COMPMID-1539 Implement YOLOLayer on CL

Change-Id: I332c0703e1399fca0c5b724529b54a28f49c88da
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/146842
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

9 files changed, 496 insertions, 73 deletions

diff --git a/src/core/CL/CLHelpers.cpp b/src/core/CL/CLHelpers.cpp
index 9703b0fe16..5c435ddc22 100644
--- a/src/core/CL/CLHelpers.cpp
+++ b/src/core/CL/CLHelpers.cpp
@@ -64,6 +64,36 @@ std::string get_cl_type_from_data_type(const DataType &dt)
     }
 }
 
+std::string get_cl_select_type_from_data_type(const DataType &dt)
+{
+    switch(dt)
+    {
+        case DataType::U8:
+            return "uchar";
+        case DataType::S8:
+            return "char";
+        case DataType::QASYMM8:
+            return "uchar";
+        case DataType::U16:
+            return "ushort";
+        case DataType::F16:
+        case DataType::S16:
+            return "short";
+        case DataType::U32:
+            return "uint";
+        case DataType::F32:
+        case DataType::S32:
+            return "int";
+        case DataType::U64:
+            return "ulong";
+        case DataType::S64:
+            return "long";
+        default:
+            ARM_COMPUTE_ERROR("Unsupported input data type.");
+            return "";
+    }
+}
+
 std::string get_data_size_from_data_type(const DataType &dt)
 {
     switch(dt)
diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 7cc586bff1..75ff2482c8 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -428,6 +428,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "winograd_output_transform_4x4_5x5_nhwc", "winograd_output_transform.cl" },
     { "winograd_output_transform_4x1_5x1_nhwc", "winograd_output_transform.cl" },
     { "winograd_output_transform_1x4_1x5_nhwc", "winograd_output_transform.cl" },
+    { "yolo_layer_nchw", "yolo_layer.cl" },
+    { "yolo_layer_nhwc", "yolo_layer.cl" },
     { "YUYV422_to_IYUV_bt709", "color_convert.cl" },
     { "YUYV422_to_NV12_bt709", "color_convert.cl" },
     { "YUYV422_to_RGB888_bt709", "color_convert.cl" },
@@ -797,6 +799,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
         "winograd_output_transform.cl",
 #include "./cl_kernels/winograd_output_transform.clembed"
     },
+    {
+        "yolo_layer.cl",
+#include "./cl_kernels/yolo_layer.clembed"
+    },
 #endif /* EMBEDDED_KERNELS */
 };
 
diff --git a/src/core/CL/cl_kernels/activation_helpers.h b/src/core/CL/cl_kernels/activation_helpers.h
new file mode 100644
index 0000000000..dfab082381
--- /dev/null
+++ b/src/core/CL/cl_kernels/activation_helpers.h
@@ -0,0 +1,99 @@
+/*
+ * 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 "helpers.h"
+
+#if defined(TYPE) && defined(SELECT_TYPE)
+
+#define CONST_ONE 1.f
+#define ABS_OP(a) fabs((a))
+#define ADD_OP(a, b) ((a) + (b))
+#define SUB_OP(a, b) ((a) - (b))
+#define MUL_OP(a, b) ((a) * (b))
+#define MLA_OP(a, b, c) ((b) * (c) + (a))
+#define DIV_OP(a, b) ((a) / (b))
+#define EXP_OP(a) exp((a))
+#define LOG_OP(a) log((a))
+#define SQRT_OP(a) sqrt((a))
+#define TANH_OP(a) tanh((a))
+
+// Logistic Activation
+inline TYPE logistic_op(TYPE x)
+{
+    return DIV_OP((TYPE)CONST_ONE, ADD_OP((TYPE)CONST_ONE, EXP_OP(-x)));
+}
+// Hyperbolic Tangent Activation
+inline TYPE tanh_op(TYPE x)
+{
+    return MUL_OP((TYPE)A_VAL, TANH_OP(MUL_OP((TYPE)B_VAL, x)));
+}
+// RELU Tangent Activation
+inline TYPE relu_op(TYPE x)
+{
+    return max((TYPE)0, x);
+}
+// Bounded RELU Activation
+inline TYPE brelu_op(TYPE x)
+{
+    return min((TYPE)A_VAL, max((TYPE)0, x));
+}
+// Lower Upper Bounded RELU Activation
+inline TYPE lu_brelu_op(TYPE x)
+{
+    return min(max(x, (TYPE)B_VAL), (TYPE)A_VAL);
+}
+// Leaky RELU Activation
+inline TYPE lrelu_op(TYPE x)
+{
+    return select(MUL_OP((TYPE)A_VAL, x), x, CONVERT(x > (TYPE)0, SELECT_TYPE));
+}
+// Soft RELU Activation
+inline TYPE srelu_op(TYPE x)
+{
+    return LOG_OP(ADD_OP((TYPE)CONST_ONE, EXP_OP(x)));
+}
+// Absolute Activation
+inline TYPE abs_op(TYPE x)
+{
+    return ABS_OP(x);
+}
+// Square Activation
+inline TYPE square_op(TYPE x)
+{
+    return MUL_OP(x, x);
+}
+// Square-root Activation
+inline TYPE sqrt_op(TYPE x)
+{
+    return SQRT_OP(x);
+}
+// Linear Activation
+inline TYPE linear_op(TYPE x)
+{
+    return MLA_OP((TYPE)B_VAL, (TYPE)A_VAL, x);
+}
+
+#define ACTIVATION_OP2(op, x) op##_op(x)
+#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x)
+
+#endif // defined(TYPE) && defined(SELECT_TYPE)
\ No newline at end of file
diff --git a/src/core/CL/cl_kernels/activation_layer.cl b/src/core/CL/cl_kernels/activation_layer.cl
index 373406a6da..cf1f434972 100644
--- a/src/core/CL/cl_kernels/activation_layer.cl
+++ b/src/core/CL/cl_kernels/activation_layer.cl
@@ -21,80 +21,10 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#include "helpers.h"
-
 #define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+#define SELECT_TYPE VEC_DATA_TYPE(SELECT_DATA_TYPE, VEC_SIZE)
 
