aboutsummaryrefslogtreecommitdiff
path: root/src/backends/cl
diff options
context:
space:
mode:
Diffstat (limited to 'src/backends/cl')
-rw-r--r--src/backends/cl/ClBackend.cpp4
-rw-r--r--src/backends/cl/test/ClCreateWorkloadTests.cpp4
-rw-r--r--src/backends/cl/test/ClImportTensorHandleTests.cpp8
-rw-r--r--src/backends/cl/workloads/ClConvolution2dWorkload.cpp19
4 files changed, 29 insertions, 6 deletions
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index 47990d87dc..bd1b94e79f 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -341,14 +341,14 @@ OptimizationViews ClBackend::OptimizeSubgraphView(const SubgraphView& subgraph,
if (baseLayer->GetParameters().m_BiasEnabled)
{
- biases = baseLayer->m_Bias->GetTensorInfo();
+ biases = baseLayer->GetInputSlot(2).GetConnectedOutputSlot()->GetTensorInfo();
}
arm_compute::Status status = ClConvolution2dWorkloadValidate(
baseLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(),
activationLayer->GetInputSlot(0).GetConnectedOutputSlot()->GetTensorInfo(),
baseLayer->GetParameters(),
- baseLayer->m_Weight->GetTensorInfo(),
+ baseLayer->GetInputSlot(1).GetConnectedOutputSlot()->GetTensorInfo(),
biases,
isFastMathEnabled,
&activationDesc);
diff --git a/src/backends/cl/test/ClCreateWorkloadTests.cpp b/src/backends/cl/test/ClCreateWorkloadTests.cpp
index 4f53b921d0..3a757f8820 100644
--- a/src/backends/cl/test/ClCreateWorkloadTests.cpp
+++ b/src/backends/cl/test/ClCreateWorkloadTests.cpp
@@ -471,6 +471,8 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "CreateConvolution2dClCompiledContext
auto tensorHandleFactory = ClWorkloadFactoryHelper::GetTensorHandleFactory(memoryManager);
std::unique_ptr<ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputInfo);
+ std::unique_ptr<armnn::ITensorHandle> weightsHandle = tensorHandleFactory.CreateTensorHandle(kernelInfo);
+ std::unique_ptr<armnn::ITensorHandle> biasHandle = tensorHandleFactory.CreateTensorHandle(biasInfo);
std::unique_ptr<ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputInfo);
@@ -487,6 +489,8 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "CreateConvolution2dClCompiledContext
queueDescriptor.m_Bias = &biasTensor;
AddInputToWorkload(queueDescriptor, workloadInfo, inputInfo, inputHandle.get());
+ AddInputToWorkload(queueDescriptor, workloadInfo, kernelInfo, weightsHandle.get());
+ AddInputToWorkload(queueDescriptor, workloadInfo, biasInfo, biasHandle.get());
AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get());
// Initialize our m_CLCompileContext using default device and context
diff --git a/src/backends/cl/test/ClImportTensorHandleTests.cpp b/src/backends/cl/test/ClImportTensorHandleTests.cpp
index e10e81ac26..9bfd1fb46d 100644
--- a/src/backends/cl/test/ClImportTensorHandleTests.cpp
+++ b/src/backends/cl/test/ClImportTensorHandleTests.cpp
@@ -11,7 +11,6 @@
#include <doctest/doctest.h>
-
#include <armnn/IRuntime.hpp>
#include <armnn/INetwork.hpp>
#include "Network.hpp"
@@ -320,10 +319,13 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportConv2dEndToEnd")
convDesc2d.m_PadTop = 1;
convDesc2d.m_PadBottom = 1;
convDesc2d.m_DataLayout = DataLayout::NHWC;
+
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d,
weights,
armnn::EmptyOptional(),
"conv");
+ ARMNN_NO_DEPRECATE_WARN_END
ARMNN_ASSERT(convLayer);
inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
@@ -876,10 +878,12 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesEndTo
convDesc2d.m_PadTop = 1;
convDesc2d.m_PadBottom = 1;
convDesc2d.m_DataLayout = DataLayout::NHWC;
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d,
weights,
armnn::EmptyOptional(),
"conv");
+ ARMNN_NO_DEPRECATE_WARN_END
ARMNN_ASSERT(convLayer);
inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
@@ -1094,10 +1098,12 @@ TEST_CASE_FIXTURE(ClContextControlFixture, "ClForceImportRepeatedInferencesInver
convDesc2d.m_PadTop = 1;
convDesc2d.m_PadBottom = 1;
convDesc2d.m_DataLayout = DataLayout::NHWC;
+ ARMNN_NO_DEPRECATE_WARN_BEGIN
armnn::IConnectableLayer* const convLayer = network->AddConvolution2dLayer(convDesc2d,
weights,
armnn::EmptyOptional(),
"conv");
+ ARMNN_NO_DEPRECATE_WARN_END
ARMNN_ASSERT(convLayer);
inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
index bf82fbf255..e3d679a773 100644
--- a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
+++ b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
@@ -28,6 +28,15 @@ arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
bool isFastMathEnabled,
const ActivationDescriptor* activationDescriptor)
{
+ // The implemented workload does support both const and non const
+ // weights. However, in the case of non const weights we'd have to call
+ // prepare or configure for each inference which we're not setup to do just yet.
+ if (!weights.IsConstant())
+ {
+ return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
+ "ArmNN ClConvolution2dWorkload does not support non constant weights."};
+ }
+
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
@@ -41,7 +50,12 @@ arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
-
+ // Same for bias as weights. We don't currently support non const.
+ if (!biases.value().IsConstant())
+ {
+ return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
+ "ArmNN ClConvolution2dWorkload does not support non constant bias."};
+ }
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
@@ -72,6 +86,7 @@ ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescrip
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload");
const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
+ m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1);
m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
@@ -85,8 +100,6 @@ ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescrip
BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
}
- m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1);
-
arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-05 11:15:42 +0000