COMPMID-2248

L2NormalizeLayer: negative axis

Change-Id: Ic164d7a9ddf1615a2e3b0e10430c34194a70f221
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1127
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 56 insertions, 34 deletions

diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
index cb2e29449c..00af590104 100644
--- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
+++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
@@ -35,29 +35,32 @@
 
 #include "support/ToolchainSupport.h"
 
-using namespace arm_compute;
-
+namespace arm_compute
+{
 CLL2NormalizeLayerKernel::CLL2NormalizeLayerKernel()
-    : _input(nullptr), _sum(nullptr), _output(nullptr), _axis(0), _epsilon(1e-12)
+    : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12)
 {
 }
 
 namespace
 {
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon)
+constexpr int max_input_tensor_dim = 3;
+
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
 {
     ARM_COMPUTE_UNUSED(epsilon);
 
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, sum, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum);
     ARM_COMPUTE_RETURN_ERROR_ON_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(axis > 2, "Axis greater than 2 is not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions");
 
     // Reduce shape on axis
     TensorShape sum_shape = input->tensor_shape();
-    sum_shape.set(axis, 1);
+    sum_shape.set(actual_axis, 1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape);
 
     if(output->total_size() != 0)
@@ -92,7 +95,7 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe
 }
 } // namespace
 
-void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, unsigned int axis, float epsilon)
+void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon));
@@ -100,7 +103,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor
     _input   = input;
     _sum     = sum;
     _output  = output;
-    _axis    = axis;
+    _actual_axis    = wrap_around(axis, max_input_tensor_dim);
     _epsilon = epsilon;
 
     const unsigned int num_elems_processed_per_iteration = 16;
@@ -113,7 +116,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor
     // Create kernel
     std::string  kernel_name;
     unsigned int idx = 0;
-    switch(axis)
+    switch(_actual_axis)
     {
         case 0:
             kernel_name = "x";
@@ -128,7 +131,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor
             idx         = num_arguments_per_3D_tensor() * 3;
             break;
         default:
-            ARM_COMPUTE_ERROR("Not supported");
+            ARM_COMPUTE_ERROR("Axis not supported");
     }
     _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("l2_normalize_" + kernel_name, build_opts));
 
@@ -149,7 +152,7 @@ void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor
     ICLKernel::configure_internal(std::get<1>(win_config));
 }
 
-Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon)
+Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon));
     ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get())));
@@ -164,7 +167,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue
 
     Window window_sum(window);
 
-    switch(_axis)
+    switch(_actual_axis)
     {
         case 0:
         {
@@ -218,3 +221,4 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue
             ARM_COMPUTE_ERROR("Not supported");
     }
 }
+} // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
index efdcc44e0e..9900446218 100644
--- a/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEL2NormalizeLayerKernel.cpp
@@ -40,6 +40,8 @@ namespace arm_compute
 {
 namespace
 {
+constexpr int max_input_tensor_dim = 3;
+
 template <typename T, int S>
 void l2_normalize_X(const ITensor *in, const ITensor *sum, ITensor *out, float epsilon, const Window &window)
 {
@@ -141,19 +143,20 @@ void l2_normalize_Z(const ITensor *in, const ITensor *sum, ITensor *out, float e
     while(window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice));
 }
 
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
 {
     ARM_COMPUTE_UNUSED(epsilon);
 
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, sum, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 2, "Axis greater than 2 is not supported");
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Normalization axis greater than max number of dimensions");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions");
 
     // Reduce shape on axis
     TensorShape sum_shape = input->tensor_shape();
-    sum_shape.set(axis, 1);
+    sum_shape.set(actual_axis, 1);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape);
 
     if(output->total_size() != 0)
@@ -167,10 +170,11 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons
     return Status{};
 }
 
-std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *sum, ITensorInfo *output, unsigned int axis)
+std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *sum, ITensorInfo *output, int axis)
 {
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
     const unsigned int num_elems_processed_per_iteration     = 16 / data_size_from_type(input->data_type());
-    const unsigned int num_elems_processed_per_iteration_sum = (axis == 0) ? 1 : num_elems_processed_per_iteration;
+    const unsigned int num_elems_processed_per_iteration_sum = (actual_axis == 0) ? 1 : num_elems_processed_per_iteration;
 
     Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
 
@@ -191,11 +195,11 @@ std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITe
 } // namespace
 
 NEL2NormalizeLayerKernel::NEL2NormalizeLayerKernel()
-    : _input(nullptr), _sum(nullptr), _output(nullptr), _axis(0), _epsilon(1e-12)
+    : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12)
 {
 }
 
-void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, unsigned int axis, float epsilon)
+void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *sum, ITensor *output, int axis, float epsilon)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output);
     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon));
@@ -203,7 +207,7 @@ void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *su
     _input   = input;
     _sum     = sum;
     _output  = output;
-    _axis    = axis;
+    _actual_axis    = wrap_around(axis, max_input_tensor_dim);
     _epsilon = epsilon;
 
     // Configure kernel window
@@ -213,7 +217,7 @@ void NEL2NormalizeLayerKernel::configure(const ITensor *input, const ITensor *su
     INEKernel::configure(std::get<1>(win_config));
 }
 
-Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, unsigned int axis, float epsilon)
+Status NEL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon)
 {
     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon));
     ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), sum->clone().get(), output->clone().get(), axis)));
@@ -227,7 +231,7 @@ void NEL2NormalizeLayerKernel::run(const Window &window, const ThreadInfo &info)
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
 
-    switch(_axis)
+    switch(_actual_axis)
     {
         case 0:
             switch(_input->info()->data_type())
diff --git a/src/runtime/CL/functions/CLL2NormalizeLayer.cpp b/src/runtime/CL/functions/CLL2NormalizeLayer.cpp
index 136cb5edef..e76e4f601e 100644
--- a/src/runtime/CL/functions/CLL2NormalizeLayer.cpp
+++ b/src/runtime/CL/functions/CLL2NormalizeLayer.cpp
@@ -34,25 +34,31 @@
 
 namespace arm_compute
 {
+namespace
+{
+constexpr int max_input_tensor_dim = 3;
+} // namespace
+    
 CLL2NormalizeLayer::CLL2NormalizeLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _reduce_func(), _normalize_kernel(), _sumsq()
 {
 }
 
-void CLL2NormalizeLayer::configure(ICLTensor *input, ICLTensor *output, unsigned int axis, float epsilon)
+void CLL2NormalizeLayer::configure(ICLTensor *input, ICLTensor *output, int axis, float epsilon)
 {
     // Manage intermediate buffers
     _memory_group.manage(&_sumsq);
 
     // Configure kernels
-    _reduce_func.configure(input, &_sumsq, axis, ReductionOperation::SUM_SQUARE);
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
+    _reduce_func.configure(input, &_sumsq, actual_axis, ReductionOperation::SUM_SQUARE);
     _normalize_kernel.configure(input, &_sumsq, output, axis, epsilon);
 
     // Allocate intermediate tensor
     _sumsq.allocator()->allocate();
 }
 
-Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon)
+Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon)
 {
     TensorShape shape(input->tensor_shape());
 
@@ -61,10 +67,11 @@ Status CLL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo
     sum_sq.set_data_type(input->data_type());
     sum_sq.set_tensor_shape(shape);
 
-    ARM_COMPUTE_RETURN_ON_ERROR(CLReductionOperation::validate(input, &sum_sq, axis, ReductionOperation::SUM_SQUARE));
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
+    ARM_COMPUTE_RETURN_ON_ERROR(CLReductionOperation::validate(input, &sum_sq, actual_axis, ReductionOperation::SUM_SQUARE));
 
     // Reduce shape on axis
-    shape.set(axis, 1);
+    shape.set(actual_axis, 1);
     sum_sq.set_tensor_shape(shape);
 
     ARM_COMPUTE_RETURN_ON_ERROR(CLL2NormalizeLayerKernel::validate(input, &sum_sq, output, axis, epsilon));
diff --git a/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp b/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp
index c9ab5c98e2..88ffdbfd08 100644
--- a/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp
+++ b/src/runtime/NEON/functions/NEL2NormalizeLayer.cpp
@@ -28,25 +28,31 @@
 
 namespace arm_compute
 {
+namespace
+{
+constexpr int max_input_tensor_dim = 3;
+} // namespace
+
 NEL2NormalizeLayer::NEL2NormalizeLayer(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _reduce_func(), _normalize_kernel(), _sumsq()
 {
 }
 
-void NEL2NormalizeLayer::configure(ITensor *input, ITensor *output, unsigned int axis, float epsilon)
+void NEL2NormalizeLayer::configure(ITensor *input, ITensor *output, int axis, float epsilon)
 {
     // Manage intermediate buffers
     _memory_group.manage(&_sumsq);
 
     // Configure Kernels
-    _reduce_func.configure(input, &_sumsq, axis, ReductionOperation::SUM_SQUARE);
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
+    _reduce_func.configure(input, &_sumsq, actual_axis, ReductionOperation::SUM_SQUARE);
     _normalize_kernel.configure(input, &_sumsq, output, axis, epsilon);
 
     // Allocate intermediate tensors
     _sumsq.allocator()->allocate();
 }
 
-Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, float epsilon)
+Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo *output, int axis, float epsilon)
 {
     TensorShape shape(input->tensor_shape());
 
@@ -55,10 +61,11 @@ Status NEL2NormalizeLayer::validate(const ITensorInfo *input, const ITensorInfo
     sum_sq.set_data_type(input->data_type());
     sum_sq.set_tensor_shape(shape);
 
-    ARM_COMPUTE_RETURN_ON_ERROR(NEReductionOperation::validate(input, &sum_sq, axis, ReductionOperation::SUM_SQUARE));
+    const uint32_t actual_axis = wrap_around(axis, max_input_tensor_dim);
+    ARM_COMPUTE_RETURN_ON_ERROR(NEReductionOperation::validate(input, &sum_sq, actual_axis, ReductionOperation::SUM_SQUARE));
 
     // Reduce shape on axis
-    shape.set(axis, 1);
+    shape.set(actual_axis, 1);
     sum_sq.set_tensor_shape(shape);
 
     ARM_COMPUTE_RETURN_ON_ERROR(NEL2NormalizeLayerKernel::validate(input, &sum_sq, output, axis, epsilon));