-#define CONST_ONE 1.f
-#define ABS_OP(a) fabs((a))
-#define ADD_OP(a, b) ((a) + (b))
-#define SUB_OP(a, b) ((a) - (b))
-#define MUL_OP(a, b) ((a) * (b))
-#define MLA_OP(a, b, c) ((b) * (c) + (a))
-#define DIV_OP(a, b) ((a) / (b))
-#define EXP_OP(a) exp((a))
-#define LOG_OP(a) log((a))
-#define SQRT_OP(a) sqrt((a))
-#define TANH_OP(a) tanh((a))
-
-// Logistic Activation
-inline TYPE logistic_op(TYPE x)
-{
-    return DIV_OP((TYPE)CONST_ONE, ADD_OP((TYPE)CONST_ONE, EXP_OP(-x)));
-}
-// Hyperbolic Tangent Activation
-inline TYPE tanh_op(TYPE x)
-{
-    return MUL_OP((TYPE)A_VAL, TANH_OP(MUL_OP((TYPE)B_VAL, x)));
-}
-// RELU Tangent Activation
-inline TYPE relu_op(TYPE x)
-{
-    return max(0, x);
-}
-// Bounded RELU Activation
-inline TYPE brelu_op(TYPE x)
-{
-    return min((TYPE)A_VAL, max(0, x));
-}
-// Lower Upper Bounded RELU Activation
-inline TYPE lu_brelu_op(TYPE x)
-{
-    return min(max(x, (TYPE)B_VAL), (TYPE)A_VAL);
-}
-// Leaky RELU Activation
-inline TYPE lrelu_op(TYPE x)
-{
-    return select(MUL_OP((TYPE)A_VAL, x), x, x > (TYPE)0);
-}
-// Soft RELU Activation
-inline TYPE srelu_op(TYPE x)
-{
-    return LOG_OP(ADD_OP((TYPE)CONST_ONE, EXP_OP(x)));
-}
-// Absolute Activation
-inline TYPE abs_op(TYPE x)
-{
-    return ABS_OP(x);
-}
-// Square Activation
-inline TYPE square_op(TYPE x)
-{
-    return MUL_OP(x, x);
-}
-// Square-root Activation
-inline TYPE sqrt_op(TYPE x)
-{
-    return SQRT_OP(x);
-}
-// Linear Activation
-inline TYPE linear_op(TYPE x)
-{
-    return MLA_OP((TYPE)B_VAL, (TYPE)A_VAL, x);
-}
-
-#define ACTIVATION_OP2(op, x) op##_op(x)
-#define ACTIVATION_OP(op, x) ACTIVATION_OP2(op, x)
+#include "activation_helpers.h"
 
