Add CLAMP operator to Dynamic Fusion interface

Add the CLAMP activation function for GPU backend with generic activation Component and TemplateWriter modules.

CLAMP is internally implemented as LU_BOUNDED_RELU activation function with the alpha and beta variables swapped.

We do NOT consider in-place computation cases in this patch.

* CLAMP operator for GPU backend.
* Activation Component and TemplateWriter for CL backend.
* TemplateWriter generates tiled kernel code.
* Supported data types: F16, F32.
* Validation tests for CLAMP operation.

Resolves: COMPMID-5519

Change-Id: Ieb097d6b1e6a7ed2b882518e88314454efb402f6
Signed-off-by: Jakub Sujak <jakub.sujak@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8762
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 378 insertions, 0 deletions

diff --git a/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp
new file mode 100644
index 0000000000..947201ff97
--- /dev/null
+++ b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp
@@ -0,0 +1,189 @@
+/*
+ * Copyright (c) 2022 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/dynamic_fusion/sketch/attributes/ClampAttributes.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuClamp.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr float                    epsilon = 1e-6f;
+constexpr AbsoluteTolerance<float> tolerance(epsilon);
+} // namespace
+
+TEST_SUITE(CL)
+TEST_SUITE(DYNAMIC_FUSION)
+TEST_SUITE(CLAMP)
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
+                framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16),
+                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Mismatching data types
+                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Mismatching shapes
+                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),    // Minimum value larger than maximum value
+                                                    }),
+                framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16),
+                                                        TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16),
+                                                        TensorInfo(TensorShape(30U, 11U, 2U), 1, DataType::F32),
+                                                        TensorInfo(TensorShape(30U, 11U, 2U), 1, DataType::F32),
+                                                    })),
+                framework::dataset::make("MinVal", { 0.2f,
+                                                     1.5f,
+                                                     0.1f,
+                                                     3.0f,
+                                                     9.0f,
+                                                    })),
+                framework::dataset::make("MaxVal", { 0.5f,
+                                                     2.0f,
+                                                     1.0f,
+                                                     4.0f,
+                                                     1.0f,
+                                                    })),
+                framework::dataset::make("Expected", { true, true, false, false, false })),
+                input_info, output_info, min_val, max_val, expected)
+{
+    // Create a new workload sketch
+    CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+    GpuWorkloadContext gpu_ctx{ &cl_compile_ctx };
+    GpuWorkloadSketch sketch{ &gpu_ctx };
+
+    // Fuse Clamp
+    const TensorInfo src_info = sketch.create_tensor_info(input_info);
+    const TensorInfo dst_info = sketch.create_tensor_info(output_info);
+
+    ClampAttributes attributes {};
+    attributes.min_val(min_val)
+              .max_val(max_val);
+
+    const bool res = static_cast<bool>(GpuClamp::validate_op(sketch, &src_info, &dst_info, attributes));
+    ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using DynamicFusionClampOpFixture = DynamicFusionClampValidationFixture<CLTensor, CLAccessor, GpuClamp, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmallOneOp,
+                       DynamicFusionClampOpFixture<half>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.1f).max_val(0.6f) })),
+                                       framework::dataset::make("Fuse", { false })),
+                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp,
+                       DynamicFusionClampOpFixture<half>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::Small5dShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.1f).max_val(0.6f) })),
+                                       framework::dataset::make("Fuse", { false })),
+                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmallTwoOps,
+                       DynamicFusionClampOpFixture<half>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.2f).max_val(0.4f) })),
+                                       framework::dataset::make("Fuse", { true })),
+                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+TEST_SUITE_END() // FP16
+
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmallOneOp,
+                       DynamicFusionClampOpFixture<float>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.3f).max_val(0.7f) })),
+                                       framework::dataset::make("Fuse", { false })),
+                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp,
+                       DynamicFusionClampOpFixture<float>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::Small5dShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.3f).max_val(0.7f) })),
+                                       framework::dataset::make("Fuse", { false })),
+                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmallTwoOps,
+                       DynamicFusionClampOpFixture<float>,
+                       framework::DatasetMode::ALL,
+                       combine(combine(combine(datasets::SmallShapes(),
+                                               framework::dataset::make("ClampAttributes", { ClampAttributes().min_val(0.1f).max_val(0.9f) })),
