COMPMID-1816: Use parallel reduction on 0 axis in CL ARG_MIN/ARG_MAX

Introducing new CLArgMinMax kernel

Change-Id: I0b8254207cc3859d19ceef9b6429cf5c1c586db0
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2202
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

3 files changed, 151 insertions, 48 deletions

diff --git a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
index fd172d5f2c..4ac6d25d75 100644
--- a/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
+++ b/src/runtime/CL/functions/CLArgMinMaxLayer.cpp
@@ -23,33 +23,145 @@
  */
 
 #include "arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h"
-#include "arm_compute/runtime/CL/functions/CLReductionOperation.h"
 
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/runtime/Utils.h"
 
 namespace arm_compute
 {
 CLArgMinMaxLayer::CLArgMinMaxLayer(std::shared_ptr<IMemoryManager> memory_manager)
-    : _reduction_function(support::cpp14::make_unique<CLReductionOperation>(std::move(memory_manager)))
+    : _memory_group(std::move(memory_manager)), _results_vector(), _not_reshaped_output(), _reduction_kernels_vector(), _reshape_kernel(), _num_of_stages(), _reduction_axis()
 {
 }
 
-void CLArgMinMaxLayer::configure(ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
+Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITensorInfo *output, const ReductionOperation &op)
 {
-    _reduction_function->configure(input, output, axis, op, false);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid reduction operation");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= static_cast<int>(TensorShape::num_max_dimensions), "Reduction axis greater than max number of dimensions");
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
+    const unsigned int num_of_stages = calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
+
+    DataType   output_data_type = DataType::S32;
+    TensorInfo not_reshaped_output;
+    const auto input_num_channles = input->num_channels();
+    const auto input_qinfo        = input->quantization_info();
+
+    if(output->total_size() != 0)
+    {
+        output_data_type                       = output->data_type();
+        const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, false));
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
+    }
+
+    auto shape_before_reshape = input->tensor_shape();
+    shape_before_reshape.set(axis, 1);
+    auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo)
+    {
+        ti.set_data_type(data_type).set_tensor_shape(shape).set_num_channels(num_channels).set_quantization_info(qinfo);
+    };
+
+    initialize_tensorinfo(not_reshaped_output, shape_before_reshape, output_data_type, input_num_channles, input_qinfo);
+
+    if(num_of_stages == 1)
+    {
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &not_reshaped_output, axis, op));
+    }
+    else
+    {
+        // Create temporary tensor infos
+        std::vector<TensorInfo> sums_vector(num_of_stages - 1);
+
+        // Create intermediate tensor info
+        TensorShape shape{ input->tensor_shape() };
+
+        for(unsigned int i = 0; i < num_of_stages - 1; i++)
+        {
+            shape.set(0, ceil(shape.x() / 128.f));
+            sums_vector[i].set_data_type(input->data_type());
+            sums_vector[i].set_tensor_shape(shape);
+            sums_vector[i].set_num_channels(input->num_channels());
+        }
+
+        // Validate ReductionOperation only on first kernel
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, nullptr, &sums_vector[0], axis, op));
+
+        // Validate ReductionOperation on intermediate stages
+        for(unsigned int i = 1; i < num_of_stages - 1; ++i)
+        {
+            ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[i - 1], &sums_vector[i], axis, op));
+        }
+
+        // Validate ReductionOperation on the last stage
+        const unsigned int last_stage = num_of_stages - 1;
+        ARM_COMPUTE_RETURN_ON_ERROR(CLArgMinMaxLayerKernel::validate(input, &sums_vector[last_stage - 1], &not_reshaped_output, axis, op));
+    }
+    ARM_COMPUTE_RETURN_ON_ERROR(CLReshapeLayerKernel::validate(&not_reshaped_output, output));
+    return Status{};
 }
 
-Status CLArgMinMaxLayer::validate(const ITensorInfo *input, int axis, const ITensorInfo *output, const ReductionOperation &op)
+void CLArgMinMaxLayer::configure(const ICLTensor *input, int axis, ICLTensor *output, const ReductionOperation &op)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG(op != ReductionOperation::ARG_IDX_MAX && op != ReductionOperation::ARG_IDX_MIN, "Invalid operation");
-    return CLReductionOperation::validate(input, output, axis, op, false);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    _num_of_stages  = calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
+    _reduction_axis = axis;
+
+    const TensorShape output_shape     = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
+    DataType          output_data_type = (output->info()->data_type() == DataType::UNKNOWN) ? DataType::S32 : output->info()->data_type();
+    auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
+
+    // Configure reduction operation kernels
+    _reduction_kernels_vector.resize(_num_of_stages);
+
+    _memory_group.manage(&_not_reshaped_output);
+    // Create temporary tensors
+    if(_num_of_stages == 1)
+    {
+        _reduction_kernels_vector[0].configure(input, nullptr, &_not_reshaped_output, axis, op);
+    }
+    else
+    {
+        _results_vector.resize(_num_of_stages - 1);
+        TensorShape shape{ input->info()->tensor_shape() };
+        for(unsigned int i = 0; i < _num_of_stages - 1; i++)
+        {
+            shape.set(0, ceil(shape.x() / 128.f));
+            _results_vector[i].allocator()->init(input->info()->clone()->set_tensor_shape(shape).set_data_type(output_data_type));
+        }
+
+        // Apply ReductionOperation only on first kernel
+        _memory_group.manage(&_results_vector[0]);
+        _reduction_kernels_vector[0].configure(input, nullptr, &_results_vector[0], axis, op);
+
+        // Apply ReductionOperation on intermediate stages
+        for(unsigned int i = 1; i < _num_of_stages - 1; ++i)
+        {
+            _memory_group.manage(&_results_vector[i]);
+            _reduction_kernels_vector[i].configure(input, &_results_vector[i - 1], &_results_vector[i], axis, op);
+            _results_vector[i - 1].allocator()->allocate();
+        }
+
+        // Apply ReductionOperation on the last stage
+        const unsigned int last_stage = _num_of_stages - 1;
+        _reduction_kernels_vector[last_stage].configure(input, &_results_vector[last_stage - 1], &_not_reshaped_output, axis, op);
+        _results_vector[last_stage - 1].allocator()->allocate();
+    }
+    _reshape_kernel.configure(&_not_reshaped_output, output);
+    _not_reshaped_output.allocator()->allocate();
 }
 
 void CLArgMinMaxLayer::run()
 {
-    _reduction_function->run();
+    MemoryGroupResourceScope scope_mg(_memory_group);
+
+    for(unsigned int i = 0; i < _num_of_stages; ++i)
+    {
+        CLScheduler::get().enqueue(_reduction_kernels_vector[i], false);
+    }
+    CLScheduler::get().enqueue(_reshape_kernel, false);
 }
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/runtime/CL/functions/CLReductionOperation.cpp b/src/runtime/CL/functions/CLReductionOperation.cpp
index 3aa5a813b6..2f9a38601d 100644
--- a/src/runtime/CL/functions/CLReductionOperation.cpp
+++ b/src/runtime/CL/functions/CLReductionOperation.cpp
@@ -33,30 +33,11 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "arm_compute/runtime/CL/CLScheduler.h"
 #include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/Utils.h"
 #include "support/ToolchainSupport.h"
 
 namespace arm_compute
 {
-namespace
-{
-unsigned int calculate_number_of_stages(const ITensorInfo *input, unsigned int axis)
-{
-    // We need only 1 stage for all axis except x-axis and x-axis for QASYMM8.
-    if(axis != 0 || (axis == 0 && is_data_type_quantized(input->data_type())))
-    {
-        return 1;
-    }
-    // Calculate number of WGs. 16 elements per thread, 8 threads per WG
-    const unsigned int num_of_wg = ceil(input->dimension(0) / 128.f);
-
-    // Calculate number of stages. First stage performs op and the rest reduction sum
-    // depending on the size of the input. Last stage should have only 1 WG.
-    const unsigned int num_of_stages = num_of_wg / 128 + 2;
-
-    return num_of_stages;
-}
-} // namespace
-
 CLReductionOperation::CLReductionOperation(std::shared_ptr<IMemoryManager> memory_manager)
     : _memory_group(std::move(memory_manager)), _results_vector(), _reduction_kernels_vector(), _border_handlers_vector(), _reshape_kernel(), _op(), _num_of_stages(), _reduction_axis(), _is_serial(),
       _is_reshape_required(false)
@@ -65,15 +46,15 @@ CLReductionOperation::CLReductionOperation(std::shared_ptr<IMemoryManager> memor
 
 Status CLReductionOperation::validate(const ITensorInfo *input, const ITensorInfo *output, unsigned int axis, ReductionOperation op, bool keep_dims)
 {
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis >= TensorShape::num_max_dimensions, "Reduction axis greater than max number of dimensions");
     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis > 3, "Unsupported reduction axis");
 
-    const unsigned int num_of_stages       = calculate_number_of_stages(input, axis);
+    const unsigned int num_of_stages       = calculate_number_of_stages_only_x_axis(input->dimension(0), axis);
     const bool         is_serial           = needs_serialized_reduction(op, input->data_type(), axis);
-    const bool         is_arg_min_max      = (op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN);
-    const bool         is_reshape_required = !keep_dims || is_arg_min_max;
+    const bool         is_reshape_required = !keep_dims;
 
