#DEFAULT_CLFRAMEWORKREVISION="branches/arm_compute_19_08" # Release 19.08
#
# For pinning to a revision use this:
-DEFAULT_CLFRAMEWORKREVISION="79f88e6d825402388bb79fc123ee2dfe01985bda" #COMPMID-2313: Implement CL INSTANCE_NORMALIZATION function
+DEFAULT_CLFRAMEWORKREVISION="94e0cf960ea6116eb57fa88d9b951f859b52c602" #COMPMID-2690 Extend Doxygen documents to include GEMM Tuner
usage() {
echo "Usage: $CMD (Use the default clframework SHA)"
DummyConvolutionLayer()
{
typename ConvolutionLayerType::DescriptorType desc;
+ desc.m_StrideX = 1;
+ desc.m_StrideY = 1;
m_Layer = dummyGraph.AddLayer<ConvolutionLayerType>(desc, "");
m_Layer->m_Weight = std::make_unique<armnn::ScopedCpuTensorHandle>(
armnn::TensorInfo(armnn::TensorShape({1,1,1,1}), armnn::DataType::Float32));
cl::Context context;
cl::CommandQueue commandQueue;
- if (arm_compute::CLScheduler::get().context()() != NULL)
+ if (arm_compute::CLScheduler::get().is_initialised() && arm_compute::CLScheduler::get().context()() != NULL)
{
// Wait for all queued CL requests to finish before reinitialising it.
arm_compute::CLScheduler::get().sync();
bool use3x3Optimisation = (weightInfo.GetShape()[2] == 3) && (weightInfo.GetShape()[3] == 3);
if (use3x3Optimisation)
{
- m_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
- static_cast<arm_compute::CLDepthwiseConvolutionLayer3x3*>(m_DepthwiseConvolutionLayer.get())->configure(
+ m_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
+ static_cast<arm_compute::CLDepthwiseConvolutionLayer*>(m_DepthwiseConvolutionLayer.get())->configure(
&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, m_Data.m_Parameters.m_Axis);
int aclAxis = boost::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis));
+ auto layer = std::make_unique<arm_compute::NEArgMinMaxLayer>();
+
if (m_Data.m_Parameters.m_Function == ArgMinMaxFunction::Max)
{
- m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MAX);
+ layer->configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MAX);
}
else
{
- m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MIN);
+ layer->configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MIN);
}
+
+ m_ArgMinMaxLayer.reset(layer.release());
}
void NeonArgMinMaxWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonArgMinMaxWorkload_Execute");
- m_ArgMinMaxLayer.run();
+ m_ArgMinMaxLayer->run();
}
} //namespace armnn
#include <backendsCommon/Workload.hpp>
#include <arm_compute/core/Error.h>
-#include <arm_compute/runtime/NEON/functions/NEArgMinMaxLayer.h>
+#include <arm_compute/runtime/IFunction.h>
+
namespace armnn
{
virtual void Execute() const override;
private:
- mutable arm_compute::NEArgMinMaxLayer m_ArgMinMaxLayer;
+ std::unique_ptr<arm_compute::IFunction> m_ArgMinMaxLayer;
};
} //namespace armnn
// Check for optimisation opportunities
arm_compute::Status optimizationStatus =
- arm_compute::NEDepthwiseConvolutionLayerOptimized::validate(inputInfo,
- kernelInfo,
- biasInfo,
- outputInfo,
- padStrideInfo,
- depthMultiplier,
- arm_compute::ActivationLayerInfo(),
- aclDilationInfo);
+ arm_compute::NEDepthwiseConvolutionLayer::validate(inputInfo,
+ kernelInfo,
+ biasInfo,
+ outputInfo,
+ padStrideInfo,
+ depthMultiplier,
+ arm_compute::ActivationLayerInfo(),
+ aclDilationInfo);
if (optimizationStatus.error_code() == arm_compute::ErrorCode::OK)
{
- m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayerOptimized>();
- static_cast<arm_compute::NEDepthwiseConvolutionLayerOptimized*>(
+ m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
+ static_cast<arm_compute::NEDepthwiseConvolutionLayer*>(
m_pDepthwiseConvolutionLayer.get())->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),