COMPMID-1367: Enable NHWC in graph examples

Change-Id: Iabc54a3a1bdcd46a9a921cda39c7c85fef672b72
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/141449
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

5 files changed, 50 insertions, 31 deletions

diff --git a/src/core/CL/cl_kernels/pooling_layer.cl b/src/core/CL/cl_kernels/pooling_layer.cl
index c38a78ce3e..080835348d 100644
--- a/src/core/CL/cl_kernels/pooling_layer.cl
+++ b/src/core/CL/cl_kernels/pooling_layer.cl
@@ -549,10 +549,10 @@ __kernel void pooling_layer_MxN_nhwc(
 
     for(int y = 0; y < POOL_SIZE_Y; ++y)
     {
-        int y1 = select(y, PAD_Y - idx_height, y + idx_height < PAD_Y || y + idx_height > MAX_HEIGHT);
+        int y1 = select(y, PAD_Y - idx_height, y + idx_height - PAD_Y < 0 || y + idx_height - PAD_Y >= MAX_HEIGHT);
         for(int x = 0; x < POOL_SIZE_X; ++x)
         {
-            int x1 = select(x, PAD_X - idx_width - 1, x + idx_width < PAD_X || x + idx_width > MAX_WIDTH);
+            int x1 = select(x, PAD_X - idx_width - 1, x + idx_width - PAD_X < 0 || x + idx_width - PAD_X >= MAX_WIDTH);
             x1     = select(x1, PAD_X - idx_width - 1, y != y1);
 
             VEC_DATA_TYPE(DATA_TYPE, 8)
diff --git a/src/core/CL/kernels/CLIm2ColKernel.cpp b/src/core/CL/kernels/CLIm2ColKernel.cpp
index b1290b8edd..a09129bba6 100644
--- a/src/core/CL/kernels/CLIm2ColKernel.cpp
+++ b/src/core/CL/kernels/CLIm2ColKernel.cpp
@@ -288,7 +288,6 @@ void CLIm2ColKernel::configure(const ICLTensor *input, ICLTensor *output, const
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_ERROR_ON(input->info()->data_layout() == DataLayout::UNKNOWN);
-    ARM_COMPUTE_ERROR_ON_MSG(output->info()->data_layout() != DataLayout::NCHW, "Special case Im2Col output layout is NCHW");
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), has_bias, dilation));
 
     _input       = input;
diff --git a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
index df01eab240..edc9e9d58c 100644
--- a/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLNormalizationLayerKernel.cpp
@@ -42,6 +42,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, N
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
 
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && norm_info.type() == NormType::IN_MAP_2D,
+                                    "Only Cross-map and 1D In-map normalization is supported for NHWC layout");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
 
     // Checks performed when output is configured
@@ -59,14 +61,15 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
     // Output tensor auto initialization if not yet initialized
     auto_init_if_empty(*output, *input->clone());
 
-    const unsigned int norm_size = norm_info.norm_size();
-    bool               is_in_map = norm_info.is_in_map();
+    const unsigned int norm_idx              = get_normalization_dimension_index(input->data_layout(), norm_info);
+    const unsigned int norm_size             = norm_info.norm_size();
+    bool               is_norm_accross_width = norm_idx == 0;
 
-    const unsigned int border_width = is_in_map ? std::min(norm_size / 2, 3U) : 0;
+    const unsigned int border_width = is_norm_accross_width ? std::min(norm_size / 2, 3U) : 0;
     const BorderSize   border_size  = BorderSize(0, border_width);
 
     const unsigned int num_elems_processed_per_iteration = 4;
-    const unsigned int num_elems_read_per_iteration      = is_in_map ? (num_elems_processed_per_iteration + 2 * (norm_size / 2)) : num_elems_processed_per_iteration;
+    const unsigned int num_elems_read_per_iteration      = is_norm_accross_width ? (num_elems_processed_per_iteration + 2 * (norm_size / 2)) : num_elems_processed_per_iteration;
 
     Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
 
@@ -84,7 +87,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLNormalizationLayerKernel::CLNormalizationLayerKernel()
-    : _input(nullptr), _output(nullptr), _border_size(0), _is_in_map(false)
+    : _input(nullptr), _output(nullptr), _border_size(0), _is_norm_across_width(false)
 {
 }
 
@@ -106,8 +109,9 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *ou
     _input  = input;
     _output = output;
 
-    _is_in_map                      = norm_info.is_in_map();
-    const unsigned int border_width = _is_in_map ? std::min(norm_info.norm_size() / 2, 3U) : 0;
+    const unsigned int norm_idx     = get_normalization_dimension_index(input->info()->data_layout(), norm_info);
+    _is_norm_across_width           = norm_idx == 0;
+    const unsigned int border_width = _is_norm_across_width ? std::min(norm_info.norm_size() / 2, 3U) : 0;
     _border_size                    = BorderSize(0, border_width);
 
     const unsigned int num_elems_processed_per_iteration = 4;
@@ -125,7 +129,7 @@ void CLNormalizationLayerKernel::configure(const ICLTensor *input, ICLTensor *ou
     build_opts.add_option_if(is_in_map_2D, "-DIN_MAP_2D");
 
