    int              _result_offset_after_shift;
+};
+} // namespace arm_compute
+#endif /*__ARM_COMPUTE_CLDIRECTCONVOLUTIONLAYEROUTPUTSTAGEKERNEL_H__ */
diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h
index 580322fb51..c27307e63b 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,8 +46,8 @@ public:
     CLGEMMMatrixVectorMultiplyKernel &operator=(CLGEMMMatrixVectorMultiplyKernel &&) = default;
     /** Set the input and output of the kernel.
      *
-     * @param[in]  input0 The reshaped input tensor. Data types supported: F16/F32
-     * @param[in]  input1 The 2D reshaped weights tensor. Data type supported: Same as @p input.
+     * @param[in]  input0 The reshaped input tensor. Data types supported: QASYMM8/F16/F32
+     * @param[in]  input1 The 2D reshaped weights tensor. Data type supported: Same as @p input, S32 for QASYMM8 input.
      * @param[out] output The output 2D tensor. Data types supported: Same as @p input
      */
     void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output);
diff --git a/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h b/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h
index 95fe916a3c..286be1acc9 100644
--- a/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h
+++ b/arm_compute/core/NEON/kernels/NEGEMMMatrixVectorMultiplyKernel.h
@@ -51,7 +51,7 @@ public:
      *
      * @param[in]  input0 First Input tensor. Data types supported: QASYMM8/F32
      * @param[in]  input1 Second Input tensor. Data types supported: same as @p input.
-     * @param[out] output Output tensor which stores the interleaved matrix. Data type supported: same as @p input.
+     * @param[out] output Output tensor which stores the interleaved matrix. Data type supported: same as @p input, S32 for QASYMM8 input.
      */
     void configure(const ITensor *input0, const ITensor *input1, ITensor *output);
 
diff --git a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
index f7899415d2..b34967d997 100644
--- a/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -28,6 +28,7 @@
 #include "arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseVectorToTensorKernel.h"
 #include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h"
+#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
 #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.h"
 #include "arm_compute/core/Types.h"
@@ -83,10 +84,10 @@ public:
     CLDepthwiseConvolutionLayer();
     /** Initialize the function's source, destination, weights and convolution information.
      *
-     * @param[in, out] input     Source tensor. Data type supported: F32. (Written to only for border filling).
+     * @param[in, out] input     Source tensor. Data type supported: QASYMM8/F32. (Written to only for border filling).
      * @param[in]      weights   Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in]      biases    (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                           Data type supported: Same as @p input.
+     *                           Data type supported: Same as @p input, S32 when input is QASYMM8.
      * @param[out]     output    Destination tensor. Data type supported: same as @p input.
      * @param[in]      conv_info Padding and stride information to use for the convolution.
      */
@@ -96,15 +97,18 @@ public:
     void run() override;
 
