APPBROWSER-377: GCConvoutionLayer support for FP16

Change-Id: I801b5e393a16a9f92c062826e6fcfd5982ca7bb3
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/116584
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

10 files changed, 281 insertions, 70 deletions

diff --git a/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
index 492f708a98..af1e34ef59 100644
--- a/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCCol2ImKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCCol2ImKernel.h"
 
+#include "arm_compute/core/AccessWindowStatic.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
 #include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
@@ -43,31 +44,50 @@ GCCol2ImKernel::GCCol2ImKernel()
 void GCCol2ImKernel::configure(const IGCTensor *input, IGCTensor    *output,
                                std::pair<unsigned int, unsigned int> convolved_dims)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+
+    TensorShape output_shape = input->info()->tensor_shape();
+    output_shape.set(0, convolved_dims.first);
+    output_shape.set(1, convolved_dims.second);
+    output_shape.set(2, input->info()->tensor_shape()[0]);
+
+    // Output auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
 
     _input          = input;
     _output         = output;
     _convolved_dims = convolved_dims;
 
+    unsigned int num_elems_processed_per_iteration = 1;
+
     // Create kernel
-    std::set<std::string>  build_opts;
-    constexpr unsigned int num_elems_processed_per_iteration = 8;
+    std::set<std::string> build_opts;
+    build_opts.emplace("#define WIDTH_OUTPUT " + support::cpp11::to_string(_convolved_dims.first));
+    std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
+    build_opts.emplace(("#define " + dt_name));
     build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(num_elems_processed_per_iteration));
-    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
-    build_opts.insert("#define COL2IM");
-    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("col2im", build_opts));
+    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(num_elems_processed_per_iteration));
 
-    // Set static kernel arguments
-    unsigned int idx = num_arguments_per_2D_tensor() + num_arguments_per_3D_tensor();
-    _kernel.set_argument(idx++, _convolved_dims.first);
+    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("col2im", build_opts));
 
     // Configure window
-    Window win = calculate_max_window(*input->info(), Steps());
+    unsigned int nums = 2;
+    Window       win  = calculate_max_window(*output->info(), Steps(nums));
 
-    // The GCCol2ImKernel doesn't need padding so update_window_and_padding() can be skipped
-    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    AccessWindowHorizontal output_access(output->info(), 0, 2);
+    const int              input_padding = ceil_to_multiple(input->info()->dimension(0), 2) - input->info()->dimension(0);
+
+    AccessWindowStatic input_access(input->info(), 0, 0, input->info()->dimension(0) + input_padding, input->info()->dimension(1) + 1);
+
+    update_window_and_padding(win, input_access,
+                              output_access);
+
+    output_access.set_valid_region(win, output->info()->valid_region());
 
     IGCKernel::configure(win);
 }
@@ -77,20 +97,25 @@ void GCCol2ImKernel::run(const Window &window)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
 
-    Window slice_in  = window.first_slice_window_2D();
-    Window slice_out = window.first_slice_window_3D();
-
     _kernel.use();
 
+    Window collapsed_window = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
+    Window slice            = collapsed_window.first_slice_window_3D();
+
+    // Set static kernel arguments
+    unsigned int idx = 2 * num_arguments_per_3D_tensor();
+    //_kernel.set_argument(idx++, _output->info()->strides_in_bytes()[3]);
+    _kernel.set_argument(idx++, uint(_output->info()->dimension(2)));
+    _kernel.set_argument(idx++, _input->info()->strides_in_bytes()[2]);
+
     do
     {
         // Set inputs
-        unsigned int idx     = 0;
-        unsigned int binding = 1;
-        add_2D_tensor_argument(idx, _input, binding++, slice_in);
-        add_3D_tensor_argument(idx, _output, binding++, slice_out);
+        unsigned int idx = 0;
+        add_2D_tensor_argument(idx, _input, 1, slice);
+        add_3D_tensor_argument(idx, _output, 2, slice);
         _kernel.update_shader_params();
-        enqueue(*this, slice_in);
+        enqueue(*this, slice);
     }
-    while(window.slide_window_slice_2D(slice_in) && window.slide_window_slice_3D(slice_out));
+    while(collapsed_window.slide_window_slice_3D(slice));
 }
diff --git a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
index 4eceab8266..7b1848c32b 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDepthConcatenateLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
index 77423fd8bc..23f1c2eada 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDirectConvolutionLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/GLES_COMPUTE/kernels/GCDropoutLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCDropoutLayerKernel.cpp
index cd06be2585..8886b84fc0 100644
--- a/src/core/GLES_COMPUTE/kernels/GCDropoutLayerKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCDropoutLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp
index ef47e7fd8c..dc86bfb2cc 100644
--- a/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,7 +44,7 @@ GCGEMMInterleave4x4Kernel::GCGEMMInterleave4x4Kernel()
 