-    if(is_reshape_required)
+    if(is_reshape_required && output->total_size() != 0)
     {
         const TensorInfo expected_output_shape = output->clone()->set_tensor_shape(arm_compute::misc::shape_calculator::compute_reduced_shape(input->tensor_shape(), axis, keep_dims));
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(&expected_output_shape, output);
@@ -86,7 +67,7 @@ Status CLReductionOperation::validate(const ITensorInfo *input, const ITensorInf
     const auto input_data_type    = input->data_type();
     const auto input_num_channles = input->num_channels();
     const auto input_qinfo        = input->quantization_info();
-    const auto output_data_type   = is_arg_min_max ? DataType::S32 : output->data_type();
+    const auto output_data_type   = output->data_type();
 
     auto initialize_tensorinfo = [](TensorInfo & ti, TensorShape shape, DataType data_type, int num_channels, QuantizationInfo qinfo)
     {
@@ -184,8 +165,7 @@ ICLTensor *CLReductionOperation::configure_intermediate_result_vector(ICLTensor
         return output;
     }
 
-    auto       intermediate_result_vector_size = _is_serial ? 1 : _num_of_stages;
-    const auto is_arg_min_max                  = (_op == ReductionOperation::ARG_IDX_MAX || _op == ReductionOperation::ARG_IDX_MIN);
+    auto intermediate_result_vector_size = _is_serial ? 1 : _num_of_stages;
 
     if(!_is_reshape_required)
     {
@@ -206,30 +186,24 @@ ICLTensor *CLReductionOperation::configure_intermediate_result_vector(ICLTensor
         v.allocator()->init(input->info()->clone()->set_tensor_shape(shape));
     }
 
-    if(is_arg_min_max)
-    {
-        _results_vector.back().info()->set_data_type(DataType::S32).set_is_resizable(true).reset_padding();
-    }
-
     return _is_reshape_required ? &_results_vector.back() : output;
 }
 
 void CLReductionOperation::configure(ICLTensor *input, ICLTensor *output, unsigned int axis, ReductionOperation op, bool keep_dims)
 {
-    _op                       = op;
-    _num_of_stages            = calculate_number_of_stages(input->info(), axis);
-    _reduction_axis           = axis;
-    _is_serial                = needs_serialized_reduction(op, input->info()->data_type(), axis);
-    const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN);
-    _is_reshape_required      = !keep_dims || is_arg_min_max;
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
+    _op                  = op;
+    _num_of_stages       = calculate_number_of_stages_only_x_axis(input->info()->dimension(0), axis);
+    _reduction_axis      = axis;
+    _is_serial           = needs_serialized_reduction(op, input->info()->data_type(), axis);
+    _is_reshape_required = !keep_dims;
 
     auto *output_internal = configure_intermediate_result_vector(input, output);
 
-    // ArgMinMax might not give initialized output tensor, so initialize here.
     if(_is_reshape_required)
     {
         const TensorShape output_shape     = arm_compute::misc::shape_calculator::compute_reduced_shape(input->info()->tensor_shape(), axis, false);
-        const auto        output_data_type = is_arg_min_max ? DataType::S32 : input->info()->data_type();
+        const auto        output_data_type = input->info()->data_type();
         auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape).set_data_type(output_data_type).reset_padding().set_is_resizable(true));
     }
 
diff --git a/src/runtime/Utils.cpp b/src/runtime/Utils.cpp
index 70494be05c..2204ec11d7 100644
--- a/src/runtime/Utils.cpp
+++ b/src/runtime/Utils.cpp
@@ -25,6 +25,7 @@
 
 #include "arm_compute/runtime/NEON/NEScheduler.h"
 
+#include <cmath>
 #include <map>
 #include <string>
 
@@ -61,4 +62,20 @@ void schedule_kernel_on_ctx(IRuntimeContext *ctx, ICPPKernel *kernel, const ISch
         NEScheduler::get().schedule(kernel, hints);
     }
 }
+
+unsigned int calculate_number_of_stages_only_x_axis(size_t input_x_dimension, unsigned int axis)
+{
+    // We need only 1 stage for all axis except x-axis
+    if(axis != 0)
+    {
+        return 1;
+    }
+    // Calculate number of WGs. 16 elements per thread, 8 threads per WG
+    const auto num_of_wg = static_cast<unsigned int>(ceil(input_x_dimension / 128.f));
+
+    // Calculate number of stages. First stage performs op and the rest reduction sum
+    // depending on the size of the input. Last stage should have only 1 WG.
+    const unsigned int num_of_stages = num_of_wg / 128 + 2;
+    return num_of_stages;
+}
 } // namespace arm_compute