COMPMID-1728 CL: Implement ArgMax/ArgMin

Change-Id: I7eae2e55cc0b0b7bbebb7617299daaca6f75f40c
Reviewed-on: https://review.mlplatform.org/292
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 338 insertions, 14 deletions

diff --git a/tests/validation/CL/ArgMinMax.cpp b/tests/validation/CL/ArgMinMax.cpp
new file mode 100644
index 0000000000..0b873945d3
--- /dev/null
+++ b/tests/validation/CL/ArgMinMax.cpp
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/CL/CLTensorAllocator.h"
+#include "arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/SplitDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/ArgMinMaxFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+TEST_SUITE(CL)
+TEST_SUITE(ArgMinMax)
+
+// *INDENT-OFF*
+// clang-format off
+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) // Invalid operation
+        }),
+        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::U32),
+                                                 TensorInfo(TensorShape(32U, 16U, 1U, 2U), 1, DataType::F32)
+        })),
+        framework::dataset::make("Axis", { 4, 0, 2, 0 })),
+        framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MAX, ReductionOperation::ARG_IDX_MAX, ReductionOperation::ARG_IDX_MAX, ReductionOperation::MEAN_SUM })),
+        framework::dataset::make("Expected", { false, false, true, false })),
+        input_info, output_info, axis, operation, expected)
+{
+    const Status status = CLArgMinMaxLayer::validate(&input_info.clone()->set_is_resizable(false), axis, &output_info.clone()->set_is_resizable(false), operation);
+    ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+DATA_TEST_CASE(Configuration,
+               framework::DatasetMode::ALL,
+               combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F16, DataType::F32 })),
+               shape, data_type)
+{
+    // Create tensors
+    CLTensor ref_src = create_tensor<CLTensor>(shape, data_type);
+    CLTensor dst;
+
+    // Create and Configure function
+    CLArgMinMaxLayer arg_min_max_layer;
+    arg_min_max_layer.configure(&ref_src, 1, &dst, ReductionOperation::ARG_IDX_MAX);
+
+    // Validate valid region
+    TensorShape output_shape = shape;
+    output_shape.set(1, 1);
+    const ValidRegion valid_region = shape_to_valid_region(output_shape);
+    validate(dst.info()->valid_region(), valid_region);
+}
+
+template <typename T>
+using CLArgMinMaxValidationFixture = ArgMinMaxValidationFixture<CLTensor, CLAccessor, CLArgMinMaxLayer, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLArgMinMaxValidationFixture<half>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLArgMinMaxValidationFixture<half>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FP16
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+                       CLArgMinMaxValidationFixture<float>,
+                       framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+                       CLArgMinMaxValidationFixture<float>,
+                       framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // ArgMinMax
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/ArgMinMaxFixture.h b/tests/validation/fixtures/ArgMinMaxFixture.h
new file mode 100644
index 0000000000..5f5f85c104
--- /dev/null
+++ b/tests/validation/fixtures/ArgMinMaxFixture.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef ARM_COMPUTE_TEST_ARG_MIN_MAX_FIXTURE
+#define ARM_COMPUTE_TEST_ARG_MIN_MAX_FIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "tests/AssetsLibrary.h"
+#include "tests/Globals.h"
+#include "tests/IAccessor.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/Helpers.h"
+#include "tests/validation/reference/ReductionOperation.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ArgMinMaxValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, DataType data_type, int axis, ReductionOperation op)
+    {
+        _target    = compute_target(shape, data_type, axis, op);
+        _reference = compute_reference(shape, data_type, axis, op);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor)
+    {
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, 0);
+    }
+
+    TensorType compute_target(TensorShape &src_shape, DataType data_type, int axis, ReductionOperation op)
+    {
+        // Create tensors
+        TensorType src = create_tensor<TensorType>(src_shape, data_type, 1);
+        TensorType dst;
+
+        // Create and configure function
+        FunctionType arg_min_max_layer;
+        arg_min_max_layer.configure(&src, axis, &dst, op);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src));
+
+        // Compute function
+        arg_min_max_layer.run();
+
+        return dst;
+    }
+
+    SimpleTensor<T> compute_reference(TensorShape &src_shape, DataType data_type, int axis, ReductionOperation op)
+    {
+        // Create reference
+        SimpleTensor<T> src{ src_shape, data_type, 1 };
+
+        // Fill reference
+        fill(src);
+
+        TensorShape output_shape = src_shape;
+        output_shape.set(axis, 1);
+        return reference::reduction_operation<T>(src, output_shape, axis, op);
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_ARG_MIN_MAX_FIXTURE */
diff --git a/tests/validation/reference/ReductionOperation.cpp b/tests/validation/reference/ReductionOperation.cpp
index 2f103a6f65..37a9be86c0 100644
--- a/tests/validation/reference/ReductionOperation.cpp
+++ b/tests/validation/reference/ReductionOperation.cpp
@@ -38,10 +38,10 @@ namespace reference
 {
 namespace
 {
-template <typename T>
-T reduce_operation(T *ptr, int reduce_elements, ReductionOperation op, int stride)
+template <typename T, typename OT>
+OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, int stride)
 {
-    using type = typename std::remove_cv<T>::type;
+    using type = typename std::remove_cv<OT>::type;
     auto res   = type(0);
 