 #if defined(ACT)
 
diff --git a/src/core/CL/cl_kernels/yolo_layer.cl b/src/core/CL/cl_kernels/yolo_layer.cl
new file mode 100644
index 0000000000..2240d7c637
--- /dev/null
+++ b/src/core/CL/cl_kernels/yolo_layer.cl
@@ -0,0 +1,176 @@
+/*
+ * 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.
+ */
+#if defined(DATA_TYPE) && defined(SELECT_DATA_TYPE) && defined(ACT) && defined(NUM_CLASSES) && defined(VEC_SIZE)
+
+#if VEC_SIZE != 1
+#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+#define SELECT_TYPE VEC_DATA_TYPE(SELECT_DATA_TYPE, VEC_SIZE)
+
+#include "activation_helpers.h"
+
+/** This performs a YOLO partial activation function for NCHW data layout
+ *
+ * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
+ * @note The number of classes should be given as a preprocessor argument using -DNUM_CLASSES=num. e.g. -DNUM_CLASSES=80
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] output_ptr                           (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ */
+__kernel void yolo_layer_nchw(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get pixels pointer
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    const int  box_ch_id = get_global_id(2) % (NUM_CLASSES + 5);
+    const bool activate  = box_ch_id != 2 && box_ch_id != 3;
+
+    if(activate)
+    {
+        // Load data
+        TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+        data      = ACTIVATION_OP(ACT, data); // select(1.0f, ACTIVATION_OP(ACT, data), (SELECT_TYPE)activate);
+
+        // Store result
+        VSTORE(VEC_SIZE)
+        (data, 0, (__global DATA_TYPE *)output.ptr);
+    }
+#ifndef IN_PLACE
+    else
+    {
+        // Load data
+        TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+
+        // Store result
+        VSTORE(VEC_SIZE)
+        (data, 0, (__global DATA_TYPE *)output.ptr);
+    }
+#endif // IN_PLACE
+}
+
+#else // VEC_SIZE != 1
+
+#define TYPE DATA_TYPE
+#define SELECT_TYPE SELECT_DATA_TYPE
+
+#include "activation_helpers.h"
+
+/** This performs a YOLO partial activation function for NCHW data layout
+ *
+ * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=1
+ * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH
+ * @note The number of classes should be given as a preprocessor argument using -DNUM_CLASSES=num. e.g. -DNUM_CLASSES=80
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
+ *
+ * @param[in]  input_ptr                            Pointer to the source tensor. Supported data types: F16/F32
+ * @param[in]  input_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] output_ptr                           (Optional) Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination tensor
+ */
+__kernel void yolo_layer_nhwc(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get pixels pointer
+    Tensor3D input  = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    const int  box_ch_id = get_global_id(0) % (NUM_CLASSES + 5);
+    const bool activate  = box_ch_id != 2 && box_ch_id != 3;
+
+    if(activate)
+    {
+        // Load data
+        DATA_TYPE data = *((__global DATA_TYPE *)input.ptr);
+        data           = select(data, ACTIVATION_OP(ACT, data), (SELECT_TYPE)activate);
+
+        // Store result
+        *((__global DATA_TYPE *)output.ptr) = data;
+    }
+#ifndef IN_PLACE
+    else
+    {
+        // Load data
+        DATA_TYPE data = *((__global DATA_TYPE *)input.ptr);
+
+        // Store result
+        *((__global DATA_TYPE *)output.ptr) = data;
+    }
+#endif // IN_PLACE
+}
+
+#endif // VEC_SIZE != 1
+#endif // defined(DATA_TYPE) && defined(SELECT_DATA_TYPE) && defined(ACT) && defined(NUM_CLASSES) && defined(VEC_SIZE)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp
index a15e99b8d4..73a4d7d2c6 100644
--- a/src/core/CL/kernels/CLActivationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp
@@ -133,6 +133,7 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act
     std::set<std::string> build_opts;
     build_opts.emplace(("-DACT=" + lower_string(string_from_activation_func(act_info.activation()))));
     build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
+    build_opts.emplace(("-DSELECT_DATA_TYPE=" + get_cl_select_type_from_data_type(dt)));
     build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
 
     if(is_data_type_quantized(dt))
diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
index 1fa5c8521f..07bcb75a6a 100644
--- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
@@ -159,6 +159,7 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out
     // Set build options
     CLBuildOptions build_opts;
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
+    build_opts.add_option("-DSELECT_DATA_TYPE=" + get_cl_select_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.add_option_if(act_info.enabled(), "-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
     build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index cc8384c81b..d56ac01a83 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -173,7 +173,6 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
     _border_size                        = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
 
     const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : (8 / input->info()->element_size());
-    ;
 
