aboutsummaryrefslogtreecommitdiff
path: root/src/backends/cl
diff options
context:
space:
mode:
Diffstat (limited to 'src/backends/cl')
-rw-r--r--src/backends/cl/CMakeLists.txt2
-rw-r--r--src/backends/cl/ClBackend.cpp24
-rw-r--r--src/backends/cl/ClBackend.hpp4
-rw-r--r--src/backends/cl/ClBackendContext.cpp17
-rw-r--r--src/backends/cl/ClBackendModelContext.cpp45
-rw-r--r--src/backends/cl/ClBackendModelContext.hpp23
-rw-r--r--src/backends/cl/ClLayerSupport.cpp33
-rw-r--r--src/backends/cl/ClLayerSupport.hpp9
-rw-r--r--src/backends/cl/backend.mk1
-rw-r--r--src/backends/cl/test/ClOptimizedNetworkTests.cpp30
-rw-r--r--src/backends/cl/workloads/ClConvolution2dWorkload.cpp7
-rw-r--r--src/backends/cl/workloads/ClConvolution2dWorkload.hpp3
12 files changed, 173 insertions, 25 deletions
diff --git a/src/backends/cl/CMakeLists.txt b/src/backends/cl/CMakeLists.txt
index f9f69f7729..4b5890af50 100644
--- a/src/backends/cl/CMakeLists.txt
+++ b/src/backends/cl/CMakeLists.txt
@@ -10,6 +10,8 @@ if(ARMCOMPUTECL)
ClBackendContext.cpp
ClBackendContext.hpp
ClBackendId.hpp
+ ClBackendModelContext.cpp
+ ClBackendModelContext.hpp
ClContextControl.cpp
ClContextControl.hpp
ClLayerSupport.cpp
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index f9a8993baf..49636d9b08 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -5,6 +5,7 @@
#include "ClBackend.hpp"
#include "ClBackendId.hpp"
+#include "ClBackendModelContext.hpp"
#include "ClWorkloadFactory.hpp"
#include "ClBackendContext.hpp"
#include "ClLayerSupport.hpp"
@@ -69,8 +70,7 @@ void ClBackend::RegisterTensorHandleFactories(TensorHandleFactoryRegistry& regis
registry.RegisterFactory(std::make_unique<ClTensorHandleFactory>(mgr));
}
-IBackendInternal::IBackendContextPtr
-ClBackend::CreateBackendContext(const IRuntime::CreationOptions& options) const
+IBackendInternal::IBackendContextPtr ClBackend::CreateBackendContext(const IRuntime::CreationOptions& options) const
{
return IBackendContextPtr{new ClBackendContext{options}};
}
@@ -86,9 +86,27 @@ IBackendInternal::Optimizations ClBackend::GetOptimizations() const
return Optimizations{};
}
+IBackendInternal::IBackendSpecificModelContextPtr ClBackend::CreateBackendSpecificModelContext(
+ const ModelOptions& modelOptions) const
+{
+ return IBackendSpecificModelContextPtr{new ClBackendModelContext{modelOptions}};
+}
+
IBackendInternal::ILayerSupportSharedPtr ClBackend::GetLayerSupport() const
{
- static ILayerSupportSharedPtr layerSupport{new ClLayerSupport};
+ static ILayerSupportSharedPtr layerSupport
+ {
+ new ClLayerSupport(IBackendInternal::IBackendSpecificModelContextPtr{})
+ };
+ return layerSupport;
+}
+
+IBackendInternal::ILayerSupportSharedPtr ClBackend::GetLayerSupport(const ModelOptions& modelOptions) const
+{
+ static ILayerSupportSharedPtr layerSupport
+ {
+ new ClLayerSupport(CreateBackendSpecificModelContext(modelOptions))
+ };
return layerSupport;
}
diff --git a/src/backends/cl/ClBackend.hpp b/src/backends/cl/ClBackend.hpp
index e85c616505..108124cac9 100644
--- a/src/backends/cl/ClBackend.hpp
+++ b/src/backends/cl/ClBackend.hpp
@@ -36,8 +36,12 @@ public:
IBackendInternal::Optimizations GetOptimizations() const override;
IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
+ IBackendInternal::ILayerSupportSharedPtr GetLayerSupport(const ModelOptions& modelOptions) const override;
OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph) const override;
+
+ IBackendInternal::IBackendSpecificModelContextPtr CreateBackendSpecificModelContext(
+ const ModelOptions& modelOptions) const override;
};
} // namespace armnn
diff --git a/src/backends/cl/ClBackendContext.cpp b/src/backends/cl/ClBackendContext.cpp
index 42f42b3023..22a4ceabd3 100644
--- a/src/backends/cl/ClBackendContext.cpp
+++ b/src/backends/cl/ClBackendContext.cpp
@@ -99,7 +99,6 @@ bool ParseBoolean(const BackendOptions::Var& value, bool defaultValue)
{
return value.AsBool();
}
-
return defaultValue;
}
@@ -112,22 +111,6 @@ std::string ParseFile(const BackendOptions::Var& value, std::string defaultValue
return defaultValue;
}
-template <typename F>
-void ParseOptions(const std::vector<BackendOptions>& options, BackendId backend, F f)
-{
- for (auto optionsGroup : options)
- {
- if (optionsGroup.GetBackendId() == backend)
- {
- for (size_t i=0; i < optionsGroup.GetOptionCount(); i++)
- {
- const BackendOptions::BackendOption option = optionsGroup.GetOption(i);
- f(option.GetName(), option.GetValue());
- }
- }
- }
-}
-
void ConfigureTuner(arm_compute::CLTuner &tuner, TuningLevel level)
{
tuner.set_tune_new_kernels(true); // Turn on tuning initially.