     if(std::is_integral<type>::value)
@@ -50,7 +50,31 @@ T reduce_operation(T *ptr, int reduce_elements, ReductionOperation op, int strid
         for(int i = 0; i < reduce_elements; ++i)
         {
             auto elem = static_cast<uint32_t>(*(ptr + stride * i));
-            int_res += (op == ReductionOperation::SUM_SQUARE) ? elem * elem : elem;
+
+            switch(op)
+            {
+                case ReductionOperation::ARG_IDX_MIN:
+                    if(static_cast<uint32_t>(*(ptr + stride * static_cast<uint32_t>(res))) > elem)
+                    {
+                        res = static_cast<uint32_t>(i);
+                    }
+                    break;
+                case ReductionOperation::ARG_IDX_MAX:
+                    if(static_cast<uint32_t>(*(ptr + stride * static_cast<uint32_t>(res))) < elem)
+                    {
+                        res = static_cast<uint32_t>(i);
+                    }
+                    break;
+                case ReductionOperation::SUM_SQUARE:
+                    int_res += elem * elem;
+                    break;
+                case ReductionOperation::MEAN_SUM:
+                case ReductionOperation::SUM:
+                    int_res += elem;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Operation not supported");
+            }
         }
         if(op == ReductionOperation::MEAN_SUM && reduce_elements > 0)
         {
@@ -63,7 +87,30 @@ T reduce_operation(T *ptr, int reduce_elements, ReductionOperation op, int strid
         for(int i = 0; i < reduce_elements; ++i)
         {
             auto elem = *(ptr + stride * i);
-            res += (op == ReductionOperation::SUM_SQUARE) ? elem * elem : elem;
+            switch(op)
+            {
+                case ReductionOperation::ARG_IDX_MIN:
+                    if(*(ptr + stride * static_cast<uint32_t>(res)) > elem)
+                    {
+                        res = static_cast<uint32_t>(i);
+                    }
+                    break;
+                case ReductionOperation::ARG_IDX_MAX:
+                    if(*(ptr + stride * static_cast<uint32_t>(res)) < elem)
+                    {
+                        res = static_cast<uint32_t>(i);
+                    }
+                    break;
+                case ReductionOperation::SUM_SQUARE:
+                    res += elem * elem;
+                    break;
+                case ReductionOperation::MEAN_SUM:
+                case ReductionOperation::SUM:
+                    res += elem;
+                    break;
+                default:
+                    ARM_COMPUTE_ERROR("Operation not supported");
+            }
         }
         if(op == ReductionOperation::MEAN_SUM && reduce_elements > 0)
         {
@@ -79,7 +126,9 @@ template <typename T>
 SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op)
 {
     // Create reference
-    SimpleTensor<T>    dst{ dst_shape, src.data_type(), 1, src.quantization_info() };
+    const bool         is_arg_min_max   = (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX);
+    DataType           output_data_type = is_arg_min_max ? DataType::U32 : src.data_type();
+    SimpleTensor<T>    dst{ dst_shape, output_data_type, 1, src.quantization_info() };
     const unsigned int src_width    = src.shape().x();
     const unsigned int src_height   = src.shape().y();
     const unsigned int src_depth    = src.shape().z();
@@ -94,8 +143,14 @@ SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShap
             for(unsigned int du = 0; du < upper_dims; ++du)
             {
                 const T *src_row_ptr = src.data() + du * reduce_elems;
-                auto     res         = reduce_operation(src_row_ptr, reduce_elems, op, 1);
-                dst[du]              = res;
+                if(is_arg_min_max)
+                {
+                    dst[du] = reduce_operation<T, uint32_t>(src_row_ptr, reduce_elems, op, 1);
+                }
+                else
+                {
+                    dst[du] = reduce_operation<T, T>(src_row_ptr, reduce_elems, op, 1);
+                }
             }
         }
         break;
@@ -109,8 +164,15 @@ SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShap
                     const int in_offset   = du * src_height * src_width + x;
                     const int out_offset  = du * src_width + x;
                     const T *src_row_ptr = src.data() + in_offset;
-                    auto      res         = reduce_operation(src_row_ptr, reduce_elems, op, src_width);
-                    dst[out_offset]       = res;
+
+                    if(is_arg_min_max)
+                    {
+                        dst[out_offset] = reduce_operation<T, uint32_t>(src_row_ptr, reduce_elems, op, src_width);
+                    }
+                    else
+                    {
+                        dst[out_offset] = reduce_operation<T, T>(src_row_ptr, reduce_elems, op, src_width);
+                    }
                 }
             }
         }
@@ -127,8 +189,15 @@ SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShap
                         const int in_offset   = du * src_depth * src_height * src_width + y * src_width + x;
                         const int out_offset  = du * src_width * src_height + y * src_width + x;
                         const T *src_row_ptr = src.data() + in_offset;
-                        auto      res         = reduce_operation(src_row_ptr, reduce_elems, op, src_height * src_width);
-                        dst[out_offset]       = res;
+
+                        if(is_arg_min_max)
+                        {
+                            dst[out_offset] = reduce_operation<T, uint32_t>(src_row_ptr, reduce_elems, op, src_height * src_width);
+                        }
+                        else
+                        {
+                            dst[out_offset] = reduce_operation<T, T>(src_row_ptr, reduce_elems, op, src_height * src_width);
+                        }
                     }
                 }
             }
@@ -148,8 +217,14 @@ SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShap
                             const int in_offset   = du * src_batch * src_depth * src_height * src_width + z * src_width * src_height + y * src_width + x;
                             const int out_offset  = du * src_depth * src_height * src_width + z * src_width * src_height + y * src_width + x;
                             const T *src_row_ptr = src.data() + in_offset;
-                            auto      res         = reduce_operation(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth);
-                            dst[out_offset]       = res;
+                            if(is_arg_min_max)
+                            {
+                                dst[out_offset] = reduce_operation<T, uint32_t>(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth);
+                            }
+                            else
+                            {
+                                dst[out_offset] = reduce_operation<T, T>(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth);
+                            }
                         }
                     }
                 }