aboutsummaryrefslogtreecommitdiff
path: root/src/runtime/NEON/functions/NEReduceMean.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/runtime/NEON/functions/NEReduceMean.cpp')
-rw-r--r--src/runtime/NEON/functions/NEReduceMean.cpp97
1 files changed, 57 insertions, 40 deletions
diff --git a/src/runtime/NEON/functions/NEReduceMean.cpp b/src/runtime/NEON/functions/NEReduceMean.cpp
index 0b145f034d..96ec8b8587 100644
--- a/src/runtime/NEON/functions/NEReduceMean.cpp
+++ b/src/runtime/NEON/functions/NEReduceMean.cpp
@@ -24,80 +24,97 @@
#include "arm_compute/runtime/NEON/functions/NEReduceMean.h"
#include "arm_compute/core/CPP/Validate.h"
+#include "arm_compute/core/Error.h"
#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
-using namespace arm_compute;
+namespace arm_compute
+{
NEReduceMean::NEReduceMean(std::shared_ptr<IMemoryManager> memory_manager)
: _memory_group(std::move(memory_manager)), _reduction_kernels(), _reduced_outs(), _reshape(), _reduction_ops(), _keep_dims()
{
}
-Status NEReduceMean::validate(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);
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, 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());
- TensorShape out_shape = input->tensor_shape();
const unsigned int reduction_ops = reduction_axis.num_dimensions();
const int input_dims = input->num_dimensions();
Coordinates axis_local = reduction_axis;
- // Convert negative axis
- for(unsigned int i = 0; i < reduction_ops; ++i)
+ for(unsigned int i = 0; i < axis_local.num_dimensions(); ++i)
{
- axis_local[i] = wrap_around(axis_local[i], input_dims);
+ //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()));
}
- std::sort(axis_local.begin(), axis_local.begin() + reduction_ops);
- for(unsigned int i = 0; i < reduction_ops; ++i)
+ if(output->tensor_shape().total_size() != 0)
{
- 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
+ // 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)
{
- out_shape.remove_dimension(axis_local[i] - 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);
}
- const TensorInfo out_info = input->clone()->set_tensor_shape(out_shape);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &out_info);
-
return Status{};
}
+Status NEReduceMean::validate(const ITensorInfo *input, const Coordinates &reduction_axis, bool keep_dims, const ITensorInfo *output)
+{
+ return validate_config(input, reduction_axis, keep_dims, output);
+}
+
void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis, bool keep_dims, ITensor *output)
{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input);
+ // Perform validate step
+ ARM_COMPUTE_ERROR_THROW_ON(NEReduceMean::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();
- const unsigned int reduction_ops = reduction_axis.num_dimensions();
+ Coordinates axis_local = reduction_axis;
+ const int input_dims = input->info()->num_dimensions();
- // Convert negative axis
- for(unsigned int i = 0; i < reduction_ops; ++i)
- {
- axis_local[i] = wrap_around(axis_local[i], input_dims);
- }
+ convert_negative_axis(axis_local, input_dims);
// Perform reduction for every axis
- for(unsigned int i = 0; i < _reduction_ops; ++i)
+ 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);
@@ -116,7 +133,7 @@ void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis,
}
// Allocate intermediate tensors
- for(unsigned 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();
}
@@ -125,11 +142,10 @@ void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis,
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(unsigned int i = 0; i < _reduction_ops; ++i)
+ for(int i = 0; i < _reduction_ops; ++i)
{
out_shape.remove_dimension(axis_local[i] - i);
}
@@ -141,10 +157,9 @@ void NEReduceMean::configure(ITensor *input, const Coordinates &reduction_axis,
void NEReduceMean::run()
{
MemoryGroupResourceScope scope_mg(_memory_group);
-
- for(unsigned int i = 0; i < _reduction_ops; ++i)
+ for(auto &kernel : _reduction_kernels)
{
- _reduction_kernels[i].run();
+ kernel.run();
}
if(!_keep_dims)
@@ -152,3 +167,5 @@ void NEReduceMean::run()
_reshape.run();
}
}
+
+} // namespace arm_compute
generated by cgit v1.2.1 (git 2.23.0) at 2024-04-29 03:27:01 +0000