 private:
-    CLDepthwiseIm2ColKernel          _im2col_kernel;
-    CLDepthwiseWeightsReshapeKernel  _weights_reshape_kernel;
-    CLGEMMMatrixVectorMultiplyKernel _v2mm_kernel;
-    CLDepthwiseVectorToTensorKernel  _vector_to_tensor_kernel;
-    CLFillBorderKernel               _v2mm_input_fill_border;
-    CLFillBorderKernel               _v2mm_weights_fill_border;
-    CLTensor                         _input_reshaped;
-    CLTensor                         _weights_reshaped;
-    CLTensor                         _v2mm_output;
+    CLDepthwiseIm2ColKernel                   _im2col_kernel;
+    CLDepthwiseWeightsReshapeKernel           _weights_reshape_kernel;
+    CLGEMMMatrixVectorMultiplyKernel          _v2mm_kernel;
+    CLDepthwiseVectorToTensorKernel           _vector_to_tensor_kernel;
+    CLDirectConvolutionLayerOutputStageKernel _output_stage_kernel;
+    CLFillBorderKernel                        _v2mm_input_fill_border;
+    CLFillBorderKernel                        _v2mm_weights_fill_border;
+    CLTensor                                  _input_reshaped;
+    CLTensor                                  _weights_reshaped;
+    CLTensor                                  _v2mm_output;
+    CLTensor                                  _output_reshaped;
+    bool                                      _is_quantized;
 };
 }
 #endif /*__ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H__ */
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
index 1c65b3ccf5..f5805f5ad1 100644
--- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
+++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h
@@ -103,7 +103,7 @@ public:
      * @param[out]     output    Destination tensor. Data type supported: same as @p input.
      * @param[in]      weights   Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input.
      * @param[in]      biases    (Optional) Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed.
-     *                           Data type supported: Same as @p input.
+     *                           Data type supported: Same as @p input, S32 when input is QASYMM8.
      * @param[in]      conv_info Padding and stride information to use for the convolution.
      */
     void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info);
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 8693a728ba..4e090debc8 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -151,6 +151,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "activation_layer_qa8", "activation_layer_qa8.cl" },
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
+    { "batchnormalization_layer", "batchnormalization_layer.cl" },
     { "bitwise_or", "bitwise_op.cl" },
     { "bitwise_and", "bitwise_op.cl" },
     { "bitwise_xor", "bitwise_op.cl" },
@@ -219,6 +220,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "gemm_ma_qs8", "gemm.cl" },
     { "gemm_ma_qs16", "gemm.cl" },
     { "gemm_mv", "gemv.cl" },
+    { "gemm_mv_quantized", "gemv.cl" },
     { "gemm_mm_interleaved_transposed_f16", "gemm.cl" },
     { "gemm_mm_interleaved_transposed_f32_midgard", "gemm.cl" },
     { "gemm_mm_interleaved_transposed_f32_bifrost", "gemm.cl" },
@@ -284,7 +286,6 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "non_max_suppression", "nonmax.cl" },
     { "normalization_layer_cross_map", "normalization_layer.cl" },
     { "normalization_layer_in_map", "normalization_layer.cl" },
-    { "batchnormalization_layer", "batchnormalization_layer.cl" },
     { "NV12_to_IYUV_bt709", "color_convert.cl" },
     { "NV12_to_RGB888_bt709", "color_convert.cl" },
     { "NV12_to_RGBA8888_bt709", "color_convert.cl" },
@@ -293,6 +294,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "NV21_to_RGB888_bt709", "color_convert.cl" },
     { "NV21_to_RGBA8888_bt709", "color_convert.cl" },
     { "NV21_to_YUV444_bt709", "color_convert.cl" },
+    { "output_stage_quantized", "depthwise_convolution_quantized.cl" },
     { "permute_201", "permute.cl" },
     { "permute_120", "permute.cl" },
     { "permute_3201", "permute.cl" },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution.cl b/src/core/CL/cl_kernels/depthwise_convolution.cl
index ac94b693e3..861788647a 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution.cl
@@ -469,7 +469,7 @@ __kernel void depthwise_weights_reshape(
 }
 #endif //defined(SRC_WIDTH) && defined(DATA_TYPE)
 
-#if 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(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE)
+#if 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(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(DATA_TYPE) && defined(PAD_VALUE)
 /** This kernel performs a reshaping of the input tensor to a tensor used to perform depthwise convolution using vector to matrix multiplication.
  *
  * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
@@ -513,7 +513,7 @@ __kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(d
         {
             if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
             {
-                *output_ptr = 0;
+                *output_ptr = PAD_VALUE;
             }
             else
             {
@@ -526,7 +526,7 @@ __kernel void depthwise_im2col(TENSOR3D_DECLARATION(src), TENSOR3D_DECLARATION(d
 #endif // defined(HAS_BIAS)
 }
 
-#endif //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(SRC_WIDTH) && defined(DATA_TYPE)
+#endif //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(SRC_WIDTH) && defined(DATA_TYPE) && defined(PAD_VALUE)
 
