COMPMID-2855: NEReduceMean throws error for invalid configs

Change-Id: I600507d0de19d7da6c1a13edcfff0a11ea6b5264
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2254
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
diff --git a/arm_compute/core/Helpers.h b/arm_compute/core/Helpers.h
index 87b1fdf..8d526e9 100644
--- a/arm_compute/core/Helpers.h
+++ b/arm_compute/core/Helpers.h
@@ -766,6 +766,20 @@
     return x >= 0 ? x % m : (x % m + m) % m;
 }
 
+/** Convert negative coordinates to positive in the range [0, num_dims_input]
+ *
+ * @param[out] coords    Array of coordinates to be converted.
+ * @param[in]  max_value Maximum value to be used when wrapping the negative values in coords
+ */
+inline Coordinates &convert_negative_axis(Coordinates &coords, int max_value)
+{
+    for(unsigned int i = 0; i < coords.num_dimensions(); ++i)
+    {
+        coords[i] = wrap_around(coords[i], max_value);
+    }
+    return coords;
+}
+
 /** Given an integer value, this function returns the next power of two
  *
  * @param[in] x Input value
diff --git a/arm_compute/runtime/NEON/functions/NEReduceMean.h b/arm_compute/runtime/NEON/functions/NEReduceMean.h
index fdd8edf..245f757 100644
--- a/arm_compute/runtime/NEON/functions/NEReduceMean.h
+++ b/arm_compute/runtime/NEON/functions/NEReduceMean.h
@@ -72,7 +72,7 @@
     std::vector<NEReductionOperation> _reduction_kernels;
     std::vector<Tensor>               _reduced_outs;
     NEReshapeLayer                    _reshape;
-    unsigned int                      _reduction_ops;
+    int                               _reduction_ops;
     bool                              _keep_dims;
 };
 } // namespace arm_compute
diff --git a/src/runtime/NEON/functions/NEReduceMean.cpp b/src/runtime/NEON/functions/NEReduceMean.cpp
index 0b145f0..4547a1f 100644
--- a/src/runtime/NEON/functions/NEReduceMean.cpp
+++ b/src/runtime/NEON/functions/NEReduceMean.cpp
@@ -24,6 +24,7 @@
 #include "arm_compute/runtime/NEON/functions/NEReduceMean.h"
 
 #include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
@@ -34,49 +35,64 @@
 {
 }
 
-Status NEReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
+Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
 {
     ARM_COMPUTE_UNUSED(keep_dims);
-    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() < 1);
     ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() > input->num_dimensions());
 
-    TensorShape        out_shape     = input->tensor_shape();
     const unsigned int reduction_ops = reduction_axis.num_dimensions();
     const int          input_dims    = input->num_dimensions();
     Coordinates        axis_local    = reduction_axis;
 
-    // Convert negative axis
-    for(unsigned int i = 0; i < reduction_ops; ++i)
+    for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
     {
-        axis_local[i] = wrap_around(axis_local[i], input_dims);
+        //axis: The dimensions to reduce. Must be in the range [-rank(input_tensor), rank(input_tensor)).
+        ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] < (-static_cast<int>(input->num_dimensions())));
+        ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] >= static_cast<int>(input->num_dimensions()));
     }
 
-    std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
-    for(unsigned int i = 0; i < reduction_ops; ++i)
+    if(output->tensor_shape().total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
-        ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
-        if(output->total_size() > 0 && keep_dims)
+        // Only validate if not using auto_init for the output tensor
+        TensorShape out_shape = input->tensor_shape();
+        // Validate output_shape only if not using auto_init
+        convert_negative_axis(axis_local, input_dims);
+        std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
+        for(unsigned int i = 0; i < reduction_ops; ++i)
         {
-            ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
+            ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
+            ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
+            if(output->total_size() > 0 && keep_dims)
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
+            }
+            if(keep_dims)
+            {
+                out_shape.set(axis_local[i], 1);
+            }
+            else
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(i > static_cast<unsigned int>(axis_local[i]));
+                const unsigned int remove_index = axis_local[i] - i;
+                ARM_COMPUTE_RETURN_ERROR_ON(remove_index >= out_shape.num_dimensions());
+                out_shape.remove_dimension(remove_index);
+            }
         }
-        if(keep_dims)
-        {
-            out_shape.set(axis_local[i], 1);
-        }
-        else
-        {
-            out_shape.remove_dimension(axis_local[i] - i);
-        }
+        const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
     }
-    const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
-
     return Status{};
 }
 
+Status NEReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
+{
+    return validate_config(input, reduction_axis, keep_dims, output);
+}
+
 void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis, bool keep_dims, ITensor *output)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
@@ -86,18 +102,13 @@
     _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0));
     _keep_dims = keep_dims;
 
