diff options
Diffstat (limited to 'src/runtime/CL/functions/CLReduceMean.cpp')
-rw-r--r-- | src/runtime/CL/functions/CLReduceMean.cpp | 109 |
1 files changed, 76 insertions, 33 deletions
diff --git a/src/runtime/CL/functions/CLReduceMean.cpp b/src/runtime/CL/functions/CLReduceMean.cpp index b761dc2f99..bef8d887fd 100644 --- a/src/runtime/CL/functions/CLReduceMean.cpp +++ b/src/runtime/CL/functions/CLReduceMean.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -27,6 +27,8 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" + +#include "src/common/utils/Log.h" #include "src/core/CL/CLValidate.h" #include "src/core/CL/kernels/CLFillBorderKernel.h" #include "src/core/CL/kernels/CLReductionOperationKernel.h" @@ -36,12 +38,14 @@ namespace arm_compute { namespace { -Status validate_config(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output) +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_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()); @@ -49,29 +53,36 @@ Status validate_config(const ITensorInfo *input, const Coordinates &reduction_ax const int input_dims = input->num_dimensions(); Coordinates axis_local = reduction_axis; - for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i) + 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) + 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); + +// Suppress warning produced by a compiler bug in GCC +// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104165 +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Warray-bounds" std::sort(axis_local.begin(), axis_local.begin() + reduction_ops); - for(unsigned int i = 0; i < reduction_ops; ++i) +#pragma GCC diagnostic pop + + 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) + if (output->total_size() > 0 && keep_dims) { ARM_COMPUTE_RETURN_ERROR_ON(output->dimension(axis_local[i]) != 1); } - if(keep_dims) + if (keep_dims) { out_shape.set(axis_local[i], 1); } @@ -80,13 +91,14 @@ Status validate_config(const ITensorInfo *input, const Coordinates &reduction_ax 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); + out_shape.remove_dimension(remove_index, false); } } const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info); - const bool requant = is_data_type_quantized(input->data_type()) && input->quantization_info() != output->quantization_info(); - if(requant) + const bool requant = + is_data_type_quantized(input->data_type()) && input->quantization_info() != output->quantization_info(); + if (requant) { TensorInfo input_no_quant(input->clone()->set_data_type(DataType::F32)); CLDequantizationLayer::validate(input, &input_no_quant); @@ -96,10 +108,19 @@ Status validate_config(const ITensorInfo *input, const Coordinates &reduction_ax } return Status{}; } -} +} // namespace CLReduceMean::CLReduceMean(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _dequant(), _requant(), _reduction_ops(), _keep_dims(), _do_requant(), _input_no_quant(), + : _memory_group(std::move(memory_manager)), + _reduction_kernels(), + _reduced_outs(), + _reshape(), + _dequant(), + _requant(), + _reduction_ops(), + _keep_dims(), + _do_requant(), + _input_no_quant(), _output_no_quant() { } @@ -109,15 +130,23 @@ void CLReduceMean::configure(ICLTensor *input, const Coordinates &reduction_axis configure(CLKernelLibrary::get().get_compile_context(), input, reduction_axis, keep_dims, output); } -void CLReduceMean::configure(const CLCompileContext &compile_context, ICLTensor *input, const Coordinates &reduction_axis, bool keep_dims, ICLTensor *output) +void CLReduceMean::configure(const CLCompileContext &compile_context, + ICLTensor *input, + const Coordinates &reduction_axis, + bool keep_dims, + ICLTensor *output) { // Perform validate step ARM_COMPUTE_ERROR_THROW_ON(CLReduceMean::validate(input->info(), reduction_axis, keep_dims, output->info())); + ARM_COMPUTE_LOG_PARAMS(input, reduction_axis, keep_dims, output); + // Output auto inizialitation if not yet initialized - const TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_reduce_mean_shape(input->info(), reduction_axis, keep_dims); + const TensorShape output_shape = + arm_compute::misc::shape_calculator::calculate_reduce_mean_shape(input->info(), reduction_axis, keep_dims); auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(output_shape)); - _do_requant = is_data_type_quantized(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info(); + _do_requant = is_data_type_quantized(input->info()->data_type()) && + input->info()->quantization_info() != output->info()->quantization_info(); _reduction_ops = reduction_axis.num_dimensions(); _reduction_kernels.resize(_reduction_ops); _reduced_outs.resize(_reduction_ops - (keep_dims ? 