 #if defined(CONV_WIDTH) && defined(CONV_HEIGHT) && defined(DATA_TYPE)
 
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
index 450342ddfc..cd7f4f83d5 100644
--- a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -254,5 +254,64 @@ __kernel void depthwise_convolution_3x3_quantized(
 
     vstore8(pixels, 0, dst.ptr);
 }
-
 #endif //defined(CONV_STRIDE_X)
+
+/** This function computes the output stage of a depthwise convolution.
+ *
+ * @param[in] src_ptr                            Pointer to the source image. Supported data types: QASYMM8
+ * @param[in] src_stride_x                       Stride of the source image 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 image in Y dimension (in bytes)
+ * @param[in] src_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] src_offset_first_element_in_bytes  The offset of the first element in the source image
+ * @param[in] src_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in] src_step_z                         src_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_ptr                            Pointer to the destination tensor. Supported data types: QASYMM8
+ * @param[in] dst_stride_x                       Stride of the destination tensor in X dimension (in bytes)
+ * @param[in] dst_step_x                         dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] dst_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in] dst_step_y                         dst_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_stride_z                       Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in] dst_step_z                         dst_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in] dst_offset_first_element_in_bytes  The offset of the first element in the destination tensor
+ * @param[in] bias_ptr                           (Optional) Pointer to the biases vector. Supported data types: S32
+ * @param[in] bias_stride_x                      (Optional) Stride of the biases vector in X dimension (in bytes)
+ * @param[in] bias_step_x                        (Optional) bias_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] bias_offset_first_element_in_bytes (Optional) The offset of the first element in the biases vector
+ * @param[in] output_offset                      Quantized offset of zero point of the output tensor data range
+ * @param[in] output_multiplier                  Output scale multiplier
+ * @param[in] output_shift                       Output scale divisor exponent
+ */
+
+__kernel void output_stage_quantized(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(bias),
+#endif //defined(HAS_BIAS)
+    int output_offset,
+    int output_multiplier,
+    int output_shift)
+{
+    Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+#if defined(HAS_BIAS)
+    Vector bias = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bias);
+#endif //defined(HAS_BIAS)
+
+    // Load input
+    int16 vals = vload16(0, (__global int *)(src.ptr));
+
+#if defined(HAS_BIAS)
+    // Load and add bias
+    int bias_value = *((__global int *)(vector_offset(&bias, get_global_id(2))));
+    vals += (int16)(bias_value);
+#endif //defined(HAS_BIAS)
+
+    vals = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(vals, output_multiplier, output_shift, 16);
+    vals = vals + output_offset;
+    vals = clamp(vals, 0, 255);
+
+    // Store result in dst
+    vstore16(convert_uchar16(vals), 0, (__global uchar *)dst.ptr);
+}
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/gemv.cl b/src/core/CL/cl_kernels/gemv.cl
index 3e38c735fe..811aa1b865 100644
--- a/src/core/CL/cl_kernels/gemv.cl
+++ b/src/core/CL/cl_kernels/gemv.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,7 @@
  */
 #include "helpers.h"
 
