Add support for 2d and 3d indices for axis 0

* Partially resolves COMPMID-5055

Change-Id: Id05374b8c69e6b9ab4c2790a4de93d7172063b71

Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Change-Id: Ic6e2c2d1d34abbf6222c8d56859514e267447266
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7488
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

6 files changed, 176 insertions, 52 deletions

diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index df907c106e..aa51ad209a 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -1496,13 +1496,24 @@ inline TensorShape compute_pool3d_shape(const TensorShape &src, Pooling3dLayerIn
 
 inline TensorShape compute_gather_shape(const TensorShape &input_shape, const TensorShape &indices_shape, uint32_t actual_axis)
 {
-    ARM_COMPUTE_ERROR_ON(indices_shape.num_dimensions() > 1);
     ARM_COMPUTE_ERROR_ON(input_shape.num_dimensions() > 4);
     ARM_COMPUTE_ERROR_ON(actual_axis >= input_shape.num_dimensions());
 
-    TensorShape output_shape  = input_shape;
-    output_shape[actual_axis] = indices_shape[0];
-
+    TensorShape output_shape = input_shape;
+    if(indices_shape.num_dimensions() == 1u)
+    {
+        output_shape[actual_axis] = indices_shape[0];
+    }
+    else
+    {
+        const auto inddims{ indices_shape.num_dimensions() };
+        output_shape.shift_right(indices_shape.num_dimensions() - 1);
+        output_shape[0] = input_shape[0];
+        for(size_t idx(1); (idx - 1) < inddims; ++idx)
+        {
+            output_shape.set(actual_axis + idx, indices_shape[idx - 1], false);
+        }
+    }
     return output_shape;
 }
 } // namespace shape_calculator
diff --git a/src/core/NEON/kernels/NEGatherKernel.cpp b/src/core/NEON/kernels/NEGatherKernel.cpp
index 7090da8015..8d86a22b7e 100644
--- a/src/core/NEON/kernels/NEGatherKernel.cpp
+++ b/src/core/NEON/kernels/NEGatherKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -47,16 +47,21 @@ namespace
 template <typename U>
 void validate_indices(const ITensor *indices)
 {
-    for(size_t i = 0; i < indices->info()->tensor_shape()[0]; ++i)
+    auto *indices_ptr = (reinterpret_cast<const U *>(  indices->buffer() + indices->info()->offset_first_element_in_bytes() ));
+    for(size_t i = 0; i < indices->info()->total_size(); ++i)
     {
-        ARM_COMPUTE_ERROR_ON(*(reinterpret_cast<U *>(indices->ptr_to_element(Coordinates(i)))) < 0);
+        const U index_value = indices_ptr[i];
+        ARM_COMPUTE_UNUSED(index_value);
+        if(index_value < 0)
+        {
+            ARM_COMPUTE_ERROR_ON(index_value < 0);
+        }
     }
 }
 
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices, const ITensorInfo *output, int axis)
 {
     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, indices, output);
-    ARM_COMPUTE_RETURN_ERROR_ON(indices->num_dimensions() > 1);
     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
 
     if(axis < 0)
@@ -65,6 +70,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *indices,
     }
 
     ARM_COMPUTE_RETURN_ERROR_ON(0 > axis || axis >= static_cast<int32_t>(input->num_dimensions()));
+    ARM_COMPUTE_RETURN_ERROR_ON(axis != 0 && indices->num_dimensions() > 1);
     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
 
     if(output->total_size() != 0)
@@ -87,6 +93,40 @@ NEGatherKernel::NEGatherKernel()
 }
 
 template <typename U>
+inline void NEGatherKernel::gather_dims_0_axis(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_ERROR_ON(_indices->info()->num_dimensions() < 2);
+    validate_indices<U>(_indices);
+
+    Window output_window{ window };
+    output_window.set(Window::DimX, Window::Dimension(0, 1, 1));
+    Iterator output_it(_output, output_window);
+
+    const uint8_t *const in_ptr_start    = _input->buffer() + _input->info()->offset_first_element_in_bytes();
+    const uint32_t       input_stride_y  = _input->info()->strides_in_bytes()[1];
+    const uint32_t       output_stride_y = _output->info()->strides_in_bytes()[1];
+
+    const U *const dex_ptr_start = reinterpret_cast<const U *const>(_indices->buffer() + _indices->info()->offset_first_element_in_bytes());
+    execute_window_loop(output_window, [&](const Coordinates & id)
+    {
+        const auto        new_index = *(dex_ptr_start + id.y() + id.z() * _output->info()->tensor_shape()[1] + id[3] * _indices->info()->tensor_shape()[1] * _indices->info()->tensor_shape()[0]);
+        U                *out_ptr   = reinterpret_cast<U *>(output_it.ptr());
+        const char *const in_ptr    = reinterpret_cast<const char *const>(in_ptr_start + new_index * input_stride_y);
+        memcpy(out_ptr, in_ptr, output_stride_y);
+    },
+    output_it);
+}
+
+template <typename U>
+inline void NEGatherKernel::gather_dims_n_axis(const Window &window, const ThreadInfo &info)
+{
+    ARM_COMPUTE_UNUSED(info);
+    ARM_COMPUTE_UNUSED(window);
+    ARM_COMPUTE_ERROR("NOT_SUPPORTED!");
+}
+
+template <typename U>
 inline void NEGatherKernel::gather_0_axis(const Window &window, const ThreadInfo &info)
 {
     ARM_COMPUTE_UNUSED(info);
@@ -147,38 +187,75 @@ void NEGatherKernel::configure(const ITensor *input, const ITensor *indices, ITe
     }
     ARM_COMPUTE_ERROR_ON(0 > _axis || _axis >= static_cast<int32_t>(input->info()->num_dimensions()));
 
