COMPMID-791: Adds support of QASYMM8 in NEDepthwiseConvolution3x3

Change-Id: I1a9ed6c3420ddf8978aeaad48d9915333b006b49
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116374
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 288 insertions, 176 deletions

diff --git a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
index 40a8601aaa..bc2f1ed266 100644
--- a/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
+++ b/src/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.cpp
@@ -42,72 +42,18 @@ using namespace arm_compute;
 using namespace arm_compute::detail;
 using namespace arm_compute::misc::shape_calculator;
 
-NEDepthwiseConvolutionLayer3x3Kernel::NEDepthwiseConvolutionLayer3x3Kernel()
-    : _border_size(0), _input(), _output(), _weights(), _conv_info()
-{
-}
-
-BorderSize NEDepthwiseConvolutionLayer3x3Kernel::border_size() const
+namespace
 {
-    return _border_size;
-}
-
-void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
-{
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
-    ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
-
-    // Get convolved dimensions
-    const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info);
-
-    // 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());
-
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
-
-    _input                           = input;
-    _output                          = output;
-    _weights                         = weights;
-    _conv_info                       = conv_info;
-    const unsigned int conv_stride_x = conv_info.stride().first;
-    const unsigned int conv_stride_y = conv_info.stride().second;
-    const unsigned int conv_pad_left = conv_info.pad_left();
-    const unsigned int conv_pad_top  = conv_info.pad_top();
-
-    ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 3);
-
-    const unsigned int num_elems_written_per_iteration = 16 >> conv_stride_x;
-    _border_size                                       = BorderSize(conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), conv_pad_left);
-
-    // Configure kernel window
-    Window win = calculate_max_window(*output->info(), Steps(num_elems_written_per_iteration));
-
-    const unsigned int num_x_steps               = (output_shape.x() + num_elems_written_per_iteration - 1) / num_elems_written_per_iteration;
-    const int          input_num_elems_processed = get_input_num_elems_processed(num_elems_written_per_iteration, conv_stride_x);
-
-    AccessWindowStatic input_access(input->info(), -conv_pad_left, -conv_pad_top, (num_x_steps - 1) * input_num_elems_processed + 12, conv_stride_y * (output_shape.y() - 1) + 2);
-    AccessWindowStatic weights_access(weights->info(), 0, 0, weights->info()->dimension(0), weights->info()->dimension(1));
-    AccessWindowStatic output_access(output->info(), 0, 0, num_x_steps * num_elems_written_per_iteration, output_shape.y());
-
-    update_window_and_padding(win, input_access, weights_access, output_access);
-    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
-
-    INEKernel::configure(win);
-}
-
-template <unsigned int stridex>
+template <typename T1, typename T2, unsigned int stridex>
 class convolver_3x3
 {
 public:
     static void convolve(const Window &window, unsigned int num_elems_written_per_iteration,
                          const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
     {
+        const int input_offset   = -input->info()->quantization_info().offset;
+        const int weights_offset = -weights->info()->quantization_info().offset;
+
         const int          input_stride_x  = input->info()->strides_in_bytes().x();
         const int          input_stride_y  = input->info()->strides_in_bytes().y();
         const int          output_stride_y = output->info()->strides_in_bytes().y();
@@ -117,8 +63,8 @@ public:
         const int          output_h        = output->info()->dimension(1);
         const int          delta_input     = get_input_num_elems_processed<stridex>(num_elems_written_per_iteration);
         const unsigned int conv_stride_y   = std::get<1>(conv_info.stride());
-        const unsigned int conv_pad_x      = std::get<0>(conv_info.pad());
-        const unsigned int conv_pad_y      = std::get<1>(conv_info.pad());
+        const unsigned int conv_pad_x      = conv_info.pad_left();
+        const unsigned int conv_pad_y      = conv_info.pad_top();
 