 void GCGEMMInterleave4x4Kernel::configure(const IGCTensor *input, IGCTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape output_shape = input->info()->tensor_shape();
diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixAccumulateBiasesKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixAccumulateBiasesKernel.cpp
index 7ee39346f8..6d856e98c3 100644
--- a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixAccumulateBiasesKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixAccumulateBiasesKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp
index a78446e074..5d9f9c2d3e 100644
--- a/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,7 +40,7 @@ using namespace arm_compute;
 
 void GCGEMMTranspose1xWKernel::configure(const IGCTensor *input, IGCTensor *output)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_ERROR_ON_NULLPTR(output);
 
     TensorShape  output_shape{ input->info()->tensor_shape() };
diff --git a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
index e849891c7c..4ab6f3e89d 100644
--- a/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCIm2ColKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,6 +21,7 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCIm2ColKernel.h"
 
 #include "arm_compute/core/AccessWindowStatic.h"
@@ -30,6 +31,7 @@
 #include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
 #include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
 #include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Size2D.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
 #include "support/ToolchainSupport.h"
@@ -39,20 +41,40 @@
 
 using namespace arm_compute;
 
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
+
+    // Checks performed when output is configured
+    if(output->total_size() != 0)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
+    }
+
+    return Status{};
+}
+} // namespace
+
 GCIm2ColKernel::GCIm2ColKernel()
-    : _input(nullptr), _output(nullptr), _convolved_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr)
+    : _input(nullptr), _output(nullptr), _convolved_dims(), _kernel_dims(), _num_elems_processed_per_iteration(1), _run_func(nullptr)
 {
 }
 
-void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, std::pair<unsigned int, unsigned int> kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
-    ARM_COMPUTE_UNUSED(kernel_dims);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+
+    // Perform validation step
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info()));
 
     _input  = input;
     _output = output;
 
+    // Create kernel
     std::set<std::string> build_opts;
     std::string           dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
     build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
@@ -65,48 +87,52 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, std::p
         build_opts.emplace("#define HAS_BIAS");
     }
 
-    int pad_x    = 0;
-    int pad_y    = 0;
     int stride_x = 0;
     int stride_y = 0;
-    std::tie(pad_x, pad_y)       = conv_info.pad();
+
     std::tie(stride_x, stride_y) = conv_info.stride();
+    _kernel_dims = std::make_pair(kernel_dims.width, kernel_dims.height);
 
     const bool run_img2col_reduced = (output->info()->dimension(0) == (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))) && (TensorShape::num_max_dimensions >= 4)
                                      && (std::equal(input->info()->tensor_shape().cbegin() + 3,
                                                     input->info()->tensor_shape().cend(),
                                                     output->info()->tensor_shape().cbegin() + 1))
-                                     && ((stride_x == 1) && (stride_y == 1) && (pad_x == 0) && (pad_y == 0));
+                                     && ((stride_x == 1) && (stride_y == 1) && !conv_info.has_padding());
 