+                                       framework::dataset::make("Fuse", { true })),
+                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance);
+}
+
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+TEST_SUITE_END() // CLAMP
+TEST_SUITE_END() // DYNAMIC_FUSION
+TEST_SUITE_END() // CL
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h
new file mode 100644
index 0000000000..dbac29fd22
--- /dev/null
+++ b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h
@@ -0,0 +1,189 @@
+/*
+ * Copyright (c) 2022 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 TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE
+#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE
+
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
+
+#include "tests/framework/Fixture.h"
+#include "tests/validation/reference/ActivationLayer.h"
+
+using namespace arm_compute::experimental::dynamic_fusion;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionClampValidationFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, ClampAttributes attributes, bool fuse, DataType data_type)
+    {
+        // CLAMP is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped.
+        ActivationLayerInfo act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() };
+
+        _fuse       = fuse;
+        _attributes = attributes;
+        _data_type  = data_type;
+        _target     = compute_target(shape, attributes);
+        _reference  = compute_reference(shape, act_info);
+    }
+
+protected:
+    std::vector<T> get_boundary_values(T min, T max)
+    {
+        // This function will return a vector filled with the following values that can
+        // represent two partitions derived from equivalent partitioning.
+        // * Lower partition: min, min + delta, lower quarter (nominal), center - delta
+        // * Upper partition: center, center + delta, upper quarter (nominal), max - delta, max
+        const auto delta         = is_data_type_float(_data_type) ? T(0.1f) : T(1);
+        const auto center_value  = (min + max) / 2;
+        const auto lower_quarter = (min + center_value) / 2;
+        const auto upper_quarter = (center_value + max) / 2;
+
+        std::vector<T> boundary_values{};
+
+        // To ensure all the inserted values are within the given range after subtracing/adding delta
+        auto insert_values = [&boundary_values, &min, &max](const std::initializer_list<T> &new_values)
+        {
+            for(auto &v : new_values)
+            {
+                if(v >= min && v <= max)
+                {
+                    boundary_values.emplace_back(v);
+                }
+            }
+        };
+
+        insert_values({ min, static_cast<T>(min + delta), static_cast<T>(lower_quarter), static_cast<T>(center_value - delta) });                               // lower partition
+        insert_values({ static_cast<T>(center_value), static_cast<T>(center_value + delta), static_cast<T>(upper_quarter), static_cast<T>(max - delta), max }); // upper partition
+
+        return boundary_values;
+    }
+
+    template <typename U>
+    void fill(U &&tensor)
+    {
+        float min_bound                = 0;
+        float max_bound                = 0;
+        std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<T>(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, _data_type);
+        library->fill_static_values(tensor, get_boundary_values(static_cast<T>(min_bound), static_cast<T>(max_bound)));
+    }
+
+    TensorType compute_target(const TensorShape &shape, ClampAttributes attributes)
+    {
+        // Create a new workload sketch
+        CLCompileContext   cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+        GpuWorkloadContext gpu_ctx{ &cl_compile_ctx };
+        GpuWorkloadSketch  sketch{ &gpu_ctx };
+
+        // Create sketch tensors
+        TensorInfo src_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type));
+        TensorInfo dst_0_info = sketch.create_tensor_info(TensorInfo(shape, 1, _data_type));
+        TensorInfo dst_1_info;
+
+        FunctionType::create_op(sketch, &src_info, &dst_0_info, attributes);
+        if(_fuse)
+        {
+            dst_1_info = sketch.create_tensor_info(shape, 1, _data_type);
+            FunctionType::create_op(sketch, &dst_0_info, &dst_1_info, attributes);
+        }
+
+        // Configure runtime
+        ClWorkloadRuntime runtime;
+        runtime.configure(sketch);
+
+        // Construct user tensors
+        TensorType t_src{};
+        TensorType t_dst_0{};
+        TensorType t_dst_1{};
+
+        // Initialize user tensors
+        t_src.allocator()->init(src_info);
+        t_dst_0.allocator()->init(dst_0_info);
+        if(_fuse)
+        {
+            t_dst_1.allocator()->init(dst_1_info);
+        }
+
+        // Allocate and fill user tensors
+        t_src.allocator()->allocate();
+        t_dst_0.allocator()->allocate();
+        if(_fuse)
+        {
+            t_dst_1.allocator()->allocate();
+        }
+
+        fill(AccessorType(t_src));
+
+        // Run runtime
+        if(_fuse)
+        {
+            runtime.run({ &t_src, &t_dst_1 });
+        }
+        else
+        {
+            runtime.run({ &t_src, &t_dst_0 });
+        }
+
+        if(_fuse)
+        {
+            return t_dst_1;
+        }
+
+        return t_dst_0;
+    }
+
+    SimpleTensor<T> compute_reference(const TensorShape &shape, ActivationLayerInfo act_info)
+    {
+        // Create reference
+        SimpleTensor<T> src{ shape, _data_type, 1, _quantization_info };
+
+        // Fill reference
+        fill(src);
+
+        auto dst = reference::activation_layer<T>(src, act_info, _quantization_info);
+        return dst;
+    }
+
+protected:
+    QuantizationInfo _quantization_info{};
+    ClampAttributes  _attributes{};
+    bool             _fuse{ false };
+    DataType         _data_type{};
+    TensorType       _target{};
+    SimpleTensor<T>  _reference{};
+};
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE */