COMPMID-3387: Support memory injection in CLActivationLayer

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I31f9620607b372fc3340c71e748a5ea177d9da62
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3520
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 95 insertions, 51 deletions

diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp
index d40e9a15be..e030177549 100644
--- a/src/core/CL/kernels/CLActivationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp
@@ -34,6 +34,7 @@
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Window.h"
 #include "arm_compute/core/utils/helpers/float_ops.h"
+#include "arm_compute/core/utils/misc/Cast.h"
 #include "support/StringSupport.h"
 
 #include <cmath>
@@ -116,16 +117,11 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLActivationLayerKernel::CLActivationLayerKernel()
-    : _input(nullptr), _output(nullptr), _run_in_place(false)
+    : _run_in_place(false)
 {
 }
 
-void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
-{
-    configure(CLKernelLibrary::get().get_compile_context(), input, output, act_info);
-}
-
-void CLActivationLayerKernel::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
+void CLActivationLayerKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input, ITensorInfo *output, ActivationLayerInfo act_info)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
 
@@ -134,14 +130,13 @@ void CLActivationLayerKernel::configure(const CLCompileContext &compile_context,
     if(output != nullptr)
     {
         // Output auto inizialitation if not yet initialized
-        auto_init_if_empty(*output->info(),
-                           *input->info()->clone());
+        auto_init_if_empty(*output, *input->clone());
     }
 
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (output != nullptr) ? output->info() : nullptr, act_info));
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, (output != nullptr) ? output : nullptr, act_info));
 
-    const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
-    const DataType     dt                                = input->info()->data_type();
+    const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
+    const DataType     dt                                = input->data_type();
     float              a_const                           = act_info.a();
     float              b_const                           = act_info.b();
 
@@ -163,7 +158,7 @@ void CLActivationLayerKernel::configure(const CLCompileContext &compile_context,
     // Set quantization info build options
     if(is_quantized)
     {
-        const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
+        const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
 
         if(!perform_activation_in_float)
         {
@@ -214,7 +209,7 @@ void CLActivationLayerKernel::configure(const CLCompileContext &compile_context,
         // Set scale and offset of the input and output if they have different quantization info
         if(output != nullptr)
         {
-            const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
+            const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
 
             if(iq_info != oq_info)
             {
@@ -233,12 +228,8 @@ void CLActivationLayerKernel::configure(const CLCompileContext &compile_context,
     // Create kernel
     _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
 
-    // Make sure _kernel is initialized before calling the parent's configure
-    _input  = input;
-    _output = output;
-
     // Configure kernel window
-    auto win_config = validate_and_configure_window(input->info(), (_run_in_place) ? nullptr : output->info());
+    auto win_config = validate_and_configure_window(input, (_run_in_place) ? nullptr : output);
     ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
     ICLKernel::configure_internal(win_config.second);
 
@@ -246,9 +237,9 @@ void CLActivationLayerKernel::configure(const CLCompileContext &compile_context,
     _config_id = "activation_layer_";
     _config_id += lower_string(string_from_data_type(dt));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(0));
+    _config_id += support::cpp11::to_string(input->dimension(0));
     _config_id += "_";
-    _config_id += support::cpp11::to_string(input->info()->dimension(1));
+    _config_id += support::cpp11::to_string(input->dimension(1));
 }
 
 Status CLActivationLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
@@ -260,21 +251,25 @@ Status CLActivationLayerKernel::validate(const ITensorInfo *input, const ITensor
     return Status{};
 }
 
-void CLActivationLayerKernel::run(const Window &window, cl::CommandQueue &queue)
+void CLActivationLayerKernel::run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs, const Window &window, cl::CommandQueue &queue)
 {
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
 
+    const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(inputs.at(TensorType::ACL_SRC));
+    auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(outputs.at(TensorType::ACL_DST));
+    ARM_COMPUTE_ERROR_ON(_run_in_place && src != dst);
+
     Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
     Window slice     = collapsed.first_slice_window_3D();
 
     do
     {
         unsigned int idx = 0;
-        add_3D_tensor_argument(idx, _input, slice);
+        add_3D_tensor_argument(idx, src, slice);
         if(!_run_in_place)
         {
-            add_3D_tensor_argument(idx, _output, slice);
+            add_3D_tensor_argument(idx, dst, slice);
         }
         enqueue(queue, *this, slice, lws_hint());
     }
diff --git a/src/core/CL/kernels/CLReshapeLayerKernel.cpp b/src/core/CL/kernels/CLReshapeLayerKernel.cpp
index 97fde8645e..d486b06c8e 100644
--- a/src/core/CL/kernels/CLReshapeLayerKernel.cpp
+++ b/src/core/CL/kernels/CLReshapeLayerKernel.cpp
@@ -34,6 +34,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Utils.h"
 #include "arm_compute/core/Window.h"
+#include "arm_compute/core/utils/misc/Cast.h"
 
 #include <string>
 
@@ -107,8 +108,8 @@ void CLReshapeLayerKernel::run_op(const InputTensorMap &inputs, const OutputTens
     Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
     Window slice            = window_collapsed.first_slice_window_3D();
 
-    const auto src = dynamic_cast<const ICLTensor *>(inputs.at(TensorType::ACL_SRC));
-    auto       dst = dynamic_cast<ICLTensor *>(outputs.at(TensorType::ACL_DST));
+    const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(inputs.at(TensorType::ACL_SRC));
+    auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(outputs.at(TensorType::ACL_DST));
 
     // Set inputs
     unsigned int idx = 0;
diff --git a/src/runtime/CL/functions/CLActivationLayer.cpp b/src/runtime/CL/functions/CLActivationLayer.cpp
index 989603a9df..d6b80200cd 100644
--- a/src/runtime/CL/functions/CLActivationLayer.cpp
+++ b/src/runtime/CL/functions/CLActivationLayer.cpp
@@ -23,6 +23,7 @@
  */
 #include "arm_compute/runtime/CL/functions/CLActivationLayer.h"
 
+#include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/kernels/CLActivationLayerKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLRuntimeContext.h"
@@ -30,11 +31,46 @@
 
 namespace arm_compute
 {
+namespace experimental
+{
+void CLActivationLayer::configure(const CLCompileContext &compile_context, ITensorInfo *input, ITensorInfo *output, ActivationLayerInfo act_info)
+{
+    auto k = arm_compute::support::cpp14::make_unique<CLActivationLayerKernel>();
+    k->configure(compile_context, input, output, act_info);
+    _kernel = std::move(k);
+}
+
+Status CLActivationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
+{
+    return CLActivationLayerKernel::validate(input, output, act_info);
+}
+
+MemoryRequirements CLActivationLayer::workspace() const
+{
+    return MemoryRequirements{};
+}
+} // namespace experimental
+
+struct CLActivationLayer::Impl
+{
+    const ICLTensor                                 *src{ nullptr };
+    ICLTensor                                       *dst{ nullptr };
+    CLRuntimeContext                                *ctx{ nullptr };
+    std::unique_ptr<experimental::CLActivationLayer> op{ nullptr };
+};
+
 CLActivationLayer::CLActivationLayer(CLRuntimeContext *ctx)
-    : ICLSimpleFunction(ctx)
+    : _impl(support::cpp14::make_unique<Impl>())
 {
+    _impl->ctx = ctx;
 }
 
+CLActivationLayer::CLActivationLayer(CLActivationLayer &&) = default;
+
+CLActivationLayer &CLActivationLayer::operator=(CLActivationLayer &&) = default;
+
+CLActivationLayer::~CLActivationLayer() = default;
+
 void CLActivationLayer::configure(ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
 {
     configure(CLKernelLibrary::get().get_compile_context(), input, output, act_info);
@@ -42,13 +78,25 @@ void CLActivationLayer::configure(ICLTensor *input, ICLTensor *output, Activatio
 
 void CLActivationLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, ActivationLayerInfo act_info)
 {
-    auto k = arm_compute::support::cpp14::make_unique<CLActivationLayerKernel>();
-    k->configure(compile_context, input, output, act_info);
-    _kernel = std::move(k);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+
+    _impl->src = input;
+    _impl->dst = output == nullptr ? input : output;
+
+    _impl->op = arm_compute::support::cpp14::make_unique<experimental::CLActivationLayer>();
+    _impl->op->configure(compile_context, _impl->src->info(), _impl->dst->info(), act_info);
 }
 