+#if defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT)
 /** This kernel applies dot product to each plane on the input tensor and the corrispective column of the reshaped weight tensor.
  *
  * @note Datatype and source width and height should be given as a preprocessor argument using -DDATA_TYPE=type, -DSRC_WIDTH=width and -DSRC_HEIGHT=height. e.g. -DDATA_TYPE=short
@@ -109,3 +110,91 @@ __kernel void gemm_mv(TENSOR3D_DECLARATION(src), IMAGE_DECLARATION(weights), VEC
         }
     }
 }
+#endif /* defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) */
+
+#if defined(SRC_WIDTH) && defined(SRC_HEIGHT)
+/** This kernel applies dot product to each plane on the input tensor and the corresponding column of the reshaped weight tensor.
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: QASYMM8
+ * @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)
+ * @param[in]  src_step_y                            src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_stride_z                          Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                            src_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[in]  weights_ptr                           Pointer to the weights tensor. Same as @p src_ptr
+ * @param[in]  weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+ * @param[in]  weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+ * @param[in]  weights_step_y                        weights_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Same as @p src_ptr
+ * @param[in]  dst_stride_x                          Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  dst_step_x                            dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[in]  input_offset                          Input's quantization offset
+ * @param[in]  weights_offset                        Weights's quantization offset
+ */
+__kernel void gemm_mv_quantized(TENSOR3D_DECLARATION(src),
+                                IMAGE_DECLARATION(weights),
+                                VECTOR_DECLARATION(dst),
+                                const int input_offset,
+                                const int weights_offset)
+{
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+
+    int y = get_global_id(1) * 4;
+    int z = get_global_id(2);
+
+    __global uchar *current_weights = weights_ptr + weights_offset_first_element_in_bytes + z * weights_stride_y;
+    __global uchar *input_ptr       = src.ptr;
+
+    int acc0 = 0;
+    int acc1 = 0;
+    int acc2 = 0;
+    int acc3 = 0;
+
+    // This kernel handle 4 rows in per thread so that it can reuse the weights
+    for(int i = 0; i < SRC_WIDTH; i += 4)
+    {
+        int4 w = convert_int4(vload4(0, (__global uchar *)(current_weights + i * weights_stride_x))) + (int4)weights_offset;
+
+        int4 offset = (int4)i * (int4)src_stride_x + (int4)(0, 1, 2, 3) * (int4)src_stride_y;
+
+        int4 tmp0 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s0))) + (int4)input_offset;
+        int4 tmp1 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s1))) + (int4)input_offset;
+        int4 tmp2 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s2))) + (int4)input_offset;
+        int4 tmp3 = convert_int4(vload4(0, (__global uchar *)(input_ptr + offset.s3))) + (int4)input_offset;
+
+        // Accumulate
+        acc0 += tmp0.s0 * w.s0 + tmp0.s1 * w.s1 + tmp0.s2 * w.s2 + tmp0.s3 * w.s3;
+        acc1 += tmp1.s0 * w.s0 + tmp1.s1 * w.s1 + tmp1.s2 * w.s2 + tmp1.s3 * w.s3;
+        acc2 += tmp2.s0 * w.s0 + tmp2.s1 * w.s1 + tmp2.s2 * w.s2 + tmp2.s3 * w.s3;
+        acc3 += tmp3.s0 * w.s0 + tmp3.s1 * w.s1 + tmp3.s2 * w.s2 + tmp3.s3 * w.s3;
+    }
+
+    __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes + (y + z * SRC_HEIGHT) * dst_stride_x;
+
+    int rows_left = SRC_HEIGHT - (y + 4);
+
+    // This if check is used to handle the last few rows when it can't be divided by the four
+    if(rows_left >= 0)
+    {
+        vstore4((int4)(acc0, acc1, acc2, acc3), 0, (__global int *)output_ptr);
+    }
+    else
+    {
+        switch(rows_left)
+        {
+            case -1: // three rows left; one is padding
+                *((__global int *)(output_ptr + 2 * dst_stride_x)) = acc2;
+            case -2: // two rows left; two are padding
+                *((__global int *)(output_ptr + 1 * dst_stride_x)) = acc1;
+            case -3: // one row left; three are padding
+                *((__global int *)(output_ptr + 0 * dst_stride_x)) = acc0;
+                break;
+        }
+    }
+}
+#endif /* defined(SRC_WIDTH) && defined(SRC_HEIGHT) */
diff --git a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
index ad9ac0ecd6..8467b39910 100644
--- a/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseIm2ColKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,9 +44,10 @@ CLDepthwiseIm2ColKernel::CLDepthwiseIm2ColKernel()
 
 void CLDepthwiseIm2ColKernel::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::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && has_bias);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2));
     ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (kernel_dims.width * kernel_dims.height + ((has_bias) ? 1 : 0)));
 