-    Coordinates        axis_local    = reduction_axis;
-    const int          input_dims    = input->info()->num_dimensions();
-    const unsigned int reduction_ops = reduction_axis.num_dimensions();
+    Coordinates axis_local = reduction_axis;
+    const int   input_dims = input->info()->num_dimensions();
 
-    // Convert negative axis
-    for(unsigned int i = 0; i < reduction_ops; ++i)
-    {
-        axis_local[i] = wrap_around(axis_local[i], input_dims);
-    }
+    convert_negative_axis(axis_local, input_dims);
 
     // Perform reduction for every axis
-    for(unsigned int i = 0; i < _reduction_ops; ++i)
+    for(int i = 0; i < _reduction_ops; ++i)
     {
         TensorShape out_shape = i == 0 ? input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape();
         out_shape.set(axis_local[i], 1);
@@ -116,7 +127,7 @@
     }
 
     // Allocate intermediate tensors
-    for(unsigned int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
+    for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
     {
         _reduced_outs[i].allocator()->allocate();
     }
@@ -125,11 +136,10 @@
     if(!keep_dims)
     {
         TensorShape out_shape = input->info()->tensor_shape();
-
         // We have to sort the reduction axis vectors in order for remove_dimension
         // to work properly
         std::sort(axis_local.begin(), axis_local.begin() + _reduction_ops);
-        for(unsigned int i = 0; i < _reduction_ops; ++i)
+        for(int i = 0; i < _reduction_ops; ++i)
         {
             out_shape.remove_dimension(axis_local[i] - i);
         }
@@ -141,10 +151,9 @@
 void NEReduceMean::run()
 {
     MemoryGroupResourceScope scope_mg(_memory_group);
-
-    for(unsigned int i = 0; i < _reduction_ops; ++i)
+    for(auto &kernel : _reduction_kernels)
     {
-        _reduction_kernels[i].run();
+        kernel.run();
     }
 
     if(!_keep_dims)
diff --git a/tests/validation/NEON/ReduceMean.cpp b/tests/validation/NEON/ReduceMean.cpp
index 3cd7ce3..6d0caf7 100644
--- a/tests/validation/NEON/ReduceMean.cpp
+++ b/tests/validation/NEON/ReduceMean.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018 ARM Limited.
+ * Copyright (c) 2018-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -57,20 +57,26 @@
 
 // *INDENT-OFF*
 // clang-format off
-DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
         framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 3U, 16U, 2U), 1, DataType::F32), // Invalid axis
                                                 TensorInfo(TensorShape(27U, 3U, 16U, 2U), 1, DataType::F32), // Invalid output shape
-                                                TensorInfo(TensorShape(32U, 16U, 16U, 2U), 1, DataType::F32)
+                                                TensorInfo(TensorShape(32U, 16U, 16U, 2U), 1, DataType::F32),// OK
+                                                TensorInfo(TensorShape{228U, 19U, 2U, 2U}, 1, DataType::F32),// OK
+                                                TensorInfo(TensorShape{228U, 19U, 2U, 1U}, 1, DataType::F32) // Cannot support axis 3 not valid
         }),
         framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(27U, 3U, 1U, 2U), 1, DataType::F32),
                                                  TensorInfo(TensorShape(27U, 3U, 1U, 2U), 1, DataType::F32),
-                                                 TensorInfo(TensorShape(32U, 16U, 1U, 2U), 1, DataType::F32)
+                                                 TensorInfo(TensorShape(32U, 16U, 1U, 2U), 1, DataType::F32),
+                                                 TensorInfo(TensorShape(19U), 1, DataType::F32),
+                                                 TensorInfo(TensorShape(19U), 1, DataType::F32)
+
         })),
-        framework::dataset::make("Axis", { Coordinates(4), Coordinates(0,2), Coordinates(2) })),
-        framework::dataset::make("Expected", { false, false, true })),
-        input_info, output_info, axis, expected)
+        framework::dataset::make("Axis", { Coordinates(4), Coordinates(0,2), Coordinates(2), Coordinates(3,2,0), Coordinates(3,2,0) })),
+        framework::dataset::make("Keep", { true, true, true, false, false })),
+        framework::dataset::make("Expected", { false, false, true, true, false })),
+        input_info, output_info, axis, keep, expected)
 {
-    const Status status = NEReduceMean::validate(&input_info.clone()->set_is_resizable(false), axis, true, &output_info.clone()->set_is_resizable(false));
+    const Status status = NEReduceMean::validate(&input_info.clone()->set_is_resizable(false), axis, keep, &output_info.clone()->set_is_resizable(false));
     ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
 }
 // clang-format on