runtime/CL: add ReduceSum layertizen_devel

Change-Id: I6a16a2264eebcd7286eca54adf152105247b4396
Signed-off-by: Inki Dae <inki.dae@samsung.com>

4 files changed, 271 insertions, 1 deletions

diff --git a/Android.bp b/Android.bp
index 59fb270d0..fe51b7082 100644
--- a/Android.bp
+++ b/Android.bp
@@ -542,6 +542,7 @@ cc_library_static {
         "src/runtime/CL/functions/CLROIPoolingLayer.cpp",
         "src/runtime/CL/functions/CLRange.cpp",
         "src/runtime/CL/functions/CLReduceMean.cpp",
+        "src/runtime/CL/functions/CLReduceSum.cpp",
         "src/runtime/CL/functions/CLReductionOperation.cpp",
         "src/runtime/CL/functions/CLRemap.cpp",
         "src/runtime/CL/functions/CLReorgLayer.cpp",
@@ -832,4 +833,4 @@ cc_library_static {
         },
     },
     rtti: true,
-}
\ No newline at end of file
+}
diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index 007a40c65..ebb258ae1 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -123,6 +123,7 @@
 #include "arm_compute/runtime/CL/functions/CLROIPoolingLayer.h"
 #include "arm_compute/runtime/CL/functions/CLRange.h"
 #include "arm_compute/runtime/CL/functions/CLReduceMean.h"
+#include "arm_compute/runtime/CL/functions/CLReduceSum.h"
 #include "arm_compute/runtime/CL/functions/CLReductionOperation.h"
 #include "arm_compute/runtime/CL/functions/CLRemap.h"
 #include "arm_compute/runtime/CL/functions/CLReorgLayer.h"
diff --git a/arm_compute/runtime/CL/functions/CLReduceSum.h b/arm_compute/runtime/CL/functions/CLReduceSum.h
new file mode 100644
index 000000000..9506dfe9e
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLReduceSum.h
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) 2018-2020 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_CL_REDUCE_SUM_H
+#define ARM_COMPUTE_CL_REDUCE_SUM_H
+
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+#include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h"
+#include "arm_compute/runtime/CL/functions/CLReductionOperation.h"
+#include "arm_compute/runtime/CL/functions/CLReshapeLayer.h"
+#include "arm_compute/runtime/IMemoryManager.h"
+
+namespace arm_compute
+{
+// Forward Declarations
+class ICLTensor;
+
+/** Basic function to perform reduce operation */
+class CLReduceSum : public IFunction
+{
+public:
+    /** Default constructor */
+    CLReduceSum(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Configure kernel
+     *
+     * @note Supported tensor rank: up to 4
+     *
+     * @param[in]  input          Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     * @param[in]  reduction_axis Reduction axis vector.
+     * @param[in]  keep_dims      If positive, retains reduced dimensions with length 1.
+     * @param[out] output         Destination tensor. Data type supported: Same as @p input
+     */
+    void configure(ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output);
+    /** Configure kernel
+     *
+     * @note Supported tensor rank: up to 4
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  input           Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     * @param[in]  reduction_axis  Reduction axis vector.
+     * @param[in]  keep_dims       If positive, retains reduced dimensions with length 1.
+     * @param[out] output          Destination tensor. Data type supported: Same as @p input
+     */
+    void configure(const CLCompileContext &compile_context, ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output);
+
+    /** Static function to check if given info will lead to a valid configuration of @ref CLReduceSum
+     *
+     * @param[in] input          Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32
+     * @param[in] reduction_axis Reduction axis vector.
+     * @param[in] keep_dims      If positive, retains reduced dimensions with length 1.
+     * @param[in] output         Destination tensor. Data type supported: Same as @p input
+     *
+     * @return A status
+     */
+    static Status validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    MemoryGroup                       _memory_group;
+    std::vector<CLReductionOperation> _reduction_kernels;
+    std::vector<CLTensor>             _reduced_outs;
+    CLReshapeLayer                    _reshape;
+    int                               _reduction_ops;
+    bool                              _keep_dims;
+};
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_REDUCE_SUM_H */
diff --git a/src/runtime/CL/functions/CLReduceSum.cpp b/src/runtime/CL/functions/CLReduceSum.cpp
new file mode 100644
index 000000000..daac4c48b
--- /dev/null
+++ b/src/runtime/CL/functions/CLReduceSum.cpp
@@ -0,0 +1,179 @@
+/*
+ * Copyright (c) 2018-2020 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/runtime/CL/functions/CLReduceSum.h"
+
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/CL/kernels/CLReductionOperationKernel.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
+{
+    ARM_COMPUTE_UNUSED(keep_dims);
+    ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
+    ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() < 1);
+    ARM_COMPUTE_RETURN_ERROR_ON(reduction_axis.num_dimensions() > input->num_dimensions());
+
+    const unsigned int reduction_ops = reduction_axis.num_dimensions();
