IVGCVSW-2354 Caffe SqueezeNet through armnn Cl and Neon

 * Compute Softmax Acl axis for Cl and Neon
 * Add unittests for Softmax in 3D and 4D
 * Correct input and output layer names in CaffeSqueezeNet inference test

Change-Id: I2d369d9a2db19c40f2af3341039dd33f0c5637b1
Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>

3 files changed, 27 insertions, 15 deletions

diff --git a/src/backends/cl/test/ClLayerTests.cpp b/src/backends/cl/test/ClLayerTests.cpp
index 23b4dc4423..58b9ba7f76 100644
--- a/src/backends/cl/test/ClLayerTests.cpp
+++ b/src/backends/cl/test/ClLayerTests.cpp
@@ -34,11 +34,6 @@ using FactoryType = armnn::ClWorkloadFactory;
 // Activation
 ARMNN_AUTO_TEST_CASE(ConstantLinearActivation, ConstantLinearActivationTest)
 
-ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1, SimpleSoftmaxTest, 1.0f)
-ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2, SimpleSoftmaxTest, 2.0f)
-ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1Uint8, SimpleSoftmaxUint8Test, 1.0f)
-ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2Uint8, SimpleSoftmaxUint8Test, 2.0f)
-
 ARMNN_AUTO_TEST_CASE(ReLu1Uint8, BoundedReLuUint8UpperAndLowerBoundTest)
 ARMNN_AUTO_TEST_CASE(ReLu6Uint8, BoundedReLuUint8UpperBoundOnlyTest)
 
@@ -377,6 +372,22 @@ ARMNN_AUTO_TEST_CASE(GreaterUint8, GreaterUint8Test)
 ARMNN_AUTO_TEST_CASE(GreaterBroadcast1ElementUint8, GreaterBroadcast1ElementUint8Test)
 ARMNN_AUTO_TEST_CASE(GreaterBroadcast1DVectorUint8, GreaterBroadcast1DVectorUint8Test)
 
+// Softmax
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1, SimpleSoftmaxTest, 1.0f)
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2, SimpleSoftmaxTest, 2.0f)
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1Uint8, SimpleSoftmaxUint8Test, 1.0f)
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2Uint8, SimpleSoftmaxUint8Test, 2.0f)
+
+ARMNN_AUTO_TEST_CASE(Simple3dSoftmax, Simple3dSoftmaxTest, 1.0f)
+ARMNN_AUTO_TEST_CASE(Simple3dSoftmaxUint8, Simple3dSoftmaxUint8Test, 1.0f)
+
+ARMNN_AUTO_TEST_CASE(Simple4dSoftmax, Simple4dSoftmaxTest, 1.0f)
+ARMNN_AUTO_TEST_CASE(Simple4dSoftmaxUint8, Simple4dSoftmaxUint8Test, 1.0f)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta1WithReference, CompareSoftmaxTest, 1.0f)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta2WithReference, CompareSoftmaxTest, 2.0f)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxUint8, CompareSoftmaxUint8Test, 1.0f)
+
 // Space To Batch Nd
 ARMNN_AUTO_TEST_CASE(SpaceToBatchNdSimpleFloat32, SpaceToBatchNdSimpleFloat32Test)
 ARMNN_AUTO_TEST_CASE(SpaceToBatchNdMultiChannelsFloat32, SpaceToBatchNdMultiChannelsFloat32Test)
@@ -445,10 +456,6 @@ ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareNormalizationAcrossWithReference, Compar
                                  armnn::NormalizationAlgorithmChannel::Across,
                                  armnn::NormalizationAlgorithmMethod::LocalBrightness)
 
-ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta1WithReference, CompareSoftmaxTest, 1.0f)
-ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta2WithReference, CompareSoftmaxTest, 2.0f)
-ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxUint8, CompareSoftmaxUint8Test, 1.0f)
-
 ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareMaxPooling2dWithRef, ComparePooling2dTest, armnn::PoolingAlgorithm::Max)
 
 ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareAveragePooling2dWithRef, ComparePooling2dTest, armnn::PoolingAlgorithm::Average)
diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
index bb56802ba3..c78ab039ef 100644
--- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
+++ b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
@@ -4,11 +4,12 @@
 //
 
 #include "ClSoftmaxFloatWorkload.hpp"
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeUtils.hpp>
 #include <cl/ClTensorHandle.hpp>
 #include <backendsCommon/CpuTensorHandle.hpp>
 
-#include "ClWorkloadUtils.hpp"
-
 namespace armnn
 {
 
@@ -21,7 +22,9 @@ ClSoftmaxFloatWorkload::ClSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& des
 
     arm_compute::ICLTensor& input  = static_cast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ICLTensor& output = static_cast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
-    m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta);
+
+    unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+    m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
 }
 
 void ClSoftmaxFloatWorkload::Execute() const
diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
index c386e3837b..086f3751b4 100644
--- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
+++ b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
@@ -4,11 +4,12 @@
 //
 
 #include "ClSoftmaxUint8Workload.hpp"
+#include "ClWorkloadUtils.hpp"
+
+#include <aclCommon/ArmComputeUtils.hpp>
 #include <cl/ClTensorHandle.hpp>
 #include <backendsCommon/CpuTensorHandle.hpp>
 
-#include "ClWorkloadUtils.hpp"
-
 namespace armnn
 {
 
@@ -30,7 +31,8 @@ ClSoftmaxUint8Workload::ClSoftmaxUint8Workload(const SoftmaxQueueDescriptor& des
             "Invalid quantization for output. Only scale = 1.0f / 256.0f and offset = 0 supported");
     }
 
-    m_SoftmaxLayer.configure(&input, &output, descriptor.m_Parameters.m_Beta);
+    unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+    m_SoftmaxLayer.configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis);
 }
 
 void ClSoftmaxUint8Workload::Execute() const