COMPMID-631: Merge branches/gles_compute branch

Last commit: commit b25c5f68042b0c81bf611d59a1bb8535e1c42497 Author: Xinghang Zhou <xinghang.zhou@arm.com> Date: Wed Oct 25 18:48:10 2017 +0800 Synced validation's tolerances of GCSoftmax from cl side Change-Id: Ibe72054205c1c8721845d679a31af7ed0a7c5cf6 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/93283 Reviewed-by: Anthony Barbier <anthony.barbier@arm.com> Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
author: Anthony Barbier <anthony.barbier@arm.com> 2017-10-26 15:23:08 +0100
committer: Anthony Barbier <anthony.barbier@arm.com> 2018-11-02 16:35:24 +0000
commit: 7068f9900d136312318ff430aef588b14e0c87ad (patch)
tree: b57ca81231860f1d8755e6f18e5be7c959fb60c6 /src/core/GLES_COMPUTE/cs_shaders
parent: d60737592736715dcfd0520535c48190d4ac77d2 (diff)
download: ComputeLibrary-7068f9900d136312318ff430aef588b14e0c87ad.tar.gz
17 files changed, 7500 insertions, 0 deletions
diff --git a/src/core/GLES_COMPUTE/cs_shaders/absdiff.cs b/src/core/GLES_COMPUTE/cs_shaders/absdiff.cs
new file mode 100644
index 0000000000..f6113e13eb
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/absdiff.cs
@@ -0,0 +1,71 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src1);
+    IMAGE_PARAM_DECLARATION(src2);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+BUFFER_DECLARATION(src1, 1, uint, readonly);
+BUFFER_DECLARATION(src2, 2, uint, readonly);
+BUFFER_DECLARATION(dst, 3, uint, writeonly);
+
+/** Calculate the absolute difference of two input images.
+ *
+ * @param[in]  src1_ptr                           Pointer to the first source image. Supported data types: U8
+ * @param[in]  src1_stride_x                      Stride of the first source image in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the first source image in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the first source image
+ * @param[in]  src2_ptr                           Pointer to the second source image. Supported data types: Same as @p in1_ptr
+ * @param[in]  src2_stride_x                      Stride of the second source image in X dimension (in bytes)
+ * @param[in]  src2_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src2_stride_y                      Stride of the second source image in Y dimension (in bytes)
+ * @param[in]  src2_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src2_offset_first_element_in_bytes The offset of the first element in the second source image
+ * @param[out] dst_ptr                            Pointer to the destination image. Supported data types: Same as @p in1_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination image 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 image 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_offset_first_element_in_bytes  The offset of the first element in the destination image
+ */
+void main(void)
+{
+    Image src1 = CONVERT_TO_IMAGE_STRUCT(src1);
+    Image src2 = CONVERT_TO_IMAGE_STRUCT(src2);
+    Image dst  = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    uvec4 tmp1 = UNPACK(LOAD4(src1, CURRENT_OFFSET(src1)), uint, uvec4);
+    uvec4 tmp2 = UNPACK(LOAD4(src2, CURRENT_OFFSET(src2)), uint, uvec4);
+    uvec4 diff = uvec4(abs(ivec4(tmp1 - tmp2)));
+
+    STORE4(dst, CURRENT_OFFSET(dst), PACK(diff, uvec4, uint));
+}
diff --git a/src/core/GLES_COMPUTE/cs_shaders/activation_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/activation_layer.cs
new file mode 100644
index 0000000000..fc9da114f7
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/activation_layer.cs
@@ -0,0 +1,262 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+#ifdef DATA_TYPE_FP32
+precision highp float;
+#elif defined(DATA_TYPE_FP16)
+#if defined(LOGISTIC) || defined(TANH) || defined(SRELU) || defined(SQRT)
+precision highp float;
+#else  /*LOGISTIC_TANH_SRELU_SQRT*/
+precision mediump float;
+#endif /*LOGISTIC_TANH_SRELU_SQRT*/
+#endif /*DATA_TYPE_FP32*/
+
+#define ABS_OP(a) abs((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 CONST_ONE (1.f)
+
+// Logistic Activation
+float logistic_op(float x)
+{
+    return DIV_OP(CONST_ONE, ADD_OP(CONST_ONE, EXP_OP(-x)));
+}
+// Hyperbolic Tangent Activation
+float tanh_op(float x)
+{
+    float tmp = float(B_VAL) * x;
+    if(tmp > 10.f)
+    {
+        return MUL_OP(float(A_VAL), 1.f);
+    }
+    else if(tmp < -10.f)
+    {
+        return MUL_OP(float(A_VAL), -1.f);
+    }
+    else
+    {
+        return MUL_OP(float(A_VAL), tanh(tmp + 0.000001f));
+    }
+}
+// RELU Tangent Activation
+float relu_op(float x)
+{
+    return max(0.f, x);
+}
+// Bounded RELU Activation
+float brelu_op(float x)
+{
+    return min(float(A_VAL), max(float(0.0), x));
+}
+// Lower Upper Bounded RELU Activation
+float lu_brelu_op(float x)
+{
+    return min(max(x, float(B_VAL)), float(A_VAL));
+}
+// Leaky RELU Activation
+float lrelu_op(float x)
+{
+    return (x > float(0.0)) ? x : MUL_OP(float(A_VAL), x);
+}
+// Soft RELU Activation
+float srelu_op(float x)
+{
+    return LOG_OP(ADD_OP(CONST_ONE, EXP_OP(x)));
+}
+// Absolute Activation
+float abs_op(float x)
+{
+    return ABS_OP(x);
+}
+// Square Activation
+float square_op(float x)
+{
+    return MUL_OP(x, x);
+}
+// Square-root Activation
+float sqrt_op(float x)
+{
+    return SQRT_OP(x);
+}
+// Linear Activation
+float linear_op(float x)
+{
+    return MLA_OP(float(B_VAL), float(A_VAL), x);
+}
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+#ifdef DATA_TYPE_FP32
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+/** This performs an activation function floating point inputs.
+ *
+ * @note Activation function should be given as a preprocessor argument using "#define act_name". e.g. "#define TANH"
+ * @note A, B variables required by some activation functions are set using A_VAL= and B_VAL= respectively.
+ *
+ * @param[in]  src_ptr                              Pointer to the source image. Supported data types: F32
+ * @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_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the source image
+ * @param[out] dst_ptr                              Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                         Stride of the destination image 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                         ride of the destination image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                           dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes    The offset of the first element in the destination image
+ */
+void main(void)
+{
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float data     = src_ptr[src.current_offset];
+    float data_out = 0.f;
+    // Perform activation
+
+#ifdef LOGISTIC
+    data_out = logistic_op(data);
+#elif defined(TANH)     /*LOGISTIC*/
+    data_out = tanh_op(data);
+#elif defined(RELU)     /*RELU*/
+    data_out = relu_op(data);
+#elif defined(BRELU)    /*BRELU*/
+    data_out = brelu_op(data);
+#elif defined(LU_BRELU) /*LU_BRELU*/
+    data_out = lu_brelu_op(data);
+#elif defined(LRELU)    /*LRELU*/
+    data_out = lrelu_op(data);
+#elif defined(SRELU)    /*SRELU*/
+    data_out = srelu_op(data);
+#elif defined(ABS)      /*ABS*/
+    data_out = abs_op(data);
+#elif defined(SQUARE)   /*SQUARE*/
+    data_out = square_op(data);
+#elif defined(SQRT)     /*SQRT*/
+    data_out = sqrt_op(data);
+#elif defined(LINEAR)   /*LINEAR*/
+    data_out = linear_op(data);
+#else                   /*LOGISTIC*/
+#error Activation function not provided
+#endif /*LOGISTIC*/
+
+    dst_ptr[dst.current_offset] = data_out;
+}
+
+#elif defined(DATA_TYPE_FP16)
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(dst, 2, uint, writeonly);
+
+/** This performs an activation function floating point inputs.
+ *
+ * @note Activation function should be given as a preprocessor argument using "#define act_name". e.g. "#define TANH"
+ * @note A, B variables required by some activation functions are set using A_VAL= and B_VAL= respectively.
+ *
+ * @param[in]  src_ptr                              Pointer to the source image. Supported data types: F16
+ * @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_stride_z                         Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the source image
+ * @param[out] dst_ptr                              Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                         Stride of the destination image 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                         ride of the destination image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                           dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes    The offset of the first element in the destination image
+ */
+void main(void)
+{
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    uint data = src_ptr[src.current_offset >> 2];
+    // Perform activation
+    float a = unpackHalf2x16(data).x;
+    float b = unpackHalf2x16(data).y;
+    vec2  data_out;
+#ifdef LOGISTIC         /*LOGISTIC*/
+    data_out.x = logistic_op(a);
+    data_out.y = logistic_op(b);
+#elif defined(TANH)     /*TANH*/
+    data_out.x = tanh_op(a);
+    data_out.y = tanh_op(b);
+#elif defined(RELU)     /*RELU*/
+    data_out.x = relu_op(a);
+    data_out.y = relu_op(b);
+#elif defined(BRELU)    /*BRELU*/
+    data_out.x = brelu_op(a);
+    data_out.y = brelu_op(b);
+#elif defined(LU_BRELU) /*LU_BRELU*/
+    data_out.x = lu_brelu_op(a);
+    data_out.y = lu_brelu_op(b);
+#elif defined(LRELU)    /*LRELU*/
+    data_out.x = lrelu_op(a);
+    data_out.y = lrelu_op(b);
+#elif defined(SRELU)    /*SRELU*/
+    data_out.x = srelu_op(a);
+    data_out.y = srelu_op(b);
+#elif defined(ABS)      /*ABS*/
+    data_out.x = abs_op(a);
+    data_out.y = abs_op(b);
+#elif defined(SQUARE)   /*SQUARE*/
+    data_out.x = square_op(a);
+    data_out.y = square_op(b);
+#elif defined(SQRT)     /*SQRT*/
+    data_out.x = sqrt_op(a);
+    data_out.y = sqrt_op(b);
+#elif defined(LINEAR)   /*LINEAR*/
+    data_out.x = linear_op(a);
+    data_out.y = linear_op(b);
+#else                   /*LOGISTIC*/
+#error Activation function not provided
+#endif /*LOGISTIC*/
+
+    dst_ptr[dst.current_offset >> 2] = packHalf2x16(data_out);
+}
+#endif /*DATA_TYPE_FP32*/
diff --git a/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs
new file mode 100644
index 0000000000..54880926cc
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/batchnormalization_layer.cs
@@ -0,0 +1,222 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+#ifdef DATA_TYPE_FP32
+precision highp float;
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+#endif /*DATA_TYPE_FP32*/
+
+#define ADD_OP(a, b) ((a) + (b))
+#define SUB_OP(a, b) ((a) - (b))
+#define MUL_OP(a, b) ((a) * (b))
+#define INVSQRT_OP(a) inversesqrt((a))
+#define SQCVT_SAT(a) (a)
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    VECTOR_PARAM_DECLARATION(mean);
+    VECTOR_PARAM_DECLARATION(var);
+    VECTOR_PARAM_DECLARATION(beta);
+    VECTOR_PARAM_DECLARATION(gamma);
+};
+
+#ifdef DATA_TYPE_FP32
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+BUFFER_DECLARATION(mean, 3, float, readonly);
+BUFFER_DECLARATION(var, 4, float, readonly);
+BUFFER_DECLARATION(beta, 5, float, readonly);
+BUFFER_DECLARATION(gamma, 6, float, readonly);
+
+/** Apply batch normalization.
+ *
+ * @note Epsilon parameter in the batch normalization equation should be given as a preprocessor argument using "#define EPSILON". e.g. "#define EPSILON 0.1"
+ *
+ * @param[in]  src_ptr                              Pointer to the first source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                         Stride of the first 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 first 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 first source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the first source tensor
+ * @param[out] dst_ptr                              Pointer to the destination tensor. Supported data types: 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_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 Z 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]  mean_ptr                             Pointer to the mean source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mean_stride_x                        Stride of the mean source tensor in X dimension (in bytes)
+ * @param[in]  mean_step_x                          mean_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mean_offset_first_element_in_bytes   The offset of the first element in the mean source tensor
+ * @param[in]  var_ptr                              Pointer to the var tensor. Supported data types: same as @p src_ptr
+ * @param[in]  var_stride_x                         Stride of the var tensor in X dimension (in bytes)
+ * @param[in]  var_step_x                           var_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  var_offset_first_element_in_bytes    The offset of the first element in the var source tensor
+ * @param[in]  beta_ptr                             Pointer to the beta source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  beta_stride_x                        Stride of the beta source tensor in X dimension (in bytes)
+ * @param[in]  beta_step_x                          beta_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  beta_offset_first_element_in_bytes   The offset of the first element in the beta source tensor
+ * @param[in]  gamma_ptr                            Pointer to the gamma source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  gamma_stride_x                       Stride of the gamma source tensor in X dimension (in bytes)
+ * @param[in]  gamma_step_x                         gamma_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  gamma_offset_first_element_in_bytes  The offset of the first element in the gamma source tensor
+ */
+void main(void)
+{
+    Tensor3D src   = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst   = CONVERT_TO_TENSOR3D_STRUCT(dst);
+    Vector   mean  = CONVERT_TO_VECTOR_STRUCT(mean);
+    Vector   var   = CONVERT_TO_VECTOR_STRUCT(var);
+    Vector   beta  = CONVERT_TO_VECTOR_STRUCT(beta);
+    Vector   gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
+
+    float input_value = 0.f;
+    float denominator = 0.f;
+    float numerator   = 0.f;
+    float x_bar       = 0.f;
+    float gamma_param = 0.f;
+    float beta_param  = 0.f;
+
+    uint current_slice = gl_GlobalInvocationID.z;
+
+    input_value = src_ptr[src.current_offset];
+    denominator = var_ptr[var.current_offset + (current_slice * var.stride_x) >> 2];
+    denominator = INVSQRT_OP(ADD_OP(denominator, SQCVT_SAT(float(ESPILON))));
+
+    // Calculate x bar and store results
+    numerator = mean_ptr[mean.current_offset + (current_slice * mean.stride_x) >> 2];
+    numerator = SUB_OP(input_value, numerator);
+    x_bar     = MUL_OP(numerator, denominator);
+
+    gamma_param = gamma_ptr[gamma.current_offset + (current_slice * beta.stride_x) >> 2];
+    beta_param  = beta_ptr[beta.current_offset + (current_slice * beta.stride_x) >> 2];
+
+    dst_ptr[dst.current_offset] = ADD_OP(MUL_OP(gamma_param, x_bar), beta_param);
+}
+
+#elif defined(DATA_TYPE_FP16)
+BUFFER_DECLARATION(src, 1, uint, );
+BUFFER_DECLARATION(dst, 2, uint, writeonly);
+BUFFER_DECLARATION(mean, 3, uint, );
+BUFFER_DECLARATION(var, 4, uint, );
+BUFFER_DECLARATION(beta, 5, uint, );
+BUFFER_DECLARATION(gamma, 6, uint, );
+
+/** Apply batch normalization.
+ *
+ * @note Epsilon parameter in the batch normalization equation should be given as a preprocessor argument using "#define EPSILON". e.g. "#define EPSILON 0.1"
+ *
+ * @param[in]  src_ptr                              Pointer to the first source tensor. Supported data types: F16
+ * @param[in]  src_stride_x                         Stride of the first 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 first 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 first source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                           src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes    The offset of the first element in the first source tensor
+ * @param[out] dst_ptr                              Pointer to the destination tensor. Supported data types: 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_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 Z 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]  mean_ptr                             Pointer to the mean source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mean_stride_x                        Stride of the mean source tensor in X dimension (in bytes)
+ * @param[in]  mean_step_x                          mean_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mean_offset_first_element_in_bytes   The offset of the first element in the mean source tensor
+ * @param[in]  var_ptr                              Pointer to the var tensor. Supported data types: same as @p src_ptr
+ * @param[in]  var_stride_x                         Stride of the var tensor in X dimension (in bytes)
+ * @param[in]  var_step_x                           var_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  var_offset_first_element_in_bytes    The offset of the first element in the var source tensor
+ * @param[in]  beta_ptr                             Pointer to the beta source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  beta_stride_x                        Stride of the beta source tensor in X dimension (in bytes)
+ * @param[in]  beta_step_x                          beta_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  beta_offset_first_element_in_bytes   The offset of the first element in the beta source tensor
+ * @param[in]  gamma_ptr                            Pointer to the gamma source tensor. Supported data types: same as @p src_ptr
+ * @param[in]  gamma_stride_x                       Stride of the gamma source tensor in X dimension (in bytes)
+ * @param[in]  gamma_step_x                         gamma_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  gamma_offset_first_element_in_bytes  The offset of the first element in the gamma source tensor
+ */
+void main(void)
+{
+    Tensor3D src   = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst   = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+    Vector   mean  = CONVERT_TO_VECTOR_STRUCT_FP16(mean);
+    Vector   var   = CONVERT_TO_VECTOR_STRUCT_FP16(var);
+    Vector   beta  = CONVERT_TO_VECTOR_STRUCT_FP16(beta);
+    Vector   gamma = CONVERT_TO_VECTOR_STRUCT_FP16(gamma);
+
+    vec2  input_value;
+    float denominator;
+    float numerator;
+    vec2  x_bar;
+    float gamma_param;
+    float beta_param;
+
+    uint current_slice = gl_GlobalInvocationID.z;
+    if((current_slice % uint(2)) == uint(0))
+    {
+        input_value = unpackHalf2x16(src_ptr[src.current_offset >> 2]);
+        denominator = unpackHalf2x16(var_ptr[(var.current_offset + current_slice * var.stride_x) >> 2]).x;
+        denominator = INVSQRT_OP(ADD_OP(denominator, SQCVT_SAT(float(ESPILON))));
+
+        //Calculate x bar and store results
+        numerator = unpackHalf2x16(mean_ptr[(mean.current_offset + current_slice * mean.stride_x) >> 2]).x;
+        x_bar     = MUL_OP(SUB_OP(input_value, numerator), denominator);
+
+        gamma_param = unpackHalf2x16(gamma_ptr[(gamma.current_offset + current_slice * beta.stride_x) >> 2]).x;
+        beta_param  = unpackHalf2x16(beta_ptr[(beta.current_offset + current_slice * beta.stride_x) >> 2]).x;
+
+        dst_ptr[dst.current_offset >> 2] = packHalf2x16(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
+    }
+    else
+    {
+        input_value = unpackHalf2x16(src_ptr[src.current_offset >> 2]);
+        denominator = unpackHalf2x16(var_ptr[(var.current_offset + current_slice * var.stride_x) >> 2]).y;
+        denominator = INVSQRT_OP(ADD_OP(denominator, SQCVT_SAT(float(ESPILON))));
+
+        //Calculate x bar and store results
+        numerator = unpackHalf2x16(mean_ptr[(mean.current_offset + current_slice * mean.stride_x) >> 2]).y;
+        x_bar     = MUL_OP(SUB_OP(input_value, numerator), denominator);
+
+        gamma_param = unpackHalf2x16(gamma_ptr[(gamma.current_offset + current_slice * beta.stride_x) >> 2]).y;
+        beta_param  = unpackHalf2x16(beta_ptr[(beta.current_offset + current_slice * beta.stride_x) >> 2]).y;
+
+        dst_ptr[dst.current_offset >> 2] = packHalf2x16(ADD_OP(MUL_OP(gamma_param, x_bar), beta_param));
+    }
+}
+#endif /*DATA_TYPE_FP32*/
diff --git a/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs b/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs
new file mode 100644
index 0000000000..65000f2de2
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/concatenate.cs
@@ -0,0 +1,106 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#ifdef DATA_TYPE_FP32
+precision highp float;
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+/** This kernel concatenates the input tensor into the output tensor along the third dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: 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_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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    dst_ptr[dst.current_offset + uint(OFFSETS_Z >> 2)] = src_ptr[tensor3D_offset(src, -OFFSETS_X, -OFFSETS_Y, 0)];
+}
+
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+
+/** This kernel concatenates the input tensor into the output tensor along the third dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: 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_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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src = GC_CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    uvec2 packed_s;
+    GC_LOAD1_3D_OFFSET(packed_s, src, -OFFSETS_X, -OFFSETS_Y, 0);
+    dst_ptr[(dst.current_offset + uint(OFFSETS_Z)) >> 3] = packed_s;
+}
+#endif /*DATA_TYPE_FP32*/
+\ No newline at end of file
diff --git a/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs
new file mode 100644
index 0000000000..1a0c9f1d30
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/convolution_layer.cs
@@ -0,0 +1,302 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#ifdef DATA_TYPE_FP16
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(dst, 2, uint, restrict);
+#else  // DATA_TYPE_FP16
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, restrict);
+#endif // DATA_TYPE_FP16
+
+layout(std140) uniform shader_params
+{
+#ifdef IM2COL_GENERIC
+    TENSOR3D_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+    uint filter_depth;
+    uint src_stride_w;
+    uint dst_stride_w;
+#endif // IM2COL_GENERIC
+
+#ifdef IM2COL_REDUCED
+    TENSOR3D_PARAM_DECLARATION(src);
+    VECTOR_PARAM_DECLARATION(dst);
+    uint width;
+    uint height;
+#endif // IM2COL_REDUCED
+
+#ifdef COL2IM
+    IMAGE_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    uint width;
+#endif // COL2IM
+};
+
+#ifdef DATA_TYPE_FP16
+
+precision mediump float;
+
+#ifdef IM2COL_REDUCED
+/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note In case biases will be added in late stage, "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F16
+ * @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[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]  width                             The width of the input tensor
+ * @param[in]  height                            The height of the input tensor
+ */
+void main(void)
+{
+    uvec3    pos            = uvec3(gl_GlobalInvocationID.xyz);
+    uvec3    size           = uvec3(gl_WorkGroupSize.xyz);
+    Tensor3D src            = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D src_nostep     = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(src);
+    Vector   dst            = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(dst);
+    uint     image_size     = width * height;
+    uint     element_count  = src_step_x / src_stride_x;
+    uint     tmp_out_offset = dst.current_offset + ((pos.x * element_count + pos.y * width + pos.z * image_size) * dst.stride_x);
+    uint     width_fp16     = ((width + uint(1)) >> uint(1));
+    uint     tmp;
+
+    // odd width
+    if(width % uint(2) != uint(0))
+    {
+        // even row
+        if((pos.y + pos.z * height) % uint(2) == uint(0))
+        {
+            LOAD1(tmp, src, src.current_offset >> uint(2));
+            STORE1(dst, tmp_out_offset >> uint(2), tmp);
+        }
+        else
+        {
+            // special op
+            uint tmpleft  = uint(0);
+            uint tmpright = uint(0);
+            LOAD1(tmpright, src, src.current_offset >> uint(2)); // right half
+            if(pos.x == uint(0))
+            {
+                LOAD1(tmpleft, src, tensor3D_offset_fp16(src_nostep, int(width), int(pos.y) - 1, int(pos.z)) >> uint(2)); // left half
+                tmpright = (tmpleft & uint(0xffff)) + (tmpright << uint(16));
+            }
+            else
+            {
+                LOAD1(tmpleft, src, tensor3D_offset_fp16(src_nostep, (int(pos.x) - 1) * int(element_count), int(pos.y), int(pos.z)) >> uint(2)); // left half
+                tmpright = ((tmpleft >> uint(16)) + (tmpright << uint(16)));
+            }
+            STORE1(dst, tmp_out_offset >> uint(2), tmpright);
+        }
+    }
+    else
+    {
+        LOAD1(tmp, src, src.current_offset >> uint(2));
+        STORE1(dst, tmp_out_offset >> uint(2), tmp);
+    }
+
+#ifdef HAS_BIAS
+    // If it is the last thread in the 3 dimensional workgroup
+    if(pos.x == (size.x - 1) && pos.y == (size.y - 1) && pos.z == (size.z - 1))
+    {
+        tmp_out_offset += dst.stride_x;
+
+        // FIXME: need odd/even detection for tmp_out_offset?
+        mediump vec2 bias_vec = vec2(1.0f, 1.0f);
+        uint         bias_u   = packHalf2x16(bias_vec);
+        STORE1(dst, tmp_out_offset >> uint(2), bias_u);
+    }
+#endif // HAS_BIAS
+}
+#endif // IM2COL_REDUCED
+
+#elif defined(DATA_TYPE_FP32)
+
+#ifdef IM2COL_GENERIC
+/** This kernel performs a reshaping of the input tensor to a tensor used to perform convolution using GEMM.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: 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_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_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  filter_depth                      The depth of the used filter
+ * @param[in]  src_stride_w                      Stride of the source tensor in W dimension (in bytes).
+ * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes).