@@ -54,24 +55,25 @@ void CLDepthwiseIm2ColKernel::configure(const ICLTensor *input, ICLTensor *outpu
     _output = output;
 
     // Create kernel
-    std::set<std::string> build_opts;
-
-    build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.emplace("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
-    build_opts.emplace("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
-    build_opts.emplace("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
-    build_opts.emplace("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
-    build_opts.emplace("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
-    build_opts.emplace("-DPAD_BOTTOM=" + support::cpp11::to_string(conv_info.pad_bottom()));
-    build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input->info()->dimension(0)));
-    build_opts.emplace("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1)));
-    build_opts.emplace("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
-    build_opts.emplace("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
-    if(has_bias)
-    {
-        build_opts.emplace("-DHAS_BIAS");
-    }
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts));
+    CLBuildOptions build_opts;
+
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
+    build_opts.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
+    build_opts.add_option("-DPAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
+    build_opts.add_option("-DPAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
+    build_opts.add_option("-DPAD_RIGHT=" + support::cpp11::to_string(conv_info.pad_right()));
+    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("-DKERNEL_WIDTH=" + support::cpp11::to_string(kernel_dims.width));
+    build_opts.add_option("-DKERNEL_HEIGHT=" + support::cpp11::to_string(kernel_dims.height));
+    build_opts.add_option_if(has_bias, "-DHAS_BIAS");
+    build_opts.add_option_if_else(is_data_type_quantized(input->info()->data_type()),
+                                  "-DPAD_VALUE=" + support::cpp11::to_string(input->info()->quantization_info().offset),
+                                  "-DPAD_VALUE=0");
+
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_im2col", build_opts.options()));
 
     // Configure the local work size for Bifrost with a value obtained
     // via exhaustive autotuning for the MobileNets tensor shapes.
diff --git a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
index dc47bb0adc..ae35bf64aa 100644
--- a/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseVectorToTensorKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,7 +41,7 @@ CLDepthwiseVectorToTensorKernel::CLDepthwiseVectorToTensorKernel()
 
 void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTensor *output, size_t conv_w, size_t conv_h)
 {
-    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::QASYMM8, DataType::S32, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
@@ -60,12 +60,12 @@ void CLDepthwiseVectorToTensorKernel::configure(const ICLTensor *input, ICLTenso
     _output = output;
 
     // Create kernel
-    std::set<std::string> build_opts;
-    build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
-    build_opts.emplace("-DCONV_WIDTH=" + support::cpp11::to_string(conv_w));
-    build_opts.emplace("-DCONV_HEIGHT=" + support::cpp11::to_string(conv_h));
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DCONV_WIDTH=" + support::cpp11::to_string(conv_w));
+    build_opts.add_option("-DCONV_HEIGHT=" + support::cpp11::to_string(conv_h));
 
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_vector_to_tensor", build_opts));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_vector_to_tensor", build_opts.options()));
 
     // Configure  kernel window
     Window win = calculate_max_window(*input->info(), Steps());
diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
index 81dd6b42cc..26da96f9ba 100644
--- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,9 +41,10 @@ CLDepthwiseWeightsReshapeKernel::CLDepthwiseWeightsReshapeKernel()
 
 void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
 {
-    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::QASYMM8, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input->info()->data_type()) && (biases != nullptr));
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(1));
     ARM_COMPUTE_ERROR_ON(output->info()->dimension(0) != (input->info()->dimension(0) * input->info()->dimension(1) + ((biases != nullptr) ? 1 : 0)));
 