+    std::string kernel_name = "im2col_generic";
     if(!run_img2col_reduced)
     {
-        // this path is currently not used and not validated
-        build_opts.insert("#define IM2COL_GENERIC");
+        if(input->info()->data_type() == DataType::F16 && _kernel_dims == std::pair<unsigned int, unsigned int>(1, 1))
+        {
+            build_opts.emplace("#define KERNEL_1x1");
+        }
+
+        build_opts.emplace("#define IM2COL_GENERIC");
         _convolved_dims = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1),
-                                            kernel_dims.first, kernel_dims.second,
+                                            kernel_dims.width, kernel_dims.height,
                                             conv_info);
-        _num_elems_processed_per_iteration = output->info()->dimension(0);
+        _num_elems_processed_per_iteration = 2;
 
-        build_opts.emplace("#define KERNEL_WIDTH " + support::cpp11::to_string(kernel_dims.first));
-        build_opts.emplace("#define KERNEL_HEIGHT " + support::cpp11::to_string(kernel_dims.second));
+        build_opts.emplace("#define KERNEL_WIDTH " + support::cpp11::to_string(kernel_dims.width));
+        build_opts.emplace("#define KERNEL_HEIGHT " + support::cpp11::to_string(kernel_dims.height));
         build_opts.emplace("#define KERNEL_DEPTH " + support::cpp11::to_string(input->info()->dimension(2)));
         build_opts.emplace("#define CONVOLVED_WIDTH " + support::cpp11::to_string(_convolved_dims.first));
         build_opts.emplace("#define CONVOLVED_HEIGHT " + support::cpp11::to_string(_convolved_dims.second));
         build_opts.emplace("#define STRIDE_X " + support::cpp11::to_string(conv_info.stride().first));
         build_opts.emplace("#define STRIDE_Y " + support::cpp11::to_string(conv_info.stride().second));
-        build_opts.emplace("#define PAD_X " + support::cpp11::to_string(conv_info.pad().first));
-        build_opts.emplace("#define PAD_Y " + support::cpp11::to_string(conv_info.pad().second));
+        build_opts.emplace("#define PAD_LEFT " + support::cpp11::to_string(conv_info.pad_left()));
+        build_opts.emplace("#define PAD_TOP " + support::cpp11::to_string(conv_info.pad_top()));
+        build_opts.emplace("#define PAD_RIGHT " + support::cpp11::to_string(conv_info.pad_right()));
+        build_opts.emplace("#define PAD_BOTTOM " + support::cpp11::to_string(conv_info.pad_bottom()));
         build_opts.emplace("#define SRC_WIDTH " + support::cpp11::to_string(input->info()->dimension(0)));
         build_opts.emplace("#define SRC_HEIGHT " + support::cpp11::to_string(input->info()->dimension(1)));
 
-        // Create kernel
-        _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("im2col_generic", build_opts));
-
         _run_func = &GCIm2ColKernel::run_generic;
     }
     else
     {
-        build_opts.insert("#define IM2COL_REDUCED");
+        build_opts.emplace("#define IM2COL_REDUCED");
+        kernel_name = "im2col_reduced";
 
         if(input->info()->data_type() == DataType::F32)
         {
@@ -117,42 +143,47 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, std::p
             int input_width  = input->info()->dimension(0);
             int input_height = input->info()->dimension(1);
 
-            build_opts.insert("#define IMAGE_SIZE " + support::cpp11::to_string(input_width * input_height));
+            build_opts.emplace("#define IMAGE_SIZE " + support::cpp11::to_string(input_width * input_height));
             if(input_width % 8 == 0)
             {
                 _num_elems_processed_per_iteration = 8;
-                build_opts.insert("#define IM2COL_REDUCED_8X");
+                build_opts.emplace("#define IM2COL_REDUCED_8X");
             }
             else if(input_width % 4 == 0)
             {
                 _num_elems_processed_per_iteration = 4;
-                build_opts.insert("#define IM2COL_REDUCED_4X");
+                build_opts.emplace("#define IM2COL_REDUCED_4X");
             }
             else if(input_width % 2 == 0)
             {
                 _num_elems_processed_per_iteration = 2;
-                build_opts.insert("#define IM2COL_REDUCED_2X");
+                build_opts.emplace("#define IM2COL_REDUCED_2X");
             }
             else
             {
                 _num_elems_processed_per_iteration = 2;
-                build_opts.insert("#define IM2COL_REDUCED_GENERIC");
+                build_opts.emplace("#define IM2COL_REDUCED_GENERIC");
             }
         }
 