     // Create kernel
-    std::string kernel_name = _is_in_map ? "normalization_layer_in_map" : "normalization_layer_cross_map";
+    std::string kernel_name = _is_norm_across_width ? "normalization_layer_in_map" : "normalization_layer_cross_map";
     _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
@@ -159,7 +163,7 @@ void CLNormalizationLayerKernel::run(const Window &window, cl::CommandQueue &que
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
-    const int collapsed_dimension = _is_in_map ? Window::DimZ : 4;
+    const int collapsed_dimension = _is_norm_across_width ? Window::DimZ : 4;
     Window    window_collapsed    = window.collapse_if_possible(ICLKernel::window(), collapsed_dimension);
     Window    slice               = window_collapsed.first_slice_window_3D();
 
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index 246ab68130..d5ea092c78 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -154,7 +154,9 @@ std::tuple<Status, Window, CLPoolingConfig> validate_and_configure_window(ITenso
             num_elems_processed_per_iteration = 8;
             win                               = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
 
-            AccessWindowRectangle  input_access(input, 0, -pool_pad_left, num_elems_processed_per_iteration, pool_size_x);
+            AccessWindowStatic input_access(input,
+                                            0, -1,
+                                            ceil_to_multiple(input->dimension(0), num_elems_processed_per_iteration), input->dimension(1));
             AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
             window_changed = update_window_and_padding(win, input_access, output_access);
             output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
diff --git a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp
index cb1996f33e..15e8298e2d 100644
--- a/src/core/NEON/kernels/NENormalizationLayerKernel.cpp
+++ b/src/core/NEON/kernels/NENormalizationLayerKernel.cpp
@@ -43,6 +43,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *input_squ
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
 
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(input->data_layout() == DataLayout::NHWC && norm_info.type() == NormType::IN_MAP_2D,
+                                    "Only Cross-map and 1D In-map normalization is supported for NHWC layout");
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, input_squared);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, input_squared);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!(norm_info.norm_size() % 2), "Normalization size should be odd");
@@ -61,8 +63,9 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 {
     unsigned int       num_elems_processed_per_iteration = 16 / input->element_size();
     const unsigned int num_elems_read_per_iteration      = num_elems_processed_per_iteration + 2 * (norm_info.norm_size() / 2);
+    const unsigned int norm_idx                          = get_normalization_dimension_index(input->data_layout(), norm_info);
     const unsigned int num_rows                          = (norm_info.type() == NormType::IN_MAP_2D) ? norm_info.norm_size() : 1;
-    const unsigned int border_width                      = (norm_info.is_cross_map()) ? 0 : std::min<unsigned int>(norm_info.norm_size() / 2, 3U);
+    const unsigned int border_width                      = (norm_idx == 2) ? 0 : std::min<unsigned int>(norm_info.norm_size() / 2, 3U);
     BorderSize         border_size                       = BorderSize(0, border_width);
     bool               window_changed                    = false;
 
@@ -107,7 +110,8 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *
     // Perform validation step
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), input_squared->info(), output->info(), norm_info));
 
-    const unsigned int border_width = (norm_info.is_cross_map()) ? 0 : std::min<unsigned int>(norm_info.norm_size() / 2, 3U);
+    const unsigned int norm_idx     = get_normalization_dimension_index(input->info()->data_layout(), norm_info);
+    const unsigned int border_width = (norm_idx == 2) ? 0 : std::min<unsigned int>(norm_info.norm_size() / 2, 3U);
 
     _input         = input;
     _input_squared = input_squared;
@@ -119,16 +123,21 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *
     {
         case DataType::F32:
         {
-            switch(norm_info.type())
+            switch(norm_idx)
             {
-                case NormType::IN_MAP_1D:
-                    _func = &NENormalizationLayerKernel::normalize_float<DataType::F32, 0, false>;
-                    break;
-                case NormType::IN_MAP_2D:
-                    // Normalize over X and Y
-                    _func = &NENormalizationLayerKernel::normalize_float<DataType::F32, 0, true>;
+                case 0:
+                {
+                    if(norm_info.type() == NormType::IN_MAP_2D)
+                    {
+                        _func = &NENormalizationLayerKernel::normalize_float<DataType::F32, 0, true>;
+                    }
+                    else
+                    {
+                        _func = &NENormalizationLayerKernel::normalize_float<DataType::F32, 0, false>;
+                    }
                     break;
-                case NormType::CROSS_MAP:
+                }
+                case 2:
                     _func = &NENormalizationLayerKernel::normalize_float<DataType::F32, 2, false>;
                     break;
                 default:
@@ -138,16 +147,21 @@ void NENormalizationLayerKernel::configure(const ITensor *input, const ITensor *
         }
         case DataType::F16:
         {
-            switch(norm_info.type())
+            switch(norm_idx)
             {
-                case NormType::IN_MAP_1D:
-                    _func = &NENormalizationLayerKernel::normalize_float<DataType::F16, 0, false>;
-                    break;
-                case NormType::IN_MAP_2D:
-                    // Normalize over X and Y
-                    _func = &NENormalizationLayerKernel::normalize_float<DataType::F16, 0, true>;
+                case 0:
+                {
+                    if(norm_info.type() == NormType::IN_MAP_2D)
+                    {
+                        _func = &NENormalizationLayerKernel::normalize_float<DataType::F16, 0, true>;
+                    }
+                    else
+                    {
+                        _func = &NENormalizationLayerKernel::normalize_float<DataType::F16, 0, false>;
+                    }
                     break;
-                case NormType::CROSS_MAP:
+                }
+                case 2:
                     _func = &NENormalizationLayerKernel::normalize_float<DataType::F16, 2, false>;
                     break;
                 default: