diff options
Diffstat (limited to 'src/backends/neon/workloads')
-rw-r--r-- | src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp | 4 | ||||
-rw-r--r-- | src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp | 6 |
2 files changed, 7 insertions, 3 deletions
diff --git a/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp b/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp index d9c78bbd43..afc6135e0f 100644 --- a/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp +++ b/src/backends/neon/workloads/NeonSoftmaxFloatWorkload.cpp @@ -7,6 +7,7 @@ #include "NeonWorkloadUtils.hpp" +#include <aclCommon/ArmComputeUtils.hpp> #include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h> namespace armnn @@ -22,8 +23,9 @@ NeonSoftmaxFloatWorkload::NeonSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& arm_compute::ITensor& input = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); arm_compute::ITensor& output = boost::polymorphic_downcast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); + unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]); auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager); - layer->configure(&input, &output, m_Data.m_Parameters.m_Beta); + 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 f780589075..7f295d6548 100644 --- a/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp +++ b/src/backends/neon/workloads/NeonSoftmaxUint8Workload.cpp @@ -4,9 +4,10 @@ // #include "NeonSoftmaxUint8Workload.hpp" - #include "NeonWorkloadUtils.hpp" +#include <aclCommon/ArmComputeUtils.hpp> + #include <arm_compute/runtime/NEON/functions/NESoftmaxLayer.h> namespace armnn @@ -29,9 +30,10 @@ NeonSoftmaxUint8Workload::NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& throw InvalidArgumentException( "Invalid quantization for output. Only scale = 1.0f / 256.0f and offset = 0 supported"); } + unsigned int aclAxis = ComputeSoftmaxAclAxis(info.m_InputTensorInfos[0]); auto layer = std::make_unique<arm_compute::NESoftmaxLayer>(memoryManager); - layer->configure(&input, &output, descriptor.m_Parameters.m_Beta); + layer->configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis); m_SoftmaxLayer.reset(layer.release()); } |