COMPMID-2819: Retain configuration step data layout to avoid side-effects.

Configuring functions serially can lead to side-effects in tensor
attributes. One of them is the data layout changing in case functions
share same IO tensors. Retain DataLayout used during configuration.

Change-Id: I17538ce08b86df6986f0fcf21fa6544fbd5bd74b
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2470
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

6 files changed, 33 insertions, 29 deletions

diff --git a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
index db1a756229..68a99039d8 100644
--- a/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
+++ b/arm_compute/core/CL/kernels/CLPoolingLayerKernel.h
@@ -75,6 +75,7 @@ public:
     const ICLTensor *_input;
     ICLTensor       *_output;
     PoolingLayerInfo _pool_info;
+    DataLayout       _data_layout;
     BorderSize       _border_size;
     unsigned int     _num_elems_processed_per_iteration;
 };
diff --git a/arm_compute/core/CL/kernels/CLScaleKernel.h b/arm_compute/core/CL/kernels/CLScaleKernel.h
index ff72af29fc..1ada3cde85 100644
--- a/arm_compute/core/CL/kernels/CLScaleKernel.h
+++ b/arm_compute/core/CL/kernels/CLScaleKernel.h
@@ -75,6 +75,7 @@ public:
 
 public:
     InterpolationPolicy _interpolationPolicy = InterpolationPolicy::BILINEAR;
+    DataLayout          _data_layout         = DataLayout::UNKNOWN;
 };
 } // namespace arm_compute
 #endif /*__ARM_COMPUTE_CLSCALEKERNEL_H__ */
diff --git a/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h b/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
index 5f45a90cef..5b143250e9 100644
--- a/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
+++ b/arm_compute/core/NEON/kernels/NEPoolingLayerKernel.h
@@ -194,6 +194,7 @@ private:
     const ITensor   *_input;
     ITensor         *_output;
     PoolingLayerInfo _pool_info;
+    DataLayout       _data_layout;
     unsigned int     _num_elems_processed_per_iteration;
     BorderSize       _border_size;
     bool             _is_square;
diff --git a/src/core/CL/kernels/CLPoolingLayerKernel.cpp b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
index 8eaf5bf76f..032d451aad 100644
--- a/src/core/CL/kernels/CLPoolingLayerKernel.cpp
+++ b/src/core/CL/kernels/CLPoolingLayerKernel.cpp
@@ -172,7 +172,7 @@ std::tuple<Status, Window, CLPoolingConfig> validate_and_configure_window(ITenso
 } // namespace
 
 CLPoolingLayerKernel::CLPoolingLayerKernel()
-    : _input(nullptr), _output(nullptr), _pool_info(), _border_size(0), _num_elems_processed_per_iteration(1)
+    : _input(nullptr), _output(nullptr), _pool_info(), _data_layout(DataLayout::UNKNOWN), _border_size(0), _num_elems_processed_per_iteration(1)
 {
 }
 
@@ -185,13 +185,18 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
 
+    // Set instance variables
+    _input       = input;
+    _output      = output;
+    _pool_info   = pool_info;
+    _data_layout = input->info()->data_layout();
+
     int                 pool_stride_x   = 0;
     int                 pool_stride_y   = 0;
     const PoolingType   pool_type       = pool_info.pool_type();
-    DataLayout          data_layout     = input->info()->data_layout();
-    const int           idx_width       = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int           idx_height      = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const int           idx_channel     = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL);
+    const int           idx_width       = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const int           idx_height      = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+    const int           idx_channel     = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL);
     const int           pool_size_x     = pool_info.is_global_pooling() ? input->info()->dimension(idx_width) : pool_info.pool_size().width;
     const int           pool_size_y     = pool_info.is_global_pooling() ? input->info()->dimension(idx_height) : pool_info.pool_size().height;
     const PadStrideInfo pad_stride_info = pool_info.pad_stride_info();
@@ -218,11 +223,6 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     auto_init(input->info(), output->info(), pool_info);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info));
 
-    // Set instance variables
-    _input     = input;
-    _output    = output;
-    _pool_info = pool_info;
-
     const DataType data_type = input->info()->data_type();
 
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
@@ -237,7 +237,7 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     build_opts.add_option_if(data_type == DataType::F16, "-DFP16");
 
     // Create kernel
-    switch(data_layout)
+    switch(_data_layout)
     {
         case DataLayout::NCHW:
         {
@@ -286,7 +286,7 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
     ICLKernel::configure_internal(std::get<1>(win_config));
 
