From fd899966cb881f5bb1ccce7903253a32d360419d Mon Sep 17 00:00:00 2001
From: Matthew Bentham <matthew.bentham@arm.com>
Date: Mon, 31 Dec 2018 15:49:42 +0000
Subject: MLCE-82 Add Neon Mean support and unit tests

Factor out new BuildArmComputeReductionCoordinates function
from CL backend into ArmComputeTensorUtils.

Update NEON LayerSupport and WorkloadFactory objects

Change-Id: Icc975ec699199bffafbdb207323df509d35e1e04
---
 src/backends/cl/workloads/ClMeanWorkload.cpp | 60 +++-------------------------
 1 file changed, 6 insertions(+), 54 deletions(-)

(limited to 'src/backends/cl/workloads/ClMeanWorkload.cpp')

diff --git a/src/backends/cl/workloads/ClMeanWorkload.cpp b/src/backends/cl/workloads/ClMeanWorkload.cpp
index 960fca2732..470b6a883d 100644
--- a/src/backends/cl/workloads/ClMeanWorkload.cpp
+++ b/src/backends/cl/workloads/ClMeanWorkload.cpp
@@ -10,50 +10,6 @@
 
 #include "ClWorkloadUtils.hpp"
 
-namespace
-{
-
-void ConvertArmnnAxesToAclCoordinates(size_t inputDimensions,
-                                      unsigned int originalInputRank,
-                                      const std::vector<unsigned int>& armnnAxes,
-                                      arm_compute::Coordinates& outAclCoords)
-{
-    if (armnnAxes.empty())
-    {
-        // If no reduction axes were provided, then the input must be reduced along all dimensions.
-        // Since arm_compute::CLReduceMean does not accept an empty vector as the reduction dimensions, we then
-        // manually create a vector including all the input dimensions (in reversed order) as:
-        //
-        // { inputDimensions - 1, inputDimensions - 2, ..., 1, 0 }
-        //
-        outAclCoords.set_num_dimensions(inputDimensions);
-        std::generate(outAclCoords.begin(), outAclCoords.end(), [d = inputDimensions - 1] () mutable { return d--; });
-    }
-    else
-    {
-        // Create a vector of reduction dimensions (in reversed order) with the given reduction axes.
-        //
-        // Adjust the given reduction axes according to the original rank of the input tensor (before ACL applied any
-        // dimension correction).
-        // For example, if the input tensor originally had 4 dimensions, and one of the reduction axes was 2, then the
-        // new value for that reduction axis should be 1.
-        //
-        // Example:
-        // ArmNN input shape = { 1, 1, 3, 2 } -> ACL input shape = { 2, 3 }
-        // ArmNN reduction axis = { 2 }       -> ACL reduction axis = { 1 }
-        // ArmNN reduction axis = { 3 }       -> ACL reduction axis = { 0 }
-        //
-        // The transformation: ACL reduction axis index = original rank - ArmNN reduction axis index - 1
-        //
-        outAclCoords.set_num_dimensions(armnnAxes.size());
-        std::transform(armnnAxes.begin(), armnnAxes.end(),
-                       outAclCoords.begin(),
-                       [originalInputRank](unsigned int i){ return originalInputRank - i - 1; });
-    }
-}
-
-} // anonymous namespace
-
 namespace armnn
 {
 using namespace armcomputetensorutils;
@@ -65,11 +21,9 @@ arm_compute::Status ClMeanValidate(const TensorInfo& input,
     const arm_compute::TensorInfo aclInputInfo  = armcomputetensorutils::BuildArmComputeTensorInfo(input);
     const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
 
-    arm_compute::Coordinates coords;
-    ConvertArmnnAxesToAclCoordinates(aclInputInfo.num_dimensions(),
-                                     input.GetNumDimensions(),
-                                     desc.m_Axis,
-                                     coords);
+    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(),
+                                                                          input.GetNumDimensions(),
+                                                                          desc.m_Axis);
 
     return arm_compute::CLReduceMean::validate(&aclInputInfo, coords, desc.m_KeepDims, &aclOutputInfo);
 }
@@ -82,11 +36,9 @@ ClMeanWorkload::ClMeanWorkload(const MeanQueueDescriptor& descriptor, const Work
     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
-    arm_compute::Coordinates coords;
-    ConvertArmnnAxesToAclCoordinates(input.info()->num_dimensions(),
-                                     info.m_InputTensorInfos[0].GetNumDimensions(),
-                                     m_Data.m_Parameters.m_Axis,
-                                     coords);
+    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(),
+                                                                          info.m_InputTensorInfos[0].GetNumDimensions(),
+                                                                          m_Data.m_Parameters.m_Axis);
 
     m_Layer.configure(&input, coords, m_Data.m_Parameters.m_KeepDims, &output);
 }
-- 
cgit v1.2.1