diff --git a/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
new file mode 100644
index 0000000000..cbc281b6ac
--- /dev/null
+++ b/src/core/CL/kernels/CLDirectConvolutionOutputStageKernel.cpp
@@ -0,0 +1,217 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/CL/kernels/CLDirectConvolutionLayerOutputStageKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/Window.h"
+
+#include <cstddef>
+#include <cstdint>
+
+using namespace arm_compute;
+
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::S32, DataType::F16,
+                                                         DataType::F32);
+
+    if(bias != nullptr)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32, DataType::F16, DataType::F32);
+
+        if(is_data_type_quantized_asymmetric(input->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
+        }
+
+        ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_data_type_float(input->data_type()),
+                                        "Calling output stage kernel with floating point arguments");
+    }
+
+    // Checks performed on output
+    if(input->data_type() == DataType::S32)
+    {
+        // Quantized configuration checks
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8);
+    }
+    else
+    {
+        // In case of out-of-place computation (supported for non-quantized configurations)
+        if((output != nullptr) && (output->total_size() != 0))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        }
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
+{
+    bool         window_changed                    = false;
+    unsigned int num_elems_processed_per_iteration = 16 / element_size_from_data_type(input->data_type());
+
+    // Update processed elements when input is S32 (comes from quantization input)
+    if(input->data_type() == DataType::S32)
+    {
+        num_elems_processed_per_iteration = 16;
+    }
+
+    // Configure kernel window
+    Window                 win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+
+    if(output != nullptr && (output->total_size() != 0))
+    {
+        AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+
+        if(bias == nullptr)
+        {
+            window_changed = update_window_and_padding(win, input_access, output_access);
+        }
+        else
+        {
+            AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
+            window_changed = update_window_and_padding(win, input_access, output_access, bias_access);
+        }
+
+        output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
+    }
+    else
+    {
+        if(bias == nullptr)
+        {
+            window_changed = update_window_and_padding(win, input_access);
+        }
+        else
+        {
+            AccessWindowStatic bias_access(bias, 0, 0, bias->dimension(0), bias->dimension(1));
+            window_changed = update_window_and_padding(win, input_access, bias_access);
+        }
+
+        input_access.set_valid_region(win, ValidRegion(Coordinates(), input->tensor_shape()));
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLDirectConvolutionLayerOutputStageKernel::CLDirectConvolutionLayerOutputStageKernel()
+    : _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0)
+{
+}
+
+void CLDirectConvolutionLayerOutputStageKernel::configure(ICLTensor *input, const ICLTensor *bias, ICLTensor *output,
+                                                          int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    // Auto-initialize output output if required
+    if(output != nullptr)
+    {
+        // Work out expected output data type
+        const DataType output_dt = (input->info()->data_type() == DataType::S32) ? DataType::QASYMM8 : input->info()->data_type();
+        // Output tensor auto initialization if not yet initialized
+        auto_init_if_empty(*output->info(), input->info()->clone()->set_data_type(output_dt));
+    }
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info()));
+
+    _bias                         = bias;
+    _input                        = input;
+    _output                       = output;
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _result_offset_after_shift    = result_offset_after_shift;
+
+    // Create kernel
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(bias != nullptr, "-DHAS_BIAS");
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("output_stage_quantized", build_opts.options()));
+
+    // Set static kernel arguments
+    int idx = 2 * num_arguments_per_3D_tensor() + ((bias != nullptr) ? num_arguments_per_1D_tensor() : 0);
+    _kernel.setArg<int>(idx++, _result_offset_after_shift);
+    _kernel.setArg<int>(idx++, _result_fixedpoint_multiplier);
+    _kernel.setArg<int>(idx++, _result_shift);
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure(win_config.second);
+}
+
+Status CLDirectConvolutionLayerOutputStageKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), bias->clone().get(), output == nullptr ? nullptr : output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLDirectConvolutionLayerOutputStageKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
+
+    Window slice = window.first_slice_window_3D();
+
+    // Set bias vector
+    if(_bias != nullptr)
+    {
+        unsigned int idx1 = 2 * num_arguments_per_3D_tensor();
+        Window       slice_biases;
+        slice_biases.use_tensor_dimensions(_bias->info()->tensor_shape());
+        add_1D_tensor_argument(idx1, _bias, slice_biases);
+    }
+
+    // Run kernel
+    do
+    {
+        // Set arguments
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice);
+        enqueue(queue, *this, slice, _lws_hint);
+    }
+    while(window.slide_window_slice_3D(slice));
+}
diff --git a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
index 951bc144aa..cc483dc44e 100644
--- a/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMMatrixVectorMultiplyKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,23 +45,35 @@ BorderSize CLGEMMMatrixVectorMultiplyKernel::border_size() const
 
 void CLGEMMMatrixVectorMultiplyKernel::configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
+    ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(input0->info()->data_type()) && (output->info()->data_type() != DataType::S32));
     ARM_COMPUTE_ERROR_ON(input0->info()->dimension(2) != input1->info()->dimension(1));
 