         // setup output window for the iterator
         Window window_out = window;
@@ -141,29 +87,31 @@ public:
 
         execute_window_loop(window_out, [&](const Coordinates & id)
         {
-            const uint8_t *input_ptr = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y;
-            int            ih        = 0;
-            int            oh        = 0;
-
-            const uint8_t      *ptr_weights_base = weights_ptr + id.z() * kernel_stride_z;
-            const auto          ptr_weights_r0   = reinterpret_cast<const float *>(ptr_weights_base);
-            const auto          ptr_weights_r1   = reinterpret_cast<const float *>(ptr_weights_base + kernel_stride_y);
-            const auto          ptr_weights_r2   = reinterpret_cast<const float *>(ptr_weights_base + kernel_stride_y * 2);
-            const float32x4x3_t vw_r0            = load_matrix_row(ptr_weights_r0);
-            const float32x4x3_t vw_r1            = load_matrix_row(ptr_weights_r1);
-            const float32x4x3_t vw_r2            = load_matrix_row(ptr_weights_r2);
+            int ih = 0;
+            int oh = 0;
+
+            const uint8_t *input_ptr        = in.ptr() - conv_pad_x * input_stride_x - conv_pad_y * input_stride_y;
+            const uint8_t *ptr_weights_base = weights_ptr + id.z() * kernel_stride_z;
+
+            const auto ptr_weights_r0 = reinterpret_cast<const T1 *>(ptr_weights_base);
+            const auto ptr_weights_r1 = reinterpret_cast<const T1 *>(ptr_weights_base + kernel_stride_y);
+            const auto ptr_weights_r2 = reinterpret_cast<const T1 *>(ptr_weights_base + kernel_stride_y * 2);
+            const auto vw_r0          = load_matrix_row(ptr_weights_r0, weights_offset);
+            const auto vw_r1          = load_matrix_row(ptr_weights_r1, weights_offset);
+            const auto vw_r2          = load_matrix_row(ptr_weights_r2, weights_offset);
 
             for(ih = 0, oh = 0; oh < output_h; ++oh, ih += conv_stride_y)
             {
-                auto in_top = reinterpret_cast<const float *>(input_ptr + (ih + 0) * input_stride_y);
-                auto in_mid = reinterpret_cast<const float *>(input_ptr + (ih + 1) * input_stride_y);
-                auto in_low = reinterpret_cast<const float *>(input_ptr + (ih + 2) * input_stride_y);
-                auto p_out  = reinterpret_cast<float *>(out.ptr() + oh * output_stride_y);
+                auto in_top = reinterpret_cast<const T1 *>(input_ptr + (ih + 0) * input_stride_y);
+                auto in_mid = reinterpret_cast<const T1 *>(input_ptr + (ih + 1) * input_stride_y);
+                auto in_low = reinterpret_cast<const T1 *>(input_ptr + (ih + 2) * input_stride_y);
+                auto p_out  = reinterpret_cast<T2 *>(out.ptr() + oh * output_stride_y);
 
                 for(int ow = 0; ow < output_w; ow += num_elems_written_per_iteration,
-                    in_top += delta_input, in_mid += delta_input, in_low += delta_input, p_out += num_elems_written_per_iteration)
+                    in_top += delta_input, in_mid += delta_input, in_low += delta_input,
+                    p_out += num_elems_written_per_iteration)
                 {
-                    auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, 0);
+                    auto vres = convolve_3x3<stridex>(in_top, in_mid, in_low, vw_r0, vw_r1, vw_r2, 0, input_offset);
                     store_results<stridex>(p_out, vres);
                 }
             }
@@ -172,24 +120,113 @@ public:
     }
 };
 