     CLBuildOptions build_opts;
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
diff --git a/src/core/CL/kernels/CLYOLOLayerKernel.cpp b/src/core/CL/kernels/CLYOLOLayerKernel.cpp
new file mode 100644
index 0000000000..7d9dbd4ac5
--- /dev/null
+++ b/src/core/CL/kernels/CLYOLOLayerKernel.cpp
@@ -0,0 +1,181 @@
+/*
+ * 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/CLYOLOLayerKernel.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/IAccessWindow.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/Window.h"
+
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/Types.h"
+#include "support/ToolchainSupport.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info, int32_t num_classes)
+{
+    ARM_COMPUTE_UNUSED(act_info);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN);
+
+    const unsigned int channel_idx = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL);
+    ARM_COMPUTE_RETURN_ERROR_ON(num_classes <= 0);
+    ARM_COMPUTE_RETURN_ERROR_ON((input->dimension(channel_idx) % (num_classes + 5)) != 0);
+
+    // Checks performed when output is configured
+    if((output != nullptr) && (output->total_size() != 0))
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+    }
+
+    return Status{};
+}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
+{
+    if(output != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+        // Output auto inizialitation if not yet initialized
+        auto_init_if_empty(*output, *input);
+    }
+
+    const bool         is_nchw                           = input->data_layout() == DataLayout::NCHW;
+    const unsigned int num_elems_processed_per_iteration = is_nchw ? 16 / input->element_size() : 1;
+
+    Window win            = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
+    bool   window_changed = false;
+
+    if(output != nullptr)
+    {
+        AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
+        AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
+        window_changed = update_window_and_padding(win, input_access, output_access);
+        output_access.set_valid_region(win, input->valid_region());
+    }
+    else
+    {
+        window_changed = update_window_and_padding(win, AccessWindowHorizontal(input, 0, num_elems_processed_per_iteration));
+    }
+
+    Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
+    return std::make_pair(err, win);
+}
+} // namespace
+
+CLYOLOLayerKernel::CLYOLOLayerKernel()
+    : _input(nullptr), _output(nullptr), _run_in_place(false)
+{
+}
+
+void CLYOLOLayerKernel::configure(ICLTensor *input, ICLTensor *output, const ActivationLayerInfo &act_info, int32_t num_classes)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    _run_in_place = (output == nullptr) || (output == input);
+
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr, act_info, num_classes));
+
+    const bool         is_nchw                           = input->info()->data_layout() == DataLayout::NCHW;
+    const unsigned int num_elems_processed_per_iteration = is_nchw ? 16 / input->info()->element_size() : 1;
+    const DataType     dt                                = input->info()->data_type();
+    float              a_const                           = act_info.a();
+    float              b_const                           = act_info.b();
+
+    // Set build options
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DACT=" + lower_string(string_from_activation_func(act_info.activation())));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(dt));
+    build_opts.add_option("-DSELECT_DATA_TYPE=" + get_cl_select_type_from_data_type(dt));
+    build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
+    build_opts.add_option("-DA_VAL=" + float_to_string_with_full_precision(a_const));
+    build_opts.add_option("-DB_VAL=" + float_to_string_with_full_precision(b_const));
+    build_opts.add_option("-DNUM_CLASSES=" + support::cpp11::to_string(num_classes));
+    build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
+
+    // Create kernel
+    std::string kernel_name = std::string("yolo_layer_") + lower_string(string_from_data_layout(input->info()->data_layout()));
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+
+    // Make sure _kernel is initialized before calling the parent's configure
+    _input  = input;
+    _output = output;
+
+    // Configure kernel window
+    auto win_config = validate_and_configure_window(input->info(), (_run_in_place) ? nullptr : output->info());
+    ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+    ICLKernel::configure_internal(win_config.second);
+
+    // Set config_id for enabling LWS tuning
+    _config_id = "yolo_layer_";
+    _config_id += lower_string(string_from_data_type(dt));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += "_";
+    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += "_";
+    _config_id += lower_string(string_from_data_layout(input->info()->data_layout()));
+}
+
+Status CLYOLOLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info, int32_t num_classes)
+{
+    const bool run_in_place = (output == nullptr) || (output == input);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, act_info, num_classes));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), (run_in_place) ? nullptr : output->clone().get()).first);
+
+    return Status{};
+}
+
+void CLYOLOLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
+
+    Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
+    Window slice     = collapsed.first_slice_window_3D();
+
+    do
+    {
+        unsigned int idx = 0;
+        add_3D_tensor_argument(idx, _input, slice);
+        if(!_run_in_place)
+        {
+            add_3D_tensor_argument(idx, _output, slice);
+        }
+        enqueue(queue, *this, slice, lws_hint());
+    }
+    while(collapsed.slide_window_slice_3D(slice));
+}
+} // namespace arm_compute
\ No newline at end of file