IVGCVSW-863 Broadcast support in CL/NEON Arithmetic Add

Also, added instrumentation to support generic tensor broadcasting for
NEON and CL backends.

Change-Id: I1bc5747a286e1a4b464c209067581e103d473b9a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/114201
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

6 files changed, 188 insertions, 29 deletions

diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h
index 7d4f2b866d..79e052c697 100644
--- a/tests/datasets/ShapeDatasets.h
+++ b/tests/datasets/ShapeDatasets.h
@@ -117,6 +117,34 @@ public:
     }
 };
 
+/** Data set containing pairs of small tensor shapes that are broadcast compatible. */
+class SmallShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset>
+{
+public:
+    SmallShapesBroadcast()
+        : ZipDataset<ShapeDataset, ShapeDataset>(
+              ShapeDataset("Shape0",
+    {
+        TensorShape{ 9U, 9U },
+                     TensorShape{ 27U, 13U, 2U },
+                     TensorShape{ 128U, 1U, 5U, 3U },
+                     TensorShape{ 9U, 9U, 3U, 4U },
+                     TensorShape{ 27U, 13U, 2U, 4U },
+                     TensorShape{ 1U, 1U, 1U, 5U }
+    }),
+    ShapeDataset("Shape1",
+    {
+        TensorShape{ 9U, 1U, 2U },
+        TensorShape{ 1U, 13U, 2U },
+        TensorShape{ 128U, 64U, 1U, 3U },
+        TensorShape{ 9U, 1U, 3U },
+        TensorShape{ 1U },
+        TensorShape{ 9U, 9U, 3U, 5U }
+    }))
+    {
+    }
+};
+
 /** Data set containing medium tensor shapes. */
 class MediumShapes final : public ShapeDataset
 {
@@ -172,6 +200,30 @@ public:
     }
 };
 
+/** Data set containing pairs of large tensor shapes that are broadcast compatible. */
+class LargeShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset>
+{
+public:
+    LargeShapesBroadcast()
+        : ZipDataset<ShapeDataset, ShapeDataset>(
+              ShapeDataset("Shape0",
+    {
+        TensorShape{ 1921U, 541U },
+                     TensorShape{ 1U, 485U, 2U, 3U },
+                     TensorShape{ 4159U, 1U },
+                     TensorShape{ 799U }
+    }),
+    ShapeDataset("Shape1",
+    {
+        TensorShape{ 1921U, 1U, 2U },
+        TensorShape{ 641U, 1U, 2U, 3U },
+        TensorShape{ 1U, 127U, 25U },
+        TensorShape{ 799U, 595U, 1U, 4U }
+    }))
+    {
+    }
+};
+
 /** Data set containing large 1D tensor shapes. */
 class Large1DShapes final : public ShapeDataset
 {
diff --git a/tests/framework/datasets/ContainerDataset.h b/tests/framework/datasets/ContainerDataset.h
index bdca97cbac..80616c46fc 100644
--- a/tests/framework/datasets/ContainerDataset.h
+++ b/tests/framework/datasets/ContainerDataset.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -72,7 +72,8 @@ public:
     {
     }
 
-    ContainerDataset(ContainerDataset &&) = default;
+    ContainerDataset(const ContainerDataset &) = default;
+    ContainerDataset(ContainerDataset &&)      = default;
 
     /** Type of the dataset. */
     using type = std::tuple<container_value_type>;
diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp
index 787b1b986f..4c19670d50 100644
--- a/tests/validation/CL/ArithmeticAddition.cpp
+++ b/tests/validation/CL/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -259,6 +259,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixture<float>, framework::
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
+
+template <typename T>
+using CLArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
 TEST_SUITE_END()
 TEST_SUITE_END()
 
diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp
index e20e8df665..32a4ff3a4d 100644
--- a/tests/validation/NEON/ArithmeticAddition.cpp
+++ b/tests/validation/NEON/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -263,6 +263,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework::
     // Validate output
     validate(Accessor(_target), _reference);
 }
+
+template <typename T>
+using NEArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticAddition, T>;
+
+FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(),
+                       ArithmeticAdditionFP32Dataset),
+                       framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })))
+{
+    // Validate output
+    validate(Accessor(_target), _reference);
+}
 TEST_SUITE_END()
 TEST_SUITE_END()
 
diff --git a/tests/validation/fixtures/ArithmeticAdditionFixture.h b/tests/validation/fixtures/ArithmeticAdditionFixture.h
index c3a51b97d1..f3888ae565 100644
--- a/tests/validation/fixtures/ArithmeticAdditionFixture.h
+++ b/tests/validation/fixtures/ArithmeticAdditionFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -41,15 +41,14 @@ namespace test
 namespace validation
 {
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class ArithmeticAdditionValidationFixedPointFixture : public framework::Fixture
+class ArithmeticAdditionBroadcastValidationFixedPointFixture : public framework::Fixture
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
     {
-        _fractional_bits = fractional_bits;
-        _target          = compute_target(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
-        _reference       = compute_reference(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        _target    = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+        _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
     }
 
 protected:
@@ -59,12 +58,13 @@ protected:
         library->fill_tensor_uniform(tensor, i);
     }
 