-    if(data_layout == DataLayout::NCHW)
+    if(_data_layout == DataLayout::NCHW)
     {
         CLPoolingConfig pooling_config     = std::get<2>(win_config);
         _num_elems_processed_per_iteration = pooling_config.first;
@@ -302,7 +302,7 @@ void CLPoolingLayerKernel::configure(const ICLTensor *input, ICLTensor *output,
     _config_id = "pooling_layer_";
     _config_id += lower_string(string_from_data_type(data_type));
     _config_id += "_";
-    _config_id += lower_string(string_from_data_layout(data_layout));
+    _config_id += lower_string(string_from_data_layout(_data_layout));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(idx_width));
     _config_id += "_";
@@ -333,7 +333,7 @@ void CLPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue)
     // Collapse window
     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
 
-    switch(_input->info()->data_layout())
+    switch(_data_layout)
     {
         case DataLayout::NCHW:
         {
diff --git a/src/core/CL/kernels/CLScaleKernel.cpp b/src/core/CL/kernels/CLScaleKernel.cpp
index 488313fd12..82c5c8a446 100644
--- a/src/core/CL/kernels/CLScaleKernel.cpp
+++ b/src/core/CL/kernels/CLScaleKernel.cpp
@@ -160,11 +160,12 @@ const ICLTensor *CLScaleKernel::output() const
 
 void CLScaleKernel::configure(const ICLTensor *input, ICLTensor *output, InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy)
 {
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy));
+
     _input               = input;
     _output              = output;
     _interpolationPolicy = policy;
-
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), policy));
+    _data_layout         = input->info()->data_layout();
 
     float wr = 0.f;
     float hr = 0.f;
@@ -172,10 +173,9 @@ void CLScaleKernel::configure(const ICLTensor *input, ICLTensor *output, Interpo
 
     const bool call_quantized_kernel = is_data_type_quantized_asymmetric(input->info()->data_type()) && policy == InterpolationPolicy::BILINEAR;
 
-    DataLayout data_layout = input->info()->data_layout();
-    const int  idx_width   = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH);
-    const int  idx_height  = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT);
-    const bool is_nhwc     = data_layout == DataLayout::NHWC;
+    const int  idx_width  = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH);
+    const int  idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+    const bool is_nhwc    = _data_layout == DataLayout::NHWC;
 
     // Compute the ratio between source width/height and destination width/height
     const unsigned int input_width   = input->info()->dimension(idx_width);
@@ -215,7 +215,7 @@ void CLScaleKernel::configure(const ICLTensor *input, ICLTensor *output, Interpo
     std::transform(interpolation_name.begin(), interpolation_name.end(), interpolation_name.begin(), ::tolower);
     std::string kernel_name = "scale_" + interpolation_name;
     kernel_name += call_quantized_kernel ? "_quantized_" : "_";
-    kernel_name += lower_string(string_from_data_layout(data_layout));
+    kernel_name += lower_string(string_from_data_layout(_data_layout));
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     unsigned int idx = is_nhwc ? 2 * num_arguments_per_4D_tensor() : 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters
@@ -249,7 +249,7 @@ void CLScaleKernel::run(const Window &window, cl::CommandQueue &queue)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
-    switch(_input->info()->data_layout())
+    switch(_data_layout)
     {
         case DataLayout::NCHW:
         {
diff --git a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
index 5be6f5b6f0..14de4a19d8 100644
--- a/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEPoolingLayerKernel.cpp
@@ -321,7 +321,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 NEPoolingLayerKernel::NEPoolingLayerKernel()
-    : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _num_elems_processed_per_iteration(0), _border_size(0), _is_square(false)
+    : _func(nullptr), _input(nullptr), _output(nullptr), _pool_info(), _data_layout(DataLayout::UNKNOWN), _num_elems_processed_per_iteration(0), _border_size(0), _is_square(false)
 {
 }
 
@@ -364,14 +364,15 @@ void NEPoolingLayerKernel::configure(const ITensor *input, ITensor *output, cons
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), pool_info, pooled_w, pooled_h));
 
     // Set instance variables
-    _input     = input;
-    _output    = output;
-    _pool_info = pool_info;
-    _is_square = (pool_size.x() == pool_size.y());
+    _input       = input;
+    _output      = output;
+    _pool_info   = pool_info;
+    _data_layout = input->info()->data_layout();
+    _is_square   = (pool_size.x() == pool_size.y());
 
     // Get data type
     const DataType data_type = input->info()->data_type();
-    const bool     is_nchw   = data_layout == DataLayout::NCHW;
+    const bool     is_nchw   = _data_layout == DataLayout::NCHW;
 
     if(data_type == DataType::QASYMM8)
     {
@@ -1840,7 +1841,7 @@ void NEPoolingLayerKernel::run(const Window &window, const ThreadInfo &info)
     const bool         exclude_padding = _pool_info.exclude_padding();
 
     Window window_input(window);
-    if(_input->info()->data_layout() == DataLayout::NCHW)
+    if(_data_layout == DataLayout::NCHW)
     {
         // Set step for input in x and y direction for the input
         unsigned int window_x_inc = 0;