diff --git a/src/backends/cl/ClBackendModelContext.cpp b/src/backends/cl/ClBackendModelContext.cpp
new file mode 100644
index 0000000000..0ef26b64d2
--- /dev/null
+++ b/src/backends/cl/ClBackendModelContext.cpp
@@ -0,0 +1,45 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "ClBackendModelContext.hpp"
+
+namespace
+{
+
+bool ParseBool(const armnn::BackendOptions::Var& value, bool defaultValue)
+{
+ if (value.IsBool())
+ {
+ return value.AsBool();
+ }
+ return defaultValue;
+}
+
+} // namespace anonymous
+
+namespace armnn
+{
+
+ClBackendModelContext::ClBackendModelContext(const ModelOptions& modelOptions)
+ : m_IsFastMathEnabled(false)
+{
+ if (!modelOptions.empty())
+ {
+ ParseOptions(modelOptions, "GpuAcc", [&](std::string name, const BackendOptions::Var& value)
+ {
+ if (name == "FastMathEnabled")
+ {
+ m_IsFastMathEnabled |= ParseBool(value, false);
+ }
+ });
+ }
+}
+
+bool ClBackendModelContext::IsFastMathEnabled() const
+{
+ return m_IsFastMathEnabled;
+}
+
+} // namespace armnn \ No newline at end of file
diff --git a/src/backends/cl/ClBackendModelContext.hpp b/src/backends/cl/ClBackendModelContext.hpp
new file mode 100644
index 0000000000..59f7f8ff92
--- /dev/null
+++ b/src/backends/cl/ClBackendModelContext.hpp
@@ -0,0 +1,23 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#pragma once
+
+#include <armnn/backends/IBackendContext.hpp>
+
+namespace armnn
+{
+
+class ClBackendModelContext : public IBackendModelContext
+{
+public:
+ ClBackendModelContext(const ModelOptions& modelOptions);
+
+ bool IsFastMathEnabled() const;
+
+private:
+ bool m_IsFastMathEnabled;
+};
+
+} // namespace armnn \ No newline at end of file
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp
index 1af5c91fc5..7c1466e0e1 100644
--- a/src/backends/cl/ClLayerSupport.cpp
+++ b/src/backends/cl/ClLayerSupport.cpp
@@ -5,14 +5,17 @@
#include "ClLayerSupport.hpp"
#include "ClBackendId.hpp"
+#include "ClBackendModelContext.hpp"
-#include <armnn/utility/IgnoreUnused.hpp>
#include <armnn/Descriptors.hpp>
#include <armnn/BackendRegistry.hpp>
#include <InternalTypes.hpp>
#include <LayerSupportCommon.hpp>
+#include <armnn/utility/IgnoreUnused.hpp>
+#include <armnn/utility/PolymorphicDowncast.hpp>
+
#if defined(ARMCOMPUTECL_ENABLED)
#include <aclCommon/ArmComputeUtils.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
@@ -155,6 +158,16 @@ bool IsSupportedForDataTypeCl(Optional<std::string&> reasonIfUnsupported,
}
} // anonymous namespace
+ClLayerSupport::ClLayerSupport(const IBackendInternal::IBackendSpecificModelContextPtr& modelContextPtr)
+ : m_ModelContextPtr(modelContextPtr)
+{
+}
+
+ClLayerSupport::ClLayerSupport()
+ : m_ModelContextPtr(nullptr)
+{
+}
+
bool ClLayerSupport::IsAbsSupported(const TensorInfo& input,
const TensorInfo& output,
Optional<std::string&> reasonIfUnsupported) const
@@ -322,13 +335,29 @@ bool ClLayerSupport::IsConvolution2dSupported(const TensorInfo& input,
const Optional<TensorInfo>& biases,
Optional<std::string&> reasonIfUnsupported) const
{
+ bool isFastMathEnabled = false;
+#if defined(ARMCOMPUTECL_ENABLED)
+ if (m_ModelContextPtr)
+ {
+ if (m_ModelContextPtr.get() != nullptr)
+ {
+ auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get());
+ if (modelOptions)
+ {
+ isFastMathEnabled = modelOptions->IsFastMathEnabled();
+ }
+ }
+ }
+#endif
+
FORWARD_WORKLOAD_VALIDATE_FUNC(ClConvolution2dWorkloadValidate,
reasonIfUnsupported,
input,
output,
descriptor,
weights,
- biases);
+ biases,
+ isFastMathEnabled);
}
bool ClLayerSupport::IsDequantizeSupported(const TensorInfo& input,
diff --git a/src/backends/cl/ClLayerSupport.hpp b/src/backends/cl/ClLayerSupport.hpp
index ed0486e33b..d7e2553278 100644
--- a/src/backends/cl/ClLayerSupport.hpp
+++ b/src/backends/cl/ClLayerSupport.hpp
@@ -4,6 +4,8 @@
//
#pragma once
+#include <armnn/backends/IBackendInternal.hpp>
+
#include <backendsCommon/LayerSupportBase.hpp>
