IVGCVSW-3620 Fix Hal 1.2 Softmax test failures on GpuAcc and CpuAcc
The following NeuralNetworkTests tests were failing on GpuAcc and CpuAcc:
GeneratedTests.softmax_v1_2_relaxed
GeneratedTests.softmax_v1_2_quant8
GeneratedTests.softmax_v1_2_2
GeneratedTests.softmax_v1_2_relaxed_2
GeneratedTests.softmax_v1_2_quant8_2
The default value for Softmax axis parameter in Android is -1 but is 1 in ACL.
Detect and handle this in ArmComputeUtils.ComputeSoftmaxAclAxis.
Signed-off-by: Colm Donelan <Colm.Donelan@arm.com>
Change-Id: Ibb0660e4cb0dc6bd4c804c4397fbd61f38acdd9c
diff --git a/src/backends/aclCommon/ArmComputeUtils.hpp b/src/backends/aclCommon/ArmComputeUtils.hpp
index 0f56160..4d69090 100644
--- a/src/backends/aclCommon/ArmComputeUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeUtils.hpp
@@ -135,9 +135,15 @@
}
}
-inline unsigned int ComputeSoftmaxAclAxis(const armnn::TensorInfo& tensor)
+inline unsigned int ComputeSoftmaxAclAxis(const SoftmaxDescriptor& softmaxDesc, const armnn::TensorInfo& tensor)
{
- unsigned int dim = tensor.GetNumDimensions();
+ // Detect the Android default value of -1 and return the ACL default value of 1.
+ if (softmaxDesc.m_Axis == -1)
+ {
+ return 1;
+ }
+
+ unsigned int dim = tensor.GetNumDimensions();
BOOST_ASSERT(dim != 0);
diff --git a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp
index 2f6d380..a355ba0 100644
--- a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp
+++ b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp
@@ -20,7 +20,7 @@
const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(input);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input);
return arm_compute::CLSoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis);
}
diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
index f2f8d17..adb4872 100644
--- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
+++ b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp
@@ -23,7 +23,7 @@
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();
- unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
}
diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
index ce2a9e6..f14ea11 100644
--- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
+++ b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp
@@ -33,7 +33,7 @@
"Invalid quantization for output. Only scale = 1.0f / 256.0f and offset = 0 supported");
}
- unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
m_SoftmaxLayer.configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis);
}
diff --git a/src/backends/neon/workloads/NeonSoftmaxBaseWorkload.cpp b/src/backends/neon/workloads/NeonSoftmaxBaseWorkload.cpp
index 8acb775..41ebfb9 100644
--- a/src/backends/neon/workloads/NeonSoftmaxBaseWorkload.cpp
+++ b/src/backends/neon/workloads/NeonSoftmaxBaseWorkload.cpp
@@ -20,7 +20,7 @@
const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(input);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input);
return arm_compute::NESoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis);
}
diff --git a/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp b/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp
index 4dc9131..152d19c 100644
--- a/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp
+++ b/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp
@@ -24,7 +24,7 @@
arm_compute::ITensor& output = boost::polymorphic_downcast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
layer->configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis);
m_SoftmaxLayer.reset(layer.release());
}
diff --git a/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp b/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp
index 8e83914..15a7066 100644
--- a/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp
+++ b/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp
@@ -34,7 +34,7 @@
}
auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
- unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]);
+ unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]);
layer->configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis);
m_SoftmaxLayer.reset(layer.release());
}