 Status CLActivationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info)
 {
-    return CLActivationLayerKernel::validate(input, output, act_info);
+    return experimental::CLActivationLayer::validate(input, output, act_info);
+}
+
+void CLActivationLayer::run()
+{
+    const InputTensorMap  src{ { TensorType::ACL_SRC, _impl->src } };
+    const OutputTensorMap dst{ { TensorType::ACL_DST, _impl->dst } };
+
+    _impl->op->run(src, dst, {});
 }
 } // namespace arm_compute
diff --git a/src/runtime/CL/functions/CLLSTMLayer.cpp b/src/runtime/CL/functions/CLLSTMLayer.cpp
index 56f22e2fe0..e63a9cceb0 100644
--- a/src/runtime/CL/functions/CLLSTMLayer.cpp
+++ b/src/runtime/CL/functions/CLLSTMLayer.cpp
@@ -499,7 +499,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
                                                                               RoundingPolicy::TO_NEAREST_EVEN));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&forget_gate, forget_gate_bias, &forget_gate, ConvertPolicy::SATURATE));
     }
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&forget_gate, &forget_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate input gate
     if(!lstm_params.has_cifg_opt())
@@ -534,7 +534,7 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
             ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&input_gate, lstm_params.input_layer_norm_weights(), &input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
             ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&input_gate, lstm_params.input_gate_bias(), &input_gate, ConvertPolicy::SATURATE));
         }
-        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&input_gate, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
     }
     else
     {
@@ -552,14 +552,14 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
                                                                               RoundingPolicy::TO_NEAREST_EVEN));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, cell_bias, &cell_state_tmp, ConvertPolicy::SATURATE));
     }
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, nullptr, activation_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, activation_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &input_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &forget_gate, &cell_state_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&cell_state_tmp, &cell_state_tmp, &cell_state_tmp, ConvertPolicy::SATURATE));
     if(cell_threshold != 0.f)
     {
-        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold,
-                                                                                                                    cell_threshold)));
+        ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold,
+                                                                                                              cell_threshold)));
     }
 
     std::vector<const ITensorInfo *> in_out_weights;
@@ -584,18 +584,18 @@ Status CLLSTMLayer::validate(const ITensorInfo *input,
                                                                               RoundingPolicy::TO_NEAREST_EVEN));
         ARM_COMPUTE_RETURN_ON_ERROR(CLArithmeticAddition::validate(&output_gate_tmp, output_gate_bias, &output_gate_tmp, ConvertPolicy::SATURATE));
     }
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&output_gate_tmp, nullptr, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)));
 
     // Validate output state
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&cell_state_tmp, &cell_state_tmp, activation_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLPixelWiseMultiplicationKernel::validate(&cell_state_tmp, &output_gate_tmp, &output_gate_tmp, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN));
     if(lstm_params.has_projection())
     {
         ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(&output_gate_tmp, lstm_params.projection_weights(), lstm_params.projection_bias(), output_state_out));
         if(projection_threshold != 0.f)
         {
-            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(output_state_out, output_state_out,
-                                                                          ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
+            ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output_state_out, output_state_out,
+                                                                    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)));
         }
     }
 
@@ -638,7 +638,7 @@ void CLLSTMLayer::run()
         CLScheduler::get().enqueue(_pixelwise_mul_forget_gate_coeff);
         CLScheduler::get().enqueue(_accum_forget_gate_bias);
     }
-    CLScheduler::get().enqueue(_activation_forget_gate);
+    _activation_forget_gate.run();
 
     if(_run_cifg_opt)
     {
@@ -661,7 +661,7 @@ void CLLSTMLayer::run()
             CLScheduler::get().enqueue(_pixelwise_mul_input_gate_coeff);
             CLScheduler::get().enqueue(_accum_input_gate_bias);
         }
-        CLScheduler::get().enqueue(_activation_input_gate);
+        _activation_input_gate.run();
     }
 
     _fully_connected_cell_state.run();
@@ -674,14 +674,14 @@ void CLLSTMLayer::run()
         CLScheduler::get().enqueue(_pixelwise_mul_cell_gate_coeff);
         CLScheduler::get().enqueue(_accum_cell_gate_bias);
     }
-    CLScheduler::get().enqueue(_activation_cell_state);
+    _activation_cell_state.run();
     CLScheduler::get().enqueue(_pixelwise_mul_cell_state1);
     CLScheduler::get().enqueue(_pixelwise_mul_cell_state2);
     CLScheduler::get().enqueue(_accum_cell_state2);
 
     if(_perform_cell_clipping)
     {
-        CLScheduler::get().enqueue(_cell_clip);
+        _cell_clip.run();
     }
 
     _fully_connected_output.run();
@@ -697,9 +697,9 @@ void CLLSTMLayer::run()
         CLScheduler::get().enqueue(_pixelwise_mul_output_gate_coeff);
         CLScheduler::get().enqueue(_accum_output_gate_bias);
     }
-    CLScheduler::get().enqueue(_activation_output);
+    _activation_output.run();
 
-    CLScheduler::get().enqueue(_activation_output_state);
+    _activation_output_state.run();
     CLScheduler::get().enqueue(_pixelwise_mul_output_state2);
 
     if(_has_projection_weights)
@@ -707,7 +707,7 @@ void CLLSTMLayer::run()
         _fully_connected_output_state.run();
         if(_perform_projection_clipping)
         {
-            CLScheduler::get().enqueue(_projection_clip);
+            _projection_clip.run();
         }
     }
 
diff --git a/src/runtime/CL/functions/CLRNNLayer.cpp b/src/runtime/CL/functions/CLRNNLayer.cpp
index 57b8d70089..075f4a4ebd 100644
--- a/src/runtime/CL/functions/CLRNNLayer.cpp
+++ b/src/runtime/CL/functions/CLRNNLayer.cpp
@@ -35,7 +35,7 @@ using namespace arm_compute;
 using namespace arm_compute::misc::shape_calculator;
 
 CLRNNLayer::CLRNNLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_kernel(), _activation_kernel(), _fully_connected_kernel(), _copy_kernel(), _fully_connected_out(), _gemm_output(), _add_output(),
+    : _memory_group(std::move(memory_manager)), _gemm_state_f(), _add_kernel(), _activation(), _fully_connected_kernel(), _copy_kernel(), _fully_connected_out(), _gemm_output(), _add_output(),
       _is_prepared(false)
 {
 }
@@ -60,7 +60,7 @@ Status CLRNNLayer::validate(const ITensorInfo *input, const ITensorInfo *weights
     ARM_COMPUTE_RETURN_ON_ERROR(CLFullyConnectedLayer::validate(input, weights, bias, &shape_info));
     ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(hidden_state, recurrent_weights, nullptr, &shape_info, 1.f, 0.f));
     ARM_COMPUTE_RETURN_ON_ERROR(CLSaturatedArithmeticOperationKernel::validate(ArithmeticOperation::ADD, &shape_info, &shape_info, &shape_info, ConvertPolicy::SATURATE));
-    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayerKernel::validate(&shape_info, &shape_info, info));
+    ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(&shape_info, &shape_info, info));
 
     return Status{};
 }
@@ -101,7 +101,7 @@ void CLRNNLayer::configure(const CLCompileContext &compile_context, const ICLTen
     _fully_connected_out.allocator()->allocate();
     _gemm_output.allocator()->allocate();
 
-    _activation_kernel.configure(compile_context, &_add_output, hidden_state, info);
+    _activation.configure(compile_context, &_add_output, hidden_state, info);
     _add_output.allocator()->allocate();
 
     _copy_kernel.configure(compile_context, hidden_state, output);
@@ -116,7 +116,7 @@ void CLRNNLayer::run()
     _fully_connected_kernel.run();
     _gemm_state_f.run();
     CLScheduler::get().enqueue(_add_kernel);
-    CLScheduler::get().enqueue(_activation_kernel);
+    _activation.run();
 
     // copy hidden out to output
     CLScheduler::get().enqueue(_copy_kernel);