     _input0 = input0;
     _input1 = input1;
     _output = output;
 
+    // Check if is a quantized operation
+    bool is_quantized = is_data_type_quantized_asymmetric(_input0->info()->data_type());
+
     // Create kernel
-    std::set<std::string> build_opts;
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(!is_quantized, "-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
+    build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input0->info()->dimension(0)));
+    build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input0->info()->dimension(1)));
 
-    build_opts.emplace("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type()));
-    build_opts.emplace("-DSRC_WIDTH=" + support::cpp11::to_string(input0->info()->dimension(0)));
-    build_opts.emplace("-DSRC_HEIGHT=" + support::cpp11::to_string(input0->info()->dimension(1)));
+    std::string kernel_name = is_quantized ? std::string("gemm_mv_quantized") : std::string("gemm_mv");
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemm_mv", build_opts));
+    // Add static arguments
+    if(is_quantized)
+    {
+        unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor() + num_arguments_per_1D_tensor();
+        _kernel.setArg<int>(idx++, -_input0->info()->quantization_info().offset);
+        _kernel.setArg<int>(idx++, -_input1->info()->quantization_info().offset);
+    }
 
     // Configure the local work size for Bifrost with a value obtained
     // via exhaustive autotuning for the MobileNets tensor shapes.
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
index 02273fe08b..fcf3969515 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,10 +25,12 @@
 
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "support/ToolchainSupport.h"
 
 using namespace arm_compute;
+using namespace arm_compute::misc;
 
 CLDepthwiseConvolutionLayer3x3::CLDepthwiseConvolutionLayer3x3()
     : _kernel(), _border_handler()
@@ -59,14 +61,14 @@ void CLDepthwiseConvolutionLayer3x3::run()
 }
 
 CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer()
-    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(), _weights_reshaped(),
-      _v2mm_output()
+    : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(),
+      _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_quantized(false)
 {
 }
 
 void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != weights->info()->dimension(2));
 
@@ -74,15 +76,17 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     const size_t weights_h = weights->info()->dimension(1);
     const size_t weights_z = weights->info()->dimension(2);
 
-    const bool      has_bias   = (biases != nullptr);
-    const GPUTarget gpu_target = CLScheduler::get().target();
+    _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+
+    bool            append_bias = (biases != nullptr) && !_is_quantized;
+    const GPUTarget gpu_target  = CLScheduler::get().target();
 
     unsigned int conv_w = 0;
     unsigned int conv_h = 0;
     std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), weights_w, weights_h, conv_info);
 
     // Set up intermediate tensors
-    const size_t patch_size = weights_w * weights_h + ((has_bias) ? 1 : 0);
+    const size_t patch_size = weights_w * weights_h + ((append_bias) ? 1 : 0);
     const size_t conv_size  = conv_w * conv_h;
 
     // Im2Col configuration
@@ -90,33 +94,49 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w
     shape_im2col.set(0, patch_size);
     shape_im2col.set(1, conv_size);
     shape_im2col.set(2, weights_z);
-    const TensorInfo info_im2col(shape_im2col, 1, input->info()->data_type(), input->info()->fixed_point_position());
-    _input_reshaped.allocator()->init(info_im2col);
+    _input_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
     _im2col_kernel.set_target(gpu_target);
-    _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, has_bias);
+    _im2col_kernel.configure(input, &_input_reshaped, Size2D(weights_w, weights_h), conv_info, append_bias);
 