-void NEDepthwiseConvolutionLayer3x3Kernel::run(const Window &window, const ThreadInfo &info)
+template <typename T1, typename T2>
+inline void convolve_3x3(const Window &window, unsigned int num_elems_written_per_iteration,
+                         const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
-    ARM_COMPUTE_UNUSED(info);
-
-    const unsigned int conv_stride_x                   = _conv_info.stride().first;
-    const unsigned int num_elems_written_per_iteration = 16 >> conv_stride_x;
-
+    const unsigned int conv_stride_x = std::get<0>(conv_info.stride());
     switch(conv_stride_x)
     {
         case 1:
-            convolver_3x3<1>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            convolver_3x3<T1, T2, 1>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info);
             break;
         case 2:
-            convolver_3x3<2>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            convolver_3x3<T1, T2, 2>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info);
             break;
         case 3:
-            convolver_3x3<3>::convolve(window, num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            convolver_3x3<T1, T2, 3>::convolve(window, num_elems_written_per_iteration, input, weights, output, conv_info);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Not implemented");
+    }
+}
+} // namespace
+
+NEDepthwiseConvolutionLayer3x3Kernel::NEDepthwiseConvolutionLayer3x3Kernel()
+    : _border_size(0), _input(), _output(), _weights(), _conv_info(), _num_elems_written_per_iteration(0)
+{
+}
+
+BorderSize NEDepthwiseConvolutionLayer3x3Kernel::border_size() const
+{
+    return _border_size;
+}
+
+void NEDepthwiseConvolutionLayer3x3Kernel::configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info)
+{
+    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(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
+
+    // Get convolved dimensions
+    const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info);
+    const DataType    output_dt    = (input->info()->data_type() == DataType::QASYMM8) ? DataType::S32 : input->info()->data_type();
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(),
+                       input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(output_shape).set_data_type(output_dt));
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
+
+    _input                           = input;
+    _output                          = output;
+    _weights                         = weights;
+    _conv_info                       = conv_info;
+    const unsigned int conv_stride_x = conv_info.stride().first;
+    const unsigned int conv_stride_y = conv_info.stride().second;
+    const unsigned int conv_pad_left = conv_info.pad_left();
+    const unsigned int conv_pad_top  = conv_info.pad_top();
+
+    ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 3);
+
+    unsigned int num_elems_read_per_iteration = 0;
+    switch(input->info()->data_type())
+    {
+        case DataType::QASYMM8:
+            num_elems_read_per_iteration     = 16;
+            _num_elems_written_per_iteration = 16 >> conv_stride_x;
+            break;
+        case DataType::F32:
+            num_elems_read_per_iteration     = 12;
+            _num_elems_written_per_iteration = 16 >> conv_stride_x;
+            break;
+        default:
+            ARM_COMPUTE_ERROR("Data type not supported.");
+    }
+    _border_size = BorderSize(conv_pad_top, conv_info.pad_right(), conv_info.pad_bottom(), conv_pad_left);
+
+    // Configure kernel window
+    Window win = calculate_max_window(*output->info(), Steps(_num_elems_written_per_iteration));
+
+    const unsigned int num_x_steps               = (output_shape.x() + _num_elems_written_per_iteration - 1) / _num_elems_written_per_iteration;
+    const int          input_num_elems_processed = get_input_num_elems_processed(_num_elems_written_per_iteration, conv_stride_x);
+
+    AccessWindowStatic input_access(input->info(),
+                                    -conv_pad_left,
+                                    -conv_pad_top,
+                                    (num_x_steps - 1) * input_num_elems_processed + num_elems_read_per_iteration,
+                                    conv_stride_y * (output_shape.y() - 1) + 2);
+    AccessWindowStatic weights_access(weights->info(), 0, 0, weights->info()->dimension(0), weights->info()->dimension(1));
+    AccessWindowStatic output_access(output->info(), 0, 0, num_x_steps * _num_elems_written_per_iteration, output_shape.y());
+
+    update_window_and_padding(win, input_access, weights_access, output_access);
+    output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
+
+    INEKernel::configure(win);
+}
+
+void NEDepthwiseConvolutionLayer3x3Kernel::run(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_UNUSED(info);
+
+    switch(_input->info()->data_type())
+    {
+        case DataType::F32:
+            convolve_3x3<float, float>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
+            break;
+        case DataType::QASYMM8:
+            convolve_3x3<uint8_t, int32_t>(window, _num_elems_written_per_iteration, _input, _weights, _output, _conv_info);
             break;
         default:
             ARM_COMPUTE_ERROR("Not implemented");
diff --git a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
index 40abdb1672..52880a378f 100644
--- a/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
+++ b/src/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -27,6 +27,7 @@
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
+#include "arm_compute/core/NEON/NEAsymm.h"
 #include "arm_compute/core/NEON/NEFixedPoint.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
@@ -43,24 +44,26 @@ 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::QS8, DataType::QS16, DataType::F16, DataType::QS32, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8,
+                                                         DataType::QS16, DataType::F16,
+                                                         DataType::QS32, DataType::S32, DataType::F32);
 
     if(bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::QS32, DataType::F32);
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::QS32, DataType::S32, DataType::F32);
 
-        if(is_data_type_quantized(input->data_type()))
+        if(is_data_type_fixed_point(input->data_type()))
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS8 && bias->data_type() != DataType::QS8, "Wrong data type for bias");
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS16 && bias->data_type() != DataType::QS8, "Wrong data type for bias");
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS32 && bias->data_type() != DataType::QS16, "Wrong data type for bias");
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, bias);
         }
         else
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
         }
 
