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>

9 files changed, 425 insertions, 43 deletions

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.