+ */
+void main(void)
+{
+    uint xc    = gl_GlobalInvocationID.x;                // x coordinate in the convolved tensor
+    uint yc    = gl_GlobalInvocationID.y;                // y coordinate in the convolved tensor
+    uint ch    = gl_GlobalInvocationID.z % filter_depth; // input feature map
+    uint batch = gl_GlobalInvocationID.z / filter_depth; // the batch
+
+    // Calculate input indeces
+    uint xi           = xc * uint(STRIDE_X) - uint(PAD_X);
+    uint yi           = yc * uint(STRIDE_Y) - uint(PAD_Y);
+    uint input_offset = (src_offset_first_element_in_bytes + (ch * src_stride_z) + (batch * src_stride_w)) >> uint(2);
+
+    // Calculate output indeces
+    uint xo            = ch * uint(KERNEL_WIDTH) * uint(KERNEL_HEIGHT);
+    uint yo            = xc + yc * uint(CONVOLVED_WIDTH); // Index of the convolution
+    uint output_offset = (dst_offset_first_element_in_bytes + (yo * dst_stride_y) + (batch * dst_stride_w) + xo) >> uint(2);
+
+    // Linearize convolution elements
+    for(uint y = yi, y_e = yi + uint(KERNEL_HEIGHT); y < y_e; ++y)
+    {
+        for(uint x = xi, x_e = xi + uint(KERNEL_WIDTH); x < x_e; ++x)
+        {
+#if PAD_X == 0 && PAD_Y == 0
+            output_offset = input_offset + ((x * src_stride_x + y * src_stride_y) >> uint(2));
+            STORE4(dst, output_offset, LOAD4(src, input_offset));
+#else  // PAD_X == 0 && PAD_Y == 0
+            if(x < 0 || x >= SRC_WIDTH || y < 0 || y >= SRC_HEIGHT)
+            {
+                STORE4(dst, output_offset, 0.0f);
+            }
+            else
+            {
+                output_offset = input_offset + (x * src_stride_x + y * src_stride_y) >> uint(2));
+                STORE4(dst, output_offset, LOAD4(src, input_offset));
+            }
+#endif // PAD_X == 0 && PAD_Y == 0
+        }
+    }
+
+#ifdef HAS_BIAS
+    if(ch == (uint(KERNEL_DEPTH) - 1))
+    {
+        STORE4(dst, output_offset, 1.0f);
+    }
+#endif // HAS_BIAS
+}
+#endif // IM2COL_GENERIC
+
+#ifdef IM2COL_REDUCED
+/** This kernel reshapes the tensor's low three dimensions to single row for GEMM operation
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note In case biases will be added in late stage, "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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[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]  width                             The width of the input tensor
+ * @param[in]  height                            The height of the input tensor
+ */
+void main(void)
+{
+    uvec3    pos            = uvec3(gl_GlobalInvocationID.xyz);
+    uvec3    size           = uvec3(gl_WorkGroupSize.xyz);
+    Tensor3D src            = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Vector   dst            = CONVERT_TO_VECTOR_STRUCT_NO_STEP(dst);
+    uint     image_size     = width * height;
+    uint     tmp_out_offset = dst.current_offset + (((pos.x + pos.y * width + pos.z * image_size) * dst.stride_x) >> 2);
+
+    STORE4(dst, tmp_out_offset, LOAD4(src, src.current_offset));
+
+#ifdef HAS_BIAS
+    // If it is the last thread in the 3 dimensional workgroup
+    if(pos.x == (size.x - 1) && pos.y == (size.y - 1) && pos.z == (size.z - 1))
+    {
+        tmp_out_offset += (dst.stride_x >> uint(2));
+        STORE4(dst, tmp_out_offset, 1.f);
+    }
+#endif // HAS_BIAS
+}
+#endif // IM2COL_REDUCED
+
+#ifdef COL2IM
+/** This kernel performs a reshaping of the output of the convolution layer.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: 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_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 Z 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]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes)
+ */
+void main(void)
+{
+    uvec2    pos = uvec2(gl_GlobalInvocationID.xy);
+    Image    src = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    uint idx            = pos.x * dst.stride_z + (pos.y / width) * dst.stride_y + (pos.y % width) * dst.stride_x;
+    uint tmp_out_offset = dst.current_offset + (idx >> 2);
+
+    STORE4(dst, tmp_out_offset, LOAD4(src, src.current_offset));
+}
+#endif // COL2IM
+
+#else // DATA_TYPE_FP16
+#error Data type not supported
+#endif // DATA_TYPE_FP16
diff --git a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs
new file mode 100644
index 0000000000..3a31cb80a7
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution1x1.cs
@@ -0,0 +1,275 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+    VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+    uint weights_stride_w;
+    uint weights_depth;
+};
+
+#if defined(DATA_TYPE_FP32)
+precision highp float;
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note The convolution stride x must be passed at compile time using "#define STRIDE_X" e.g. "#define STRIDE_X 1"
+ * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    float pixels  = CONVERT(0, float);
+    uint  z_index = gl_GlobalInvocationID.z;
+    weights.current_offset += z_index * weights_stride_w >> 2;
+    float temp;
+    float temp_weight;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        temp        = LOAD4(src, CURRENT_OFFSET(src));
+        temp_weight = LOAD4(weights, CURRENT_OFFSET(weights));
+        pixels += temp * temp_weight;
+
+        src.current_offset += (src_stride_z >> 2);
+        weights.current_offset += (weights_stride_z >> 2);
+    }
+
+#ifdef BIAS
+    pixels += LOAD4(biases, vector_offset(biases, int(z_index)));
+#endif /* BIAS */
+
+    STORE4(dst, CURRENT_OFFSET(dst), pixels);
+}
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec4, readonly);
+BUFFER_DECLARATION(dst, 2, uvec4, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 2
+#define CONVOLVE(s, w) convolve_stride2(s, w)
+#elif STRIDE_X == 1 /* STRIDE_X == 1 */
+#define CONVOLVE(s, w) convolve_stride1(s, w)
+#else /* STRIDE_X not equals 1 or 2 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 2 */
+
+vec4[2] convolve_stride1(Image src, float w)
+{
+    uvec4 packed_s;
+    vec4  s[2];
+
+    GC_LOAD1_2D_OFFSET(packed_s, src, 0, 0);
+
+    s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y));
+    s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w));
+
+    s[0] *= w;
+    s[1] *= w;
+
+    return s;
+}
+
+vec4[2] convolve_stride2(Image src, float w)
+{
+    uvec4 packed_s;
+    vec4  s[2];
+    vec4  r[2];
+
+    GC_LOAD1_2D_OFFSET(packed_s, src, 0, 0);
+    s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y));
+    s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w));
+
+    r[0] = vec4(s[0].xz, s[1].xz);
+
+    GC_LOAD1_2D_OFFSET(packed_s, src, 8, 0);
+    s[0] = vec4(unpackHalf2x16(packed_s.x), unpackHalf2x16(packed_s.y));
+    s[1] = vec4(unpackHalf2x16(packed_s.z), unpackHalf2x16(packed_s.w));
+
+    r[1] = vec4(s[0].xz, s[1].xz);
+
+    r[0] *= w;
+    r[1] *= w;
+
+    return r;
+}
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note The convolution stride x must be passed at compile time using "#define STRIDE_X" e.g. "#define STRIDE_X 1"
+ * @note In case biases will be added to the convolution "#define HAS_BIAS" has to be passed to append the final matrix with 1 in each row.
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = GC_CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector   biases  = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    vec4 pixels[2];
+    pixels[0] = vec4(0.f);
+    pixels[1] = vec4(0.f);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    uint  packed_w;
+    float w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        GC_LOAD1_3D_OFFSET(packed_w, weights, 0, 0, 0);
+        w = unpackHalf2x16(packed_w).x;
+
+        vec4 r[2] = CONVOLVE(src, w);
+        pixels[0] += r[0];
+        pixels[1] += r[1];
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+
+    GC_LOAD1_1D_OFFSET(packed_b, biases, z_index);
+
+    if(z_index % uint(2) == uint(0))
+    {
+        b = unpackHalf2x16(packed_b).x;
+    }
+    else
+    {
+        b = unpackHalf2x16(packed_b).y;
+    }
+
+    pixels[0] += vec4(b);
+    pixels[1] += vec4(b);
+#endif /* BIAS */
+
+    uvec4 packed_d;
+    packed_d = uvec4(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw),
+                     packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw));
+    GC_STORE1_3D_OFFSET(packed_d, dst, 0, 0, 0);
+}
+#else  /* DATA_TYPE_FP32 */
+#error Data type not supported
+#endif /* DATA_TYPE_FP32 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution3x3.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution3x3.cs
new file mode 100644
index 0000000000..67b92cb8cf
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution3x3.cs
@@ -0,0 +1,1583 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+    VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+    uint weights_stride_w;
+    uint weights_depth;
+};
+
+#define LOAD12(r, name, offset)          \
+    r.x = LOAD4(name, offset);           \
+    r.y = LOAD4(name, offset + uint(1)); \
+    r.z = LOAD4(name, offset + uint(2))
+
+#define LOAD3X3(r, name)                                \
+    r[0] = LOAD4(name, tensor3D_offset(name, 0, 0, 0)); \
+    r[1] = LOAD4(name, tensor3D_offset(name, 1, 0, 0)); \
+    r[2] = LOAD4(name, tensor3D_offset(name, 2, 0, 0)); \
+    r[3] = LOAD4(name, tensor3D_offset(name, 0, 1, 0)); \
+    r[4] = LOAD4(name, tensor3D_offset(name, 1, 1, 0)); \
+    r[5] = LOAD4(name, tensor3D_offset(name, 2, 1, 0)); \
+    r[6] = LOAD4(name, tensor3D_offset(name, 0, 2, 0)); \
+    r[7] = LOAD4(name, tensor3D_offset(name, 1, 2, 0)); \
+    r[8] = LOAD4(name, tensor3D_offset(name, 2, 2, 0))
+
+#if defined(PROCESS_1_ELEMENT)
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    float pixels = CONVERT(0, float);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w >> 2;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        vec3 temp;
+        vec3 w;
+
+        LOAD12(temp, src, offset(src, 0, 0));
+        LOAD12(w, weights, tensor3D_offset(weights, 0, 0, 0));
+
+        pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2];
+
+        LOAD12(temp, src, offset(src, 0, 1));
+        LOAD12(w, weights, tensor3D_offset(weights, 0, 1, 0));
+
+        pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2];
+
+        LOAD12(temp, src, offset(src, 0, 2));
+        LOAD12(w, weights, tensor3D_offset(weights, 0, 2, 0));
+
+        pixels += temp.x * w[0] + temp.y * w[1] + temp.z * w[2];
+
+        src.current_offset += src_stride_z >> 2;
+        weights.current_offset += weights_stride_z >> 2;
+    }
+
+#ifdef BIAS
+    pixels += LOAD4(biases, vector_offset(biases, int(z_index)));
+#endif /* BIAS */
+
+    STORE4(dst, CURRENT_OFFSET(dst), pixels);
+}
+#elif defined(PROCESS_8_ELEMENT)
+BUFFER_DECLARATION(src, 1, vec4, readonly);
+BUFFER_DECLARATION(dst, 2, vec4, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 2
+#define CONVOLVE1x3(offset, w) convolve1x3_stride2(offset, w)
+#elif STRIDE_X == 1 /* STRIDE_X == 1 */
+#define CONVOLVE1x3(offset, w) convolve1x3_stride1(offset, w)
+#else /* STRIDE_X not equals 1 or 2 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 2 */
+
+vec4[2] convolve1x3_stride1(uint offset, vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 tmp[3];
+    vec4 r[2];
+
+    LOAD3(tmp, src, offset);
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r[0] = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    middle = vec4(tmp[1].yzw, tmp[2].x);
+    right  = vec4(tmp[1].zw, tmp[2].xy);
+
+    r[1] = tmp[1] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[2] convolve1x3_stride2(uint offset, vec3 w)
+{
+    vec4 left;
+    vec4 middle;
+    vec4 right;
+    vec4 tmp[3];
+    vec4 r[2];
+
+    LOAD3(tmp, src, offset);
+
+    left   = vec4(tmp[0].xz, tmp[1].xz);
+    middle = vec4(tmp[0].yw, tmp[1].yw);
+    right  = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+
+    r[0] = left * w[0] + middle * w[1] + right * w[2];
+
+    LOAD2(tmp, src, offset + ((uint(3) * src_stride_x) >> 2));
+
+    left   = vec4(tmp[2].xz, tmp[0].xz);
+    middle = vec4(tmp[2].yw, tmp[0].yw);
+    right  = vec4(tmp[2].z, tmp[0].xz, tmp[1].x);
+
+    r[1] = left * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 8 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    vec4 pixels[2];
+    pixels[0] = vec4(0);
+    pixels[1] = vec4(0);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w >> 2;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        vec3 w;
+        vec4 r[2];
+
+        // first line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 0, 0));
+
+        r = CONVOLVE1x3(src.current_offset >> uint(2), w);
+        pixels[0] += r[0];
+        pixels[1] += r[1];
+
+        // second line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 1, 0));
+
+        r = CONVOLVE1x3((src.current_offset + (src_stride_y >> 2)) >> uint(2), w);
+        pixels[0] += r[0];
+        pixels[1] += r[1];
+
+        // third line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 2, 0));
+
+        r = CONVOLVE1x3((src.current_offset + (src_stride_y >> 1)) >> uint(2), w);
+        pixels[0] += r[0];
+        pixels[1] += r[1];
+
+        src.current_offset += src_stride_z >> 2;
+        weights.current_offset += weights_stride_z >> 2;
+    }
+
+#ifdef BIAS
+    float b;
+    LOAD1(b, biases, vector_offset(biases, int(z_index)));
+    pixels[0] += vec4(b);
+    pixels[1] += vec4(b);
+#endif /* BIAS */
+
+    STORE2(dst, dst.current_offset >> uint(2), pixels);
+}
+#elif defined(PROCESS_4_ELEMENT)
+BUFFER_DECLARATION(src, 1, vec4, readonly);
+BUFFER_DECLARATION(dst, 2, vec4, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 2
+#define CONVOLVE1x3(offset, w) convolve1x3_stride2(offset, w)
+#elif STRIDE_X == 1 /* STRIDE_X == 1 */
+#define CONVOLVE1x3(offset, w) convolve1x3_stride1(offset, w)
+#else /* STRIDE_X not equals 1 or 2 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 2 */
+
+vec4 convolve1x3_stride1(uint offset, vec3 w)
+{
+    vec4 tmp[2];
+    vec4 middle;
+    vec4 right;
+
+    LOAD2(tmp, src, offset);
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    tmp[1] = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    return tmp[1];
+}
+
+vec4 convolve1x3_stride2(uint offset, vec3 w)
+{
+    vec4 left;
+    vec4 middle;
+    vec4 right;
+
+    vec4 tmp[3];
+
+    LOAD3(tmp, src, offset);
+
+    left   = vec4(tmp[0].xz, tmp[1].xz);
+    middle = vec4(tmp[0].yw, tmp[1].yw);
+    right  = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+
+    tmp[0] = left * w[0] + middle * w[1] + right * w[2];
+
+    return tmp[0];
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    vec4 pixels;
+    pixels = vec4(0);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w >> 2;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        vec3 w;
+
+        // first line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 0, 0));
+
+        pixels += CONVOLVE1x3(src.current_offset >> uint(2), w);
+
+        // second line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 1, 0));
+
+        pixels += CONVOLVE1x3((src.current_offset + (src_stride_y >> 2)) >> uint(2), w);
+
+        // third line
+        LOAD3(w, weights, tensor3D_offset(weights, 0, 2, 0));
+
+        pixels += CONVOLVE1x3((src.current_offset + (src_stride_y >> 1)) >> uint(2), w);
+
+        src.current_offset += src_stride_z >> 2;
+        weights.current_offset += weights_stride_z >> 2;
+    }
+
+#ifdef BIAS
+    float b;
+    LOAD1(b, biases, vector_offset(biases, int(z_index)));
+    pixels += vec4(b);
+#endif /* BIAS */
+
+    STORE1(dst, dst.current_offset >> uint(2), pixels);
+}
+#elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS)
+BUFFER_DECLARATION(src, 1, vec4, readonly);
+BUFFER_DECLARATION(dst, 2, vec4, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+#define CONVOLVE1x3(left, middle, right, w) convolve1x3_stride1(left, middle, right, w)
+
+vec4 convolve1x3_stride1(vec4 left, vec4 middle, vec4 right, vec3 w)
+{
+    vec4 r;
+
+    r = left * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    vec4 pixels[3];
+    pixels[0] = vec4(0);
+    pixels[1] = vec4(0);
+    pixels[2] = vec4(0);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w >> 2;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        vec3 w[3];
+
+        LOAD3(w[0], weights, tensor3D_offset(weights, 0, 0, 0));
+        LOAD3(w[1], weights, tensor3D_offset(weights, 0, 1, 0));
+        LOAD3(w[2], weights, tensor3D_offset(weights, 0, 2, 0));
+
+        vec4 s[2];
+        vec4 middle;
+        vec4 right;
+        // first line
+        LOAD2(s, src, src.current_offset >> uint(2));
+        middle = vec4(s[0].yzw, s[1].x);
+        right  = vec4(s[0].zw, s[1].xy);
+        pixels[0] += CONVOLVE1x3(s[0], middle, right, w[0]);
+
+        // second line
+        LOAD2(s, src, (src.current_offset + (src_stride_y >> 2)) >> uint(2));
+        middle = vec4(s[0].yzw, s[1].x);
+        right  = vec4(s[0].zw, s[1].xy);
+        pixels[0] += CONVOLVE1x3(s[0], middle, right, w[1]);
+        pixels[1] += CONVOLVE1x3(s[0], middle, right, w[0]);
+
+        // third line
+        LOAD2(s, src, (src.current_offset + (src_stride_y >> 1)) >> uint(2));
+        middle = vec4(s[0].yzw, s[1].x);
+        right  = vec4(s[0].zw, s[1].xy);
+        pixels[0] += CONVOLVE1x3(s[0], middle, right, w[2]);
+        pixels[1] += CONVOLVE1x3(s[0], middle, right, w[1]);
+        pixels[2] += CONVOLVE1x3(s[0], middle, right, w[0]);
+
+        // forth line
+        LOAD2(s, src, (src.current_offset + (uint(3) * (src_stride_y >> 2))) >> uint(2));
+        middle = vec4(s[0].yzw, s[1].x);
+        right  = vec4(s[0].zw, s[1].xy);
+        pixels[1] += CONVOLVE1x3(s[0], middle, right, w[2]);
+        pixels[2] += CONVOLVE1x3(s[0], middle, right, w[1]);
+
+        // fifth line
+        LOAD2(s, src, (src.current_offset + (src_stride_y)) >> uint(2));
+        middle = vec4(s[0].yzw, s[1].x);
+        right  = vec4(s[0].zw, s[1].xy);
+        pixels[2] += CONVOLVE1x3(s[0], middle, right, w[2]);
+
+        src.current_offset += src_stride_z >> 2;
+        weights.current_offset += weights_stride_z >> 2;
+    }
+
+#ifdef BIAS
+    float b;
+    LOAD1(b, biases, vector_offset(biases, int(z_index)));
+
+    pixels[0] += vec4(b);
+    pixels[1] += vec4(b);
+    pixels[2] += vec4(b);
+#endif /* BIAS */
+
+    STORE1(dst, dst.current_offset >> uint(2), pixels[0]);
+    STORE1(dst, (dst.current_offset + (dst_stride_y >> 2)) >> uint(2), pixels[1]);
+    STORE1(dst, (dst.current_offset + (dst_stride_y >> 1)) >> uint(2), pixels[2]);
+}
+#elif defined(PROCESS_X_8ELEMENTS_Y_3ELEMENTS_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec4, readonly);
+BUFFER_DECLARATION(dst, 2, uvec4, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w)
+
+vec4[2] convolve1x3_stride1(vec4 tmp[3], vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 r[2];
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r[0] = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    middle = vec4(tmp[1].yzw, tmp[2].x);
+    right  = vec4(tmp[1].zw, tmp[2].xy);
+
+    r[1] = tmp[1] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[3] load_and_unpack(uint offset)
+{
+    uvec4 packed_s[2];
+    vec4  s[3];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+    ;
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[0].z), unpackHalf2x16(packed_s[0].w));
+    s[2] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    return s;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 8x3 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT_FP16(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases);
+#endif /* BIAS */
+
+    uvec2 packed_d[2];
+    uvec4 vd;
+
+    vec4 pixels[3][2];
+    int  i, j;
+    for(i = 0; i < 3; i++)
+    {
+        for(j = 0; j < 2; j++)
+        {
+            pixels[i][j] = vec4(0);
+        }
+    }
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        uvec2 packed_w[3];
+
+        LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2);
+        LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2);
+        LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2);
+
+        vec3 w[3];
+        w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x);
+        w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x);
+        w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x);
+
+        uvec4 packed_s[2];
+        vec4  s[3];
+        vec4  r[2];
+        uint  offset;
+        // first line
+        offset = src.current_offset >> uint(4);
+        s      = load_and_unpack(offset);
+
+        r = CONVOLVE1x3(s, w[0]);
+        pixels[0][0] += r[0];
+        pixels[0][1] += r[1];
+
+        // second line
+        offset = (src.current_offset + src_stride_y) >> uint(4);
+        s      = load_and_unpack(offset);
+
+        r = CONVOLVE1x3(s, w[1]);
+        pixels[0][0] += r[0];
+        pixels[0][1] += r[1];
+        r = CONVOLVE1x3(s, w[0]);
+        pixels[1][0] += r[0];
+        pixels[1][1] += r[1];
+
+        // third line
+        offset = (src.current_offset + (src_stride_y << 1)) >> uint(4);
+        s      = load_and_unpack(offset);
+
+        r = CONVOLVE1x3(s, w[2]);
+        pixels[0][0] += r[0];
+        pixels[0][1] += r[1];
+        r = CONVOLVE1x3(s, w[1]);
+        pixels[1][0] += r[0];
+        pixels[1][1] += r[1];
+        r = CONVOLVE1x3(s, w[0]);
+        pixels[2][0] += r[0];
+        pixels[2][1] += r[1];
+
+        // forth line
+        offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(4);
+        s      = load_and_unpack(offset);
+
+        r = CONVOLVE1x3(s, w[2]);
+        pixels[1][0] += r[0];
+        pixels[1][1] += r[1];
+        r = CONVOLVE1x3(s, w[1]);
+        pixels[2][0] += r[0];
+        pixels[2][1] += r[1];
+
+        // fifth line
+        offset = (src.current_offset + (src_stride_y << 2)) >> uint(4);
+        s      = load_and_unpack(offset);
+
+        r = CONVOLVE1x3(s, w[2]);
+        pixels[2][0] += r[0];
+        pixels[2][1] += r[1];
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+    LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2);
+
+    if(z_index % uint(2) == uint(0))
+    {
+        b = unpackHalf2x16(packed_b).x;
+    }
+    else
+    {
+        b = unpackHalf2x16(packed_b).y;
+    }
+
+    for(i = 0; i < 3; i++)
+    {
+        for(j = 0; j < 2; j++)
+        {
+            pixels[i][j] += vec4(b);
+        }
+    }
+#endif /* BIAS */
+
+    packed_d[0] = uvec2(packHalf2x16(pixels[0][0].xy), packHalf2x16(pixels[0][0].zw));
+    packed_d[1] = uvec2(packHalf2x16(pixels[0][1].xy), packHalf2x16(pixels[0][1].zw));
+    vd          = uvec4(packed_d[0], packed_d[1]);
+    STORE1(dst, dst.current_offset >> uint(4), vd);
+
+    packed_d[0] = uvec2(packHalf2x16(pixels[1][0].xy), packHalf2x16(pixels[1][0].zw));
+    packed_d[1] = uvec2(packHalf2x16(pixels[1][1].xy), packHalf2x16(pixels[1][1].zw));
+    vd          = uvec4(packed_d[0], packed_d[1]);
+    STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(4), vd);
+
+    packed_d[0] = uvec2(packHalf2x16(pixels[2][0].xy), packHalf2x16(pixels[2][0].zw));
+    packed_d[1] = uvec2(packHalf2x16(pixels[2][1].xy), packHalf2x16(pixels[2][1].zw));
+    vd          = uvec4(packed_d[0], packed_d[1]);
+    STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(4), vd);
+}
+#elif defined(PROCESS_X_4ELEMENTS_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 2
+#define CONVOLVE1x3(s, w) convolve1x3_stride2(s, w)
+#define LOAD_AND_UNPACK(offset) load_and_unpack_stride2(offset)
+#elif STRIDE_X == 1 /* STRIDE_X == 1 */
+#define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w)
+#define LOAD_AND_UNPACK(offset) load_and_unpack_stride1(offset)
+#else /* STRIDE_X not equals 1 or 2 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 2 */
+
+vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 r;
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4 convolve1x3_stride2(vec4 tmp[3], vec3 w)
+{
+    vec4 left;
+    vec4 middle;
+    vec4 right;
+    vec4 r;
+
+    left   = vec4(tmp[0].xz, tmp[1].xz);
+    middle = vec4(tmp[0].yw, tmp[1].yw);
+    right  = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+
+    r = left * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[2] load_and_unpack_stride1(uint offset)
+{
+    uvec2 packed_s[2];
+    vec4  s[2];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    return s;
+}
+
+vec4[3] load_and_unpack_stride2(uint offset)
+{
+    uvec2 packed_s[3];
+    vec4  s[3];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+    LOAD1(packed_s[2], src, offset + uint(2));
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+    s[2] = vec4(unpackHalf2x16(packed_s[2].x), unpackHalf2x16(packed_s[2].y));
+
+    return s;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT_FP16(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases);
+#endif /* BIAS */
+
+    uvec2 packed_d;
+
+    vec4 pixels = vec4(0);
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        uvec2 packed_w[3];
+
+        LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2);
+        LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2);
+        LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2);
+
+        vec3 w[3];
+        w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x);
+        w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x);
+        w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x);
+
+#if STRIDE_X == 2
+        vec4 s[3];
+#elif STRIDE_X == 1 /* STRIDE_X == 1 */
+        vec4 s[2];
+#else               /* STRIDE_X not equals 1 or 2 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 2 */
+        vec4 r;
+        uint offset;
+        // first line
+        offset = src.current_offset >> uint(3);
+        s      = LOAD_AND_UNPACK(offset);
+
+        pixels += CONVOLVE1x3(s, w[0]);
+
+        // second line
+        offset = (src.current_offset + src_stride_y) >> uint(3);
+        s      = LOAD_AND_UNPACK(offset);
+
+        pixels += CONVOLVE1x3(s, w[1]);
+
+        // third line
+        offset = (src.current_offset + (src_stride_y << 1)) >> uint(3);
+        s      = LOAD_AND_UNPACK(offset);
+
+        pixels += CONVOLVE1x3(s, w[2]);
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+    LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2);
+
+    if(z_index % uint(2) == uint(0))
+    {
+        b = unpackHalf2x16(packed_b).x;
+    }
+    else
+    {
+        b = unpackHalf2x16(packed_b).y;
+    }
+
+    pixels += vec4(b);
+#endif /* BIAS */
+
+    packed_d = uvec2(packHalf2x16(pixels.xy), packHalf2x16(pixels.zw));
