COMPMID-809: Add NHWC data format on CLGEMMConvolutionLayer.

Change-Id: I50e4f5e7d47e21c300f754bee2c216863075b5cf
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/136191
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

3 files changed, 20 insertions, 18 deletions

diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index 4e444206f1..64e6a0b7d8 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -140,23 +140,25 @@ void CLCol2ImKernel::run(const Window &window, cl::CommandQueue &queue)
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
-    // The collapse method rely on the assumption that the third dimension of input buffer is 1
-    ARM_COMPUTE_ERROR_ON(window.z().end() != 1);
+
+    Window out_window;
+    out_window.use_tensor_dimensions(_output->info()->tensor_shape());
 
     Window collapsed_window = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
-    Window slice            = collapsed_window.first_slice_window_3D();
+    Window slice            = collapsed_window.first_slice_window_2D();
+    Window slice_out        = out_window.first_slice_window_3D();
 
     // Set static kernel arguments
-    unsigned int idx = 2 * num_arguments_per_3D_tensor();
+    unsigned int idx = num_arguments_per_2D_tensor() + num_arguments_per_3D_tensor();
     _kernel.setArg<cl_uint>(idx++, _output->info()->strides_in_bytes()[3]);
 
     do
     {
         // Set inputs
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
-        add_3D_tensor_argument(idx, _output, slice);
+        add_2D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, _output, slice_out);
         enqueue(queue, *this, slice, _lws_hint);
     }
-    while(collapsed_window.slide_window_slice_3D(slice));
+    while(collapsed_window.slide_window_slice_2D(slice) && out_window.slide_window_slice_3D(slice_out));
 }
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index 328b39681b..21deb9217c 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -143,7 +143,7 @@ CLIm2ColKernel::configure_window(const ICLTensor *input, ICLTensor *output, cons
                 {
                     case 1:
                         // Optimized im2col1x1 if stride_x = 1 and conv_info.has_padding() = false
-                        if(conv_info.stride().first == 1 && !conv_info.has_padding())
+                        if(conv_info.stride().first == 1 && !conv_info.has_padding() && data_layout == DataLayout::NCHW)
                         {
                             // Set hint for LWS
                             _lws_hint                          = cl::NDRange(1, 1, 8);
@@ -350,11 +350,14 @@ void CLIm2ColKernel::run_generic(const Window &window, cl::CommandQueue &queue)
     // Change the Z dimension's step back to 1
     window_collapsed.set_dimension_step(Window::DimZ, 1);
 
+    Window window_output;
+    window_output.use_tensor_dimensions(_output->info()->tensor_shape());
+
     const Window first_slice_3d = window_collapsed.first_slice_window_3D();
 
     Window slice     = first_slice_3d;
     Window slice_in  = first_slice_3d;
-    Window slice_out = first_slice_3d;
+    Window slice_out = window_output.first_slice_window_2D();
 
     const bool out_dim_not_same_input_dim = _convolved_dims.first != _input->info()->dimension(width_idx) || _convolved_dims.second != _input->info()->dimension(height_idx);
 
@@ -386,21 +389,16 @@ void CLIm2ColKernel::run_generic(const Window &window, cl::CommandQueue &queue)
     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), _kernel_dims.area()));
-    slice_out.set(Window::DimY, Window::Dimension(0, _output->info()->dimension(1), _output->info()->dimension(1)));
-    slice_out.set(Window::DimZ, Window::Dimension(0, 1, 1));
-
     do
     {
         unsigned int idx = 0;
         add_3D_tensor_argument(idx, _input, slice_in);
         add_2D_tensor_argument(idx, _output, slice_out);
         _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input->info()->strides_in_bytes()[3]));
-        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[3]));
+        _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
         enqueue(queue, *this, slice, _lws_hint);
     }
-    while(window_collapsed.slide_window_slice_3D(slice) && window_collapsed.slide_window_slice_3D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
+    while(window_collapsed.slide_window_slice_3D(slice) && window_output.slide_window_slice_2D(slice_out) && window_collapsed.slide_window_slice_3D(slice_in));
 }
 
 void CLIm2ColKernel::run_reduced(const Window &window, cl::CommandQueue &queue)
diff --git a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
index c0a4517ad3..b012d58d59 100644
--- a/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLWeightsReshapeKernel.cpp
@@ -85,7 +85,8 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *
                                                   (biases != nullptr) ? biases->info() : nullptr,
                                                   output->info()));
 
-    const DataType data_type = input->info()->data_type();
+    const DataType   data_type   = input->info()->data_type();
+    const DataLayout data_layout = input->info()->data_layout();
 
     _biases = biases;
     _output = output;
@@ -98,7 +99,8 @@ void CLWeightsReshapeKernel::configure(const ICLTensor *input, const ICLTensor *
     build_opts.add_option_if(is_data_type_fixed_point(data_type), "-DFIXED_POINT_POSITION=" + support::cpp11::to_string(input->info()->fixed_point_position()));
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("reshape_to_columns", build_opts.options()));
+    std::string kernel_name = std::string("reshape_to_columns_") + lower_string(string_from_data_layout(data_layout));
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Set static arguments
     unsigned int idx = num_arguments_per_3D_tensor() + num_arguments_per_2D_tensor();