From aec513c52f88b21e01a92d3e1f20f4e2b7bd9f01 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 15 Sep 2017 19:36:30 +0100 Subject: COMPMID-417: Fix potential memory leak in CLReduction Kernel. Change-Id: I1c285b2fdac5ebf154731e8e34e0549a7f92525f Reviewed-on: http://mpd-gerrit.cambridge.arm.com/87939 Tested-by: Kaizen Reviewed-by: Moritz Pflanzer --- src/runtime/CL/functions/CLReductionOperation.cpp | 29 +++++++++++++++-------- 1 file changed, 19 insertions(+), 10 deletions(-) (limited to 'src/runtime/CL/functions/CLReductionOperation.cpp') diff --git a/src/runtime/CL/functions/CLReductionOperation.cpp b/src/runtime/CL/functions/CLReductionOperation.cpp index 6643c9bd46..d02afb4e90 100644 --- a/src/runtime/CL/functions/CLReductionOperation.cpp +++ b/src/runtime/CL/functions/CLReductionOperation.cpp @@ -49,6 +49,9 @@ void CLReductionOperation::configure(ICLTensor *input, ICLTensor *output, unsign // depending on the size of the input. Last stage should have only 1 WG. _num_of_stages = num_of_wg / 128 + 2; + // Create temporary tensors + _sums_vector = arm_compute::support::cpp14::make_unique(_num_of_stages - 1); + // Configure reduction operation kernels _reduction_kernels_vector = arm_compute::support::cpp14::make_unique(_num_of_stages); _border_handlers_vector = arm_compute::support::cpp14::make_unique(_num_of_stages); @@ -57,22 +60,28 @@ void CLReductionOperation::configure(ICLTensor *input, ICLTensor *output, unsign for(unsigned int i = 0; i < _num_of_stages - 1; i++) { shape.set(0, ceil(shape.x() / 128.f)); - auto *tensor = new CLTensor; - tensor->allocator()->init(TensorInfo(shape, input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position())); - _memory_group.manage(tensor); - _sums_vector.push_back(tensor); + _sums_vector[i].allocator()->init(TensorInfo(shape, input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position())); } // Apply ReductionOperation only on first kernel - _reduction_kernels_vector[0].configure(input, _sums_vector.at(0), axis, op); + _memory_group.manage(_sums_vector.get()); + _reduction_kernels_vector[0].configure(input, _sums_vector.get(), axis, op); _border_handlers_vector[0].configure(input, _reduction_kernels_vector[0].border_size(), BorderMode::CONSTANT, PixelValue(0)); - for(unsigned int i = 1; i < _num_of_stages; i++) + + // Apply ReductionOperation on intermediate stages + for(unsigned int i = 1; i < _num_of_stages - 1; ++i) { - // Last sum vector is the output vector - _reduction_kernels_vector[i].configure(_sums_vector.at(i - 1), i == _num_of_stages - 1 ? output : _sums_vector.at(i), axis, ReductionOperation::SUM); - _border_handlers_vector[i].configure(_sums_vector.at(i - 1), _reduction_kernels_vector[i].border_size(), BorderMode::CONSTANT, PixelValue(0)); - _sums_vector.at(i - 1)->allocator()->allocate(); + _memory_group.manage(_sums_vector.get() + i); + _reduction_kernels_vector[i].configure(_sums_vector.get() + i - 1, _sums_vector.get() + i, axis, ReductionOperation::SUM); + _border_handlers_vector[i].configure(_sums_vector.get() + i - 1, _reduction_kernels_vector[i].border_size(), BorderMode::CONSTANT, PixelValue(0)); + _sums_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(_sums_vector.get() + last_stage - 1, output, axis, ReductionOperation::SUM); + _border_handlers_vector[last_stage].configure(_sums_vector.get() + last_stage - 1, _reduction_kernels_vector[last_stage].border_size(), BorderMode::CONSTANT, PixelValue(0)); + _sums_vector[last_stage - 1].allocator()->allocate(); } void CLReductionOperation::run() -- cgit v1.2.1