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, bias);
         ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
     }
     else
@@ -71,18 +74,22 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
     // Checks performed when output is configured
     if((output != nullptr) && (output->total_size() != 0))
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::QS16, DataType::F32);
-        if(is_data_type_quantized(input->data_type()))
+        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F32);
+        if(is_data_type_fixed_point(input->data_type()))
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS8 && output->data_type() != DataType::QS8, "Wrong data type for output");
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS16 && output->data_type() != DataType::QS8, "Wrong data type for output");
             ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::QS32 && output->data_type() != DataType::QS16, "Wrong data type for output");
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output);
+        }
+        else if(is_data_type_quantized_asymmetric(output->data_type()))
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_type() == DataType::S32 && output->data_type() != DataType::QASYMM8, "Wrong data type for bias");
         }
         else
         {
             ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
         }
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output);
     }
 
     return Status{};
@@ -90,8 +97,14 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, con
 
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *bias, ITensorInfo *output)
 {
-    bool               window_changed                    = false;
-    const unsigned int num_elems_processed_per_iteration = 16 / element_size_from_data_type(input->data_type());
+    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));
@@ -145,7 +158,6 @@ inline qint16x8_t internal_vld1q(const qint16_t *in)
 {
     return vld1q_qs16(in);
 }
-
 inline qint32x4_t internal_vld1q(const qint32_t *in)
 {
     return vld1q_s32(in);
@@ -168,7 +180,6 @@ inline void internal_vst1q(qint16_t *p, const qint16x8_t &v)
 {
     vst1q_qs16(p, v);
 }
-
 inline void internal_vst1q(qint32_t *p, const qint32x4_t &v)
 {
     vst1q_s32(p, v);
@@ -192,7 +203,6 @@ inline qint16x8_t internal_vdupq_n(qint16_t v)
 {
     return vdupq_n_qs16(v);
 }
-
 inline qint32x4_t internal_vdupq_n(qint32_t v)
 {
     return vdupq_n_qs32(v);
@@ -236,8 +246,13 @@ inline float16x8_t internal_vqaddq(const float16x8_t &x, const float16x8_t &y)
 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
 
 template <typename T1, typename T2, bool in_place, bool has_bias>
-void output_stage(ITensor *input, const ITensor *bias, const Window window, ITensor *output)
+void output_stage(ITensor *input, const ITensor *bias, const Window &window, ITensor *output,
+                  int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
 {
+    ARM_COMPUTE_UNUSED(result_fixedpoint_multiplier);
+    ARM_COMPUTE_UNUSED(result_shift);
+    ARM_COMPUTE_UNUSED(result_offset_after_shift);
+
     Iterator in(input, window);
 
     if(in_place) // In place accumulate
@@ -283,31 +298,112 @@ void output_stage(ITensor *input, const ITensor *bias, const Window window, ITen
         in, out);
     }
 }
+
+// QASYMM8 specializations
+template <>
+void output_stage<int32_t, uint8_t, false, true>(ITensor *input, const ITensor *bias, const Window &window, ITensor *output,
+                                                 int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
+{
+    const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift);
+    uint8x16_t      min                           = vdupq_n_u8(0);
+    uint8x16_t      max                           = vdupq_n_u8(255);
+
+    Iterator in(input, window);
+    Iterator out(output, window);
+
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        // Get bias and pointer to input
+        const auto  in_ptr = reinterpret_cast<int32_t *>(in.ptr());
+        int32x4x4_t v_in =
+        {
+            {
+                vld1q_s32(in_ptr),
+                vld1q_s32(in_ptr + 4),
+                vld1q_s32(in_ptr + 8),
+                vld1q_s32(in_ptr + 12)
+            }
+        };
+
+        // Accumulate bias
+        const auto vb = vdupq_n_s32(*reinterpret_cast<const int32_t *>(bias->ptr_to_element(Coordinates(id.z()))));
+        v_in =
+        {
+            {
+                vaddq_s32(v_in.val[0], vb),
+                vaddq_s32(v_in.val[1], vb),
+                vaddq_s32(v_in.val[2], vb),
+                vaddq_s32(v_in.val[3], vb)
+            }
+        };
+
+        const auto out_ptr = reinterpret_cast<uint8_t *>(out.ptr());
+        vst1q_u8(out_ptr, finalize_quantization<false>(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, min, max));
+    },
+    in, out);
+}
+template <>
+void output_stage<int32_t, uint8_t, false, false>(ITensor *input, const ITensor *bias, const Window &window, ITensor *output,
+                                                  int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift)
+{
+    ARM_COMPUTE_UNUSED(bias);
+
+    const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(result_offset_after_shift);
+    uint8x16_t      min                           = vdupq_n_u8(0);
+    uint8x16_t      max                           = vdupq_n_u8(255);
+
+    Iterator in(input, window);
+    Iterator out(output, window);
+    execute_window_loop(window, [&](const Coordinates & id)
+    {
+        // Get bias and pointer to input
+        const auto  in_ptr = reinterpret_cast<int32_t *>(in.ptr());
+        int32x4x4_t v_in =
+        {
+            {
+                vld1q_s32(in_ptr),
+                vld1q_s32(in_ptr + 4),
+                vld1q_s32(in_ptr + 8),
+                vld1q_s32(in_ptr + 12)
+            }
+        };
+
+        const auto out_ptr = reinterpret_cast<uint8_t *>(out.ptr());
+        vst1q_u8(out_ptr, finalize_quantization<false>(v_in, result_fixedpoint_multiplier, result_shift, result_offset_after_shift_s32, min, max));
+    },
+    in, out);
+}
 } // namespace
 
