From be88a57579a9a848efe13e6c524b5b104b871733 Mon Sep 17 00:00:00 2001
From: Sadik Armagan <sadik.armagan@arm.com>
Date: Thu, 30 Apr 2020 11:39:37 +0100
Subject: IVGCVSW-4753 Fix CpuAcc Hal 1.3 Softmax Failures

* Refactor Neon Softmax workload to accept supported data types

Signed-off-by: Sadik Armagan <sadik.armagan@arm.com>
Change-Id: I54aa72d5cbb862cafcc1eabe48f6a00d61050cd7
---
 .../neon/workloads/NeonSoftmaxWorkload.cpp         | 53 ++++++++++++++++++++++
 1 file changed, 53 insertions(+)
 create mode 100644 src/backends/neon/workloads/NeonSoftmaxWorkload.cpp

(limited to 'src/backends/neon/workloads/NeonSoftmaxWorkload.cpp')

diff --git a/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp b/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp
new file mode 100644
index 0000000000..149804bdd6
--- /dev/null
+++ b/src/backends/neon/workloads/NeonSoftmaxWorkload.cpp
@@ -0,0 +1,53 @@
+//
+// Copyright © 2020 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "NeonSoftmaxWorkload.hpp"
+#include "NeonWorkloadUtils.hpp"
+
+#include <armnn/utility/PolymorphicDowncast.hpp>
+
+#include <aclCommon/ArmComputeUtils.hpp>
+#include <aclCommon/ArmComputeTensorUtils.hpp>
+
+#include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h>
+
+namespace armnn
+{
+
+arm_compute::Status NeonSoftmaxWorkloadValidate(const TensorInfo& input,
+                                                const TensorInfo& output,
+                                                const SoftmaxDescriptor& descriptor)
+{
+    const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input);
+    const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output);
+
+    unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input);
+    return arm_compute::NESoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis);
+}
+
+NeonSoftmaxWorkload::NeonSoftmaxWorkload(const SoftmaxQueueDescriptor& descriptor,
+    const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager)
+    : BaseWorkload<SoftmaxQueueDescriptor>(descriptor, info)
+{
+    m_Data.ValidateInputsOutputs("NeonSoftmaxWorkload", 1, 1);
+
+    // The ArmCompute softmax layer uses 2D input/output tensors, so flatten the first three dimensions.
+    arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
+    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
+
+    auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager);
+    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());
+}
+
+void NeonSoftmaxWorkload::Execute() const
+{
+    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonSoftmaxWorkload_Execute");
+    m_SoftmaxLayer->run();
+}
+
+} //namespace armnn
+
-- 
cgit v1.2.1