-    TensorType compute_target(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position)
+    TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+                              int fixed_point_position)
     {
         // Create tensors
-        TensorType ref_src1 = create_tensor<TensorType>(shape, data_type0, 1, fixed_point_position);
-        TensorType ref_src2 = create_tensor<TensorType>(shape, data_type1, 1, fixed_point_position);
-        TensorType dst      = create_tensor<TensorType>(shape, output_data_type, 1, fixed_point_position);
+        TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, fixed_point_position);
+        TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, fixed_point_position);
+        TensorType dst      = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, fixed_point_position);
 
         // Create and configure function
         FunctionType add;
@@ -93,11 +93,12 @@ protected:
         return dst;
     }
 
-    SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position)
+    SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy,
+                                      int fixed_point_position)
     {
         // Create reference
-        SimpleTensor<T> ref_src1{ shape, data_type0, 1, fixed_point_position };
-        SimpleTensor<T> ref_src2{ shape, data_type1, 1, fixed_point_position };
+        SimpleTensor<T> ref_src1{ shape0, data_type0, 1, fixed_point_position };
+        SimpleTensor<T> ref_src2{ shape1, data_type1, 1, fixed_point_position };
 
         // Fill reference
         fill(ref_src1, 0);
@@ -108,14 +109,36 @@ protected:
 
     TensorType      _target{};
     SimpleTensor<T> _reference{};
-    int             _fractional_bits{};
 };
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    {
+        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArithmeticAdditionValidationFixedPointFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits)
+    {
+        ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits);
+    }
+};
+
 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
 class ArithmeticAdditionValidationFixture : public ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
 {
 public:
     template <typename...>
-    void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
+    void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy)
     {
         ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type0, data_type1, output_data_type, convert_policy, 0);
     }
diff --git a/tests/validation/reference/ArithmeticAddition.cpp b/tests/validation/reference/ArithmeticAddition.cpp
index 82dd1437cd..17020a6277 100644
--- a/tests/validation/reference/ArithmeticAddition.cpp
+++ b/tests/validation/reference/ArithmeticAddition.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,27 +35,72 @@ namespace validation
 {
 namespace reference
 {
+namespace
+{
 template <typename T>
-SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+T add(T src1, T src2, ConvertPolicy convert_policy)
 {
-    SimpleTensor<T> result(src1.shape(), dst_data_type);
-
     using intermediate_type = typename common_promoted_signed_type<T>::intermediate_type;
 
-    for(int i = 0; i < src1.num_elements(); ++i)
+    intermediate_type val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2);
+
+    T result = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
+
+    return result;
+}
+
+template <size_t dim>
+struct BroadcastUnroll
+{
+    template <typename T>
+    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+                       ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
     {
-        intermediate_type val = static_cast<intermediate_type>(src1[i]) + static_cast<intermediate_type>(src2[i]);
-        result[i]             = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val);
+        const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]);
+        const bool src2_is_broadcast = (src2.shape()[dim - 1] != dst.shape()[dim - 1]);
+
+        id_src1.set(dim - 1, 0);
+        id_src2.set(dim - 1, 0);
+        id_dst.set(dim - 1, 0);
+
+        for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1])
+        {
+            BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+
+            id_src1[dim - 1] += !src1_is_broadcast;
+            id_src2[dim - 1] += !src2_is_broadcast;
+        }
     }
+};
 
-    return result;
+template <>
+struct BroadcastUnroll<0>
+{
+    template <typename T>
+    static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst,
+                       ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst)
+    {
+        dst[coord2index(dst.shape(), id_dst)] = add(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], convert_policy);
+    }
+};
+} // namespace
+
+template <typename T>
+SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy)
+{
+    SimpleTensor<T> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type);
+
+    Coordinates id_src1, id_src2, id_dst;
+
+    BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst);
+
+    return dst;
 }
 
 template SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int16_t> arithmetic_addition(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<int8_t> arithmetic_addition(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
-template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type,
-                                                ConvertPolicy convert_policy);
+template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 template SimpleTensor<float> arithmetic_addition(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy);
 } // namespace reference
 } // namespace validation