 NEDirectConvolutionLayerOutputStageKernel::NEDirectConvolutionLayerOutputStageKernel()
-    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr)
+    : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0)
 {
 }
 
-void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const ITensor *bias, ITensor *output)
+void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const ITensor *bias, ITensor *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());
+        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()));
 
-    _func   = nullptr;
-    _bias   = bias;
-    _input  = input;
-    _output = output;
+    _func                         = nullptr;
+    _bias                         = bias;
+    _input                        = input;
+    _output                       = output;
+    _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
+    _result_shift                 = result_shift;
+    _result_offset_after_shift    = result_offset_after_shift;
 
     // Configure kernel window
     auto win_config = validate_and_configure_window(input->info(), (bias == nullptr) ? nullptr : bias->info(), (output == nullptr) ? nullptr : output->info());
@@ -350,6 +446,9 @@ void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const
             _func = (output == nullptr) ? &output_stage<qint32_t, qint16_t, true, true> : &output_stage<qint32_t, qint16_t, false, true>;
             break;
         }
+        case DataType::S32:
+            _func = (bias == nullptr) ? &output_stage<int32_t, uint8_t, false, false> : &output_stage<int32_t, uint8_t, false, true>;
+            break;
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
         case DataType::F16:
         {
@@ -365,7 +464,6 @@ void NEDirectConvolutionLayerOutputStageKernel::configure(ITensor *input, const
         default:
         {
             ARM_COMPUTE_ERROR("Unsupported combination of types among the inputs.");
-            break;
         }
     }
 }
@@ -385,5 +483,5 @@ void NEDirectConvolutionLayerOutputStageKernel::run(const Window &window, const
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
     ARM_COMPUTE_ERROR_ON(_func == nullptr);
 
-    (*_func)(_input, _bias, window, _output);
+    (*_func)(_input, _bias, window, _output, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift);
 }
diff --git a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
index 8b3f2383ab..5e14e1a95d 100644
--- a/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
+++ b/src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -96,57 +96,11 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
     return std::make_pair(err, win);
 }
+} // namespace
 