+    STORE1(dst, dst.current_offset >> uint(3), packed_d);
+}
+#elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w)
+
+vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 r;
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[2] load_and_unpack(uint offset)
+{
+    uvec2 packed_s[2];
+    vec4  s[2];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    return s;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT_FP16(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases);
+#endif /* BIAS */
+
+    uvec2 packed_d;
+
+    vec4 pixels[3];
+    int  i;
+
+    for(i = 0; i < 3; i++)
+    {
+        pixels[i] = vec4(0);
+    }
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        uvec2 packed_w[3];
+
+        LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2);
+        LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2);
+        LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2);
+
+        vec3 w[3];
+        w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x);
+        w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x);
+        w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x);
+
+        vec4 s[2];
+        vec4 r;
+        uint offset;
+        // first line
+        offset = src.current_offset >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[0]);
+
+        // second line
+        offset = (src.current_offset + src_stride_y) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[1]);
+        pixels[1] += CONVOLVE1x3(s, w[0]);
+
+        // third line
+        offset = (src.current_offset + (src_stride_y << 1)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[2]);
+        pixels[1] += CONVOLVE1x3(s, w[1]);
+        pixels[2] += CONVOLVE1x3(s, w[0]);
+
+        // forth line
+        offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[1] += CONVOLVE1x3(s, w[2]);
+        pixels[2] += CONVOLVE1x3(s, w[1]);
+
+        // fifth line
+        offset = (src.current_offset + (src_stride_y << 2)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[2] += CONVOLVE1x3(s, w[2]);
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+    LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2);
+
+    if(z_index % uint(2) == uint(0))
+    {
+        b = unpackHalf2x16(packed_b).x;
+    }
+    else
+    {
+        b = unpackHalf2x16(packed_b).y;
+    }
+
+    for(i = 0; i < 3; i++)
+    {
+        pixels[i] += vec4(b);
+    }
+#endif /* BIAS */
+
+    packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw));
+    STORE1(dst, dst.current_offset >> uint(3), packed_d);
+
+    packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw));
+    STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d);
+
+    packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw));
+    STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d);
+}
+#elif defined(PROCESS_X_4ELEMENTS_Y_4ELEMENTS_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w)
+
+vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 r;
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[2] load_and_unpack(uint offset)
+{
+    uvec2 packed_s[2];
+    vec4  s[2];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    return s;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x4 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT_FP16(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases);
+#endif /* BIAS */
+
+    uvec2 packed_d;
+
+    vec4 pixels[4];
+    int  i;
+
+    for(i = 0; i < 4; i++)
+    {
+        pixels[i] = vec4(0);
+    }
+
+    uint z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        // load 3 weights once
+        uvec2 packed_w[3];
+
+        LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2);
+        LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2);
+        LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2);
+
+        vec3 w[3];
+        w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x);
+        w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x);
+        w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x);
+
+        vec4 s[2];
+        vec4 r;
+        uint offset;
+        // first line
+        offset = src.current_offset >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[0]);
+
+        // second line
+        offset = (src.current_offset + src_stride_y) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[1]);
+        pixels[1] += CONVOLVE1x3(s, w[0]);
+
+        // third line
+        offset = (src.current_offset + (src_stride_y << 1)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[0] += CONVOLVE1x3(s, w[2]);
+        pixels[1] += CONVOLVE1x3(s, w[1]);
+        pixels[2] += CONVOLVE1x3(s, w[0]);
+
+        // forth line
+        offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[1] += CONVOLVE1x3(s, w[2]);
+        pixels[2] += CONVOLVE1x3(s, w[1]);
+        pixels[3] += CONVOLVE1x3(s, w[0]);
+
+        // fifth line
+        offset = (src.current_offset + (src_stride_y << 2)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[2] += CONVOLVE1x3(s, w[2]);
+        pixels[3] += CONVOLVE1x3(s, w[1]);
+
+        // sixth line
+        offset = (src.current_offset + uint(5) * (src_stride_y)) >> uint(3);
+        s      = load_and_unpack(offset);
+
+        pixels[3] += CONVOLVE1x3(s, w[2]);
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+    LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2);
+
+    if(z_index % uint(2) == uint(0))
+    {
+        b = unpackHalf2x16(packed_b).x;
+    }
+    else
+    {
+        b = unpackHalf2x16(packed_b).y;
+    }
+
+    for(i = 0; i < 4; i++)
+    {
+        pixels[i] += vec4(b);
+    }
+#endif /* BIAS */
+
+    packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw));
+    STORE1(dst, dst.current_offset >> uint(3), packed_d);
+
+    packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw));
+    STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d);
+
+    packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw));
+    STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d);
+
+    packed_d = uvec2(packHalf2x16(pixels[3].xy), packHalf2x16(pixels[3].zw));
+    STORE1(dst, (dst.current_offset + uint(3) * (dst_stride_y)) >> uint(3), packed_d);
+}
+#elif defined(PROCESS_X_4ELEMENTS_Y_3ELEMENTS_Z_2ELEMENTS_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#define CONVOLVE1x3(s, w) convolve1x3_stride1(s, w)
+
+vec4 convolve1x3_stride1(vec4 tmp[2], vec3 w)
+{
+    vec4 middle;
+    vec4 right;
+    vec4 r;
+
+    middle = vec4(tmp[0].yzw, tmp[1].x);
+    right  = vec4(tmp[0].zw, tmp[1].xy);
+
+    r = tmp[0] * w[0] + middle * w[1] + right * w[2];
+
+    return r;
+}
+
+vec4[2] load_and_unpack(uint offset)
+{
+    uvec2 packed_s[2];
+    vec4  s[2];
+
+    LOAD1(packed_s[0], src, offset);
+    LOAD1(packed_s[1], src, offset + uint(1));
+
+    s[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    s[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    return s;
+}
+
+/** An optimized direct convolution 3x3 OpenGL ES compute shader for process 4x3x2 elements at once
+ *
+ * @note This OpenGL ES shader works with stride_x = 1 and 2
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT_FP16(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+#ifdef BIAS
+    Vector   biases  = CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(biases);
+#endif /* BIAS */
+
+    uvec2 packed_d;
+
+    vec4 pixels[3];
+    int  i;
+
+    uint z_base_index = gl_GlobalInvocationID.z << 1;
+
+    // store orginal src current offset
+    uint s_offset = src.current_offset;
+
+    weights.current_offset += z_base_index * weights_stride_w;
+
+    for(int z = 0; z < 2; ++z)
+    {
+        uint z_index = z_base_index + uint(z);
+
+        src.current_offset = s_offset;
+        //weights.current_offset = z_index * weights_stride_w;
+
+        for(i = 0; i < 3; i++)
+        {
+            pixels[i] = vec4(0);
+        }
+
+        for(int d = 0; d < int(weights_depth); ++d)
+        {
+            // load 3 weights once
+            uvec2 packed_w[3];
+
+            LOAD2(packed_w[0], weights, tensor3D_offset_fp16(weights, 0, 0, 0) >> 2);
+            LOAD2(packed_w[1], weights, tensor3D_offset_fp16(weights, 0, 1, 0) >> 2);
+            LOAD2(packed_w[2], weights, tensor3D_offset_fp16(weights, 0, 2, 0) >> 2);
+
+            vec3 w[3];
+            w[0] = vec3(unpackHalf2x16(packed_w[0].x), unpackHalf2x16(packed_w[0].y).x);
+            w[1] = vec3(unpackHalf2x16(packed_w[1].x), unpackHalf2x16(packed_w[1].y).x);
+            w[2] = vec3(unpackHalf2x16(packed_w[2].x), unpackHalf2x16(packed_w[2].y).x);
+
+            vec4 s[2];
+            vec4 r;
+            uint offset;
+            // first line
+            offset = src.current_offset >> uint(3);
+            s      = load_and_unpack(offset);
+
+            pixels[0] += CONVOLVE1x3(s, w[0]);
+
+            // second line
+            offset = (src.current_offset + src_stride_y) >> uint(3);
+            s      = load_and_unpack(offset);
+
+            pixels[0] += CONVOLVE1x3(s, w[1]);
+            pixels[1] += CONVOLVE1x3(s, w[0]);
+
+            // third line
+            offset = (src.current_offset + (src_stride_y << 1)) >> uint(3);
+            s      = load_and_unpack(offset);
+
+            pixels[0] += CONVOLVE1x3(s, w[2]);
+            pixels[1] += CONVOLVE1x3(s, w[1]);
+            pixels[2] += CONVOLVE1x3(s, w[0]);
+
+            // forth line
+            offset = (src.current_offset + uint(3) * (src_stride_y)) >> uint(3);
+            s      = load_and_unpack(offset);
+
+            pixels[1] += CONVOLVE1x3(s, w[2]);
+            pixels[2] += CONVOLVE1x3(s, w[1]);
+
+            // fifth line
+            offset = (src.current_offset + (src_stride_y << 2)) >> uint(3);
+            s      = load_and_unpack(offset);
+
+            pixels[2] += CONVOLVE1x3(s, w[2]);
+
+            src.current_offset += src_stride_z;
+            weights.current_offset += weights_stride_z;
+        }
+
+#ifdef BIAS
+        uint  packed_b;
+        float b;
+        LOAD1(packed_b, biases, vector_offset_fp16(biases, int(z_index)) >> 2);
+
+        if(z_index % uint(2) == uint(0))
+        {
+            b = unpackHalf2x16(packed_b).x;
+        }
+        else
+        {
+            b = unpackHalf2x16(packed_b).y;
+        }
+
+        for(i = 0; i < 3; i++)
+        {
+            pixels[i] += vec4(b);
+        }
+#endif /* BIAS */
+
+        packed_d = uvec2(packHalf2x16(pixels[0].xy), packHalf2x16(pixels[0].zw));
+        STORE1(dst, dst.current_offset >> uint(3), packed_d);
+
+        packed_d = uvec2(packHalf2x16(pixels[1].xy), packHalf2x16(pixels[1].zw));
+        STORE1(dst, (dst.current_offset + dst_stride_y) >> uint(3), packed_d);
+
+        packed_d = uvec2(packHalf2x16(pixels[2].xy), packHalf2x16(pixels[2].zw));
+        STORE1(dst, (dst.current_offset + (dst_stride_y << 1)) >> uint(3), packed_d);
+
+        dst.current_offset += dst_stride_z;
+    }
+}
+#endif /* PROCESS_1_ELEMENT */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs
new file mode 100644
index 0000000000..4fdbf0d19e
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/direct_convolution5x5.cs
@@ -0,0 +1,313 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    TENSOR3D_PARAM_DECLARATION(weights);
+#ifdef BIAS
+    VECTOR_PARAM_DECLARATION(biases);
+#endif /* BIAS */
+    uint weights_stride_w;
+    uint weights_depth;
+};
+
+#ifdef DATA_TYPE_FP32
+
+precision highp float;
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+BUFFER_DECLARATION(weights, 3, float, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, float, readonly);
+#endif /* BIAS */
+
+#define LOAD20(r, name, offset)           \
+    r[0] = LOAD4(name, offset);           \
+    r[1] = LOAD4(name, offset + uint(1)); \
+    r[2] = LOAD4(name, offset + uint(2)); \
+    r[3] = LOAD4(name, offset + uint(3)); \
+    r[4] = LOAD4(name, offset + uint(4))
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                          Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                          Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    float pixels  = CONVERT(0, float);
+    uint  z_index = gl_GlobalInvocationID.z;
+    weights.current_offset += z_index * weights_stride_w >> 2;
+    float temp[5];
+    float temp_weight[5];
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        LOAD20(temp, src, offset(src, 0, 0));
+        LOAD20(temp_weight, weights, tensor3D_offset(weights, 0, 0, 0));
+        pixels += temp[0] * temp_weight[0] + temp[1] * temp_weight[1] + temp[2] * temp_weight[2] + temp[3] * temp_weight[3] + temp[4] * temp_weight[4];
+
+        LOAD20(temp, src, offset(src, 0, 1));
+        LOAD20(temp_weight, weights, tensor3D_offset(weights, 0, 1, 0));
+        pixels += temp[0] * temp_weight[0] + temp[1] * temp_weight[1] + temp[2] * temp_weight[2] + temp[3] * temp_weight[3] + temp[4] * temp_weight[4];
+
+        LOAD20(temp, src, offset(src, 0, 2));
+        LOAD20(temp_weight, weights, tensor3D_offset(weights, 0, 2, 0));
+        pixels += temp[0] * temp_weight[0] + temp[1] * temp_weight[1] + temp[2] * temp_weight[2] + temp[3] * temp_weight[3] + temp[4] * temp_weight[4];
+
+        LOAD20(temp, src, offset(src, 0, 3));
+        LOAD20(temp_weight, weights, tensor3D_offset(weights, 0, 3, 0));
+        pixels += temp[0] * temp_weight[0] + temp[1] * temp_weight[1] + temp[2] * temp_weight[2] + temp[3] * temp_weight[3] + temp[4] * temp_weight[4];
+
+        LOAD20(temp, src, offset(src, 0, 4));
+        LOAD20(temp_weight, weights, tensor3D_offset(weights, 0, 4, 0));
+        pixels += temp[0] * temp_weight[0] + temp[1] * temp_weight[1] + temp[2] * temp_weight[2] + temp[3] * temp_weight[3] + temp[4] * temp_weight[4];
+
+        src.current_offset += (src_stride_z >> 2);
+        weights.current_offset += (weights_stride_z >> 2);
+    }
+
+#ifdef BIAS
+    pixels += LOAD4(biases, vector_offset(biases, int(z_index)));
+#endif /* BIAS */
+
+    STORE4(dst, CURRENT_OFFSET(dst), pixels);
+}
+
+#elif defined(DATA_TYPE_FP16)
+
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+BUFFER_DECLARATION(weights, 3, uint, readonly);
+#ifdef BIAS
+BUFFER_DECLARATION(biases, 4, uint, readonly);
+#endif /* BIAS */
+
+#if STRIDE_X == 1
+#define LOAD_SRC(src, row) load_src_stride1(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride1(src, weight)
+#elif STRIDE_X == 2 /* STRIDE_X == 1 */
+#define LOAD_SRC(src, row) load_src_stride2(src, row)
+#define CONVOLVE1x5(src, weight) convolve1x5_stride2(src, weight)
+#else /* STRDIDE_X == 1 */
+#error STRIDE_X larger than 2 is not supported
+#endif /* STRIDE_X == 1 */
+
+vec4[2] load_src_stride1(Image src, int row)
+{
+    uvec2 packed[2];
+    vec4  ret[2];
+
+    GC_LOAD2_2D_OFFSET(packed, src, 0, row);
+
+    ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+    ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+
+    return ret;
+}
+
+vec4[3] load_src_stride2(Image src, int row)
+{
+    uvec2 packed[3];
+    vec4  ret[3];
+
+    GC_LOAD3_2D_OFFSET(packed, src, 0, row);
+
+    ret[0] = vec4(unpackHalf2x16(packed[0].x), unpackHalf2x16(packed[0].y));
+    ret[1] = vec4(unpackHalf2x16(packed[1].x), unpackHalf2x16(packed[1].y));
+    ret[2] = vec4(unpackHalf2x16(packed[2].x), unpackHalf2x16(packed[2].y));
+
+    return ret;
+}
+
+vec2[3] load_weight(Tensor3D weights, int row)
+{
+    uvec3 packed_w;
+    vec2  ret[3];
+
+    GC_LOAD3_3D_OFFSET(packed_w, weights, 0, row, 0);
+
+    ret[0] = vec2(unpackHalf2x16(packed_w[0]));
+    ret[1] = vec2(unpackHalf2x16(packed_w[1]));
+    ret[2] = vec2(unpackHalf2x16(packed_w[2]));
+
+    return ret;
+}
+
+// output 4 element per thread
+vec4 convolve1x5_stride1(vec4 tmp[2], vec2 w[3])
+{
+    vec4 src0 = tmp[0];
+    vec4 src1 = vec4(tmp[0].yzw, tmp[1].x);
+    vec4 src2 = vec4(tmp[0].zw, tmp[1].xy);
+    vec4 src3 = vec4(tmp[0].w, tmp[1].xyz);
+    vec4 src4 = tmp[1];
+    vec4 ret  = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+    return ret;
+}
+
+vec4 convolve1x5_stride2(vec4 tmp[3], vec2 w[3])
+{
+    vec4 src0 = vec4(tmp[0].xz, tmp[1].xz);
+    vec4 src1 = vec4(tmp[0].yw, tmp[1].yw);
+    vec4 src2 = vec4(tmp[0].z, tmp[1].xz, tmp[2].x);
+    vec4 src3 = vec4(tmp[0].w, tmp[1].yw, tmp[2].y);
+    vec4 src4 = vec4(tmp[1].x, tmp[1].z, tmp[2].xz);
+    vec4 ret  = src0 * w[0].x + src1 * w[0].y + src2 * w[1].x + src3 * w[1].y + src4 * w[2].x;
+
+    return ret;
+}
+
+/** This kernel performs a direct convolution to convolve the low three dimensions.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ * @note If biases are used then "define HAS_BIAS" has to be passed at compile time
+ *
+ * @param[in]  src_ptr                               Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes     The offset of the first element in the source tensor
+ * @param[out] dst_ptr                               Pointer to the destination tensor. Supported data types: 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_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                            dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+ * @param[out] weights_ptr                           Pointer to the weights tensor. Supported data types: 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_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+ * @param[in]  weights_step_z                        weights_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+ * @param[in]  biases_ptr                            Pointer to the biases tensor. Same as @p src_ptr
+ * @param[in]  biases_stride_x                       Stride of the biases tensor in X dimension (in bytes)
+ * @param[in]  biases_step_x                         biases_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  biases_offset_first_element_in_bytes  The offset of the first element in the biases tensor
+ * @param[in]  weights_stride_w                      Stride of the weights tensor in the 4th dimension
+ * @param[in]  weights_depth                         The third dimensions of the weights tensors
+ */
+void main()
+{
+    Image    src     = GC_CONVERT_TO_IMAGE_STRUCT(src);
+    Tensor3D weights = GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(weights);
+    Tensor3D dst     = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+#ifdef BIAS
+    Vector   biases  = GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif /* BIAS */
+
+    vec4  res = vec4(0);
+    vec2  w[3];
+    vec4  s[STRIDE_X + 1];
+    uvec2 packed_d;
+    uint  z_index = gl_GlobalInvocationID.z;
+
+    weights.current_offset += z_index * weights_stride_w;
+
+    for(int d = 0; d < int(weights_depth); ++d)
+    {
+        for(int row = 0; row < 5; row++)
+        {
+            w = load_weight(weights, row);
+            s = LOAD_SRC(src, row);
+            res += CONVOLVE1x5(s, w);
+        }
+
+        src.current_offset += src_stride_z;
+        weights.current_offset += weights_stride_z;
+    }
+
+#ifdef BIAS
+    uint  packed_b;
+    float b;
+
+    GC_LOAD1_1D_OFFSET(packed_b, biases, z_index);
+    b = (z_index % uint(2) == uint(0)) ? unpackHalf2x16(packed_b).x : unpackHalf2x16(packed_b).y;
+    res += vec4(b);
+#endif /* BIAS */
+
+    packed_d = uvec2(packHalf2x16(res.xy), packHalf2x16(res.zw));
+    GC_STORE1_3D_OFFSET(packed_d, dst, 0, 0, 0);
+}
+
+#else /* DATA_TYPE_FP16 */
+#error Data type not supported
+#endif /* DATA_TYPE_FP16 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/dropout.cs b/src/core/GLES_COMPUTE/cs_shaders/dropout.cs
new file mode 100644
index 0000000000..54e08b1306
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/dropout.cs
@@ -0,0 +1,204 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(mask);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+uint hash(uint x)
+{
+    x += (x << 10u);
+    x ^= (x >> 6u);
+    x += (x << 3u);
+    x ^= (x >> 11u);
+    x += (x << 15u);
+    return x;
+}
+
+uint hash(uvec3 v)
+{
+    return hash(v.x ^ hash(v.y) ^ hash(v.z));
+}
+
+float float_construct(uint m)
+{
+    const uint ieee_mantissa = 0x007FFFFFu;
+    const uint ieee_one      = 0x3F800000u;
+
+    m &= ieee_mantissa;
+    m |= ieee_one;
+
+    float f = uintBitsToFloat(m);
+    return f - 1.0;
+}
+
+float rand(vec3 v, float seed)
+{
+    return float_construct(hash(floatBitsToUint(v + seed)));
+}
+
+#ifdef DATA_TYPE_FP32
+
+precision highp float;
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(mask, 2, float, );
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+
+/** Dropout is used to improve over-fit on neural networks.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP32"
+ *
+ * @param[in]  src_ptr                            Pointer to the source tensor. Supported data types: F32
+ * @param[in]  src_stride_x                       Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                         src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                       Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] mask_ptr                           Pointer to the mask tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mask_stride_x                      Stride of the mask tensor in X dimension (in bytes)
+ * @param[in]  mask_step_x                        mask_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mask_stride_y                      Stride of the mask tensor in Y dimension (in bytes)
+ * @param[in]  mask_step_y                        mask_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  mask_stride_z                      Stride of the mask tensor in Z dimension (in bytes)
+ * @param[in]  mask_step_z                        mask_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  mask_offset_first_element_in_bytes The offset of the first element in the mask tensor
+ * @param[out] dst_ptr                            Pointer to the destination tensor. Supported data types: 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_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src  = GC_CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D mask = GC_CONVERT_TO_TENSOR3D_STRUCT(mask);
+    Tensor3D dst  = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float random  = 0.f;
+    float inputv  = 0.f;
+    float maskv   = 0.f;
+    float outputv = 0.f;
+
+#ifdef FORWARD
+    random = rand(vec3(gl_GlobalInvocationID.xyz), SEED);
+    maskv  = (random > RATIO) ? 1.f : 0.f;
+    GC_STORE1_3D_OFFSET(maskv, mask, 0, 0, 0);
+#else  /* FORWARD */
+    GC_LOAD1_3D_OFFSET(maskv, mask, 0, 0, 0);
+#endif /* FORWARD */
+
+    GC_LOAD1_3D_OFFSET(inputv, src, 0, 0, 0);
+    outputv = maskv * inputv * float(SCALE);
+    GC_STORE1_3D_OFFSET(outputv, dst, 0, 0, 0);
+}
+
+#elif defined(DATA_TYPE_FP16)
+
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(mask, 2, uint, );
+BUFFER_DECLARATION(dst, 3, uint, writeonly);
+
+/** Dropout is used to improve over-fit on neural networks.
+ *
+ * @note The data type must be passed at compile time using "#define DATA_TYPE_FP16"
+ *
+ * @param[in]  src_ptr                            Pointer to the source tensor. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes  The offset of the first element in the source tensor
+ * @param[out] mask_ptr                           Pointer to the mask tensor. Supported data types: same as @p src_ptr
+ * @param[in]  mask_stride_x                      Stride of the mask tensor in X dimension (in bytes)
+ * @param[in]  mask_step_x                        mask_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mask_stride_y                      Stride of the mask tensor in Y dimension (in bytes)
+ * @param[in]  mask_step_y                        mask_stride_y * number of elements along y processed per workitem(in bytes)
+ * @param[in]  mask_stride_z                      Stride of the mask tensor in Z dimension (in bytes)
+ * @param[in]  mask_step_z                        mask_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  mask_offset_first_element_in_bytes The offset of the first element in the mask tensor
+ * @param[out] dst_ptr                            Pointer to the destination tensor. Supported data types: 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_stride_y                       Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_stride_y * number of elements along Z 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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes  The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src  = GC_CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D mask = GC_CONVERT_TO_TENSOR3D_STRUCT(mask);
+    Tensor3D dst  = GC_CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float random1    = 0.f;
+    float random2    = 0.f;
+    uint  inputv     = uint(0);
+    uint  outputv    = uint(0);
+    uint  maskv      = uint(0);
+    vec2  input_vec  = vec2(0, 0);
+    vec2  output_vec = vec2(0, 0);
+    vec2  mask_vec   = vec2(0, 0);
+
+#ifdef FORWARD
+    random1          = rand(vec3(gl_GlobalInvocationID.xyz), SEED);
+    random2          = rand(vec3(float(gl_GlobalInvocationID.x) + 0.5f, gl_GlobalInvocationID.yz), SEED);
+    mask_vec.x       = (random1 > RATIO) ? 1.f : 0.f;
+    mask_vec.y       = (random2 > RATIO) ? 1.f : 0.f;
+    maskv            = packHalf2x16(mask_vec);
+    GC_STORE1_3D_OFFSET(maskv, mask, 0, 0, 0);
+#else  /* FORWARD */
+    GC_LOAD1_3D_OFFSET(maskv, mask, 0, 0, 0);
+    mask_vec = unpackHalf2x16(maskv);
+#endif /* FORWARD */
+
+    GC_LOAD1_3D_OFFSET(inputv, src, 0, 0, 0);
+
+    input_vec  = unpackHalf2x16(inputv);
+    output_vec = mask_vec * input_vec * float(SCALE);
+    outputv    = packHalf2x16(output_vec);
+
+    GC_STORE1_3D_OFFSET(outputv, dst, 0, 0, 0);
+}
+
+#else /* DATA_TYPE_FP32 */
+
+#endif /* DATA_TYPE_FP32 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/fill_border.cs b/src/core/GLES_COMPUTE/cs_shaders/fill_border.cs
new file mode 100644
index 0000000000..01a39866c7
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/fill_border.cs
@@ -0,0 +1,553 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#if defined(DATA_TYPE_FP32)
+#ifdef FILL_IMAGE_BORDERS_REPLICATE
+BUFFER_DECLARATION(buf, 1, float, restrict);
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(buf);
+    uint width;
+    uint height;
+    int  start_pos_x;
+    int  start_pos_y;
+};
+
+/** Fill N pixel of the padding edge of a single channel image by replicating the closest valid pixel.
+ *
+ * @attention  The border size for top, bottom, left, right needs to be passed at the compile time.
