diff options
| author | Teresa Charlin <teresa.charlinreyes@arm.com> | 2023-08-17 18:44:58 +0100 |
|---|---|---|
| committer | Teresa Charlin <teresa.charlinreyes@arm.com> | 2023-08-28 12:37:25 +0100 |
| commit | 9145e38edf49fa4862008c163c34590141eecb14 (patch) | |
| tree | 64706ef579f548b804d5b674b33f6b239c638d0f /src/backends/neon/NeonLayerSupport.cpp | |
| parent | e40cc8359b02a7786908294300c45b672cf6b0e4 (diff) | |
| download | armnn-9145e38edf49fa4862008c163c34590141eecb14.tar.gz | |
IVGCVSW-7505 Create FusedLayer and NeonFusedWorkload for AddMulAdd Neon kernel
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: Ic778d35b001474b44fb1e433a6fe276e4ec9f565
Diffstat (limited to 'src/backends/neon/NeonLayerSupport.cpp')
| -rw-r--r-- | src/backends/neon/NeonLayerSupport.cpp | 54 |
1 files changed, 41 insertions, 13 deletions
diff --git a/src/backends/neon/NeonLayerSupport.cpp b/src/backends/neon/NeonLayerSupport.cpp index b491ba8493..ef1d21835a 100644 --- a/src/backends/neon/NeonLayerSupport.cpp +++ b/src/backends/neon/NeonLayerSupport.cpp @@ -4,7 +4,6 @@ // #include "NeonLayerSupport.hpp" -#include "NeonBackendId.hpp" #include "NeonBackendModelContext.hpp" #include <armnn/Exceptions.hpp> @@ -12,7 +11,6 @@ #include <armnn/Types.hpp> #include <armnn/BackendRegistry.hpp> -#include <InternalTypes.hpp> #include <LayerSupportCommon.hpp> #include <armnn/utility/IgnoreUnused.hpp> #include <armnn/utility/PolymorphicDowncast.hpp> @@ -39,8 +37,13 @@ #include "workloads/NeonDepthToSpaceWorkload.hpp" #include "workloads/NeonDepthwiseConvolutionWorkload.hpp" #include "workloads/NeonDequantizeWorkload.hpp" +#include "workloads/NeonDivisionWorkload.hpp" #include "workloads/NeonElementwiseBinaryWorkload.hpp" #include "workloads/NeonExpWorkload.hpp" +#include "workloads/NeonFullyConnectedWorkload.hpp" +#include "workloads/NeonFusedWorkload.hpp" +#include "workloads/NeonGatherWorkload.hpp" +#include "workloads/NeonGatherNdWorkload.hpp" #include "workloads/NeonInstanceNormalizationWorkload.hpp" #include "workloads/NeonL2NormalizationFloatWorkload.hpp" #include "workloads/NeonLogWorkload.hpp" @@ -53,12 +56,8 @@ #include "workloads/NeonMeanWorkload.hpp" #include "workloads/NeonMinimumWorkload.hpp" #include "workloads/NeonMultiplicationWorkload.hpp" -#include "workloads/NeonDivisionWorkload.hpp" #include "workloads/NeonNegWorkload.hpp" #include "workloads/NeonNormalizationFloatWorkload.hpp" -#include "workloads/NeonFullyConnectedWorkload.hpp" -#include "workloads/NeonGatherWorkload.hpp" -#include "workloads/NeonGatherNdWorkload.hpp" #include "workloads/NeonPadWorkload.hpp" #include "workloads/NeonPermuteWorkload.hpp" #include "workloads/NeonPooling2dWorkload.hpp" @@ -128,13 +127,13 @@ bool IsSupportedForDataTypeNeon(Optional<std::string&> reasonIfUnsupported, { return IsNeonBackendSupported(reasonIfUnsupported) && IsSupportedForDataTypeGeneric(reasonIfUnsupported, - dataType, - floatFuncPtr, - floatFuncPtr, - uint8FuncPtr, - &FalseFunc<>, - &FalseFunc<>, - std::forward<Params>(params)...); + dataType, + floatFuncPtr, + floatFuncPtr, + uint8FuncPtr, + &FalseFunc<>, + &FalseFunc<>, + std::forward<Params>(params)...); } #if defined(ARMCOMPUTENEON_ENABLED) @@ -430,6 +429,22 @@ bool IsLayerTypeSupported(const LayerType& type, *(PolymorphicDowncast<const FullyConnectedDescriptor*>(&descriptor)), reasonIfUnsupported); + case LayerType::Fused: + { + auto fusedDescriptor = *(PolymorphicDowncast<const FusedDescriptor*>(&descriptor)); + if (fusedDescriptor.m_NumInputSlots + fusedDescriptor.m_NumOutputSlots != infos.size()) + { + throw InvalidArgumentException("Invalid number of FusedLayer TensorInfos."); + } + + std::vector<TensorInfo> inputInfos(infos.begin(), infos.begin() + fusedDescriptor.m_NumInputSlots); + std::vector<TensorInfo> outputInfos(infos.begin() + fusedDescriptor.m_NumInputSlots, infos.end()); + + return support.IsFusedSupported({inputInfos.begin(), inputInfos.end()}, + {outputInfos.begin(), outputInfos.end()}, + fusedDescriptor, + reasonIfUnsupported); + } case LayerType::Gather: return support.IsGatherSupported(infos[0], infos[1], @@ -1155,6 +1170,19 @@ bool NeonLayerSupport::IsFullyConnectedSupported(const TensorInfo& input, nullptr); } +bool NeonLayerSupport::IsFusedSupported(const std::vector<std::reference_wrapper<TensorInfo>>& inputs, + const std::vector<std::reference_wrapper<TensorInfo>>& outputs, + const FusedDescriptor& descriptor, + Optional<std::string&> reasonIfUnsupported) const +{ + FORWARD_WORKLOAD_VALIDATE_FUNC(NeonFusedWorkloadValidate, + reasonIfUnsupported, + inputs, + outputs, + descriptor, + nullptr); +} + bool NeonLayerSupport::IsGatherSupported(const TensorInfo& input0, const TensorInfo& input1, const TensorInfo& output, |