-        // Create kernel
-        _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("im2col_reduced", build_opts));
-
         _run_func = &GCIm2ColKernel::run_reduced;
     }
 
+    // Create kernel
+    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel(kernel_name, build_opts));
+
     // Configure kernel window
     Window win = calculate_max_window(*input->info(), Steps(_num_elems_processed_per_iteration));
 
     if(input->info()->data_type() == DataType::F16)
     {
         // Calculate input right and bottom border
-        AccessWindowHorizontal input_access(input->info(), 0, _num_elems_processed_per_iteration);
+        const int input_width         = input->info()->dimension(0);
+        const int input_height        = input->info()->dimension(1);
+        int       input_total_width   = input->info()->padding().left + input_width + input->info()->padding().right;
+        int       input_padding_right = ceil_to_multiple(input_total_width, _num_elems_processed_per_iteration) - input_total_width;
+        input_total_width             = input_width + input_padding_right + input->info()->padding().right;
+        AccessWindowStatic input_access(input->info(), 0, 0, input_total_width, input_height);
 
         // Calculate output right and bottom border
         const int          output_width         = output->info()->dimension(0);
@@ -174,6 +205,15 @@ void GCIm2ColKernel::configure(const IGCTensor *input, IGCTensor *output, std::p
     IGCKernel::configure(win);
 }
 
+Status GCIm2ColKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
+{
+    ARM_COMPUTE_UNUSED(kernel_dims);
+    ARM_COMPUTE_UNUSED(conv_info);
+    ARM_COMPUTE_UNUSED(has_bias);
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output));
+    return Status{};
+}
+
 void GCIm2ColKernel::run(const Window &window)
 {
     ARM_COMPUTE_ERROR_ON(_run_func == nullptr);
@@ -187,6 +227,7 @@ void GCIm2ColKernel::run_generic(const Window &window)
 
     // Get initial windows
     Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ);
+
     // Change the Z dimension's step back to 1
     window_collapsed.set_dimension_step(Window::DimZ, 1);
 
@@ -198,17 +239,18 @@ void GCIm2ColKernel::run_generic(const Window &window)
     slice.set(Window::DimX, Window::Dimension(0, static_cast<int>(_convolved_dims.first), 1));
     slice.set(Window::DimY, Window::Dimension(0, static_cast<int>(_convolved_dims.second), 1));
 
-    // Setup input slice
-    // The first three dimensions of the input are increased by the inner loops
-    slice_in.set(Window::DimX, Window::Dimension(0, 0, 0));
-    slice_in.set(Window::DimY, Window::Dimension(0, 0, 0));
-    slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0));
-
     // Setup output slice
     slice_out.set(Window::DimX, Window::Dimension(0, _output->info()->dimension(0), _num_elems_processed_per_iteration));
     slice_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), 1));
     slice_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
 
+    // we need top/left pad to be included in valid region
+    if(_input->info()->data_type() == DataType::F16)
+    {
+        (dynamic_cast<TensorInfo *>(_input->info()))->init(_input->info()->tensor_shape(), _input->info()->num_channels(), _input->info()->data_type(), _input->info()->strides_in_bytes(), 0,
+                                                           _input->info()->total_size(), _input->info()->fixed_point_position());
+    }
+
     _kernel.use();
 