+ * e.g. BORDER_SIZE_TOP=0 BORDER_SIZE_BOTTOM=2 BORDER_SIZE_LEFT=0 BORDER_SIZE_RIGHT=2
+ *
+ * @param[in,out] buf_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in]     buf_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]     buf_step_x                        buf_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     buf_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]     buf_step_y                        buf_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     buf_stride_z                      Stride between images if batching images (in bytes)
+ * @param[in]     buf_step_z                        buf_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     buf_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]     width                             Width of the valid region of the image
+ * @param[in]     height                            Height of the valid region of the image
+ * @param[in]     start_pos_x                       X coordinate indicating the start point of the valid region
+ * @param[in]     start_pos_y                       Y coordinate indicating the start point of the valid region
+ */
+void main()
+{
+    Image buf = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(buf);
+
+    // Update pointer to point to the starting point of the valid region
+    buf.current_offset = uint(int(buf.current_offset) + ((start_pos_y * int(buf_stride_y) + start_pos_x * int(buf_stride_x)) >> 2));
+
+    int total_width = BORDER_SIZE_LEFT + int(width) + BORDER_SIZE_RIGHT;
+    int gid0        = int(gl_GlobalInvocationID.x);
+    int gidH        = gid0 - total_width;
+    int gidW        = gid0 - BORDER_SIZE_LEFT;
+
+    if(gidH >= 0)
+    {
+        // Handle left border
+        float left_val = LOAD4(buf, offset(buf, 0, gidH));
+        for(int i = -BORDER_SIZE_LEFT; i < 0; ++i)
+        {
+            STORE4(buf, offset(buf, i, gidH), left_val);
+        }
+        // Handle right border
+        float right_val = LOAD4(buf, offset(buf, int(width) - 1, gidH));
+        for(int i = 0; i < BORDER_SIZE_RIGHT; ++i)
+        {
+            STORE4(buf, offset(buf, int(width) + i, gidH), right_val);
+        }
+    }
+    else
+    {
+        // Get value for corners
+        int val_idx = gidW;
+        if(gidW < 0 || gidW > (int(width) - 1))
+        {
+            val_idx = gidW < 0 ? 0 : int(width) - 1;
+        }
+
+        // Handle top border
+        float top_val = LOAD4(buf, offset(buf, val_idx, 0));
+        for(int i = -BORDER_SIZE_TOP; i < 0; ++i)
+        {
+            STORE4(buf, offset(buf, gidW, i), top_val);
+        }
+        // Handle bottom border
+        float bottom_val = LOAD4(buf, offset(buf, val_idx, int(height) - 1));
+        for(int i = 0; i < BORDER_SIZE_BOTTOM; ++i)
+        {
+            STORE4(buf, offset(buf, gidW, int(height) + i), bottom_val);
+        }
+    }
+}
+#endif /* FILL_IMAGE_BORDERS_REPLICATE */
+
+#ifdef FILL_IMAGE_BORDERS_CONSTANT
+BUFFER_DECLARATION(buf, 1, float, writeonly);
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(buf);
+    uint  width;
+    uint  height;
+    int   start_pos_x;
+    int   start_pos_y;
+    float constant_value;
+};
+
+/** Fill N pixels of the padding edge of a single channel image with a constant value.
+ *
+ * @attention  The border size for top, bottom, left, right needs to be passed at the compile time.
+ * e.g. BORDER_SIZE_TOP=0 BORDER_SIZE_BOTTOM=2 BORDER_SIZE_LEFT=0 BORDER_SIZE_RIGHT=2
+ *
+ * @param[out] buf_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in]  buf_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  buf_step_x                        buf_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  buf_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  buf_step_y                        buf_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  buf_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]  width                             Width of the valid region of the image
+ * @param[in]  height                            Height of the valid region of the image
+ * @param[in]  start_pos_x                       X coordinate indicating the start point of the valid region
+ * @param[in]  start_pos_y                       Y coordinate indicating the start point of the valid region
+ * @param[in]  constant_value                    Constant value to use to fill the edges
+ */
+void main()
+{
+    Image buf = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(buf);
+
+    // Update pointer to point to the starting point of the valid region
+    buf.current_offset = uint(int(buf.current_offset) + ((start_pos_y * int(buf_stride_y) + start_pos_x * int(buf_stride_x)) >> 2));
+
+    int total_width = BORDER_SIZE_LEFT + int(width) + BORDER_SIZE_RIGHT;
+    int gid0        = int(gl_GlobalInvocationID.x);
+    int gidH        = gid0 - total_width;
+    int gidW        = gid0 - BORDER_SIZE_LEFT;
+
+    if(gidH >= 0)
+    {
+        // Handle left border
+        for(int i = -BORDER_SIZE_LEFT; i < 0; ++i)
+        {
+            STORE1(buf, offset(buf, i, gidH), constant_value);
+        }
+        // Handle right border
+        for(int i = 0; i < BORDER_SIZE_RIGHT; ++i)
+        {
+            STORE1(buf, offset(buf, int(width) + i, gidH), constant_value);
+        }
+    }
+    else
+    {
+        // Handle top border
+        for(int i = -BORDER_SIZE_TOP; i < 0; ++i)
+        {
+            STORE1(buf, offset(buf, gidW, i), constant_value);
+        }
+        // Handle bottom border
+        for(int i = 0; i < BORDER_SIZE_BOTTOM; ++i)
+        {
+            STORE1(buf, offset(buf, gidW, int(height) + i), constant_value);
+        }
+    }
+}
+#endif /* FILL_IMAGE_BORDERS_CONSTANT */
+
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+
+#ifdef FILL_IMAGE_BORDERS_REPLICATE
+BUFFER_DECLARATION(buf, 1, uint, restrict);
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(buf);
+    uint width;
+    uint height;
+    int  start_pos_x;
+    int  start_pos_y;
+};
+
+void set_replicate(uint offset, int pos, uint replicate_value)
+{
+    uint packed_b;
+    LOAD1(packed_b, buf, offset);
+
+    vec2 b = unpackHalf2x16(packed_b);
+    vec2 c = unpackHalf2x16(replicate_value);
+
+    if(pos % 2 == 0)
+    {
+        b.x = c.y;
+    }
+    else
+    {
+        b.y = c.x;
+    }
+
+    packed_b = packHalf2x16(b);
+
+    STORE1(buf, offset, packed_b);
+}
+
+/** Fill N pixel of the padding edge of a single channel image by replicating the closest valid pixel.
+ *
+ * @attention  The border size for top, bottom, left, right needs to be passed at the compile time.
+ * e.g. BORDER_SIZE_TOP=0 BORDER_SIZE_BOTTOM=2 BORDER_SIZE_LEFT=0 BORDER_SIZE_RIGHT=2
+ *
+ * @param[in,out] buf_ptr                           Pointer to the source image. Supported data types: F16
+ * @param[in]     buf_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]     buf_step_x                        buf_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]     buf_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]     buf_step_y                        buf_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]     buf_stride_z                      Stride between images if batching images (in bytes)
+ * @param[in]     buf_step_z                        buf_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]     buf_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]     width                             Width of the valid region of the image
+ * @param[in]     height                            Height of the valid region of the image
+ * @param[in]     start_pos_x                       X coordinate indicating the start point of the valid region
+ * @param[in]     start_pos_y                       Y coordinate indicating the start point of the valid region
+ */
+void main()
+{
+    Image buf = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP_FP16(buf);
+
+    // Update pointer to point to the starting point of the valid region
+    buf.current_offset = uint(buf.current_offset + uint(start_pos_y) * buf_stride_y + uint(start_pos_x) * buf_stride_x);
+
+    int total_width = BORDER_SIZE_LEFT + int(width) + BORDER_SIZE_RIGHT;
+    int gid0        = int(gl_GlobalInvocationID.x);
+    int gidH        = gid0 - total_width;
+    int gidW        = gid0 - BORDER_SIZE_LEFT;
+
+    if(gidH >= 0)
+    {
+        // Handle left border
+        uint left_val;
+        LOAD1(left_val, buf, offset_fp16(buf, 0, gidH) >> uint(2));
+        for(int i = -BORDER_SIZE_LEFT; i < 0; ++i)
+        {
+            uint offset = offset_fp16(buf, i, gidH) >> 2;
+            int  pos    = i + BORDER_SIZE_LEFT;
+            if(i == -1)
+            {
+                if(pos % 2 == 0)
+                {
+                    set_replicate(offset, pos, left_val);
+                }
+            }
+            else
+            {
+                if(pos % 2 == 0)
+                {
+                    vec2 a = unpackHalf2x16(left_val);
+                    uint b = packHalf2x16(a.xx);
+                    STORE1(buf, offset, b);
+                }
+            }
+        }
+        // Handle right border
+        uint right_val;
+        LOAD1(right_val, buf, offset_fp16(buf, int(width) - 1, gidH) >> uint(2));
+        for(int i = 0; i < BORDER_SIZE_RIGHT; ++i)
+        {
+            uint offset = offset_fp16(buf, int(width) + i, gidH) >> 2;
+            int  pos    = i + BORDER_SIZE_LEFT + int(width);
+
+            if(i == 0)
+            {
+                if(pos % 2 == 0)
+                {
+                    vec2 a = unpackHalf2x16(right_val);
+                    uint b = packHalf2x16(a.yy);
+                    STORE1(buf, offset, b);
+                }
+                else
+                {
+                    set_replicate(offset, pos, right_val);
+                }
+            }
+            else
+            {
+                if(pos % 2 == 0)
+                {
+                    vec2 a = unpackHalf2x16(right_val);
+                    uint b = packHalf2x16(a.yy);
+                    STORE1(buf, offset, b);
+                }
+            }
+        }
+    }
+    else
+    {
+        // Get value for corners
+        int val_idx = gidW;
+        if(gidW < 0 || (gidW > (int(width) - 1)))
+        {
+            val_idx = gidW < 0 ? 0 : (int(width) - 1);
+        }
+
+        // Handle top border
+        uint top_val;
+        LOAD1(top_val, buf, offset_fp16(buf, val_idx, 0) >> uint(2));
+        for(int i = -BORDER_SIZE_TOP; i < 0; ++i)
+        {
+            uint offset = offset_fp16(buf, gidW, i) >> 2;
+
+            if(gid0 % 2 == 0)
+            {
+                if(gidW == (int(width) - 1))
+                {
+                    vec2 a = unpackHalf2x16(top_val);
+                    uint b = packHalf2x16(a.xx);
+                    STORE1(buf, offset, b);
+                }
+                else
+                {
+                    if(gidW < 0)
+                    {
+                        vec2 a = unpackHalf2x16(top_val);
+                        uint b;
+                        if(BORDER_SIZE_LEFT % 2 == 0)
+                        {
+                            b = packHalf2x16(a.xx);
+                        }
+                        else
+                        {
+                            b = packHalf2x16(a.yy);
+                        }
+                        STORE1(buf, offset, b);
+                    }
+                    else if(gidW >= int(width))
+                    {
+                        vec2 a = unpackHalf2x16(top_val);
+                        uint b;
+                        if((BORDER_SIZE_LEFT + int(width)) % 2 == 0)
+                        {
+                            b = packHalf2x16(a.yy);
+                        }
+                        STORE1(buf, offset, b);
+                    }
+                    else
+                    {
+                        STORE1(buf, offset, top_val);
+                    }
+                }
+            }
+        }
+        // Handle bottom border
+        uint bottom_val;
+        LOAD1(bottom_val, buf, offset_fp16(buf, val_idx, int(height) - 1) >> uint(2));
+        for(int i = 0; i < BORDER_SIZE_BOTTOM; ++i)
+        {
+            uint offset = offset_fp16(buf, gidW, int(height) + i) >> 2;
+
+            if(gid0 % 2 == 0)
+            {
+                if(gidW == (int(width) - 1))
+                {
+                    vec2 a = unpackHalf2x16(bottom_val);
+                    uint b = packHalf2x16(a.xx);
+                    STORE1(buf, offset, b);
+                }
+                else
+                {
+                    if(gidW < 0)
+                    {
+                        vec2 a = unpackHalf2x16(bottom_val);
+                        uint b;
+                        if(BORDER_SIZE_LEFT % 2 == 0)
+                        {
+                            b = packHalf2x16(a.xx);
+                        }
+                        else
+                        {
+                            b = packHalf2x16(a.yy);
+                        }
+                        STORE1(buf, offset, b);
+                    }
+                    else if(gidW >= int(width))
+                    {
+                        vec2 a = unpackHalf2x16(bottom_val);
+                        uint b;
+                        if((BORDER_SIZE_LEFT + int(width)) % 2 == 0)
+                        {
+                            b = packHalf2x16(a.yy);
+                        }
+                        STORE1(buf, offset, b);
+                    }
+                    else
+                    {
+                        STORE1(buf, offset, bottom_val);
+                    }
+                }
+            }
+        }
+    }
+}
+#endif /* FILL_IMAGE_BORDERS_REPLICATE */
+
+#ifdef FILL_IMAGE_BORDERS_CONSTANT
+BUFFER_DECLARATION(buf, 1, uint, restrict);
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(buf);
+    uint  width;
+    uint  height;
+    int   start_pos_x;
+    int   start_pos_y;
+    float constant_value;
+};
+
+void set_constant(uint offset, int pos)
+{
+    uint packed_b;
+    LOAD1(packed_b, buf, offset);
+
+    vec2 b = unpackHalf2x16(packed_b);
+
+    if(pos % 2 == 0)
+    {
+        b.x = constant_value;
+    }
+    else
+    {
+        b.y = constant_value;
+    }
+
+    packed_b = packHalf2x16(b);
+
+    STORE1(buf, offset, packed_b);
+}
+
+/** Fill N pixels of the padding edge of a single channel image with a constant value.
+ *
+ * @attention  The border size for top, bottom, left, right needs to be passed at the compile time.
+ * e.g. BORDER_SIZE_TOP=0 BORDER_SIZE_BOTTOM=2 BORDER_SIZE_LEFT=0 BORDER_SIZE_RIGHT=2
+ *
+ * @param[out] buf_ptr                           Pointer to the source image. Supported data types: F16
+ * @param[in]  buf_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  buf_step_x                        buf_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  buf_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  buf_step_y                        buf_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  buf_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]  width                             Width of the valid region of the image
+ * @param[in]  height                            Height of the valid region of the image
+ * @param[in]  start_pos_x                       X coordinate indicating the start point of the valid region
+ * @param[in]  start_pos_y                       Y coordinate indicating the start point of the valid region
+ * @param[in]  constant_value                    Constant value to use to fill the edges
+ */
+void main()
+{
+    Image buf = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP_FP16(buf);
+
+    int total_width = BORDER_SIZE_LEFT + int(width) + BORDER_SIZE_RIGHT;
+    int gid0        = int(gl_GlobalInvocationID.x);
+    int gidH        = gid0 - total_width;
+    int gidW        = gid0 - BORDER_SIZE_LEFT;
+
+    // Update pointer to point to the starting point of the valid region
+    buf.current_offset = uint(int(buf.current_offset) + ((start_pos_y * int(buf_stride_y) + start_pos_x * int(buf_stride_x))));
+
+    vec2 b = vec2(constant_value, constant_value);
+
+    uint packed_b = packHalf2x16(b);
+
+    if(gidH >= 0)
+    {
+        // Handle left border
+        for(int i = -BORDER_SIZE_LEFT; i < 0; ++i)
+        {
+            uint offset = offset_fp16(buf, i, gidH) >> 2;
+            int  pos    = i + BORDER_SIZE_LEFT;
+
+            if(i == -1)
+            {
+                if(pos % 2 == 0)
+                {
+                    set_constant(offset, pos);
+                }
+            }
+            else
+            {
+                if(pos % 2 == 0)
+                {
+                    STORE1(buf, offset, packed_b);
+                }
+            }
+        }
+        // Handle right border
+        for(int i = 0; i < BORDER_SIZE_RIGHT; ++i)
+        {
+            uint offset = offset_fp16(buf, int(width) + i, gidH) >> 2;
+            int  pos    = i + BORDER_SIZE_LEFT + int(width);
+
+            if(i == 0)
+            {
+                if(pos % 2 == 0)
+                {
+                    STORE1(buf, offset, packed_b);
+                }
+                else
+                {
+                    set_constant(offset, pos);
+                }
+            }
+            else
+            {
+                if(pos % 2 == 0)
+                {
+                    STORE1(buf, offset, packed_b);
+                }
+            }
+        }
+    }
+    else
+    {
+        // Handle top border
+        for(int i = -BORDER_SIZE_TOP; i < 0; ++i)
+        {
+            uint offset = offset_fp16(buf, gidW, i) >> 2;
+
+            if(gid0 % 2 == 0)
+            {
+                STORE1(buf, offset, packed_b);
+            }
+        }
+        // Handle bottom border
+        for(int i = 0; i < BORDER_SIZE_BOTTOM; ++i)
+        {
+            uint offset = offset_fp16(buf, gidW, int(height) + i) >> 2;
+
+            if(gid0 % 2 == 0)
+            {
+                STORE1(buf, offset, packed_b);
+            }
+        }
+    }
+}
+#endif /* FILL_IMAGE_BORDERS_CONSTANT */
+#endif /* DATA_TYPE_FP32 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/gemm.cs b/src/core/GLES_COMPUTE/cs_shaders/gemm.cs
new file mode 100755
index 0000000000..3313b88718
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/gemm.cs
@@ -0,0 +1,623 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#if defined(DATA_TYPE_FP32)
+#define LOAD8(r, name, offset) \
+    r.x = LOAD4(name, offset); \
+    r.y = LOAD4(name, offset + uint(1))
+
+#define LOAD16(r, name, offset)          \
+    r.x = LOAD4(name, offset);           \
+    r.y = LOAD4(name, offset + uint(1)); \
+    r.z = LOAD4(name, offset + uint(2)); \
+    r.w = LOAD4(name, offset + uint(3))
+
+#define STORE16(name, offset, r)         \
+    STORE4(name, offset, r.x);           \
+    STORE4(name, offset + uint(1), r.y); \
+    STORE4(name, offset + uint(2), r.z); \
+    STORE4(name, offset + uint(3), r.w)
+
+#ifdef GEMM_TRANSPOSE1xW
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel computes the "vector" 1x4 transposition of input matrix
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source matrix 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 matrix 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 matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * 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 matrix
+ */
+void main(void)
+{
+    /* Compute address for Matrix B - source */
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Compute address for Matrix B transposed - destination. X and Y are swapped */
+    uint dst_addr_in_bytes = (gl_GlobalInvocationID.y * uint(16) + gl_GlobalInvocationID.x * dst.stride_y + dst.offset_first_element_in_bytes) >> 2;
+    vec4 b0;
+    LOAD16(b0, src, offset(src, 0, 0));
+    STORE16(dst, dst_addr_in_bytes, b0);
+}
+#endif /* GEMM_TRANSPOSE1xW */
+
+#ifdef GEMM_INTERLEAVE4x4
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGLES kernel reshapes the input matrix interleaving the values
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source matrix 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 matrix 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 matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * 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 matrix
+ */
+void main(void)
+{
+    /* Compute source and destination addresses */
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    int i;
+    int j;
+
+    for(i = 0; i < 4; ++i)
+    {
+        for(j = 0; j < 4; ++j)
+        {
+            float res    = LOAD4(src, offset(src, i, j));
+            uint  ofset0 = CURRENT_OFFSET(dst) + uint(i * 4 + j);
+            STORE4(dst, ofset0, res);
+        }
+    }
+}
+#endif /* GEMM_INTERLEAVE4x4 */
+
+#ifdef GEMM_ACCUMULATE_BIASES
+BUFFER_DECLARATION(accum, 1, float, restrict);
+BUFFER_DECLARATION(biases, 2, float, readonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(accum);
+    VECTOR_PARAM_DECLARATION(biases);
+};
+
+/** This kernel accumulates each row with the biases vector
+ *
+ * @param[in, out] accum_ptr                            Pointer to the accumulate tensor. Supported data type: F32
+ * @param[in]      accum_stride_x                       Stride of the accmulate tensor in X dimension (in bytes)
+ * @param[in]      accum_step_x                         accum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]      accum_stride_y                       Stride of the accumlulate tensor in Y dimension (in bytes)
+ * @param[in]      accum_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]      accum_offset_first_element_in_bytes  The offset of the first element in the accumulate tensor
+ * @param[in]      biases_ptr                           Pointer to the biases vector. Same as @p accum_ptr
+ * @param[in]      biases_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]      biases_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]      biases_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Image  accum  = CONVERT_TO_IMAGE_STRUCT(accum);
+    Vector biases = CONVERT_TO_VECTOR_STRUCT(biases);
+
+    for(int i = 0; i < 16; ++i)
+    {
+        float accum_value  = LOAD4(accum, CURRENT_OFFSET(accum) + uint(i));
+        float biases_value = LOAD4(biases, CURRENT_OFFSET(biases) + uint(i));
+        accum_value        = biases_value + accum_value;
+
+        // Store result in the accummulate buffer
+        STORE4(accum, CURRENT_OFFSET(accum) + uint(i), accum_value);
+    }
+}
+#endif /* GEMM_ACCUMULATE_BIASES */
+
+#ifdef GEMM_MM_INTERLEAVED_TRANSPOSED /* unvalidate */
+BUFFER_DECLARATION(src0, 1, float, readonly);
+BUFFER_DECLARATION(src1, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src0);
+    IMAGE_PARAM_DECLARATION(src1);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel is optimised for Midgard. It computes the matrix multiplication between matrix A (src0) and matrix B (src1)
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication
+ *
+ * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
+ * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_gx_stride_y * 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 matrix
+ */
+void main()
+{
+    Image src0 = CONVERT_TO_IMAGE_STRUCT(src0);
+    Image src1 = CONVERT_TO_IMAGE_STRUCT(src1);
+    Image dst  = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Compute address for matrix A and B */
+    src0.current_offset = (src0.offset_first_element_in_bytes + (uint(gl_GlobalInvocationID.y) * uint(src0.stride_y))) >> uint(2);
+    src1.current_offset = (src1.offset_first_element_in_bytes + (uint(gl_GlobalInvocationID.x) * uint(src1.stride_y))) >> uint(2);
+
+    /* Compute end row address for matrix B */
+    int end_row_mtx_b = int(src1.current_offset) + int(COLS_B);
+
+    /* Reset accumulators */
+    vec4 c00 = vec4(0.0f);
+    vec4 c10 = vec4(0.0f);
+    vec4 c20 = vec4(0.0f);
+    vec4 c30 = vec4(0.0f);
+
+    // FIXME: loop unrolling really needed for GLES?