-    if(0 == _axis)
+    if(indices->info()->num_dimensions() == 1u)
     {
-        switch(_indices->info()->data_type())
+        if(0 == _axis)
+        {
+            switch(_indices->info()->data_type())
+            {
+                case DataType::U32:
+                    _func = &NEGatherKernel::gather_0_axis<uint32_t>;
+                    break;
+                case DataType::S32:
+                    _func = &NEGatherKernel::gather_0_axis<int32_t>;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Not supported");
+                    break;
+            }
+        }
+        else
         {
-            case DataType::U32:
-                _func = &NEGatherKernel::gather_0_axis<uint32_t>;
-                break;
-            case DataType::S32:
-                _func = &NEGatherKernel::gather_0_axis<int32_t>;
-                break;
-            default:
-                ARM_COMPUTE_ERROR("Not supported");
-                break;
+            switch(_indices->info()->data_type())
+            {
+                case DataType::U32:
+                    _func = &NEGatherKernel::gather_n_axis<uint32_t>;
+                    break;
+                case DataType::S32:
+                    _func = &NEGatherKernel::gather_n_axis<int32_t>;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Not supported");
+                    break;
+            }
         }
     }
     else
     {
-        switch(_indices->info()->data_type())
+        if(0 == _axis)
         {
-            case DataType::U32:
-                _func = &NEGatherKernel::gather_n_axis<uint32_t>;
-                break;
-            case DataType::S32:
-                _func = &NEGatherKernel::gather_n_axis<int32_t>;
-                break;
-            default:
-                ARM_COMPUTE_ERROR("Not supported");
-                break;
+            switch(_indices->info()->data_type())
+            {
+                case DataType::U32:
+                    _func = &NEGatherKernel::gather_dims_0_axis<uint32_t>;
+                    break;
+                case DataType::S32:
+                    _func = &NEGatherKernel::gather_dims_0_axis<int32_t>;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Not supported");
+                    break;
+            }
+        }
+        else
+        {
+            switch(_indices->info()->data_type())
+            {
+                case DataType::U32:
+                    _func = &NEGatherKernel::gather_dims_n_axis<uint32_t>;
+                    break;
+                case DataType::S32:
+                    _func = &NEGatherKernel::gather_dims_n_axis<int32_t>;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Not supported");
+                    break;
+            }
         }
     }
+
     // Output auto initialization if not yet initialized
-    TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->info()->tensor_shape(), indices->info()->tensor_shape(), _axis);
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_gather_shape(input->info()->tensor_shape(), indices->info()->tensor_shape(), _axis);
     auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
 
     // Create window
diff --git a/src/core/NEON/kernels/NEGatherKernel.h b/src/core/NEON/kernels/NEGatherKernel.h
index 0711f8190b..fc0e67854b 100644
--- a/src/core/NEON/kernels/NEGatherKernel.h
+++ b/src/core/NEON/kernels/NEGatherKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited.
+ * Copyright (c) 2019-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -85,18 +85,23 @@ private:
      *
      * For gather on the 0 axis an element by element copy is performed.
      *
-     * @param[in] window Region on which to execute the kernel. (Must be a region of the window returned by window())
-     * @param[in] info   Info about executing thread and CPU.
+     * @param[in] window Region on which to run the kernel. (Must be a region of the window returned by window())
+     * @param[in] info   Info about running thread and CPU.
      */
     template <typename U>
     void gather_0_axis(const Window &window, const ThreadInfo &info);
 
+    template <typename U>
+    void gather_dims_0_axis(const Window &window, const ThreadInfo &info);
+
+    template <typename U>
+    void gather_dims_n_axis(const Window &window, const ThreadInfo &info);
     /** Implementation of the gather operation.
      *
      * For 1<=axis a row-wise copy is taking place.
      *
-     * @param[in] window Region on which to execute the kernel. (Must be a region of the window returned by window())
-     * @param[in] info   Info about executing thread and CPU.
+     * @param[in] window Region on which to run the kernel. (Must be a region of the window returned by window())
+     * @param[in] info   Info about running thread and CPU.
      */
     template <typename U>
     void gather_n_axis(const Window &window, const ThreadInfo &info);
diff --git a/tests/datasets/GatherDataset.h b/tests/datasets/GatherDataset.h
index 29a99d5239..f4ad7c9497 100644
--- a/tests/datasets/GatherDataset.h
+++ b/tests/datasets/GatherDataset.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019, 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -106,6 +106,18 @@ private:
     std::vector<int>         _axis{};
 };
 