namespace armnn
@@ -12,6 +14,10 @@ namespace armnn
class ClLayerSupport : public LayerSupportBase
{
public:
+ explicit ClLayerSupport(const IBackendInternal::IBackendSpecificModelContextPtr& modelContextPtr);
+ ClLayerSupport();
+ ~ClLayerSupport() {}
+
ARMNN_DEPRECATED_MSG("Use IsElementwiseUnarySupported instead")
bool IsAbsSupported(const TensorInfo& input,
const TensorInfo& output,
@@ -318,6 +324,9 @@ public:
const TransposeDescriptor& descriptor,
Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+private:
+ const IBackendInternal::IBackendSpecificModelContextPtr m_ModelContextPtr;
+
};
} // namespace armnn
diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk
index 269057a952..9cbe21edca 100644
--- a/src/backends/cl/backend.mk
+++ b/src/backends/cl/backend.mk
@@ -16,6 +16,7 @@ ifeq ($(ARMNN_COMPUTE_CL_ENABLED),1)
BACKEND_SOURCES := \
ClBackend.cpp \
ClBackendContext.cpp \
+ ClBackendModelContext.cpp \
ClContextControl.cpp \
ClLayerSupport.cpp \
ClRegistryInitializer.cpp \
diff --git a/src/backends/cl/test/ClOptimizedNetworkTests.cpp b/src/backends/cl/test/ClOptimizedNetworkTests.cpp
index c2a8005e8a..2797080360 100644
--- a/src/backends/cl/test/ClOptimizedNetworkTests.cpp
+++ b/src/backends/cl/test/ClOptimizedNetworkTests.cpp
@@ -100,4 +100,34 @@ BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnGpuAcc)
BOOST_TEST(GraphHasNamedLayer(graph, "output layer"));
}
+BOOST_AUTO_TEST_CASE(FastMathEnabledTestOnGpuAcc)
+{
+ armnn::INetworkPtr net(armnn::INetwork::Create());
+
+ armnn::IConnectableLayer* input = net->AddInputLayer(0);
+ armnn::IConnectableLayer* output = net->AddOutputLayer(0);
+
+ input->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+ input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
+
+ armnn::IRuntime::CreationOptions options;
+ armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+ std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
+ armnn::OptimizerOptions optimizerOptions;
+ armnn::BackendOptions modelOptions("GpuAcc", {{"FastMathEnabled", true}});
+ optimizerOptions.m_ModelOptions.push_back(modelOptions);
+
+ armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(
+ *net, backends, runtime->GetDeviceSpec(), optimizerOptions);
+
+ BOOST_CHECK(optimizedNet);
+
+ auto modelOptionsOut = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetModelOptions();
+
+ BOOST_TEST(modelOptionsOut.size() == 1);
+ BOOST_TEST(modelOptionsOut[0].GetOption(0).GetName() == "FastMathEnabled");
+ BOOST_TEST(modelOptionsOut[0].GetOption(0).GetValue().AsBool() == true);
+}
+
BOOST_AUTO_TEST_SUITE_END();
diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
index 73ec95ce9f..42c9903dc4 100644
--- a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
+++ b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp
@@ -24,7 +24,8 @@ arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
- const Optional<TensorInfo>& biases)
+ const Optional<TensorInfo>& biases,
+ bool isFastMathEnabled)
{
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
@@ -52,7 +53,9 @@ arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
&aclOutputInfo,
layerInfo,
arm_compute::WeightsInfo(),
- aclDilationInfo);
+ aclDilationInfo,
+ arm_compute::ActivationLayerInfo(),
+ isFastMathEnabled);
}
ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor,
diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp
index 6d7e9f3ea1..8b0afada36 100644
--- a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp
+++ b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp
@@ -22,7 +22,8 @@ arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
- const Optional<TensorInfo>& biases);
+ const Optional<TensorInfo>& biases,
+ bool isFastMathEnabled = false);
class ClConvolution2dWorkload : public BaseWorkload<Convolution2dQueueDescriptor>
{
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-29 23:23:46 +0000