+    const int          input_dims    = input->num_dimensions();
+    Coordinates        axis_local    = reduction_axis;
+
+    for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
+    {
+        //axis: The dimensions to reduce. Must be in the range [-rank(input_tensor), rank(input_tensor)).
+        ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] < (-static_cast<int>(input->num_dimensions())));
+        ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] >= static_cast<int>(input->num_dimensions()));
+    }
+
+    if(output->tensor_shape().total_size() != 0)
+    {
+        // Only validate if not using auto_init for the output tensor
+        TensorShape out_shape = input->tensor_shape();
+        // Validate output_shape only if not using auto_init
+        convert_negative_axis(axis_local, input_dims);
+        std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
+        for(unsigned int i = 0; i < reduction_ops; ++i)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON(axis_local[i] > 3);
+            ARM_COMPUTE_RETURN_ERROR_ON(static_cast<unsigned int>(axis_local[i]) > input->num_dimensions() - 1);
+            if(output->total_size() > 0 && keep_dims)
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1);
+            }
+            if(keep_dims)
+            {
+                out_shape.set(axis_local[i], 1);
+            }
+            else
+            {
+                ARM_COMPUTE_RETURN_ERROR_ON(i > static_cast<unsigned int>(axis_local[i]));
+                const unsigned int remove_index = axis_local[i] - i;
+                ARM_COMPUTE_RETURN_ERROR_ON(remove_index >= out_shape.num_dimensions());
+                out_shape.remove_dimension(remove_index);
+            }
+        }
+        const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
+    }
+    return Status{};
+}
+}
+CLReduceSum::CLReduceSum(std::shared_ptr<IMemoryManager> memory_manager)
+    : _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _reduction_ops(), _keep_dims()
+{
+}
+void CLReduceSum::configure(ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
+{
+    configure(CLKernelLibrary::get().get_compile_context(), input, reduction_axis, keep_dims, output);
+}
+
+void CLReduceSum::configure(const CLCompileContext &compile_context, ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output)
+{
+    // Perform validate step
+    ARM_COMPUTE_ERROR_THROW_ON(CLReduceSum::validate(input->info(), reduction_axis, keep_dims, output->info()));
+    // Output auto inizialitation if not yet initialized
+    const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_reduce_mean_shape(input, reduction_axis, keep_dims);
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape));
+
+    _reduction_ops = reduction_axis.num_dimensions();
+    _reduction_kernels.resize(_reduction_ops);
+    _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0));
+    _keep_dims = keep_dims;
+
+    Coordinates axis_local = reduction_axis;
+    const int   input_dims = input->info()->num_dimensions();
+
+    convert_negative_axis(axis_local, input_dims);
+
+    // Perform reduction for every axis
+    for(int i = 0; i < _reduction_ops; ++i)
+    {
+        TensorShape out_shape = i == 0 ? input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape();
+        out_shape.set(axis_local[i], 1);
+        auto in = (i == 0) ? input : (&_reduced_outs[i - 1]);
+
+        if(i == _reduction_ops - 1 && keep_dims)
+        {
+            _reduction_kernels[i].configure(compile_context, in, output, axis_local[i], ReductionOperation::SUM);
+        }
+        else
+        {
+            _reduced_outs[i].allocator()->init(TensorInfo(out_shape, input->info()->num_channels(), input->info()->data_type(), input->info()->quantization_info()));
+            _memory_group.manage(&_reduced_outs[i]);
+            _reduction_kernels[i].configure(compile_context, in, &_reduced_outs[i], axis_local[i], ReductionOperation::SUM);
+        }
+    }
+
+    // Allocate intermediate tensors
+    for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i)
+    {
+        _reduced_outs[i].allocator()->allocate();
+    }
+
+    // Configure reshape layer if we want to drop the dimensions
+    if(!keep_dims)
+    {
+        TensorShape out_shape = input->info()->tensor_shape();
+
+        // We have to sort the reduction axis vectors in order for remove_dimension
+        // to work properly
+        std::sort(axis_local.begin(), axis_local.begin() + _reduction_ops);
+        for(int i = 0; i < _reduction_ops; ++i)
+        {
+            out_shape.remove_dimension(axis_local[i] - i);
+        }
+        auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(out_shape));
+        _reshape.configure(compile_context, &_reduced_outs[_reduction_ops - 1], output);
+    }
+}
+
+Status CLReduceSum::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
+{
+    return validate_config(input, reduction_axis, keep_dims, output);
+}
+
+void CLReduceSum::run()
+{
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    for(auto &kernel : _reduction_kernels)
+    {
+        kernel.run();
+    }
+
+    if(!_keep_dims)
+    {
+        _reshape.run();
+    }
+}
+} // namespace arm_compute