     do
@@ -216,8 +258,6 @@ void GCIm2ColKernel::run_generic(const Window &window)
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, 1, slice_in);
         add_2D_tensor_argument(idx, _output, 2, slice_out);
-
-        _kernel.set_argument(idx++, static_cast<unsigned int>(_input->info()->dimension(2)));
         _kernel.set_argument(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
         _kernel.set_argument(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
         _kernel.update_shader_params();
diff --git a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp
index b3a7a90931..bda08e4238 100644
--- a/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp
+++ b/src/core/GLES_COMPUTE/kernels/GCTransposeKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, 2018 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp
new file mode 100644
index 0000000000..4c08873dcf
--- /dev/null
+++ b/src/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.cpp
@@ -0,0 +1,146 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.h"
+
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
+#include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h"
+#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h"
+#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Validate.h"
+
+#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h"
+
+using namespace arm_compute;
+using namespace arm_compute::gles_compute;
+
+GCWeightsReshapeKernel::GCWeightsReshapeKernel()
+    : _input(nullptr), _biases(nullptr), _output(nullptr)
+{
+}
+
+void GCWeightsReshapeKernel::configure(const IGCTensor *input, const IGCTensor *biases, IGCTensor *output)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(output);
+
+    // Calculate output shape
+    TensorShape output_shape{ input->info()->tensor_shape() };
+    output_shape.collapse(3);
+    const size_t tmp_dim = output_shape[0];
+    output_shape.set(0, output_shape[1]);
+    output_shape.set(1, tmp_dim + (biases != nullptr ? 1 : 0));
+
+    // Output tensor auto inizialitation if not yet initialized
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
+
+    if(biases != nullptr)
+    {
+        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+        ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (biases->info()->num_dimensions() != 1));
+        ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (biases->info()->num_dimensions() != 2));
+        ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 4) && (biases->info()->dimension(0) != input->info()->tensor_shape()[3]));
+        ARM_COMPUTE_ERROR_ON((input->info()->num_dimensions() == 5) && (biases->info()->dimension(0) != input->info()->tensor_shape()[3] || biases->info()->dimension(1) != input->info()->tensor_shape()[4]));
+    }
+
+    _biases = biases;
+    _output = output;
+    _input  = input;
+
+    // Create build options
+    std::set<std::string> build_opts;
+    std::string           dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16";
+    build_opts.emplace("#define " + dt_name);
+    build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1));
+    build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1));
+    build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1));
+    build_opts.emplace("#define RESHAPE_TO_COLUMNS");
+    if(biases != nullptr)
+    {
+        build_opts.emplace("#define HAS_BIAS");
+    }
+
+    // Create kernel
+    _kernel = static_cast<GCKernel>(GCKernelLibrary::get().create_kernel("reshape_to_columns", build_opts));
+
+    // Set static arguments
+    unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();
+    idx += (biases != nullptr) ? num_arguments_per_1D_tensor() : 0;
+    _kernel.set_argument(idx++, _input->info()->dimension(0));
+    _kernel.set_argument(idx++, _input->info()->dimension(1));
+    _kernel.set_argument(idx++, _input->info()->dimension(2));
+    _kernel.set_argument(idx++, _input->info()->dimension(3));
+
+    // Configure window
+    Window win = calculate_max_window(*input->info(), Steps());
+
+    // The GCWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+    output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
+    IGCKernel::configure(win);
+}
+
+void GCWeightsReshapeKernel::run(const Window &window)
+{
+    ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(IGCKernel::window(), window);
+
+    Window out_window;
+    out_window.use_tensor_dimensions(_output->info()->tensor_shape());
+
+    Window in_slice  = window.first_slice_window_3D();
+    Window out_slice = out_window.first_slice_window_2D();
+
+    Window biases_window;
+    Window biases_slice;
+
+    if(_biases != nullptr)
+    {
+        biases_window.use_tensor_dimensions(_biases->info()->tensor_shape());
+        biases_slice = biases_window.first_slice_window_1D();
+    }
+
+    _kernel.use();
+
+    do
+    {
+        // Set arguments
+        unsigned idx = 0;
+        add_3D_tensor_argument(idx, _input, 1, in_slice);
+        add_2D_tensor_argument(idx, _output, 2, out_slice);
+        if(_biases != nullptr)
+        {
+            add_1D_tensor_argument(idx, _biases, 3, biases_slice);
+            biases_window.slide_window_slice_1D(biases_slice);
+        }
+
+        _kernel.update_shader_params();
+        // Run kernel
+        enqueue(*this, in_slice);
+    }
+    while(window.slide_window_slice_4D(in_slice) && out_window.slide_window_slice_2D(out_slice));
+}