-template <bool    is_bounded_relu>
-inline uint8x16_t finalize_quantization(int32x4x4_t &in_s32, int result_fixedpoint_multiplier, int32_t result_shift, int32x4_t result_offset_after_shift_s32, uint8x16_t min_u8,
-                                        uint8x16_t max_u8)
+namespace arm_compute
 {
-    const static int32x4_t zero_s32 = vdupq_n_s32(0);
-
-    // Fixed point multiplication with vector saturating rounding doubling multiply high with scalar
-    in_s32.val[0] = vqrdmulhq_n_s32(in_s32.val[0], result_fixedpoint_multiplier);
-    in_s32.val[1] = vqrdmulhq_n_s32(in_s32.val[1], result_fixedpoint_multiplier);
-    in_s32.val[2] = vqrdmulhq_n_s32(in_s32.val[2], result_fixedpoint_multiplier);
-    in_s32.val[3] = vqrdmulhq_n_s32(in_s32.val[3], result_fixedpoint_multiplier);
-
-    // Round to the nearest division by a power-of-two using result_shift_s32
-    in_s32.val[0] = rounding_divide_by_pow2(in_s32.val[0], result_shift);
-    in_s32.val[1] = rounding_divide_by_pow2(in_s32.val[1], result_shift);
-    in_s32.val[2] = rounding_divide_by_pow2(in_s32.val[2], result_shift);
-    in_s32.val[3] = rounding_divide_by_pow2(in_s32.val[3], result_shift);
-
-    // Add the offset terms
-    in_s32.val[0] = vaddq_s32(in_s32.val[0], result_offset_after_shift_s32);
-    in_s32.val[1] = vaddq_s32(in_s32.val[1], result_offset_after_shift_s32);
-    in_s32.val[2] = vaddq_s32(in_s32.val[2], result_offset_after_shift_s32);
-    in_s32.val[3] = vaddq_s32(in_s32.val[3], result_offset_after_shift_s32);
-
-    // Saturate negative values
-    in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32);
-    in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32);
-    in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32);
-    in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32);
-
-    // Convert S32 to S16
-    const int16x8x2_t in_s16 =
-    {
-        {
-            vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])),
-            vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3]))
-        }
-    };
-
-    // Convert S16 to U8
-    uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1]));
-
-    if(is_bounded_relu)
-    {
-        out_u8 = vmaxq_u8(out_u8, min_u8);
-        out_u8 = vminq_u8(out_u8, max_u8);
-    }
-
-    return out_u8;
-}
+class Coordinates;
 
 /* Function used by the left-over for loop to perform the quantization */
 template <bool is_bounded_relu>
@@ -178,11 +132,6 @@ inline uint8_t finalize_quantization(int32x4_t in_s32, int result_fixedpoint_mul
 
     return out_u8;
 }
-} // namespace
-
-namespace arm_compute
-{
-class Coordinates;
 } // namespace arm_compute
 
 template <bool is_bounded_relu>
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
index 298101a09d..2d08b45210 100644
--- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,28 +26,56 @@
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/ITensor.h"
 #include "arm_compute/core/PixelValue.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 #include "support/ToolchainSupport.h"
 
 using namespace arm_compute;
 
 NEDepthwiseConvolutionLayer3x3::NEDepthwiseConvolutionLayer3x3()
-    : _kernel(), _output_stage_kernel(), _border_handler(), _has_bias(false)
+    : _kernel(), _output_stage_kernel(), _border_handler(), _accumulator(), _has_bias(false), _is_quantized(false)
 {
 }
 
 void NEDepthwiseConvolutionLayer3x3::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *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);
 
-    // Configure kernels
-    _kernel.configure(input, weights, output, conv_info);
-    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast<float>(0.f)));
-    if(biases != nullptr)
+    PixelValue zero_value(0.f);
+
+    _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
+    _has_bias     = biases != nullptr;
+
+    // Allocate the intermediate accumulator tensor in case of fixed point input
+    if(_is_quantized)
+    {
+        _accumulator.allocator()->init(TensorInfo(output->info()->tensor_shape(), 1, DataType::S32));
+        _accumulator.info()->set_quantization_info(input->info()->quantization_info());
+        zero_value = PixelValue(static_cast<uint32_t>(input->info()->quantization_info().offset));
+    }
+
+    // Configure depthwise convolution kernel
+    _kernel.configure(input, weights, (_is_quantized) ? &_accumulator : output, conv_info);
+
+    // Configure border handler
+    _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, zero_value);
+
+    // Configure biases accumulation
+    if(_has_bias || _is_quantized)
     {
-        _output_stage_kernel.configure(output, biases);
-        _has_bias = true;
+        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(&_accumulator, biases, output, output_multiplier, output_shift, output->info()->quantization_info().offset);
+            _accumulator.allocator()->allocate();
+        }
+        else
+        {
+            _output_stage_kernel.configure(output, biases);
+        }
     }
 }
 
@@ -55,7 +83,7 @@ void NEDepthwiseConvolutionLayer3x3::run()
 {
     NEScheduler::get().schedule(&_border_handler, Window::DimX);
     NEScheduler::get().schedule(&_kernel, Window::DimX);
-    if(_has_bias)
+    if(_has_bias || _is_quantized)
     {
         NEScheduler::get().schedule(&_output_stage_kernel, Window::DimX);
     }