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>

1 files changed, 158 insertions, 0 deletions

diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp
new file mode 100644
index 0000000000..ffef6115d6
--- /dev/null
+++ b/src/dynamic_fusion/sketch/gpu/operators/GpuClamp.cpp
@@ -0,0 +1,158 @@
+/*
+ * 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/dynamic_fusion/sketch/gpu/operators/GpuClamp.h"
+
+#include "arm_compute/core/experimental/Types.h"
+
+#include "src/core/helpers/AutoConfiguration.h"
+#include "src/dynamic_fusion/sketch/ArgumentPack.h"
+#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
+#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.h"
+#include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.h"
+
+#include "src/common/utils/Log.h"
+
+namespace arm_compute
+{
+namespace experimental
+{
+namespace dynamic_fusion
+{
+namespace
+{
+constexpr GpuOperatorType operator_type = GpuOperatorType::Simple;
+} // namespace
+
+Status GpuClamp::is_supported_op(const GpuWorkloadContext &context,
+                                 const ITensorInfo        *src,
+                                 const ITensorInfo        *dst,
+                                 const ClampAttributes    &attributes)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(attributes.max_val() < attributes.min_val(), "Maximum clamp value cannot be lower than minimum value");
+
+    // Auto initialize dst tensor info
+    TensorInfo dst_info_to_validate = *dst;
+    auto_init_if_empty(dst_info_to_validate, *src->clone());
+
+    // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped
+    const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() };
+
+    // Check components
+    if(context.gpu_language() == GpuLanguage::OpenCL)
+    {
+        // Validate Activation Component
+        const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
+
+        ArgumentPack<ITensorInfo> arguments;
+        arguments.add_const_tensor(ACL_SRC, src);
+        arguments.add_const_tensor(ACL_DST, &dst_info_to_validate);
+        ARM_COMPUTE_RETURN_ON_ERROR(ClComponentActivation::validate(properties, arguments, act_info));
+    }
+    else
+    {
+        ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
+    }
+    return Status{};
+}
+
+Status GpuClamp::validate_op(const GpuWorkloadSketch &sketch,
+                             const ITensorInfo       *src,
+                             const ITensorInfo       *dst,
+                             const ClampAttributes   &attributes)
+{
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
+
+    // Check if tensors have valid id, i.e. they are created from a sketch
+    ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !dst->has_valid_id());
+
+    // Auto initialize dst tensor info
+    TensorInfo dst_info_to_validate = *dst;
+    auto_init_if_empty(dst_info_to_validate, *src->clone());
+
+    // Perform fusion test to check if the operator meets fusion constraints
+    ArgumentPack<ITensorInfo> tensors;
+    tensors.add_const_tensor(ACL_SRC, src);
+    tensors.add_const_tensor(ACL_DST, &dst_info_to_validate);
+    const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
+                                    "Operator fusion test failed. This operator cannot be fused into the workload");
+
+    // Check if configuration is supported
+    return is_supported_op(*sketch.gpu_context(), src, &dst_info_to_validate, attributes);
+}
+
+void GpuClamp::create_op(GpuWorkloadSketch     &sketch,
+                         ITensorInfo           *src,
+                         ITensorInfo           *dst,
+                         const ClampAttributes &attributes)
+{
+    // Assert validation
+    ARM_COMPUTE_ERROR_THROW_ON(GpuClamp::validate_op(sketch, src, dst, attributes));
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    ARM_COMPUTE_LOG_PARAMS(src, dst, attributes);
+
+    // Auto initialize dst tensor
+    auto_init_if_empty(*dst, *src->clone());
+
+    // Translate into components and add to component graph
+    GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph();
+
+    // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped
+    const ClComponentActivation::Attributes act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() };
+
+    const auto *const sketch_ctx = sketch.implementation().context();
+
+    if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
+    {
+        // Add Activation Component
+        auto properties = IGpuKernelComponent::Properties();
+        properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
+
+        ArgumentPack<ITensorInfo> arguments;
+        arguments.add_const_tensor(ACL_SRC, src);
+        arguments.add_const_tensor(ACL_DST, dst);
+        comp_graph.add_new_component<ClComponentActivation>(properties, arguments, act_info);
+    }
+    else
+    {
+        ARM_COMPUTE_ERROR("Unimplemented Gpu language");
+    }
+
+    // Set up fusion test by adding to the Operator Group
+    // Note this has to be performed after all the components have been successfully added to the component graph
+
+    // Pack tensor infos
+    ArgumentPack<ITensorInfo> tensors;
+    tensors.add_const_tensor(ACL_SRC, src);
+    tensors.add_const_tensor(ACL_DST, dst);
+
+    const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
+    sketch.implementation().operator_group().add_operator(op);
+}
+
+} // namespace dynamic_fusion
+} // namespace experimental
+} // namespace arm_compute