+    for(; int(src1.current_offset) <= (end_row_mtx_b - 8); src0.current_offset += uint(8), src1.current_offset += uint(8))
+    {
+        /* Load values from matrix A (interleaved) and matrix B (transposed) */
+        vec4 a0;
+        vec4 b0;
+        LOAD16(a0, src0, src0.current_offset);
+        LOAD16(b0, src1, src1.current_offset);
+
+        c00 += vec4(a0.x) * b0;
+        c10 += vec4(a0.y) * b0;
+        c20 += vec4(a0.z) * b0;
+        c30 += vec4(a0.w) * b0;
+
+        /* Load values from matrix A (interleaved) and matrix B (transposed) */
+        LOAD16(a0, src0, src0.current_offset + uint(4));
+        LOAD16(b0, src1, src1.current_offset + uint(4));
+
+        c00 += vec4(a0.x) * b0;
+        c10 += vec4(a0.y) * b0;
+        c20 += vec4(a0.z) * b0;
+        c30 += vec4(a0.w) * b0;
+    }
+
+    for(; int(src1.current_offset) < end_row_mtx_b; src0.current_offset += uint(4), src1.current_offset += uint(4))
+    {
+        /* Load values from matrix A (interleaved) and matrix B (transposed) */
+        vec4 a0;
+        vec4 b0;
+        LOAD16(a0, src0, src0.current_offset);
+        LOAD16(b0, src1, src1.current_offset);
+
+        c00 += vec4(a0.x) * b0;
+        c10 += vec4(a0.y) * b0;
+        c20 += vec4(a0.z) * b0;
+        c30 += vec4(a0.w) * b0;
+    }
+
+    /* Multiply by the weight of matrix product */
+    c00 = c00 * vec4(ALPHA);
+    c10 = c10 * vec4(ALPHA);
+    c20 = c20 * vec4(ALPHA);
+    c30 = c30 * vec4(ALPHA);
+
+    /* Store 4x4 block */
+    STORE16(dst, offset(dst, 0, 0), c00);
+    STORE16(dst, offset(dst, 0, 1), c10);
+    STORE16(dst, offset(dst, 0, 2), c20);
+    STORE16(dst, offset(dst, 0, 3), c30);
+}
+#endif /* GEMM_MM_INTERLEAVED_TRANSPOSED */
+
+#ifdef GEMM_MM_FLOATING_POINT
+BUFFER_DECLARATION(src0, 1, float, readonly);
+BUFFER_DECLARATION(src1, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src0);
+    IMAGE_PARAM_DECLARATION(src1);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication
+ *
+ * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
+ * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_gx_stride_y * 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 matrix
+ */
+void main()
+{
+    Image src0 = CONVERT_TO_IMAGE_STRUCT(src0);
+    Image src1 = CONVERT_TO_IMAGE_STRUCT(src1);
+    Image dst  = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X);
+    /* Compute the address for the vector A and matrix B */
+    src0.current_offset = (src0_offset_first_element_in_bytes + uint(gl_GlobalInvocationID.y) * src0_stride_y * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y)) >> uint(2);
+    src1.current_offset = (src1_offset_first_element_in_bytes + uint(idx * 4)) >> uint(2);
+
+    /* Compute end row address for matrix A */
+    int end_row_vec_a = int(src0.current_offset) + ((COLS_A * 4) >> 2);
+
+    /* Reset accumulators */
+    vec4 acc0 = vec4(0.0f);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+    vec4 acc1 = vec4(0.0f);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+    vec4 acc2 = vec4(0.0f);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+    vec4 acc3 = vec4(0.0f);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+
+    for(; int(src0.current_offset) <= (end_row_vec_a - 2); src0.current_offset += uint(2), src1.current_offset += uint((2 * int(src1_stride_y)) >> 2))
+    {
+        vec2 a0;
+        LOAD8(a0, src0, src0.current_offset);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+        vec2 a1;
+        LOAD8(a1, src0, src0.current_offset + (src0_stride_y >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+        vec2 a2;
+        LOAD8(a2, src0, src0.current_offset + ((uint(2) * src0_stride_y) >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+        vec2 a3;
+        LOAD8(a3, src0, src0.current_offset + ((uint(3) * src0_stride_y) >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+
+        vec4 b0;
+        vec4 b1;
+        LOAD16(b0, src1, src1.current_offset);
+        LOAD16(b1, src1, src1.current_offset + (src1_stride_y >> uint(2)));
+
+        acc0 += b0 * vec4(a0.x);
+        acc0 += b1 * vec4(a0.y);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+        acc1 += b0 * vec4(a1.x);
+        acc1 += b1 * vec4(a1.y);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+        acc2 += b0 * vec4(a2.x);
+        acc2 += b1 * vec4(a2.y);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+        acc3 += b0 * vec4(a3.x);
+        acc3 += b1 * vec4(a3.y);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+    }
+
+    for(; int(src0.current_offset) < end_row_vec_a; src0.current_offset += uint(1), src1.current_offset += uint(int(src1_stride_y) >> 2))
+    {
+        // Load values from matrix A
+        float a0;
+        a0 = LOAD4(src0, src0.current_offset);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+        float a1;
+        a1 = LOAD4(src0, src0.current_offset + ((uint(1) * src0_stride_y) >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+        float a2;
+        a2 = LOAD4(src0, src0.current_offset + ((uint(2) * src0_stride_y) >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+        float a3;
+        a3 = LOAD4(src0, src0.current_offset + ((uint(3) * src0_stride_y) >> uint(2)));
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+
+        vec4 b0;
+        LOAD16(b0, src1, src1.current_offset);
+
+        acc0 += b0 * vec4(a0);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+        acc1 += b0 * vec4(a1);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+        acc2 += b0 * vec4(a2);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+        acc3 += b0 * vec4(a3);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+    }
+
+    /* Multiply by the weight of vector-matrix product */
+    acc0 = acc0 * vec4(ALPHA);
+    STORE16(dst, offset(dst, 0, 0), acc0);
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+    acc1 = acc1 * vec4(ALPHA);
+    STORE16(dst, offset(dst, 0, 1), acc1);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 1
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+    acc2 = acc2 * vec4(ALPHA);
+    STORE16(dst, offset(dst, 0, 2), acc2);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 2
+#if NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+    acc3 = acc3 * vec4(ALPHA);
+    STORE16(dst, offset(dst, 0, 3), acc3);
+#endif // NUM_ELEMS_PROCESSED_PER_THREAD_Y > 3
+}
+#endif /* GEMM_MM_FLOATING_POINT */
+
+#ifdef GEMM_MATRIXADDITION
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, restrict);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel performs the in-place matrix addition between 2 matrices taking into account that the second matrix might be weighted by a scalar value beta:
+ *
+ * @attention The beta's value need to be passed at compile time using BETA
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source matrix 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 matrix 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 matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * 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 matrix
+ */
+void main(void)
+{
+    /* Compute source and destination addresses */
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    /* Load values from A x B */
+    vec4 alpha_ab;
+    vec4 c;
+    vec4 out1;
+
+    LOAD16(alpha_ab, dst, dst.current_offset);
+    LOAD16(c, src, src.current_offset);
+
+    /* Computes alpha * axb + beta * c */
+    out1 = alpha_ab + vec4(BETA * c);
+
+    /* Store final result in axb matrix */
+    STORE16(dst, dst.current_offset, out1);
+}
+#endif /* GEMM_MATRIXADDITION */
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+#ifdef GEMM_MM_FLOATING_POINT
+BUFFER_DECLARATION(src0, 1, uint, readonly);
+BUFFER_DECLARATION(src1, 2, uvec2, readonly);
+BUFFER_DECLARATION(dst, 3, uvec2, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src0);
+    IMAGE_PARAM_DECLARATION(src1);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel computes the matrix multiplication between matrix A (src0) and matrix B (src1)
+ *  Matrix A and matrix B must be reshaped respectively with @ref gemm_interleave4x4_32bit and @ref gemm_transpose1x4 before running the matrix multiplication
+ *
+ * @attention The width of matrix B and the alpha's value need to be passed at compile time using WIDTH_MATRIX_B and ALPHA
+ *
+ * @param[in]  src0_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src0_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src0_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src0_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src0_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src0_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[in]  src1_ptr                           Pointer to the source matrix. Supported data types: same as @p src0_ptr
+ * @param[in]  src1_stride_x                      Stride of the source matrix in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source matrix in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source matrix
+ * @param[out] dst_ptr                            Pointer to the destination matrix Supported data types: same as @p src0_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                         dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                       Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                         dst_gx_stride_y * 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 matrix
+ */
+void main()
+{
+    Image src0 = GC_CONVERT_TO_IMAGE_STRUCT(src0);
+    Image src1 = GC_CONVERT_TO_IMAGE_STRUCT(src1);
+    Image dst  = GC_CONVERT_TO_IMAGE_STRUCT(dst);
+
+    int idx = int(gl_GlobalInvocationID.x) * int(NUM_ELEMS_PROCESSED_PER_THREAD_X);
+    /* Compute the address for the vector A and matrix B */
+    src0.current_offset = (src0_offset_first_element_in_bytes + uint(gl_GlobalInvocationID.y) * src0_stride_y * uint(NUM_ELEMS_PROCESSED_PER_THREAD_Y));
+    src1.current_offset = src1_offset_first_element_in_bytes + uint(idx) * src1_stride_x;
+
+    /* Compute end row address for matrix A */
+    uint end_row_vec_a = src0.current_offset + uint(COLS_A << 1);
+
+    /* Reset accumulators */
+    vec4 acc0 = vec4(0.0f);
+
+    for(; src0.current_offset < (end_row_vec_a - uint(2)); src0.current_offset += uint(2 * 2), src1.current_offset += uint(2) * src1_stride_y)
+    {
+        uint packed_a0;
+        vec2 a0;
+
+        GC_LOAD1_2D_OFFSET(packed_a0, src0, 0, 0);
+        a0 = vec2(unpackHalf2x16(packed_a0));
+
+        uvec2 packed_b0;
+        uvec2 packed_b1;
+        vec4  b0;
+        vec4  b1;
+
+        GC_LOAD1_2D_OFFSET(packed_b0, src1, 0, 0);
+        GC_LOAD1_2D_OFFSET(packed_b1, src1, 0, 1);
+
+        b0 = vec4(unpackHalf2x16(packed_b0.x), unpackHalf2x16(packed_b0.y));
+        b1 = vec4(unpackHalf2x16(packed_b1.x), unpackHalf2x16(packed_b1.y));
+
+        acc0 += b0 * vec4(a0.x);
+        acc0 += b1 * vec4(a0.y);
+    }
+
+    for(; src0.current_offset < end_row_vec_a; src0.current_offset += uint(2 * 2), src1.current_offset += src1_stride_y)
+    {
+        uint packed_a0;
+        vec2 a0;
+
+        GC_LOAD1_2D_OFFSET(packed_a0, src0, 0, 0);
+        a0 = vec2(unpackHalf2x16(packed_a0));
+
+        uvec2 packed_b0;
+        vec4  b0;
+
+        GC_LOAD1_2D_OFFSET(packed_b0, src1, 0, 0);
+
+        b0 = vec4(unpackHalf2x16(packed_b0.x), unpackHalf2x16(packed_b0.y));
+
+        acc0 += b0 * (a0.x);
+    }
+
+    /* Multiply by the weight of vector-matrix product */
+    acc0 = acc0 * vec4(ALPHA);
+
+    uvec2 packed_d;
+    packed_d = uvec2(packHalf2x16(acc0.xy), packHalf2x16(acc0.zw));
+    GC_STORE1_2D_OFFSET(packed_d, dst, 0, 0);
+}
+#endif /* GEMM_MM_FLOATING_POINT */
+
+#ifdef GEMM_ACCUMULATE_BIASES
+BUFFER_DECLARATION(accum, 1, uvec2, restrict);
+BUFFER_DECLARATION(biases, 2, uvec2, readonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(accum);
+    VECTOR_PARAM_DECLARATION(biases);
+};
+
+/** This kernel accumulates each row with the biases vector
+ *
+ * @param[in, out] accum_ptr                            Pointer to the accumulate tensor. Supported data type: F16
+ * @param[in]      accum_stride_x                       Stride of the accmulate tensor in X dimension (in bytes)
+ * @param[in]      accum_step_x                         accum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]      accum_stride_y                       Stride of the accumlulate tensor in Y dimension (in bytes)
+ * @param[in]      accum_step_y                         src_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]      accum_offset_first_element_in_bytes  The offset of the first element in the accumulate tensor
+ * @param[in]      biases_ptr                           Pointer to the biases vector. Same as @p accum_ptr
+ * @param[in]      biases_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]      biases_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]      biases_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Image  accum  = GC_CONVERT_TO_IMAGE_STRUCT(accum);
+    Vector biases = GC_CONVERT_TO_VECTOR_STRUCT(biases);
+
+    vec4  u[2];
+    uvec2 packed_s[2];
+    GC_LOAD1_2D_OFFSET(packed_s[0], accum, 0, 0);
+    GC_LOAD1_1D_OFFSET(packed_s[1], biases, 0);
+    u[0] = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    u[1] = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+
+    vec4 tmp;
+    tmp         = u[0] + u[1];
+    packed_s[0] = uvec2(packHalf2x16(tmp.xy), packHalf2x16(tmp.zw));
+    GC_STORE1_2D_OFFSET(packed_s[0], accum, 0, 0);
+}
+#endif /* GEMM_ACCUMULATE_BIASES */
+#else  /* DATA_TYPE_F32 */
+#error Data type not supported
+#endif /* DATA_TYPE_F32 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/helpers.h b/src/core/GLES_COMPUTE/cs_shaders/helpers.h
new file mode 100644
index 0000000000..86dedf5a9c
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/helpers.h
@@ -0,0 +1,582 @@
+/*
+ * Copyright (c) 2017 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#ifndef ARM_COMPUTE_HELPER_H
+#define ARM_COMPUTE_HELPER_H
+
+#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val)
+
+#define VEC_DATA_TYPE_STR(type, size) type##size
+#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size)
+
+#define CONVERT(x, type) type(x)
+
+#define PACK(value, stype, dtype) \
+    pack_##stype##_##dtype(value)
+
+#define UNPACK(value, stype, dtype) \
+    unpack_##stype##_##dtype(value)
+
+#define BUFFER_DECLARATION(name, location, type, access)          \
+    layout(std430, binding = location) access buffer name##Buffer \
+    {                                                             \
+        type name##_ptr[];                                        \
+    }
+
+#define VECTOR_PARAM_DECLARATION(name)         \
+    uint name##_stride_x;                      \
+    uint name##_step_x;                        \
+    uint name##_offset_first_element_in_bytes; \
+    uint name##_buffer_data_type_size
+
+#define IMAGE_PARAM_DECLARATION(name)          \
+    uint name##_stride_x;                      \
+    uint name##_step_x;                        \
+    uint name##_stride_y;                      \
+    uint name##_step_y;                        \
+    uint name##_offset_first_element_in_bytes; \
+    uint name##_buffer_data_type_size
+
+#define TENSOR3D_PARAM_DECLARATION(name)       \
+    uint name##_stride_x;                      \
+    uint name##_step_x;                        \
+    uint name##_stride_y;                      \
+    uint name##_step_y;                        \
+    uint name##_stride_z;                      \
+    uint name##_step_z;                        \
+    uint name##_offset_first_element_in_bytes; \
+    uint name##_buffer_data_type_size
+
+/** Structure to hold Vector information */
+struct Vector
+{
+    uint current_offset;                /**< Current offset of vector */
+    uint offset_first_element_in_bytes; /**< The offset of the first element in the source image */
+    uint stride_x;                      /**< Stride of the image in X dimension (in bytes) */
+};
+
+/** Structure to hold Image information */
+struct Image
+{
+    uint current_offset;                /**< Current offset of image */
+    uint offset_first_element_in_bytes; /**< The offset of the first element in the source image */
+    uint stride_x;                      /**< Stride of the image in X dimension (in bytes) */
+    uint stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
+};
+
+/** Structure to hold 3D tensor information */
+struct Tensor3D
+{
+    uint current_offset;                /**< Current offset of tensor */
+    uint offset_first_element_in_bytes; /**< The offset of the first element in the source image */
+    uint stride_x;                      /**< Stride of the image in X dimension (in bytes) */
+    uint stride_y;                      /**< Stride of the image in Y dimension (in bytes) */
+    uint stride_z;                      /**< Stride of the image in Z dimension (in bytes) */
+};
+
+/////////////////////////////////////////////////////////////
+// TODO: old to be removed
+
+#define CONVERT_TO_VECTOR_STRUCT(name) \
+    update_vector_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)
+
+#define CONVERT_TO_VECTOR_STRUCT_FP16(name) \
+    update_vector_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)
+
+#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
+    update_vector_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0))
+
+#define CONVERT_TO_VECTOR_STRUCT_NO_STEP_FP16(name) \
+    update_vector_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, uint(0))
+
+#define CONVERT_TO_IMAGE_STRUCT(name) \
+    update_image_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)
+
+#define CONVERT_TO_IMAGE_STRUCT_FP16(name) \
+    update_image_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)
+
+#define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
+    update_image_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0))
+
+#define CONVERT_TO_IMAGE_STRUCT_NO_STEP_FP16(name) \
+    update_image_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0))
+
+#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
+    update_image_from_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, name##_step_z)
+
+#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP_FP16(name) \
+    update_image_from_tensor3D_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, name##_step_z)
+
+#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
+    update_image_from_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)
+
+#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_FP16(name) \
+    update_image_from_tensor3D_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)
+
+#define CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                  \
+    update_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
+                                    name##_stride_z, name##_step_z)
+
+#define CONVERT_TO_TENSOR3D_STRUCT_FP16(name)                                                                                                  \
+    update_tensor3D_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
+                                         name##_stride_z, name##_step_z)
+
+#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
+    update_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, uint(0))
+
+#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP_FP16(name) \
+    update_tensor3D_workitem_offset_fp16(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, uint(0))
+
+// FIXME: Redesign the macros if different data types are supported.
+#define LOAD4(name, offset) \
+    name##_ptr[offset]
+
+#define STORE4(name, offset, value) \
+    name##_ptr[offset] = value
+
+// Load 1 element, which size is determined by ssbo type.
+#define LOAD1(r, name, offset) \
+    r = name##_ptr[offset]
+
+#define STORE1(name, offset, value) \
+    name##_ptr[offset] = value
+
+#define LOAD2(r, name, offset) \
+    LOAD1(r[0], name, offset); \
+    LOAD1(r[1], name, (offset) + uint(1))
+
+#define STORE2(name, offset, value)            \
+    name##_ptr[offset]             = value[0]; \
+    name##_ptr[(offset) + uint(1)] = value[1]
+
+#define LOAD3(r, name, offset)             \
+    LOAD1(r[0], name, offset);             \
+    LOAD1(r[1], name, (offset) + uint(1)); \
+    LOAD1(r[2], name, (offset) + uint(2))
+
+#define CURRENT_OFFSET(name) \
+    name.current_offset
+
+/** Wrap vector information into an Vector structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
+ * @param[in] stride_x                      Stride of the vector in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ *
+ * @return An vector object
+ */
+Vector update_vector_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x)
+{
+    Vector vector;
+    vector.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    vector.stride_x                      = stride_x;
+    vector.current_offset                = (vector.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x) >> 2;
+
+    return vector;
+}
+
+/** Wrap vector information into an Vector structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source vector
+ * @param[in] stride_x                      Stride of the vector in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ *
+ * @return An vector object
+ */
+Vector update_vector_workitem_offset_fp16(uint offset_first_element_in_bytes, uint stride_x, uint step_x)
+{
+    Vector vector;
+    vector.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    vector.stride_x                      = stride_x;
+    vector.current_offset                = vector.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x;
+
+    return vector;
+}
+
+/** Wrap image information into an Image structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ *
+ * @return An image object
+ */
+Image update_image_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = (img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y) >> 2;
+
+    return img;
+}
+
+/** Wrap image information into an Image structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ *
+ * @return An image object
+ */
+Image update_image_workitem_offset_fp16(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y;
+
+    return img;
+}
+
+/** Wrap 3D tensor information into an image structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
+ * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
+ *
+ * @return A 2D Image object
+ */
+Image update_image_from_tensor3D_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = (img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z) >> 2;
+
+    return img;
+}
+
+/** Wrap 3D tensor information into an image structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
+ * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
+ *
+ * @return A 2D Image object
+ */
+Image update_image_from_tensor3D_workitem_offset_fp16(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z;
+
+    return img;
+}
+
+/** Wrap 3D tensor information into an tensor structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
+ * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
+ *
+ * @return A 3D tensor object
+ */
+Tensor3D update_tensor3D_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Tensor3D tensor;
+    tensor.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    tensor.stride_x                      = stride_x;
+    tensor.stride_y                      = stride_y;
+    tensor.stride_z                      = stride_z;
+    tensor.current_offset                = (tensor.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z) >> 2;
+
+    return tensor;
+}
+
+/** Wrap 3D tensor information into an tensor structure, and make the offset to be this workitem's position.
+ *
+ * @param[in] offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in] stride_x                      Stride of the image in X dimension (in bytes)
+ * @param[in] step_x                        stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in] stride_y                      Stride of the image in Y dimension (in bytes)
+ * @param[in] step_y                        stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in] stride_z                      Stride of the image in Z dimension (in bytes)
+ * @param[in] step_z                        stride_z * number of elements along Z processed per workitem(in bytes)
+ *
+ * @return A 3D tensor object
+ */
+Tensor3D update_tensor3D_workitem_offset_fp16(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Tensor3D tensor;
+    tensor.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    tensor.stride_x                      = stride_x;
+    tensor.stride_y                      = stride_y;
+    tensor.stride_z                      = stride_z;
+    tensor.current_offset                = tensor.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z;
+
+    return tensor;
+}
+
+/** Get the pointer position of a Vector
+ *
+ * @param[in] vec Pointer to the starting position of the buffer
+ * @param[in] x   Relative X position
+ */
+uint vector_offset(Vector vec, int x)
+{
+    return CONVERT(CONVERT(vec.current_offset << 2, int) + x * CONVERT(vec.stride_x, int), uint) >> 2;
+}
+
+/** Get the pointer position of a Vector
+ *
+ * @param[in] vec Pointer to the starting position of the buffer
+ * @param[in] x   Relative X position
+ */
+uint vector_offset_fp16(Vector vec, int x)
+{
+    return CONVERT(CONVERT(vec.current_offset, int) + x * CONVERT(vec.stride_x, int), uint);
+}
+
+/** Get the pointer position of a Image
+ *
+ * @param[in] img Pointer to the starting position of the buffer
+ * @param[in] x   Relative X position
+ * @param[in] y   Relative Y position
+ */
+uint offset(Image img, int x, int y)
+{
+    return CONVERT(CONVERT(img.current_offset << 2, int) + x * CONVERT(img.stride_x, int) + y * CONVERT(img.stride_y, int), uint) >> 2;
+}
+
+/** Get the pointer position of a Image
+ *
+ * @param[in] img Pointer to the starting position of the buffer
+ * @param[in] x   Relative X position
+ * @param[in] y   Relative Y position
+ */
+uint offset_fp16(Image img, int x, int y)
+{
+    return CONVERT(CONVERT(img.current_offset, int) + x * CONVERT(img.stride_x, int) + y * CONVERT(img.stride_y, int), uint);
+}
+
+/** Get the pointer position of a Tensor3D
+ *
+ * @param[in] tensor Pointer to the starting postion of the buffer
+ * @param[in] x      Relative X position
+ * @param[in] y      Relative Y position
+ * @param[in] z      Relative Z position
+ */
+uint tensor3D_offset(Tensor3D tensor, int x, int y, int z)
+{
+    return CONVERT(CONVERT(tensor.current_offset << 2, int) + x * CONVERT(tensor.stride_x, int) + y * CONVERT(tensor.stride_y, int) + z * CONVERT(tensor.stride_z, int), uint) >> 2;
+}
+
+/** Get the pointer position of a Tensor3D
+ *
+ * @param[in] tensor Pointer to the starting postion of the buffer
+ * @param[in] x      Relative X position
+ * @param[in] y      Relative Y position
+ * @param[in] z      Relative Z position
+ */
+uint tensor3D_offset_fp16(Tensor3D tensor, int x, int y, int z)
+{
+    return CONVERT(CONVERT(tensor.current_offset, int) + x * CONVERT(tensor.stride_x, int) + y * CONVERT(tensor.stride_y, int) + z * CONVERT(tensor.stride_z, int), uint);
+}
+
+/////////////////////////////////////////////////////////////
+// new one
+
+#define GC_CONVERT_TO_VECTOR_STRUCT(name) \
+    gc_update_vector_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x)
+
+#define GC_CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \
+    gc_update_vector_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0))
+
+#define GC_CONVERT_TO_IMAGE_STRUCT(name) \
+    gc_update_image_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y)
+
+#define GC_CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \
+    gc_update_image_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0))
+
+#define GC_CONVERT_TO_TENSOR3D_STRUCT(name)                                                                                                  \
+    gc_update_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \
+                                       name##_stride_z, name##_step_z)
+
+#define GC_CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \
+    gc_update_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, uint(0))
+
+#define GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \
+    gc_update_image_from_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z)
+
+#define GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \
+    gc_update_image_from_tensor3D_workitem_offset(name##_offset_first_element_in_bytes, name##_stride_x, uint(0), name##_stride_y, uint(0), name##_stride_z, name##_step_z)
+
+Vector gc_update_vector_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x)
+{
+    Vector vector;
+    vector.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    vector.stride_x                      = stride_x;
+    vector.current_offset                = vector.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x;
+
+    return vector;
+}
+
+Image gc_update_image_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y;
+
+    return img;
+}
+
+Tensor3D gc_update_tensor3D_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Tensor3D tensor;
+    tensor.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    tensor.stride_x                      = stride_x;
+    tensor.stride_y                      = stride_y;
+    tensor.stride_z                      = stride_z;
+    tensor.current_offset                = tensor.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z;
+
+    return tensor;
+}
+
+Image gc_update_image_from_tensor3D_workitem_offset(uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z)
+{
+    Image img;
+    img.offset_first_element_in_bytes = offset_first_element_in_bytes;
+    img.stride_x                      = stride_x;
+    img.stride_y                      = stride_y;
+    img.current_offset                = img.offset_first_element_in_bytes + gl_GlobalInvocationID.x * step_x + gl_GlobalInvocationID.y * step_y + gl_GlobalInvocationID.z * step_z;
+
+    return img;
+}
+
+#define GC_CURRENT_OFFSET(name) \
+    name.current_offset
+
+uint gc_vector_offset(Vector vec, int x)
+{
+    return CONVERT(CONVERT(vec.current_offset, int) + x * CONVERT(vec.stride_x, int), uint);
+}
+
+uint gc_image_offset(Image img, int x, int y)
+{
+    return CONVERT(CONVERT(img.current_offset, int) + x * CONVERT(img.stride_x, int) + y * CONVERT(img.stride_y, int), uint);
+}
+
+uint gc_tensor3D_offset(Tensor3D tensor, int x, int y, int z)
+{
+    return CONVERT(CONVERT(tensor.current_offset, int) + x * CONVERT(tensor.stride_x, int) + y * CONVERT(tensor.stride_y, int) + z * CONVERT(tensor.stride_z, int), uint);
+}
+
+// load/store number of element depends on buffer type
+#define GC_LOAD1(r, name, offset) \
+    r = name##_ptr[offset]
+
+#define GC_LOAD2(r, name, offset) \
+    GC_LOAD1(r[0], name, offset); \
+    GC_LOAD1(r[1], name, (offset) + uint(1))
+
+#define GC_LOAD3(r, name, offset)             \
+    GC_LOAD1(r[0], name, offset);             \
+    GC_LOAD1(r[1], name, (offset) + uint(1)); \
+    GC_LOAD1(r[2], name, (offset) + uint(2))
+
+#define GC_STORE1(value, name, offset) \
+    name##_ptr[offset] = value
+
+#define GC_STORE2(value, name, offset) \
+    GC_STORE1(value[0], name, offset); \
+    GC_STORE1(value[1], name, (offset) + uint(1))
+
+#define GC_STORE3(value, name, offset)             \
+    GC_STORE1(value[0], name, offset);             \
+    GC_STORE1(value[1], name, (offset) + uint(1)); \
+    GC_STORE1(value[2], name, (offset) + uint(2))
+
+// has to manually expand them since not supported by compiler
+#define GC_LOAD1_1D_OFFSET(r, name, x) \
+    GC_LOAD1(r, name, gc_vector_offset(name, int(x)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD1_2D_OFFSET(r, name, x, y) \
+    GC_LOAD1(r, name, gc_image_offset(name, int(x), int(y)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD1_3D_OFFSET(r, name, x, y, z) \
+    GC_LOAD1(r, name, gc_tensor3D_offset(name, int(x), int(y), int(z)) >> name##_buffer_data_type_size)
+
+#define GC_STORE1_1D_OFFSET(value, name, x) \
+    GC_STORE1(value, name, gc_vector_offset(name, int(x)) >> name##_buffer_data_type_size)
+
+#define GC_STORE1_2D_OFFSET(value, name, x, y) \
+    GC_STORE1(value, name, gc_image_offset(name, int(x), int(y)) >> name##_buffer_data_type_size)
+
+#define GC_STORE1_3D_OFFSET(value, name, x, y, z) \
+    GC_STORE1(value, name, gc_tensor3D_offset(name, int(x), int(y), int(z)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD2_1D_OFFSET(r, name, x) \
+    GC_LOAD2(r, name, gc_vector_offset(name, int(x)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD2_2D_OFFSET(r, name, x, y) \
+    GC_LOAD2(r, name, gc_image_offset(name, int(x), int(y)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD2_3D_OFFSET(r, name, x, y, z) \
+    GC_LOAD2(r, name, gc_tensor3D_offset(name, int(x), int(y), int(z)) >> name##_buffer_data_type_size)
+
+#define GC_STORE2_1D_OFFSET(value, name, x) \
+    GC_STORE2(value, name, gc_vector_offset(name, int(x)) >> name##_buffer_data_type_size)
+
+#define GC_STORE2_2D_OFFSET(value, name, x, y) \
+    GC_STORE2(value, name, gc_image_offset(name, int(x), int(y)) >> name##_buffer_data_type_size)
+
+#define GC_STORE2_3D_OFFSET(value, name, x, y, z) \
+    GC_STORE2(value, name, gc_tensor3D_offset(name, int(x), int(y), int(z)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD3_1D_OFFSET(r, name, x) \
+    GC_LOAD3(r, name, gc_vector_offset(name, int(x)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD3_2D_OFFSET(r, name, x, y) \
+    GC_LOAD3(r, name, gc_image_offset(name, int(x), int(y)) >> name##_buffer_data_type_size)
+
+#define GC_LOAD3_3D_OFFSET(r, name, x, y, z) \
+    GC_LOAD3(r, name, gc_tensor3D_offset(name, int(x), int(y), int(z)) >> name##_buffer_data_type_size)
+
+/////////////////////////////////////////////////////////////
+
+#endif // _HELPER_H
diff --git a/src/core/GLES_COMPUTE/cs_shaders/normalization_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/normalization_layer.cs
new file mode 100755
index 0000000000..5699340c14
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/normalization_layer.cs
@@ -0,0 +1,157 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src1);
+    TENSOR3D_PARAM_DECLARATION(src2);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+BUFFER_DECLARATION(src1, 1, float, readonly);
+BUFFER_DECLARATION(src2, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+
+#ifdef CROSS_MAP
+/** Apply cross map normalization.
+ *
+ * @note Alpha parameter / norm_size should be given as a preprocessor argument using "#define COEFF x"
+ * @note BETA parameter in the normalization equation should be given as a preprocessor argument using "#define BETA x"
+ * @note KAPPA parameter in the normalization equation should be given as a preprocessor argument using "#define KAPPA x"
+ * @note Number of elements on the right or left side to normalize across should be given as a preprocessor argument using "#define RADIUS x"
+ *
+ * @param[in]  src1_ptr                                    Pointer to the first source tensor. Supported data types: F32
+ * @param[in]  src1_stride_x                               Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  src1_step_x                                 src1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                               Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in]  src1_step_y                                 src1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_stride_z                               Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  src1_step_z                                 src1_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes          The offset of the first element in the first source tensor
+ * @param[in]  src2_ptr                                    Pointer to the second source tensor. Supported data types: Same as @p src1_ptr
+ * @param[in]  src2_stride_x                               Stride of the second source tensor in X dimension (in bytes)
+ * @param[in]  src2_step_x                                 src2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src2_stride_y                               Stride of the second source tensor in Y dimension (in bytes)
+ * @param[in]  src2_step_y                                 src2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src2_stride_z                               Stride of the second source tensor in Z dimension (in bytes)
+ * @param[in]  src2_step_z                                 src2_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src2_offset_first_element_in_bytes          The offset of the second element in the second source tensor
+ * @param[out] dst_ptr                                     Pointer to the destination tensor. Supported data types: Same as @p src1_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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes           The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src1 = CONVERT_TO_TENSOR3D_STRUCT(src1);
+    Tensor3D src2 = CONVERT_TO_TENSOR3D_STRUCT(src2);
+    Tensor3D dst  = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float acc = 0.0;
+
+    int num_of_slices = int(gl_NumWorkGroups.z * gl_WorkGroupSize.z);
+    int current_slice = int(gl_GlobalInvocationID.z);
+
+    int left_slice  = max(current_slice - int(RADIUS), int(0));
+    int right_slice = min(current_slice + int(RADIUS), int(num_of_slices - 1));
+
+    for(int i = left_slice; i <= right_slice; i++)
+    {
+        acc += src2_ptr[tensor3D_offset(src2, 0, 0, i - current_slice)];
+    }
+
+    float normalized = pow(float(KAPPA) + float(COEFF) * acc, float(BETA));
+
+    float normalized_pixel = (src1_ptr[src1.current_offset]) / normalized;
+
+    dst_ptr[dst.current_offset] = normalized_pixel;
+}
+
+#elif defined(IN_MAP_1D)
+/** Apply in map normalization.
+ *
+ * @note Alpha parameter / norm_size should be given as a preprocessor argument using "#define COEFF x"
+ * @note BETA parameter in the normalization equation should be given as a preprocessor argument using "#define BETA x"
+ * @note KAPPA parameter in the normalization equation should be given as a preprocessor argument using "#define KAPPA x"
+ * @note Number of elements on the right or left side to normalize across should be given as a preprocessor argument using "#define RADIUS x"
+ *
+ * @param[in]  src1_ptr                                    Pointer to the first source tensor. Supported data types: F32
+ * @param[in]  src1_stride_x                               Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  src1_step_x                                 src1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                               Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in]  src1_step_y                                 src1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_stride_z                               Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  src1_step_z                                 src1_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes          The offset of the first element in the first source tensor
+ * @param[in]  src2_ptr                                    Pointer to the second source tensor. Supported data types: Same as @p src1_ptr
+ * @param[in]  src2_stride_x                               Stride of the second source tensor in X dimension (in bytes)
+ * @param[in]  src2_step_x                                 src2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src2_stride_y                               Stride of the second source tensor in Y dimension (in bytes)
+ * @param[in]  src2_step_y                                 src2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src2_stride_z                               Stride of the second source tensor in Z dimension (in bytes)
+ * @param[in]  src2_step_z                                 src2_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src2_offset_first_element_in_bytes          The offset of the second element in the second source tensor
+ * @param[out] dst_ptr                                     Pointer to the destination tensor. Supported data types: Same as @p src1_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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes           The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Tensor3D src1 = CONVERT_TO_TENSOR3D_STRUCT(src1);
+    Tensor3D src2 = CONVERT_TO_TENSOR3D_STRUCT(src2);
+    Tensor3D dst  = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    float acc = 0.0;
+
+    int num_of_items_x = int(gl_NumWorkGroups.x * gl_WorkGroupSize.x);
+    int current_pos    = int(gl_GlobalInvocationID.x);
+
+    int left_pos  = max(current_pos - int(RADIUS), int(0));
+    int right_pos = min(current_pos + int(RADIUS), int(num_of_items_x + -1));
+
+    for(int i = left_pos; i <= right_pos; i++)
+    {
+        acc += src2_ptr[tensor3D_offset(src2, i - current_pos, 0, 0)];
+    }
+
+    float normalized = pow(float(KAPPA) + float(COEFF) * acc, float(BETA));
+
+    float normalized_pixel = (src1_ptr[src1.current_offset]) / normalized;
+
+    dst_ptr[dst.current_offset] = normalized_pixel;
+}
+#endif /*CROSS_MAP*/
diff --git a/src/core/GLES_COMPUTE/cs_shaders/pixelwise_mul_float.cs b/src/core/GLES_COMPUTE/cs_shaders/pixelwise_mul_float.cs
new file mode 100644
index 0000000000..031687af0c
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/pixelwise_mul_float.cs
@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src1);
+    TENSOR3D_PARAM_DECLARATION(src2);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+BUFFER_DECLARATION(src1, 1, float, readonly);
+BUFFER_DECLARATION(src2, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+
+/** Performs a pixelwise multiplication with float scale of either integer or float inputs.
+ *
+ * @param[in]  src1_ptr                           Pointer to the source image. Supported data types: F32
+ * @param[in]  src1_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  src1_step_x                        src1_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src1_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  src1_step_y                        src1_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  src1_step_z                        src1_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src1_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[in]  src2_ptr                           Pointer to the source image. Supported data types: Same as @p src1_ptr
+ * @param[in]  src2_stride_x                      Stride of the source image in X dimension (in bytes)
+ * @param[in]  src2_step_x                        src2_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src2_stride_y                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  src2_step_y                        src2_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src2_stride_z                      Stride of the source image in Y dimension (in bytes)
+ * @param[in]  src2_step_z                        src2_stride_z * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  src2_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                            Pointer to the destination image. Supported data types: Same as @p src1_ptr
+ * @param[in]  dst_stride_x                       Stride of the destination image 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 image 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 image in Y 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 image
+ * @param[in]  scale                              Float scaling factor. Supported data types: F32
+ */
+void main()
+{
+    // Get pixels pointer
+    Tensor3D src1 = CONVERT_TO_TENSOR3D_STRUCT(src1);
+    Tensor3D src2 = CONVERT_TO_TENSOR3D_STRUCT(src2);
+    Tensor3D dst  = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    dst_ptr[dst.current_offset] = (src1_ptr[src1.current_offset] * src2_ptr[src2.current_offset] * float(SCALE));
+}
diff --git a/src/core/GLES_COMPUTE/cs_shaders/pooling_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/pooling_layer.cs
new file mode 100644
index 0000000000..1e0fee4688
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/pooling_layer.cs
@@ -0,0 +1,1444 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#if defined(DATA_TYPE_FP32)
+
+float calculate_max(const int, Tensor3D, const int, const int, const int, const int, const int, const int);
+float calculate_avg(const int, Tensor3D, const int, const int, const int, const int, const int, const int);
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+#define LOAD8(r, name, offset) \
+    r.x = LOAD4(name, offset); \
+    r.y = LOAD4(name, offset + uint(1))
+
+#define LOAD16(r, name, offset)          \
+    r.x = LOAD4(name, offset);           \
+    r.y = LOAD4(name, offset + uint(1)); \
+    r.z = LOAD4(name, offset + uint(2)); \
+    r.w = LOAD4(name, offset + uint(3))
+
+#define STORE16(name, offset, r)         \
+    STORE4(name, offset, r.x);           \
+    STORE4(name, offset + uint(1), r.y); \
+    STORE4(name, offset + uint(2), r.z); \
+    STORE4(name, offset + uint(3), r.w)
+
+#if defined(POOL_AVG) || defined(POOL_L2)
+#define POOL_OP(res, a, b) ((res) = (a) + (b))
+#define POOL_OP_float(res, a, b) (res = a + b)
+#define POOL_OP_vec2(res, a, b) ((res) = (a) + (b))
+#else /* defined(POOL_AVG) || defined(POOL_L2) */
+#define POOL_OP(res, a, b)        \
+    (res) = (a);                  \
+    if(isnan(a.x) || (a.x < b.x)) \
+    {                             \
+        res.x = b.x;              \
+    }                             \
+    if(isnan(a.y) || (a.y < b.y)) \
+    {                             \
+        res.y = b.y;              \
+    }                             \
+    if(isnan(a.z) || (a.z < b.z)) \
+    {                             \
+        res.z = b.z;              \
+    }                             \
+    if(isnan(a.w) || (a.w < b.w)) \
+    {                             \
+        res.w = b.w;              \
+    }
+#define POOL_OP_float(res, a, b) \
+    (res) = (a);                 \
+    if(isnan(a) || (a < b))      \
+    {                            \
+        res = b;                 \
+    }
+#define POOL_OP_vec2(res, a, b)   \
+    (res) = (a);                  \
+    if(isnan(a.x) || (a.x < b.x)) \
+    {                             \
+        res.x = b.x;              \
+    }                             \
+    if(isnan(a.y) || (a.y < b.y)) \
+    {                             \
+        res.y = b.y;              \
+    }
+#endif /* defined(POOL_AVG) || defined(POOL_L2) */
+
+#if defined(POOL_L2)
+#define POW2_OP(x, vec_size) ((x) * (x))
+#else /* defined(POOL_L2) */
+#define POW2_OP(x, vec_size) (x)
+#endif /* defined(POOL_L2) */
+
+#define DIV_OP(x, y) (x * (1.f / y))
+#define SQRT_OP(x) sqrt((x))
+
+#if defined(POOL_SIZE)
+// Set the initial value for the pooling operation accordingly with the data type
+#if defined(POOL_AVG) || defined(POOL_L2)
+#define INITIAL_VALUE 0.0f
+#else /* defined(POOL_AVG) || defined(POOL_L2) */
+#define INITIAL_VALUE -3.402823466385289e+38
+#endif // POOL_AVG
+#endif //POOL_SIZE
+
+#define POOLING3x3_STRIDE1(res, input, output)                                                                     \
+    vec4 data00;                                                                                                   \
+    vec2 data01;                                                                                                   \
+    vec4 data10;                                                                                                   \
+    vec2 data11;                                                                                                   \
+    vec4 data20;                                                                                                   \
+    vec2 data21;                                                                                                   \
+    LOAD16(data00, input, tensor3D_offset(input, 0, 0, 0));                                                        \
+    LOAD8(data01, input, tensor3D_offset(input, 0, 0, 0) + uint(4));                                               \
+    LOAD16(data10, input, tensor3D_offset(input, 0, 1, 0));                                                        \
+    LOAD8(data11, input, tensor3D_offset(input, 0, 1, 0) + uint(4));                                               \
+    LOAD16(data20, input, tensor3D_offset(input, 0, 2, 0));                                                        \
+    LOAD8(data21, input, tensor3D_offset(input, 0, 2, 0) + uint(4));                                               \
+    data00 = POW2_OP(data00, 4);                                                                                   \
+    data01 = POW2_OP(data01, 2);                                                                                   \
+    data10 = POW2_OP(data10, 4);                                                                                   \
+    data11 = POW2_OP(data11, 2);                                                                                   \
+    data20 = POW2_OP(data20, 4);                                                                                   \
+    data21 = POW2_OP(data21, 2);                                                                                   \
+    \
+    vec4 values000;                                                                                                \
+    vec4 values001;                                                                                                \
+    vec4 values010;                                                                                                \
+    vec4 values100;                                                                                                \
+    vec4 values101;                                                                                                \
+    vec4 values11;                                                                                                 \
+    vec4 values200;                                                                                                \
+    vec4 values201;                                                                                                \
+    vec4 values21;                                                                                                 \
+    values000.xyzw = data00.xyzy;                                                                                  \
+    values001.xyzw = data00.zwzw;                                                                                  \
+    values010.x    = data01.x;                                                                                     \
+    values010.y    = data00.w;                                                                                     \
+    values010.zw   = data01.xy;                                                                                    \
+    values100.xyzw = data10.xyzy;                                                                                  \
+    values101.xyzw = data10.zwzw;                                                                                  \
+    values11.x     = data11.x;                                                                                     \
+    values11.y     = data10.w;                                                                                     \
+    values11.zw    = data11.xy;                                                                                    \
+    values200.xyzw = data20.xyzy;                                                                                  \
+    values201.xyzw = data20.zwzw;                                                                                  \
+    values21.x     = data21.x;                                                                                     \
+    values21.y     = data20.w;                                                                                     \
+    values21.zw    = data21.xy;                                                                                    \
+    POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw);                                                        \
+    POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw);                                                        \
+    POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw))
+
+#define POOLING3x3_STRIDE2(res, input, output)                                                                     \
+    vec4  data000;                                                                                                 \
+    vec4  data001;                                                                                                 \
+    float data010;                                                                                                 \
+    vec4  data100;                                                                                                 \
+    vec4  data101;                                                                                                 \
+    float data11;                                                                                                  \
+    vec4  data200;                                                                                                 \
+    vec4  data201;                                                                                                 \
+    float data21;                                                                                                  \
+    LOAD16(data000, input, tensor3D_offset(input, 0, 0, 0));                                                       \
+    LOAD16(data001, input, tensor3D_offset(input, 0, 0, 0) + uint(4));                                             \
+    data010 = LOAD4(input, tensor3D_offset(input, 0, 0, 0) + uint(8));                                             \
+    LOAD16(data100, input, tensor3D_offset(input, 0, 1, 0));                                                       \
+    LOAD16(data101, input, tensor3D_offset(input, 0, 1, 0) + uint(4));                                             \
+    data11 = LOAD4(input, tensor3D_offset(input, 0, 1, 0) + uint(8));                                              \
+    LOAD16(data200, input, tensor3D_offset(input, 0, 2, 0));                                                       \
+    LOAD16(data201, input, tensor3D_offset(input, 0, 2, 0) + uint(4));                                             \
+    data21  = LOAD4(input, tensor3D_offset(input, 0, 2, 0) + uint(8));                                             \
+    data000 = POW2_OP(data000, 4);                                                                                 \
+    data001 = POW2_OP(data001, 4);                                                                                 \
+    data010 = POW2_OP(data010, 1);                                                                                 \
+    data100 = POW2_OP(data100, 4);                                                                                 \
+    data101 = POW2_OP(data101, 4);                                                                                 \
+    data11  = POW2_OP(data11, 1);                                                                                  \
+    data200 = POW2_OP(data200, 4);                                                                                 \
+    data201 = POW2_OP(data201, 4);                                                                                 \
+    data21  = POW2_OP(data21, 1);                                                                                  \
+    \
+    vec4 values000;                                                                                                \
+    vec4 values001;                                                                                                \
+    vec4 values010;                                                                                                \
+    vec4 values100;                                                                                                \
+    vec4 values101;                                                                                                \
+    vec4 values11;                                                                                                 \
+    vec4 values200;                                                                                                \
+    vec4 values201;                                                                                                \
+    vec4 values21;                                                                                                 \
+    values000.xyzw = data000.xyzz;                                                                                 \
+    values001.xyzw = vec4(data000.w, data001.xxy);                                                                 \
+    values010.xyzw = vec4(data001.zzw, data010);                                                                   \
+    values100.xyzw = data100.xyzz;                                                                                 \
+    values101.xyzw = vec4(data100.w, data101.xxy);                                                                 \
+    values11.xyzw  = vec4(data101.zzw, data11);                                                                    \
+    values200.xyzw = data200.xyzz;                                                                                 \
+    values201.xyzw = vec4(data200.w, data201.xxy);                                                                 \
+    values21.xyzw  = vec4(data201.zzw, data21);                                                                    \
+    POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw);                                                        \
+    POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw);                                                        \
+    POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw))
+
+#define POOLING3x3_STRIDE3(res, input, output)                                                         \
+    vec4 data000;                                                                                      \
+    vec4 data001;                                                                                      \
+    vec4 data010;                                                                                      \
+    vec4 data100;                                                                                      \
+    vec4 data101;                                                                                      \
+    vec4 data11;                                                                                       \
+    vec4 data200;                                                                                      \
+    vec4 data201;                                                                                      \
+    vec4 data21;                                                                                       \
+    LOAD16(data000, input, tensor3D_offset(input, 0, 0, 0));                                           \
+    LOAD16(data001, input, tensor3D_offset(input, 0, 0, 0) + uint(4));                                 \
+    LOAD16(data010, input, tensor3D_offset(input, 0, 0, 0) + uint(8));                                 \
+    LOAD16(data100, input, tensor3D_offset(input, 0, 1, 0));                                           \
+    LOAD16(data101, input, tensor3D_offset(input, 0, 1, 0) + uint(4));                                 \
+    LOAD16(data11, input, tensor3D_offset(input, 0, 1, 0) + uint(8));                                  \
+    LOAD16(data200, input, tensor3D_offset(input, 0, 2, 0));                                           \
+    LOAD16(data201, input, tensor3D_offset(input, 0, 2, 0) + uint(4));                                 \
+    LOAD16(data21, input, tensor3D_offset(input, 0, 2, 0) + uint(8));                                  \
+    data000 = POW2_OP(data000, 4);                                                                     \
+    data001 = POW2_OP(data001, 4);                                                                     \
+    data010 = POW2_OP(data010, 4);                                                                     \
+    data100 = POW2_OP(data100, 4);                                                                     \
+    data101 = POW2_OP(data101, 4);                                                                     \
+    data11  = POW2_OP(data11, 4);                                                                      \
+    data200 = POW2_OP(data200, 4);                                                                     \
+    data201 = POW2_OP(data201, 4);                                                                     \
+    data21  = POW2_OP(data21, 4);                                                                      \
+    \
+    POOL_OP(data000.xyzw, data000.xyzw, data100.xyzw);                                                 \
+    POOL_OP(data001.xyzw, data001.xyzw, data101.xyzw);                                                 \
+    POOL_OP(data010.xyzw, data010.xyzw, data11.xyzw);                                                  \
+    POOL_OP(data000.xyzw, data000.xyzw, data200.xyzw);                                                 \
+    POOL_OP(data001.xyzw, data001.xyzw, data201.xyzw);                                                 \
+    POOL_OP(data010.xyzw, data010.xyzw, data21.xyzw);                                                  \
+    POOL_OP(res.xyzw, vec4(data000.xw, data001.z, data010.y), vec4(data000.y, data001.xw, data010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(data000.z, data001.y data010.xw))
+
+float calculate_max(const int pool_size, Tensor3D src, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y)
+{
+    int start_x = int(gl_GlobalInvocationID.x) * stride_x - pad_x;
+    int start_y = int(gl_GlobalInvocationID.y) * stride_y - pad_y;
+    int end_x   = int(min(start_x + pool_size, upper_bound_w));
+    int end_y   = int(min(start_y + pool_size, upper_bound_h));
+
+    float data_max;
+    data_max = LOAD4(src, tensor3D_offset(src, 0, 0, 0));
+
+    for(int i = 0; (start_x + i) < end_x; ++i)
+    {
+        for(int j = 0; (start_y + j) < end_y; ++j)
+        {
+            float data = LOAD4(src, tensor3D_offset(src, i, j, 0));
+            POOL_OP_float(data_max, data_max, data);
+        }
+    }
+
+    return data_max;
+}
+
+float calculate_avg(const int pool_size, Tensor3D src, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y)
+{
+    int start_x = int(gl_GlobalInvocationID.x) * stride_x - pad_x;
+    int start_y = int(gl_GlobalInvocationID.y) * stride_y - pad_y;
+    int end_x   = int(min(start_x + pool_size, upper_bound_w));
+    int end_y   = int(min(start_y + pool_size, upper_bound_h));
+
+    float data_total = 0.0f;
+    for(int i = 0; (start_x + i) < end_x; i++)
+    {
+        for(int j = 0; (start_y + j) < end_y; ++j)
+        {
+            float data = LOAD4(src, tensor3D_offset(src, i, j, 0));
+            if(isnan(data))
+            {
+                data = 0.0f;
+            }
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data = POW2_OP(data, 1);
+#endif /* defined(POOL_L2) */
+            data_total = data_total + data;
+        }
+    }
+
+    return data_total / float((end_y - start_y) * (end_x - start_x));
+}
+
+#ifdef POOLING_LAYER_2
+/** Performs a pooling function of pool size equal to 2.
+ *
+ * @note Supported data types are F32;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F32
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    //Load and calculate data
+    float res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    res = calculate_avg(2, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    res = calculate_max(2, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    // Store result
+    STORE4(dst, CURRENT_OFFSET(dst), res);
+}
+
+#elif defined(POOLING_LAYER_3)
+/** Performs a pooling function of pool size equal to 3.
+ *
+ * @note Supported data types are F32;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F32
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    //Load and calculate data
+    float res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    res = calculate_avg(3, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    res = calculate_max(3, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    // Store result
+    STORE4(dst, CURRENT_OFFSET(dst), res);
+}
+
+#elif defined(POOLING_LAYER_3_OPTIMIZED)
+/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3
+ *
+ * @note Supported data types are F32;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F32
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    vec4 res;
+    // Perform pooling 3x3 for 4 output elements
+#if STRIDE_X == 1
+    POOLING3x3_STRIDE1(res, src, dst);
+#elif STRIDE_X == 2
+    POOLING3x3_STRIDE2(res, src, dst);
+#elif STRIDE_X == 3
+    POOLING3x3_STRIDE3(res, src, dst);
+#endif /*STRIDE_X == 1*/
+
+    // Divide by pool region in case of average pooling
+#if defined(POOL_AVG) || defined(POOL_L2)
+    ivec4 start_x = ((ivec4(int(gl_GlobalInvocationID.x) * 4) + ivec4(0, 1, 2, 3)) * (ivec4(STRIDE_X))) - (ivec4(PAD_X));
+    int   start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y;
+    ivec4 end_x   = min((start_x + (ivec4(3))), (ivec4(MAX_WIDTH)));
+    int   end_y   = min((start_y + 3), MAX_HEIGHT);
+    res *= (vec4((1.f)) / vec4((ivec4(end_y - start_y)) * (end_x - start_x)));
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    STORE16(dst, CURRENT_OFFSET(dst), res);
+}
+
+#elif defined(POOLING_LAYER_7)
+/** Performs a pooling function of pool size equal to 7.
+ *
+ * @note Supported data types are F32;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F32
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    //Load and calculate data
+    float res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    res = calculate_avg(7, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    res = calculate_max(7, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    // Store result
+    STORE4(dst, CURRENT_OFFSET(dst), res);
+}
+
+#elif defined(POOLING_LAYER_N)
+/** Performs a pooling function of pool size equal to N
+ *
+ * @note Supported data types are F32;
+ * @note Pool size must be passed using POOL_SIZE e.g. POOL_SIZE=13;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F32
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
+
+    vec4 vdata0;
+    vdata0 = vec4(INITIAL_VALUE);
+    vec4 vdata1;
+    vdata1 = vec4(INITIAL_VALUE);
+    float sdata;
+    sdata = float(INITIAL_VALUE);
+
+    for(int y = 0; y < int(POOL_SIZE); y++)
+    {
+        int x = 0;
+        for(; x <= (int(POOL_SIZE) - 8); x += 8)
+        {
+            vec4 data2;
+            vec4 data3;
+            LOAD16(data2, src, tensor3D_offset(src, x, y, 0));
+            LOAD16(data3, src, tensor3D_offset(src, x, y, 0) + uint(4));
+
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data2 *= data2;
+            data3 *= data3;
+#endif /* defined(POOL_L2) */
+
+            POOL_OP(vdata0, vdata0, data2);
+            POOL_OP(vdata1, vdata1, data3);
+        }
+
+        // Leftover
+        for(; x < int(POOL_SIZE); ++x)
+        {
+            float data4 = LOAD4(src, tensor3D_offset(src, x, y, 0));
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data4 *= data4;
+#endif /* defined(POOL_L2) */
+            POOL_OP_float(sdata, sdata, data4);
+        }
+    }
+
+    //Reduce result
+    vec4 reduce4;
+    POOL_OP(reduce4, vdata0.xyzw, vdata1.xyzw);
+    vec2 reduce2;
+    POOL_OP_vec2(reduce2, reduce4.xy, reduce4.zw);
+    float res;
+    POOL_OP_float(res, reduce2.x, reduce2.y);
+    POOL_OP_float(res, res, sdata);
+
+#if defined(POOL_AVG) || defined(POOL_L2)
+    {
+        // Divide by pool region in case of average pooling
+        int   start_x = int(gl_GlobalInvocationID.x) * STRIDE_X - PAD_X;
+        int   start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y;
+        int   end_x   = int(min(STRIDE_X + POOL_SIZE, MAX_WIDTH));
+        int   end_y   = int(min(STRIDE_Y + POOL_SIZE, MAX_HEIGHT));
+        float res1    = float((end_y - start_y) * (end_x - start_x));
+        res           = DIV_OP(res, res1);
+    }
+#endif /* defined(POOL_AVG) || defined(POOL_L2) */
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    // Store result
+    STORE4(dst, CURRENT_OFFSET(dst), res);
+}
+#endif /* POOLING_LAYER_2 */
+
+#elif defined(DATA_TYPE_FP16)
+
+precision mediump float;
+
+vec2 load_and_unpack(Tensor3D, uint);
+vec2 calculate_max(const int, Tensor3D, const int, const int, const int, const int, const int, const int);
+vec2 calculate_avg(const int, Tensor3D, const int, const int, const int, const int, const int, const int);
+
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(dst, 2, uint, writeonly);
+
+layout(std140) uniform shader_params
+{
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+};
+
+#define LOAD2_fp16(r, name, offset) \
+    r.xy = load_and_unpack(name, offset)
+
+#define LOAD4_fp16(r, name, offset)       \
+    r.xy = load_and_unpack(name, offset); \
+    r.zw = load_and_unpack(name, offset + uint(1))
+
+#define STORE4_fp16(name, offset, r)             \
+    uint datastore1;                             \
+    uint datastore2;                             \
+    datastore1 = uint(packHalf2x16(r.xy));       \
+    datastore2 = uint(packHalf2x16(r.zw));       \
+    STORE1(name, offset << uint(1), datastore1); \
+    STORE1(name, (offset << uint(1)) + uint(1), datastore2)
+
+#if defined(POOL_AVG) || defined(POOL_L2)
+#define POOL_OP(res, a, b) ((res) = (a) + (b))
+#define POOL_OP_float(res, a, b) (res = a + b)
+#define POOL_OP_vec2(res, a, b) ((res) = (a) + (b))
+#else /* defined(POOL_AVG) || defined(POOL_L2) */
+#define POOL_OP(res, a, b)        \
+    (res) = (a);                  \
+    if(isnan(a.x) || (a.x < b.x)) \
+    {                             \
+        res.x = b.x;              \
+    }                             \
+    if(isnan(a.y) || (a.y < b.y)) \
+    {                             \
+        res.y = b.y;              \
+    }                             \
+    if(isnan(a.z) || (a.z < b.z)) \
+    {                             \
+        res.z = b.z;              \
+    }                             \
+    if(isnan(a.w) || (a.w < b.w)) \
+    {                             \
+        res.w = b.w;              \
+    }
+#define POOL_OP_float(res, a, b) \
+    (res) = (a);                 \
+    if(isnan(a) || (a < b))      \
+    {                            \
+        res = b;                 \
+    }
+#define POOL_OP_vec2(res, a, b)   \
+    (res) = (a);                  \
+    if(isnan(a.x) || (a.x < b.x)) \
+    {                             \
+        res.x = b.x;              \
+    }                             \
+    if(isnan(a.y) || (a.y < b.y)) \
+    {                             \
+        res.y = b.y;              \
+    }
+#endif /* defined(POOL_AVG) || defined(POOL_L2) */
+
+#if defined(POOL_L2)
+#define POW2_OP(x, vec_size) ((x) * (x))
+#else /* defined(POOL_L2) */
+#define POW2_OP(x, vec_size) (x)
+#endif /* defined(POOL_L2) */
+
+#define DIV_OP(x, y) (x * (1.f / y))
+#define SQRT_OP(x) sqrt((x))
+
+#if defined(POOL_SIZE)
+// Set the initial value for the pooling operation accordingly with the data type
+#if defined(POOL_AVG) || defined(POOL_L2)
+#define INITIAL_VALUE 0.0f
+#else /* defined(POOL_AVG) || defined(POOL_L2) */
+#define INITIAL_VALUE -65504.0f
+#endif //POOL_AVG
+#endif //POOL_SIZE
+
+#define POOLING3x3_STRIDE1_fp16(res, input, output)                                                                \
+    vec4 data00;                                                                                                   \
+    vec2 data01;                                                                                                   \
+    vec4 data10;                                                                                                   \
+    vec2 data11;                                                                                                   \
+    vec4 data20;                                                                                                   \
+    vec2 data21;                                                                                                   \
+    LOAD4_fp16(data00, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)));                                  \
+    LOAD2_fp16(data01, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)) + uint(2));                        \
+    LOAD4_fp16(data10, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)));                                  \
+    LOAD2_fp16(data11, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)) + uint(2));                        \
+    LOAD4_fp16(data20, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)));                                  \
+    LOAD2_fp16(data21, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)) + uint(2));                        \
+    data00 = POW2_OP(data00, 4);                                                                                   \
+    data01 = POW2_OP(data01, 2);                                                                                   \
+    data10 = POW2_OP(data10, 4);                                                                                   \
+    data11 = POW2_OP(data11, 2);                                                                                   \
+    data20 = POW2_OP(data20, 4);                                                                                   \
+    data21 = POW2_OP(data21, 2);                                                                                   \
+    \
+    vec4 values000;                                                                                                \
+    vec4 values001;                                                                                                \
+    vec4 values010;                                                                                                \
+    vec4 values100;                                                                                                \
+    vec4 values101;                                                                                                \
+    vec4 values11;                                                                                                 \
+    vec4 values200;                                                                                                \
+    vec4 values201;                                                                                                \
+    vec4 values21;                                                                                                 \
+    values000.xyzw = data00.xyzy;                                                                                  \
+    values001.xyzw = data00.zwzw;                                                                                  \
+    values010.x    = data01.x;                                                                                     \
+    values010.y    = data00.w;                                                                                     \
+    values010.zw   = data01.xy;                                                                                    \
+    values100.xyzw = data10.xyzy;                                                                                  \
+    values101.xyzw = data10.zwzw;                                                                                  \
+    values11.x     = data11.x;                                                                                     \
+    values11.y     = data10.w;                                                                                     \
+    values11.zw    = data11.xy;                                                                                    \
+    values200.xyzw = data20.xyzy;                                                                                  \
+    values201.xyzw = data20.zwzw;                                                                                  \
+    values21.x     = data21.x;                                                                                     \
+    values21.y     = data20.w;                                                                                     \
+    values21.zw    = data21.xy;                                                                                    \
+    POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw);                                                        \
+    POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw);                                                        \
+    POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw))
+
+#define POOLING3x3_STRIDE2_fp16(res, input, output)                                                                \
+    vec4  data000;                                                                                                 \
+    vec4  data001;                                                                                                 \
+    float data010;                                                                                                 \
+    vec4  data100;                                                                                                 \
+    vec4  data101;                                                                                                 \
+    float data11;                                                                                                  \
+    vec4  data200;                                                                                                 \
+    vec4  data201;                                                                                                 \
+    float data21;                                                                                                  \
+    vec2  datamiddle0;                                                                                             \
+    vec2  datamiddle1;                                                                                             \
+    vec2  datamiddle2;                                                                                             \
+    LOAD4_fp16(data000, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)));                                 \
+    LOAD4_fp16(data001, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)) + uint(2));                       \
+    datamiddle0 = load_and_unpack(input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)) + uint(4));             \
+    data010     = datamiddle0.x;                                                                                   \
+    LOAD4_fp16(data100, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)));                                 \
+    LOAD4_fp16(data101, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)) + uint(2));                       \
+    datamiddle1 = load_and_unpack(input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)) + uint(4));             \
+    data11      = datamiddle1.x;                                                                                   \
+    LOAD4_fp16(data200, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)));                                 \
+    LOAD4_fp16(data201, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)) + uint(2));                       \
+    datamiddle2 = load_and_unpack(input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)) + uint(4));             \
+    data21      = datamiddle2.x;                                                                                   \
+    data000     = POW2_OP(data000, 4);                                                                             \
+    data001     = POW2_OP(data001, 4);                                                                             \
+    data010     = POW2_OP(data010, 1);                                                                             \
+    data100     = POW2_OP(data100, 4);                                                                             \
+    data101     = POW2_OP(data101, 4);                                                                             \
+    data11      = POW2_OP(data11, 1);                                                                              \
+    data200     = POW2_OP(data200, 4);                                                                             \
+    data201     = POW2_OP(data201, 4);                                                                             \
+    data21      = POW2_OP(data21, 1);                                                                              \
+    \
+    vec4 values000;                                                                                                \
+    vec4 values001;                                                                                                \
+    vec4 values010;                                                                                                \
+    vec4 values100;                                                                                                \
+    vec4 values101;                                                                                                \
+    vec4 values11;                                                                                                 \
+    vec4 values200;                                                                                                \
+    vec4 values201;                                                                                                \
+    vec4 values21;                                                                                                 \
+    values000.xyzw = data000.xyzz;                                                                                 \
+    values001.xyzw = vec4(data000.w, data001.xxy);                                                                 \
+    values010.xyzw = vec4(data001.zzw, data010);                                                                   \
+    values100.xyzw = data100.xyzz;                                                                                 \
+    values101.xyzw = vec4(data100.w, data101.xxy);                                                                 \
+    values11.xyzw  = vec4(data101.zzw, data11);                                                                    \
+    values200.xyzw = data200.xyzz;                                                                                 \
+    values201.xyzw = vec4(data200.w, data201.xxy);                                                                 \
+    values21.xyzw  = vec4(data201.zzw, data21);                                                                    \
+    POOL_OP(values000.xyzw, values000.xyzw, values100.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values101.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values11.xyzw);                                                        \
+    POOL_OP(values000.xyzw, values000.xyzw, values200.xyzw);                                                       \
+    POOL_OP(values001.xyzw, values001.xyzw, values201.xyzw);                                                       \
+    POOL_OP(values010.xyzw, values010.xyzw, values21.xyzw);                                                        \
+    POOL_OP(res.xyzw, vec4(values000.xw, values001.z, values010.y), vec4(values000.y, values001.xw, values010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(values000.z, values001.y, values010.xw))
+
+#define POOLING3x3_STRIDE3_fp16(res, input, output)                                                    \
+    vec4 data000;                                                                                      \
+    vec4 data001;                                                                                      \
+    vec4 data010;                                                                                      \
+    vec4 data100;                                                                                      \
+    vec4 data101;                                                                                      \
+    vec4 data11;                                                                                       \
+    vec4 data200;                                                                                      \
+    vec4 data201;                                                                                      \
+    vec4 data21;                                                                                       \
+    LOAD4_fp16(data000, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)));                     \
+    LOAD4_fp16(data001, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)) + uint(2));           \
+    LOAD4_fp16(data010, input, (tensor3D_offset_fp16(input, 0, 0, 0) >> uint(2)) + uint(4));           \
+    LOAD4_fp16(data100, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)));                     \
+    LOAD4_fp16(data101, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)) + uint(2));           \
+    LOAD4_fp16(data11, input, (tensor3D_offset_fp16(input, 0, 1, 0) >> uint(2)) + uint(4));            \
+    LOAD4_fp16(data200, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)));                     \
+    LOAD4_fp16(data201, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)) + uint(2));           \
+    LOAD4_fp16(data21, input, (tensor3D_offset_fp16(input, 0, 2, 0) >> uint(2)) + uint(4));            \
+    data000 = POW2_OP(data000, 4);                                                                     \
+    data001 = POW2_OP(data001, 4);                                                                     \
+    data010 = POW2_OP(data010, 4);                                                                     \
+    data100 = POW2_OP(data100, 4);                                                                     \
+    data101 = POW2_OP(data101, 4);                                                                     \
+    data11  = POW2_OP(data11, 4);                                                                      \
+    data200 = POW2_OP(data200, 4);                                                                     \
+    data201 = POW2_OP(data201, 4);                                                                     \
+    data21  = POW2_OP(data21, 4);                                                                      \
+    \
+    POOL_OP(data000.xyzw, data000.xyzw, data100.xyzw);                                                 \
+    POOL_OP(data001.xyzw, data001.xyzw, data101.xyzw);                                                 \
+    POOL_OP(data010.xyzw, data010.xyzw, data11.xyzw);                                                  \
+    POOL_OP(data000.xyzw, data000.xyzw, data200.xyzw);                                                 \
+    POOL_OP(data001.xyzw, data001.xyzw, data201.xyzw);                                                 \
+    POOL_OP(data010.xyzw, data010.xyzw, data21.xyzw);                                                  \
+    POOL_OP(res.xyzw, vec4(data000.xw, data001.z, data010.y), vec4(data000.y, data001.xw, data010.z)); \
+    POOL_OP(res.xyzw, res.xyzw, vec4(data000.z, data001.y data010.xw))
+
+vec2 load_and_unpack(Tensor3D src, uint offset)
+{
+    uint packed_s;
+    vec2 s;
+    LOAD1(packed_s, src, offset);
+
+    s = vec2(unpackHalf2x16(packed_s));
+    return s;
+}
+
+vec2 calculate_max(const int pool_size, Tensor3D src, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y)
+{
+    int start_x1 = int(gl_GlobalInvocationID.x) * stride_x - pad_x;
+    int start_y1 = int(gl_GlobalInvocationID.y) * stride_y - pad_y;
+    int end_x1   = int(min(start_x1 + pool_size, upper_bound_w));
+    int end_y1   = int(min(start_y1 + pool_size, upper_bound_h));
+
+    int start_x2 = start_x1 + stride_x;
+    int start_y2 = start_y1;
+    int end_x2   = int(min(start_x2 + pool_size, upper_bound_w));
+    int end_y2   = int(min(start_y2 + pool_size, upper_bound_h));
+
+    //Initialize maximum
+    vec2 data_max = vec2(0);
+
+    //Load and Set initial maximum1
+    vec2 data_init1 = load_and_unpack(src, tensor3D_offset_fp16(src, 0, 0, 0) >> uint(2));
+    data_max.x      = data_init1.x;
+
+    //Load and Set initial maximum2
+    if(end_x1 < upper_bound_w)
+    {
+        if((stride_x % 2) == 0)
+        {
+            vec2 data_init2 = load_and_unpack(src, tensor3D_offset_fp16(src, stride_x, 0, 0) >> uint(2));
+            data_max.y      = data_init2.x;
+        }
+        else
+        {
+            vec2 data_init2 = load_and_unpack(src, tensor3D_offset_fp16(src, stride_x - 1, 0, 0) >> uint(2));
+            data_max.y      = data_init2.y;
+        }
+    }
+
+    for(int i = 0; (start_y1 + i) < end_y1; i++)
+        for(int j = 0; (start_x1 + j) < end_x1; j = j + 2)
+        {
+            //Calculate maximum1
+            if((start_x1 + j + 1) < end_x1)
+            {
+                vec2  data1 = load_and_unpack(src, tensor3D_offset_fp16(src, j, i, 0) >> uint(2));
+                float data_mr1;
+                POOL_OP_float(data_mr1, data1.x, data1.y);
+                POOL_OP_float(data_max.x, data_max.x, data_mr1);
+            }
+            else
+            {
+                vec2 data1 = load_and_unpack(src, tensor3D_offset_fp16(src, j, i, 0) >> uint(2));
+                POOL_OP_float(data_max.x, data_max.x, data1.x);
+            }
+
+            //Calculate maximum2
+            if((start_x2 + j) < end_x2 && end_x1 < upper_bound_w)
+            {
+                if((stride_x % 2) == 0)
+                {
+                    vec2 data2 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x), i, 0) >> uint(2)));
+
+                    if((start_x2 + j + 1) < end_x2)
+                    {
+                        float data_mr2;
+                        POOL_OP_float(data_mr2, data2.x, data2.y);
+                        POOL_OP_float(data_max.y, data_max.y, data_mr2);
+                    }
+                    else
+                    {
+                        POOL_OP_float(data_max.y, data_max.y, data2.x);
+                    }
+                }
+                else
+                {
+                    vec2 data2 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x - 1), i, 0) >> uint(2)));
+                    vec2 data3 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x + 1), i, 0) >> uint(2)));
+                    if((start_x2 + j + 1) < end_x2)
+                    {
+                        float data_mr2;
+                        POOL_OP_float(data_mr2, data3.x, data2.y);
+                        POOL_OP_float(data_max.y, data_max.y, data_mr2);
+                    }
+                    else
+                    {
+                        POOL_OP_float(data_max.y, data_max.y, data2.y);
+                    }
+                }
+            }
+        }
+    return data_max;
+}
+
+vec2 calculate_avg(const int pool_size, Tensor3D src, const int upper_bound_w, const int upper_bound_h, const int pad_x, const int pad_y, const int stride_x, const int stride_y)
+{
+    int start_x1 = int(gl_GlobalInvocationID.x) * stride_x - pad_x;
+    int start_y1 = int(gl_GlobalInvocationID.y) * stride_y - pad_y;
+    int end_x1   = int(min(start_x1 + pool_size, upper_bound_w));
+    int end_y1   = int(min(start_y1 + pool_size, upper_bound_h));
+
+    int start_x2 = start_x1 + stride_x;
+    int start_y2 = start_y1;
+    int end_x2   = int(min(start_x2 + pool_size, upper_bound_w));
+    int end_y2   = int(min(start_y2 + pool_size, upper_bound_h));
+
+    //Initialize sum
+    float data_total1 = float(0);
+    float data_total2 = float(0);
+    for(int i = 0; (start_y1 + i) < end_y1; i++)
+        for(int j = 0; (start_x1 + j) < end_x1; j = j + 2)
+        {
+            vec2 data1 = load_and_unpack(src, tensor3D_offset_fp16(src, j, i, 0) >> uint(2));
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data1 = POW2_OP(data1, 2);
+#endif /* defined(POOL_L2) */
+            //Calculate sum1
+            if((start_x1 + j + 1) < end_x1)
+            {
+                data_total1 = data_total1 + data1.x + data1.y;
+            }
+            else
+            {
+                data_total1 = data_total1 + data1.x;
+            }
+
+            //Calculate sum2
+            if((start_x2 + j) < end_x2 && end_x1 < upper_bound_w)
+            {
+                if((stride_x % 2) == 0)
+                {
+                    vec2 data2 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x + 1), i, 0) >> uint(2)));
+#if defined(POOL_L2)
+                    // Raise to power of 2 for L2 Pooling
+                    data2 = POW2_OP(data2, 2);
+#endif /* defined(POOL_L2) */
+                    if((start_x2 + j + 1) < end_x2)
+                    {
+                        data_total2 = data_total2 + data2.x + data2.y;
+                    }
+                    else
+                    {
+                        data_total2 = data_total2 + data2.x;
+                    }
+                }
+                else
+                {
+                    vec2 data2 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x - 1), i, 0) >> uint(2)));
+                    vec2 data3 = load_and_unpack(src, (tensor3D_offset_fp16(src, (j + stride_x + 1), i, 0) >> uint(2)));
+#if defined(POOL_L2)
+                    // Raise to power of 2 for L2 Pooling
+                    data2 = POW2_OP(data2, 2);
+                    data3 = POW2_OP(data3, 2);
+#endif /* defined(POOL_L2) */
+                    if((start_x2 + j + 1) < end_x2)
+                    {
+                        data_total2 = data_total2 + data3.x + data2.y;
+                    }
+                    else
+                    {
+                        data_total2 = data_total2 + data2.y;
+                    }
+                }
+            }
+        }
+    //Calculate average
+    vec2 data_avg;
+    data_avg.x = data_total1 / float((end_y1 - start_y1) * (end_x1 - start_x1));
+    data_avg.y = data_total2 / float((end_y2 - start_y2) * (end_x2 - start_x2));
+
+    return data_avg;
+}
+
+#ifdef POOLING_LAYER_2
+/** Performs a pooling function of pool size equal to 2.
+ *
+ * @note Supported data types are F16;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F16
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    //Load and calculate data
+    vec2 data;
+    uint res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    data = calculate_avg(2, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    data = calculate_max(2, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    data = SQRT_OP(data);
+#endif /* defined(POOL_L2) */
+
+    res = uint(packHalf2x16(data));
+
+    // Store result
+    STORE1(dst, CURRENT_OFFSET(dst) >> uint(2), res);
+}
+
+#elif defined(POOLING_LAYER_3)
+/** Performs a pooling function of pool size equal to 3.
+ *
+ * @note Supported data types are F16;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F16
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    //Load and calculate data
+    vec2 data;
+    uint res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    data = calculate_avg(3, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    data = calculate_max(3, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    data = SQRT_OP(data);
+#endif /* defined(POOL_L2) */
+
+    res = uint(packHalf2x16(data));
+
+    // Store result
+    STORE1(dst, CURRENT_OFFSET(dst) >> uint(2), res);
+}
+
+#elif defined(POOLING_LAYER_3_OPTIMIZED)
+/** Performs an optimized pooling function of pool size equal to 3 when the stride_x is less equal than 3
+ *
+ * @note Supported data types are F16;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F16
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    vec4 res;
+    // Perform pooling 3x3 for 4 output elements
+#if STRIDE_X == 1
+    POOLING3x3_STRIDE1_fp16(res, src, dst);
+#elif STRIDE_X == 2
+    POOLING3x3_STRIDE2_fp16(res, src, dst);
+#elif STRIDE_X == 3
+    POOLING3x3_STRIDE3_fp16(res, src, dst);
+#endif /*STRIDE_X == 1*/
+
+    // Divide by pool region in case of average pooling
+#if defined(POOL_AVG) || defined(POOL_L2)
+    ivec4 start_x = ((ivec4(int(gl_GlobalInvocationID.x) * 4) + ivec4(0, 1, 2, 3)) * (ivec4(STRIDE_X))) - (ivec4(PAD_X));
+    int   start_y = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y;
+    ivec4 end_x   = min((start_x + (ivec4(3))), (ivec4(MAX_WIDTH)));
+    int   end_y   = min((start_y + 3), MAX_HEIGHT);
+    res *= (vec4((1.f)) / vec4((ivec4(end_y - start_y)) * (end_x - start_x)));
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    res = SQRT_OP(res);
+#endif /* defined(POOL_L2) */
+
+    STORE4_fp16(dst, CURRENT_OFFSET(dst) >> uint(3), res);
+}
+
+#elif defined(POOLING_LAYER_7)
+/** Performs a pooling function of pool size equal to 7.
+ *
+ * @note Supported data types are F16;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F16
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    //Load and calculate data
+    vec2 data;
+    uint res;
+#if defined(POOL_AVG) || defined(POOL_L2)
+    data = calculate_avg(7, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#else  /*POOL_AVG*/
+    data = calculate_max(7, src, MAX_WIDTH, MAX_HEIGHT, PAD_X, PAD_Y, STRIDE_X, STRIDE_Y);
+#endif /*POOL_AVG*/
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    data = SQRT_OP(data);
+#endif /* defined(POOL_L2) */
+
+    res = uint(packHalf2x16(data));
+
+    // Store result
+    STORE1(dst, CURRENT_OFFSET(dst) >> uint(2), res);
+}
+
+#elif defined(POOLING_LAYER_N)
+/** Performs a pooling function of pool size equal to N
+ *
+ * @note Supported data types are F16;
+ * @note Pool size must be passed using POOL_SIZE e.g. POOL_SIZE=13;
+ * @note In case of average pooling the following information must be passed at compile time:
+ *       POOL_AVG must be provided otherwise max pooling will be performed.