     // Weights reshape configuration
     const TensorShape shape_weights_reshape(patch_size, weights_z);
-    const TensorInfo  info_weights_reshape(shape_weights_reshape, 1, weights->info()->data_type(), weights->info()->fixed_point_position());
-    _weights_reshaped.allocator()->init(info_weights_reshape);
-    _weights_reshape_kernel.configure(weights, &_weights_reshaped, biases);
+    _weights_reshaped.allocator()->init(weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_weights_reshape));
+    _weights_reshape_kernel.configure(weights, &_weights_reshaped, append_bias ? biases : nullptr);
 
     // GEMV configuration
+    DataType    v2mm_dt        = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
     TensorShape shape_v2mm_out = input->info()->tensor_shape();
     shape_v2mm_out.set(0, conv_size * weights_z);
     shape_v2mm_out.set(1, 1);
     shape_v2mm_out.set(2, 1);
-    const TensorInfo info_v2mm_out(shape_v2mm_out, 1, input->info()->data_type(), input->info()->fixed_point_position());
-    _v2mm_output.allocator()->init(info_v2mm_out);
+    _v2mm_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_data_type(v2mm_dt).set_tensor_shape(shape_v2mm_out));
     _v2mm_kernel.set_target(gpu_target);
     _v2mm_kernel.configure(&_input_reshaped, &_weights_reshaped, &_v2mm_output);
-    _vector_to_tensor_kernel.configure(&_v2mm_output, output, conv_w, conv_h);
+    _vector_to_tensor_kernel.configure(&_v2mm_output, (_is_quantized) ? &_output_reshaped : output, conv_w, conv_h);
+
+    // Output staged configuration
+    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);
+        _output_stage_kernel.configure(&_output_reshaped, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+        _output_reshaped.allocator()->allocate();
+    }
 
+    // Fill borders on inputs
+    PixelValue zero_in(static_cast<int32_t>(0));
+    PixelValue zero_w(static_cast<int32_t>(0));
+    if(_is_quantized)
+    {
+        zero_in = PixelValue(static_cast<int32_t>(input->info()->quantization_info().offset));
+        zero_w  = PixelValue(static_cast<int32_t>(weights->info()->quantization_info().offset));
+    }
     BorderSize border_size = _v2mm_kernel.border_size();
-    _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0));
+    _v2mm_input_fill_border.configure(&_input_reshaped, border_size, BorderMode::CONSTANT, zero_in);
 
     border_size.bottom = 0;
-    _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, PixelValue(0));
+    _v2mm_weights_fill_border.configure(&_weights_reshaped, border_size, BorderMode::CONSTANT, zero_w);
 
     // Allocate intermediate tensors
     _input_reshaped.allocator()->allocate();
@@ -135,4 +155,9 @@ void CLDepthwiseConvolutionLayer::run()
     CLScheduler::get().enqueue(_v2mm_kernel);
 
     CLScheduler::get().enqueue(_vector_to_tensor_kernel);
+
+    if(_is_quantized)
+    {
+        CLScheduler::get().enqueue(_output_stage_kernel);
+    }
 }
diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
index 43e04fbf07..d747d43b30 100644
--- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp
@@ -89,10 +89,26 @@ TEST_SUITE_END()
 TEST_SUITE_END()
 
 template <typename T>
+using CLDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T>;
+template <typename T>
 using CLDepthwiseConvolutionLayerQuantizedFixture3x3 = DepthwiseConvolutionLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer3x3, T>;
 
 TEST_SUITE(Quantized)
 TEST_SUITE(QASYMM8)
+TEST_SUITE(Generic)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+                       framework::dataset::make("DataType", DataType::QASYMM8)),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(),
+                       framework::dataset::make("DataType", DataType::QASYMM8)),
+                       framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) })))
+{
+    validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END()
 TEST_SUITE(W3x3)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(),
                        framework::dataset::make("DataType", DataType::QASYMM8)),