+
+class SmallGatherMultiDimIndicesDataset final : public GatherDataset
+{
+public:
+    SmallGatherMultiDimIndicesDataset()
+    {
+        add_config(TensorShape(15U, 15U), TensorShape(4U, 13U, 2U), 0);
+        add_config(TensorShape(15U, 15U), TensorShape(2U, 12U), 0);
+    }
+};
+
+
 class SmallGatherDataset final : public GatherDataset
 {
 public:
diff --git a/tests/validation/NEON/Gather.cpp b/tests/validation/NEON/Gather.cpp
index ca1e166bd1..1ab2668aad 100644
--- a/tests/validation/NEON/Gather.cpp
+++ b/tests/validation/NEON/Gather.cpp
@@ -100,12 +100,14 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
 template <typename T>
 using NEGatherFixture = GatherFixture<Tensor, Accessor, NEGather, T>;
 
+const auto gather_small_shapes = arm_compute::test::framework::dataset::concat(datasets::SmallGatherDataset(),datasets::SmallGatherMultiDimIndicesDataset());
+
 TEST_SUITE(Float)
 TEST_SUITE(FP16)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEGatherFixture<half>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F16)))
+                       combine( gather_small_shapes, framework::dataset::make("DataType", DataType::F16)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -125,7 +127,7 @@ TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEGatherFixture<float>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F32)))
+                       combine(gather_small_shapes, framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -146,7 +148,7 @@ TEST_SUITE(U8)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEGatherFixture<uint8_t>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U8)))
+                       combine(gather_small_shapes, framework::dataset::make("DataType", DataType::U8)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -166,7 +168,7 @@ TEST_SUITE(U16)
 FIXTURE_DATA_TEST_CASE(RunSmall,
                        NEGatherFixture<uint16_t>,
                        framework::DatasetMode::PRECOMMIT,
-                       combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U16)))
+                       combine(gather_small_shapes, framework::dataset::make("DataType", DataType::U16)))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/reference/Gather.cpp b/tests/validation/reference/Gather.cpp
index 93ac09cf95..02292fb74e 100644
--- a/tests/validation/reference/Gather.cpp
+++ b/tests/validation/reference/Gather.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2018-2019 Arm Limited.
+ * Copyright (c) 2018-2019, 2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -45,22 +45,39 @@ SimpleTensor<T> gather(const SimpleTensor<T> &src, const SimpleTensor<uint32_t>
 
     Window win;
     win.use_tensor_dimensions(dst_shape);
-    execute_window_loop(win, [&](const Coordinates & id)
+    if(indices.shape().num_dimensions() == 1u)
     {
-        Coordinates offset;
-        for(unsigned int dim = 0; dim < id.num_dimensions(); ++dim)
+        execute_window_loop(win, [&](const Coordinates & id)
         {
-            if(dim == actual_axis)
+            Coordinates offset;
+            for(unsigned int dim = 0; dim < id.num_dimensions(); ++dim)
             {
-                offset.set(dim, indices_ptr[id[dim]]);
+                if(dim == actual_axis)
+                {
+                    offset.set(dim, indices_ptr[id[dim]]);
+                }
+                else
+                {
+                    offset.set(dim, id[dim]);
+                }
             }
-            else
+            *reinterpret_cast<T *>(dst(id)) = *reinterpret_cast<const T *>(src(offset));
+        });
+    }
+    else
+    {
+        if(actual_axis == 0)
+        {
+            win.set(Window::DimX, Window::Dimension(0, 1, 1));
+            uint32_t index = 0;
+            execute_window_loop(win, [&](const Coordinates & id)
             {
-                offset.set(dim, id[dim]);
-            }
+                auto     *dst_ptr     = reinterpret_cast<T *>(dst(id));
+                const int row_to_copy = indices[index++];
+                std::copy_n(src.data() + row_to_copy * src.shape()[0], src.shape()[0], dst_ptr);
+            });
         }
-        *reinterpret_cast<T *>(dst(id)) = *reinterpret_cast<const T *>(src(offset));
-    });
+    }
 
     return dst;
 }
@@ -72,4 +89,4 @@ template SimpleTensor<uint8_t> gather(const SimpleTensor<uint8_t> &src, const Si
 } // namespace reference
 } // namespace validation
 } // namespace test
-} // namespace arm_compute
\ No newline at end of file
+} // namespace arm_compute