+ *       MAX_WIDTH and MAX_HEIGHT which are the maximum accessible indeces in x and y dimensions (width + pad)
+ *       STRIDE_X and STRIDE_Y which are the steps of the window along the x and y directions
+ *       PAD_X and PAD_Y which are the pooling paddings in x and y dimension
+ *
+ * @param[in]  src_ptr                           Pointer to the source image. Supported data types: F16
+ * @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_stride_z                      Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source image
+ * @param[out] dst_ptr                           Pointer to the destination image. Supported data types: same as @p src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination image 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 image 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 source tensor in Z dimension (in bytes)
+ * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination image
+ */
+void main(void)
+{
+    // Get pixels pointer
+    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT_FP16(src);
+    Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_FP16(dst);
+
+    vec4 vdata00;
+    vdata00 = vec4(INITIAL_VALUE);
+    vec4 vdata01;
+    vdata01 = vec4(INITIAL_VALUE);
+    vec4 vdata10;
+    vdata10 = vec4(INITIAL_VALUE);
+    vec4 vdata11;
+    vdata11 = vec4(INITIAL_VALUE);
+    vec2 sdata;
+    sdata = vec2(INITIAL_VALUE);
+
+    for(int y = 0; y < int(POOL_SIZE); y++)
+    {
+        int x = 0;
+        for(; x <= (int(POOL_SIZE) - 8); x += 8)
+        {
+            vec4 data2;
+            vec4 data3;
+            LOAD4_fp16(data2, src, (tensor3D_offset_fp16(src, x, y, 0) >> uint(2)));
+            LOAD4_fp16(data3, src, (tensor3D_offset_fp16(src, x, y, 0) >> uint(2)) + uint(2));
+
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data2 *= data2;
+            data3 *= data3;
+#endif /* defined(POOL_L2) */
+
+            POOL_OP(vdata00, vdata00, data2);
+            POOL_OP(vdata10, vdata10, data3);
+        }
+
+        // Leftover
+        for(; x < int(POOL_SIZE); x = x + 2)
+        {
+            vec2 data4middle;
+            data4middle = load_and_unpack(src, (tensor3D_offset_fp16(src, x, y, 0) >> uint(2)));
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data4middle *= data4middle;
+#endif /* defined(POOL_L2) */
+            if((x + 1) >= int(POOL_SIZE))
+            {
+                POOL_OP_float(sdata.x, sdata.x, data4middle.x);
+            }
+            else
+            {
+                float data4;
+                POOL_OP_float(data4, data4middle.x, data4middle.y);
+                POOL_OP_float(sdata.x, sdata.x, data4);
+            }
+        }
+    }
+
+    for(int y = STRIDE_X; y < int(POOL_SIZE + STRIDE_X); y++)
+    {
+        int x1 = STRIDE_X;
+        for(; x1 <= (int(POOL_SIZE + STRIDE_X) - 8); x1 += 8)
+        {
+            vec4 data2;
+            vec4 data3;
+            LOAD4_fp16(data2, src, (tensor3D_offset_fp16(src, x1, y, 0) >> uint(2)));
+            LOAD4_fp16(data3, src, (tensor3D_offset_fp16(src, x1, y, 0) >> uint(2)) + uint(2));
+
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data2 *= data2;
+            data3 *= data3;
+#endif /* defined(POOL_L2) */
+
+            POOL_OP(vdata01, vdata01, data2);
+            POOL_OP(vdata11, vdata11, data3);
+        }
+
+        // Leftover
+        for(; x1 < int(POOL_SIZE + STRIDE_X); x1 = x1 + 2)
+        {
+            vec2 data4middle;
+            data4middle = load_and_unpack(src, (tensor3D_offset_fp16(src, x1, y, 0) >> uint(2)));
+#if defined(POOL_L2)
+            // Raise to power of 2 for L2 Pooling
+            data4middle *= data4middle;
+#endif /* defined(POOL_L2) */
+            if((x1 + 1) >= int(POOL_SIZE + STRIDE_X))
+            {
+                POOL_OP_float(sdata.y, sdata.y, data4middle.x);
+            }
+            else
+            {
+                float data4;
+                POOL_OP_float(data4, data4middle.x, data4middle.y);
+                POOL_OP_float(sdata.y, sdata.y, data4);
+            }
+        }
+    }
+
+    //Reduce result
+    vec4 reduce40;
+    POOL_OP(reduce40, vdata00.xyzw, vdata10.xyzw);
+    vec2 reduce20;
+    POOL_OP_vec2(reduce20, reduce40.xy, reduce40.zw);
+    vec4 reduce41;
+    POOL_OP(reduce41, vdata01.xyzw, vdata11.xyzw);
+    vec2 reduce21;
+    POOL_OP_vec2(reduce21, reduce41.xy, reduce41.zw);
+    vec2 data;
+    POOL_OP_float(data.x, reduce20.x, reduce20.y);
+    POOL_OP_float(data.x, data.x, sdata.x);
+    POOL_OP_float(data.y, reduce21.x, reduce21.y);
+    POOL_OP_float(data.y, data.y, sdata.y);
+
+#if defined(POOL_AVG) || defined(POOL_L2)
+    {
+        // Divide by pool region in case of average pooling
+        int  start_x1 = int(gl_GlobalInvocationID.x) * STRIDE_X - PAD_X;
+        int  start_y1 = int(gl_GlobalInvocationID.y) * STRIDE_Y - PAD_Y;
+        int  end_x1   = int(min(start_x1 + POOL_SIZE, MAX_WIDTH));
+        int  end_y1   = int(min(start_y1 + POOL_SIZE, MAX_HEIGHT));
+        int  start_x2 = start_x1 + STRIDE_X;
+        int  start_y2 = start_y1;
+        int  end_x2   = int(min(start_x2 + POOL_SIZE, MAX_WIDTH));
+        int  end_y2   = int(min(start_y2 + POOL_SIZE, MAX_HEIGHT));
+        vec2 res1;
+        res1.x = float((end_y1 - start_y1) * (end_x1 - start_x1));
+        res1.y = float((end_y2 - start_y2) * (end_x2 - start_x2));
+        data.x = DIV_OP(data.x, res1.x);
+        data.y = DIV_OP(data.y, res1.y);
+    }
+#endif /* defined(POOL_AVG) || defined(POOL_L2) */
+
+#if defined(POOL_L2)
+    // Take square root of the result in L2 pooling
+    data = SQRT_OP(data);
+#endif /* defined(POOL_L2) */
+    uint res;
+    res = uint(packHalf2x16(data));
+
+    // Store result
+    STORE1(dst, CURRENT_OFFSET(dst) >> uint(2), res);
+}
+#endif /*POOLING_LAYER_2*/
+#endif /*DATA_TYPE_FP32 */
diff --git a/src/core/GLES_COMPUTE/cs_shaders/softmax_layer.cs b/src/core/GLES_COMPUTE/cs_shaders/softmax_layer.cs
new file mode 100644
index 0000000000..0bbabeaafc
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/softmax_layer.cs
@@ -0,0 +1,541 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+
+#include "helpers.h"
+
+#define MAX_OP(x, y) max((x), (y))
+#define ADD_OP(x, y) ((x) + (y))
+#define SUB_OP(x, y) ((x) - (y))
+#define DIV_OP(x, y) ((x) / (y))
+#define EXP_OP(x) exp((x))
+
+#if defined(DATA_TYPE_FP32)
+const float MINVAL   = -1.0 / 0.0;
+vec4        type_min = CONVERT(MINVAL, vec4);
+
+#define LOAD16(name, offset)            \
+    vec4(LOAD4(name, offset),           \
+         LOAD4(name, offset + uint(1)), \
+         LOAD4(name, offset + uint(2)), \
+         LOAD4(name, offset + uint(3)))
+
+#define STORE16(name, offset, value)         \
+    STORE4(name, offset, value.x);           \
+    STORE4(name, offset + uint(1), value.y); \
+    STORE4(name, offset + uint(2), value.z); \
+    STORE4(name, offset + uint(3), value.w)
+
+#ifdef SOFTMAX_LAYER_MAX
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM)
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(max, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+BUFFER_DECLARATION(sum, 4, float, writeonly);
+#elif defined(SOFTMAX_LAYER_NORM)
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(sum, 2, float, readonly);
+BUFFER_DECLARATION(dst, 3, float, writeonly);
+#endif // SOFTMAX_LAYER_MAX
+
+layout(std140) uniform shader_params
+{
+#ifdef SOFTMAX_LAYER_MAX
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    uint width;
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM)
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(max);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    TENSOR3D_PARAM_DECLARATION(sum);
+    uint width;
+#elif defined(SOFTMAX_LAYER_NORM)
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(sum);
+    TENSOR3D_PARAM_DECLARATION(dst);
+#endif // SOFTMAX_LAYER_MAX
+};
+
+#ifdef SOFTMAX_LAYER_MAX
+/** Identifies the maximum value across the 1st dimension.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP32"
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z 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]  width                             Input image width
+ */
+void main(void)
+{
+    Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+
+    // Initialize local maximum
+    vec4 max_val = CONVERT(type_min, vec4);
+
+    // Calculate max of row
+    uint width2 = width >> 2;
+    for(int i = 0; i < int(width2); i++)
+    {
+        vec4 data = LOAD16(src, offset(src, i << 2, 0));
+        max_val   = MAX_OP(data, max_val);
+    }
+
+#ifdef NON_MULTIPLE_OF_4
+    // Handle non multiple of 4
+    for(int i = int(width2 << 2); i < int(width); i++)
+    {
+        float data = LOAD4(src, offset(src, i, 0));
+        max_val.x  = MAX_OP(data, max_val.x);
+    }
+#endif /* NON_MULTIPLE_OF_4 */
+
+    // Perform max reduction
+    max_val.xy = MAX_OP(max_val.xy, max_val.zw);
+    max_val.x  = MAX_OP(max_val.x, max_val.y);
+
+    // Store result
+    STORE4(dst, CURRENT_OFFSET(dst), max_val.x);
+}
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM) // SOFTMAX_LAYER_MAX
+/** Shifts the values of the input tensor by the max calculated in softmax_layer_max kernel,
+ * then gets the exponent of each element as sums all elements across each row.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP32"
+ *
+ * @note In case the input is not multiple of 4 NON_MULTIPLE_OF_4 must be passed.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z 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]  max_ptr                           Pointer to the max values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  max_stride_x                      Stride of the max values tensor in X dimension (in bytes)
+ * @param[in]  max_step_x                        max_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  max_stride_y                      Stride of the max values tensor in Y dimension (in bytes)
+ * @param[in]  max_step_y                        max_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  max_stride_z                      Stride of the max values tensor in Z dimension (in bytes)
+ * @param[in]  max_step_z                        max_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  max_offset_first_element_in_bytes The offset of the first element in the max values tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[out] sum_ptr                           Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  sum_stride_x                      Stride of the sum values tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                      Stride of the sum values tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_stride_z                      Stride of the sum values tensor in Z dimension (in bytes)
+ * @param[in]  sum_step_z                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
+ * @param[in]  width                             Input image width
+ */
+void main(void)
+{
+    Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Image max = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(max);
+    Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(sum);
+
+    // Load max value of 1D logits vector (row)
+    vec4 max_val = CONVERT(LOAD4(max, CURRENT_OFFSET(max)), vec4);
+
+    // Set sum vector
+    vec4 sum1D = CONVERT(0, vec4);
+
+    // Shift values, exp and sum
+    uint width2 = width >> 2;
+    for(int i = 0; i < int(width2); i++)
+    {
+        vec4 data = LOAD16(src, offset(src, i << 2, 0));
+        data      = SUB_OP(data, max_val);
+        data      = EXP_OP(data);
+        STORE16(dst, offset(dst, i << 2, 0), data);
+        sum1D = ADD_OP(sum1D, data);
+    }
+
+#ifdef NON_MULTIPLE_OF_4
+    // Handle non multiple of 4
+    for(int i = int(width2 << 2); i < int(width); i++)
+    {
+        float data;
+        data = LOAD4(src, offset(src, i, 0));
+        data = SUB_OP(data, max_val.x);
+        data = EXP_OP(data);
+        STORE4(dst, offset(dst, i, 0), data);
+        sum1D.x = ADD_OP(sum1D.x, data);
+    }
+#endif                            /* NON_MULTIPLE_OF_4 */
+
+    // Perform min/max reduction
+    sum1D.xy = ADD_OP(sum1D.xy, sum1D.zw);
+    sum1D.x  = ADD_OP(sum1D.x, sum1D.y);
+
+    // Calculate and store result
+    STORE4(sum, CURRENT_OFFSET(sum), sum1D.x);
+}
+#elif defined(SOFTMAX_LAYER_NORM) // SOFTMAX_LAYER_MAX
+/** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP32"
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
+ * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
+ * @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 Z 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]  sum_ptr                           Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  sum_stride_x                      Stride of the sum values tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                      Stride of the sum values tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                        sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_stride_z                      Stride of the sum values tensor in Z dimension (in bytes)
+ * @param[in]  sum_step_z                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Image src = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Image sum = CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(sum);
+
+    // Load max value of 1D logits vector (row)
+    vec4 sum_val = CONVERT(LOAD4(sum, offset(sum, 0, int(gl_GlobalInvocationID.y))), vec4);
+    vec4 data    = LOAD16(src, CURRENT_OFFSET(src));
+    STORE16(dst, CURRENT_OFFSET(dst), DIV_OP(data, sum_val));
+}
+#endif                            // SOFTMAX_LAYER_MAX
+
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+
+const float MINVAL1   = -1.0 / 0.0;
+vec4        type_min1 = CONVERT(MINVAL1, vec4);
+
+#define GC_LOAD4_IMAGE(r, name, x, y)  \
+    load_and_unpack(r.xy, name, x, y); \
+    load_and_unpack(r.zw, name, (x + 2), y)
+
+#define GC_STORE4_IMAGE(r, name, x, y)                         \
+    GC_STORE1_2D_OFFSET(uint(packHalf2x16(r.xy)), name, x, y); \
+    GC_STORE1_2D_OFFSET(uint(packHalf2x16(r.zw)), name, (x + 2), y)
+
+#ifdef SOFTMAX_LAYER_MAX
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(dst, 2, uint, writeonly);
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM)
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(max, 2, uint, readonly);
+BUFFER_DECLARATION(dst, 3, uint, writeonly);
+BUFFER_DECLARATION(sum, 4, uint, writeonly);
+#elif defined(SOFTMAX_LAYER_NORM)
+BUFFER_DECLARATION(src, 1, uint, readonly);
+BUFFER_DECLARATION(sum, 2, uint, readonly);
+BUFFER_DECLARATION(dst, 3, uint, writeonly);
+#endif // SOFTMAX_LAYER_MAX
+
+layout(std140) uniform shader_params
+{
+#ifdef SOFTMAX_LAYER_MAX
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    uint width;
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM)
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(max);
+    TENSOR3D_PARAM_DECLARATION(dst);
+    TENSOR3D_PARAM_DECLARATION(sum);
+    uint width;
+#elif defined(SOFTMAX_LAYER_NORM)
+    TENSOR3D_PARAM_DECLARATION(src);
+    TENSOR3D_PARAM_DECLARATION(sum);
+    TENSOR3D_PARAM_DECLARATION(dst);
+#endif // SOFTMAX_LAYER_MAX
+};
+
+#define load_and_unpack(rs, names, xs, ys)           \
+    do                                               \
+    {                                                \
+        uint packed_s;                               \
+        GC_LOAD1_2D_OFFSET(packed_s, names, xs, ys); \
+        rs = vec2(unpackHalf2x16(packed_s));         \
+    } while(false)
+
+#ifdef SOFTMAX_LAYER_MAX
+/** Identifies the maximum value across the 1st dimension.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP16"
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F16
+ * @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 Z processed per workitem(in bytes)
+ * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z 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]  width                             Input image width
+ */
+void main(void)
+{
+    Image src = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+
+    // Initialize local maximum
+    vec4 max_val1 = CONVERT(type_min1, vec4);
+
+    // Calculate max of row
+    uint width2 = width >> 2;
+    for(int i = 0; i < int(width2); i++)
+    {
+        vec4 data1;
+        GC_LOAD4_IMAGE(data1, src, (i << 2), 0);
+        max_val1 = MAX_OP(data1, max_val1);
+    }
+
+#ifdef NON_MULTIPLE_OF_4
+    // Handle non multiple of 4
+    for(int i = int(width2 << 2); i < int(width); i = i + 2)
+    {
+        vec2 data;
+        load_and_unpack(data, src, i, 0);
+        max_val1.x = MAX_OP(data.x, max_val1.x);
+        if((i + 1) < int(width))
+        {
+            max_val1.x = MAX_OP(data.y, max_val1.x);
+        }
+    }
+#endif                                     /* NON_MULTIPLE_OF_4 */
+
+    // Perform max reduction
+    max_val1.xy = MAX_OP(max_val1.xy, max_val1.zw);
+    max_val1.x  = MAX_OP(max_val1.x, max_val1.y);
+    vec2 res1   = vec2(max_val1.x, 0.f);
+    uint res;
+    res = uint(packHalf2x16(res1));
+
+    // Store result
+    GC_STORE1_2D_OFFSET(res, dst, 0, 0);
+}
+#elif defined(SOFTMAX_LAYER_SHIFT_EXP_SUM) // SOFTMAX_LAYER_MAX
+/** Shifts the values of the input tensor by the max calculated in softmax_layer_max kernel,
+ * then gets the exponent of each element as sums all elements across each row.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP16"
+ *
+ * @note In case the input is not multiple of 4 NON_MULTIPLE_OF_4 must be passed.
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F16
+ * @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 Z 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]  max_ptr                           Pointer to the max values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  max_stride_x                      Stride of the max values tensor in X dimension (in bytes)
+ * @param[in]  max_step_x                        max_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  max_stride_y                      Stride of the max values tensor in Y dimension (in bytes)
+ * @param[in]  max_step_y                        max_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  max_stride_z                      Stride of the max values tensor in Z dimension (in bytes)
+ * @param[in]  max_step_z                        max_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  max_offset_first_element_in_bytes The offset of the first element in the max values tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[out] sum_ptr                           Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  sum_stride_x                      Stride of the sum values tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                      Stride of the sum values tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_stride_z                      Stride of the sum values tensor in Z dimension (in bytes)
+ * @param[in]  sum_step_z                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
+ * @param[in]  width                             Input image width
+ */
+void main(void)
+{
+    Image src = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Image max = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(max);
+    Image sum = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(sum);
+
+    // Load max value of 1D logits vector (row)
+    vec2 datamaxinit;
+    load_and_unpack(datamaxinit, max, 0, 0);
+    vec4 max_val = CONVERT(datamaxinit.x, vec4);
+
+    // Set sum vector
+    vec4 sum1D1 = CONVERT(0.f, vec4);
+
+    // Shift values, exp and sum
+    uint width2 = width >> 2;
+    for(int i = 0; i < int(width2); i++)
+    {
+        vec4 data;
+        GC_LOAD4_IMAGE(data, src, (i << 2), 0);
+        data = SUB_OP(data, max_val);
+        data = EXP_OP(data);
+        GC_STORE4_IMAGE(data, dst, (i << 2), 0);
+        sum1D1 = ADD_OP(sum1D1, data);
+    }
+
+#ifdef NON_MULTIPLE_OF_4
+    // Handle non multiple of 4
+    for(int i = int(width2 << 2); i < int(width); i = i + 2)
+    {
+        vec2  datamiddle;
+        float data1;
+        load_and_unpack(datamiddle, src, i, 0);
+        data1 = SUB_OP(datamiddle.x, max_val.x);
+        data1 = EXP_OP(data1);
+        vec2 datares1;
+        if((i + 1) < int(width))
+        {
+            float data2;
+            data2    = SUB_OP(datamiddle.y, max_val.x);
+            data2    = EXP_OP(data2);
+            datares1 = vec2(data1, data2);
+            data1    = ADD_OP(data2, data1);
+        }
+        else
+        {
+            datares1 = vec2(data1, 0.f);
+        }
+        uint datares;
+        datares = uint(packHalf2x16(datares1));
+        GC_STORE1_2D_OFFSET(datares, dst, i, 0);
+        sum1D1.x = ADD_OP(sum1D1.x, data1);
+    }
+#endif                            /* NON_MULTIPLE_OF_4 */
+
+    // Perform min/max reduction
+    sum1D1.xy = ADD_OP(sum1D1.xy, sum1D1.zw);
+    sum1D1.x  = ADD_OP(sum1D1.x, sum1D1.y);
+    vec2 res1 = vec2(sum1D1.x, 0.f);
+    uint res;
+    res = uint(packHalf2x16(res1));
+    // Calculate and store result
+    GC_STORE1_2D_OFFSET(res, sum, 0, 0);
+}
+#elif defined(SOFTMAX_LAYER_NORM) // SOFTMAX_LAYER_MAX
+/** Divides all the values of the input tensor by the sum calculated from softmax_layer_shift_exp_sum kernel.
+ *
+ * @note Datatype must be given as a preprocessor argument using "#define DATA_TYPE_FP16"
+ *
+ * @param[in]  src_ptr                           Pointer to the source tensor slice. Supported data types: F16
+ * @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 Z 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]  sum_ptr                           Pointer to the sum values tensor slice. Supported data types: same as @p src_ptr
+ * @param[in]  sum_stride_x                      Stride of the sum values tensor in X dimension (in bytes)
+ * @param[in]  sum_step_x                        sum_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  sum_stride_y                      Stride of the sum values tensor in Y dimension (in bytes)
+ * @param[in]  sum_step_y                        sum_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  sum_stride_z                      Stride of the sum values tensor in Z dimension (in bytes)
+ * @param[in]  sum_step_z                        sum_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  sum_offset_first_element_in_bytes The offset of the first element in the sum values tensor
+ * @param[out] dst_ptr                           Pointer to the destination tensor slice. Supported data types: 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_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 Z processed per workitem(in bytes)
+ * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ */
+void main(void)
+{
+    Image src = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image dst = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Image sum = GC_CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(sum);
+
+    // Load max value of 1D logits vector (row)
+    vec2 sum1;
+    load_and_unpack(sum1, sum, 0, int(gl_GlobalInvocationID.y));
+    vec4 sum_val1 = CONVERT(sum1.x, vec4);
+
+    vec4 data1;
+    GC_LOAD4_IMAGE(data1, src, 0, 0);
+    vec4 res = DIV_OP(data1, sum_val1);
+    GC_STORE4_IMAGE(res, dst, 0, 0);
+}
+#endif                            // SOFTMAX_LAYER_MAX
+#endif                            // DATA_TYPE_FP32
+\ No newline at end of file
diff --git a/src/core/GLES_COMPUTE/cs_shaders/transpose.cs b/src/core/GLES_COMPUTE/cs_shaders/transpose.cs
new file mode 100755
index 0000000000..6d020fe70d
--- /dev/null
+++ b/src/core/GLES_COMPUTE/cs_shaders/transpose.cs
@@ -0,0 +1,187 @@
+/*
+ * Copyright (c) 2017 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.
+ */
+layout(local_size_x = LOCAL_SIZE_X, local_size_y = LOCAL_SIZE_Y, local_size_z = LOCAL_SIZE_Z) in;
+#include "helpers.h"
+
+#ifdef DATA_TYPE_FP32
+precision highp float;
+
+BUFFER_DECLARATION(src, 1, float, readonly);
+BUFFER_DECLARATION(dst, 2, float, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+#define LOAD16(r, name, offset)          \
+    r.x = LOAD4(name, offset);           \
+    r.y = LOAD4(name, offset + uint(1)); \
+    r.z = LOAD4(name, offset + uint(2)); \
+    r.w = LOAD4(name, offset + uint(3))
+
+#define STORE16(name, offset, r)         \
+    STORE4(name, offset, r.x);           \
+    STORE4(name, offset + uint(1), r.y); \
+    STORE4(name, offset + uint(2), r.z); \
+    STORE4(name, offset + uint(3), r.w)
+
+/** This OpenGL ES kernel computes the matrix transposition of input matrix
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F32
+ * @param[in]  src_stride_x                      Stride of the source matrix 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 matrix 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 matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * 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 matrix
+ */
+void main(void)
+{
+    // Compute source address
+    Image src = CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = CONVERT_TO_IMAGE_STRUCT(dst);
+
+    // Load the NxN block at (x, y)
+    vec4 u0;
+    vec4 u1;
+    vec4 u2;
+    vec4 u3;
+    LOAD16(u0, src, offset(src, 0, 0));
+    LOAD16(u1, src, offset(src, 0, 1));
+    LOAD16(u2, src, offset(src, 0, 2));
+    LOAD16(u3, src, offset(src, 0, 3));
+
+    // Transpose the block
+    vec4 tmp;
+    tmp.xyz = u0.yzw;
+    u0.y    = u1.x;
+    u0.z    = u2.x;
+    u0.w    = u3.x;
+    u1.x    = tmp.x;
+    u2.x    = tmp.y;
+    u3.x    = tmp.z;
+    tmp.xy  = u1.zw;
+    u1.z    = u2.y;
+    u1.w    = u3.y;
+    u2.y    = tmp.x;
+    u3.y    = tmp.y;
+    tmp.x   = u2.w;
+    u2.w    = u3.z;
+    u3.z    = tmp.x;
+
+    // Store the block at (y, x)
+    uint dst_offset_in_bytes = uint(16) * uint(gl_GlobalInvocationID.y) + uint(4) * uint(gl_GlobalInvocationID.x) * (dst.stride_y) + (dst.offset_first_element_in_bytes);
+
+    STORE16(dst, uint((dst_offset_in_bytes + uint(0) * dst.stride_y) >> 2), u0);
+    STORE16(dst, uint((dst_offset_in_bytes + uint(1) * dst.stride_y) >> 2), u1);
+    STORE16(dst, uint((dst_offset_in_bytes + uint(2) * dst.stride_y) >> 2), u2);
+    STORE16(dst, uint((dst_offset_in_bytes + uint(3) * dst.stride_y) >> 2), u3);
+}
+
+#elif defined(DATA_TYPE_FP16)
+precision mediump float;
+
+BUFFER_DECLARATION(src, 1, uvec2, readonly);
+BUFFER_DECLARATION(dst, 2, uvec2, writeonly);
+
+layout(std140) uniform shader_params
+{
+    IMAGE_PARAM_DECLARATION(src);
+    IMAGE_PARAM_DECLARATION(dst);
+};
+
+/** This OpenGL ES kernel computes the matrix transposition of input matrix
+ *
+ * @param[in]  src_ptr                           Pointer to the source matrix. Supported data types: F16
+ * @param[in]  src_stride_x                      Stride of the source matrix 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 matrix 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 matrix
+ * @param[out] dst_ptr                           Pointer to the destination matrix Supported data type: same as src_ptr
+ * @param[in]  dst_stride_x                      Stride of the destination matrix in X dimension (in bytes)
+ * @param[in]  dst_step_x                        dst_gx_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  dst_stride_y                      Stride of the destination matrix in Y dimension (in bytes)
+ * @param[in]  dst_step_y                        dst_gx_stride_y * 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 matrix
+ */
+void main(void)
+{
+    // Compute source address
+    Image src = GC_CONVERT_TO_IMAGE_STRUCT(src);
+    Image dst = GC_CONVERT_TO_IMAGE_STRUCT(dst);
+
+    // Load the NxN block at (x, y)
+    vec4  u0;
+    vec4  u1;
+    vec4  u2;
+    vec4  u3;
+    uvec2 packed_s[4];
+    GC_LOAD1_2D_OFFSET(packed_s[0], src, 0, 0);
+    GC_LOAD1_2D_OFFSET(packed_s[1], src, 0, 1);
+    GC_LOAD1_2D_OFFSET(packed_s[2], src, 0, 2);
+    GC_LOAD1_2D_OFFSET(packed_s[3], src, 0, 3);
+    u0 = vec4(unpackHalf2x16(packed_s[0].x), unpackHalf2x16(packed_s[0].y));
+    u1 = vec4(unpackHalf2x16(packed_s[1].x), unpackHalf2x16(packed_s[1].y));
+    u2 = vec4(unpackHalf2x16(packed_s[2].x), unpackHalf2x16(packed_s[2].y));
+    u3 = vec4(unpackHalf2x16(packed_s[3].x), unpackHalf2x16(packed_s[3].y));
+
+    // Transpose the block
+    vec4 tmp;
+    tmp.xyz = u0.yzw;
+    u0.y    = u1.x;
+    u0.z    = u2.x;
+    u0.w    = u3.x;
+    u1.x    = tmp.x;
+    u2.x    = tmp.y;
+    u3.x    = tmp.z;
+    tmp.xy  = u1.zw;
+    u1.z    = u2.y;
+    u1.w    = u3.y;
+    u2.y    = tmp.x;
+    u3.y    = tmp.y;
+    tmp.x   = u2.w;
+    u2.w    = u3.z;
+    u3.z    = tmp.x;
+
+    // Store the block at (y, x)
+    uint dst_offset_in_bytes = uint(8) * uint(gl_GlobalInvocationID.y) + uint(gl_GlobalInvocationID.x) * (dst_step_y) + (dst.offset_first_element_in_bytes);
+
+    packed_s[0] = uvec2(packHalf2x16(u0.xy), packHalf2x16(u0.zw));
+    packed_s[1] = uvec2(packHalf2x16(u1.xy), packHalf2x16(u1.zw));
+    packed_s[2] = uvec2(packHalf2x16(u2.xy), packHalf2x16(u2.zw));
+    packed_s[3] = uvec2(packHalf2x16(u3.xy), packHalf2x16(u3.zw));
+    GC_STORE1(packed_s[0], dst, uint((dst_offset_in_bytes + uint(0) * dst_stride_y) >> 3));
+    GC_STORE1(packed_s[1], dst, uint((dst_offset_in_bytes + uint(1) * dst_stride_y) >> 3));
+    GC_STORE1(packed_s[2], dst, uint((dst_offset_in_bytes + uint(2) * dst_stride_y) >> 3));
+    GC_STORE1(packed_s[3], dst, uint((dst_offset_in_bytes + uint(3) * dst_stride_y) >> 3));
+}
+#endif /*ARM_COMPUTE_ENABLE_FP16*/
author	Anthony Barbier <anthony.barbier@arm.com>	2017-10-26 15:23:08 +0100
committer	Anthony Barbier <anthony.barbier@arm.com>	2018-11-02 16:35:24 +0000
commit	7068f9900d136312318ff430aef588b14e0c87ad (patch)
tree	b57ca81231860f1d8755e6f18e5be7c959fb60c6 /src/core/GLES_COMPUTE/cs_shaders
parent	d60737592736715dcfd0520535c48190d4ac77d2 (diff)
download	ComputeLibrary-7068f9900d136312318ff430aef588b14e0c87ad.tar.gz