1 : 0)); @@ -125,7 +154,7 @@ void CLReduceMean::configure(const CLCompileContext &compile_context, ICLTensor ICLTensor *tmp_input = input; ICLTensor *tmp_output = output; - if(_do_requant) + if (_do_requant) { _memory_group.manage(&_input_no_quant); _memory_group.manage(&_output_no_quant); @@ -144,46 +173,57 @@ void CLReduceMean::configure(const CLCompileContext &compile_context, ICLTensor convert_negative_axis(axis_local, input_dims); // Perform reduction for every axis - for(int i = 0; i < _reduction_ops; ++i) + for (int i = 0; i < _reduction_ops; ++i) { - TensorShape out_shape = i == 0 ? tmp_input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape(); + TensorShape out_shape = + i == 0 ? tmp_input->info()->tensor_shape() : (&_reduced_outs[i - 1])->info()->tensor_shape(); out_shape.set(axis_local[i], 1); auto in = (i == 0) ? tmp_input : (&_reduced_outs[i - 1]); - if(i == _reduction_ops - 1 && keep_dims) + if (i == _reduction_ops - 1 && keep_dims) { - _reduction_kernels[i].configure(compile_context, in, tmp_output, axis_local[i], ReductionOperation::MEAN_SUM); + _reduction_kernels[i].configure(compile_context, in, tmp_output, axis_local[i], + ReductionOperation::MEAN_SUM); } else { - _reduced_outs[i].allocator()->init(TensorInfo(out_shape, tmp_input->info()->num_channels(), tmp_input->info()->data_type(), tmp_input->info()->quantization_info())); + _reduced_outs[i].allocator()->init(TensorInfo(out_shape, tmp_input->info()->num_channels(), + tmp_input->info()->data_type(), + tmp_input->info()->quantization_info())); _memory_group.manage(&_reduced_outs[i]); - _reduction_kernels[i].configure(compile_context, in, &_reduced_outs[i], axis_local[i], ReductionOperation::MEAN_SUM); + _reduction_kernels[i].configure(compile_context, in, &_reduced_outs[i], axis_local[i], + ReductionOperation::MEAN_SUM); } } // Allocate intermediate tensors - for(int i = 0; i < _reduction_ops - (keep_dims ? 1 : 0); ++i) + 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) + if (!_keep_dims) { TensorShape out_shape = tmp_input->info()->tensor_shape(); // We have to sort the reduction axis vectors in order for remove_dimension // to work properly + +// Suppress warning produced by a compiler bug in GCC +// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104165 +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Warray-bounds" std::sort(axis_local.begin(), axis_local.begin() + _reduction_ops); - for(int i = 0; i < _reduction_ops; ++i) +#pragma GCC diagnostic pop + for (int i = 0; i < _reduction_ops; ++i) { - out_shape.remove_dimension(axis_local[i] - i); + out_shape.remove_dimension(axis_local[i] - i, false); } auto_init_if_empty(*tmp_output->info(), tmp_input->info()->clone()->set_tensor_shape(out_shape)); _reshape.configure(compile_context, &_reduced_outs[_reduction_ops - 1], tmp_output); } - if(_do_requant) + if (_do_requant) { _requant.configure(compile_context, &_output_no_quant, output); _input_no_quant.allocator()->allocate(); @@ -191,7 +231,10 @@ void CLReduceMean::configure(const CLCompileContext &compile_context, ICLTensor } } -Status CLReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output) +Status CLReduceMean::validate(const ITensorInfo *input, + const Coordinates &reduction_axis, + bool keep_dims, + const ITensorInfo *output) { return validate_config(input, reduction_axis, keep_dims, output); } @@ -200,19 +243,19 @@ void CLReduceMean::run() { MemoryGroupResourceScope scope_mg(_memory_group); - if(_do_requant) + if (_do_requant) { _dequant.run(); } - for(auto &kernel : _reduction_kernels) + for (auto &kernel : _reduction_kernels) { kernel.run(); } - if(!_keep_dims) + if (!_keep_dims) { _reshape.run(); } - if(_do_requant) + if (_do_requant) { _requant.run(); } |