IVGCVSW-2019 : replace Compute enum in the backend preferences list
Change-Id: Ie7549fd27378acfa97e68d098e338b8c9d4ea5d2
diff --git a/include/armnn/ArmNN.hpp b/include/armnn/ArmNN.hpp
index c14b958..6fe8b18 100644
--- a/include/armnn/ArmNN.hpp
+++ b/include/armnn/ArmNN.hpp
@@ -4,6 +4,7 @@
//
#pragma once
+#include "BackendId.hpp"
#include "Descriptors.hpp"
#include "Exceptions.hpp"
#include "IRuntime.hpp"
diff --git a/include/armnn/BackendId.hpp b/include/armnn/BackendId.hpp
index 711833d..72248bc 100644
--- a/include/armnn/BackendId.hpp
+++ b/include/armnn/BackendId.hpp
@@ -75,6 +75,7 @@
class BackendId final
{
public:
+ BackendId() { GetComputeDeviceAsCString(Compute::Undefined); }
BackendId(UninitializedBackendId) { GetComputeDeviceAsCString(Compute::Undefined); }
BackendId(const std::string& id) : m_Id{id} {}
BackendId(const char* id) : m_Id{id} {}
@@ -128,17 +129,21 @@
const std::string& Get() const { return m_Id; }
private:
- // backend Id mustn't be empty:
- BackendId() = delete;
std::string m_Id;
};
+inline std::ostream& operator<<(std::ostream& os, const BackendId& id)
+{
+ os << id.Get();
+ return os;
+}
+
template <template <class...> class TContainer>
inline std::ostream& operator<<(std::ostream& os,
const TContainer<BackendId>& ids)
{
os << '[';
- for (const auto& id : ids) { os << id.Get() << " "; }
+ for (const auto& id : ids) { os << id << " "; }
os << ']';
return os;
}
diff --git a/include/armnn/INetwork.hpp b/include/armnn/INetwork.hpp
index aaf49a3..7a80935 100644
--- a/include/armnn/INetwork.hpp
+++ b/include/armnn/INetwork.hpp
@@ -4,12 +4,12 @@
//
#pragma once
-#include "armnn/NetworkFwd.hpp"
-#include "armnn/DescriptorsFwd.hpp"
-#include "armnn/TensorFwd.hpp"
-#include "armnn/Optional.hpp"
+#include <armnn/NetworkFwd.hpp>
+#include <armnn/DescriptorsFwd.hpp>
+#include <armnn/TensorFwd.hpp>
+#include <armnn/Optional.hpp>
-#include "armnn/Types.hpp"
+#include <armnn/Types.hpp>
#include <memory>
#include <vector>
@@ -339,7 +339,7 @@
/// armnn::Exception if process fails.
IOptimizedNetworkPtr Optimize(const INetwork& network,
- const std::vector<armnn::Compute>& backendPreferences,
+ const std::vector<BackendId>& backendPreferences,
const IDeviceSpec& deviceSpec,
const OptimizerOptions& options = OptimizerOptions(),
Optional<std::vector<std::string>&> errMessages = EmptyOptional());
diff --git a/src/armnn/DeviceSpec.hpp b/src/armnn/DeviceSpec.hpp
index 34acbcb..af0d8f5 100644
--- a/src/armnn/DeviceSpec.hpp
+++ b/src/armnn/DeviceSpec.hpp
@@ -22,7 +22,7 @@
return std::vector<IBackendSharedPtr>();
}
- std::set<Compute> m_SupportedComputeDevices;
+ std::set<BackendId> m_SupportedComputeDevices;
};
}
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index c43f336..8c70e5d 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -94,7 +94,7 @@
}
IOptimizedNetworkPtr Optimize(const INetwork& inNetwork,
- const std::vector<armnn::Compute>& backendPreferences,
+ const std::vector<BackendId>& backendPreferences,
const IDeviceSpec& deviceSpec,
const OptimizerOptions& options,
Optional<std::vector<std::string>&> errMessages)
@@ -133,8 +133,8 @@
// determine which of the preferred backends we have available for use
// and whether we have specified CpuRef as one of those backends.
bool cpuRefUsed = false;
- std::vector<armnn::Compute> availablePreferredBackends;
- for (const armnn::Compute& backend : backendPreferences)
+ std::vector<BackendId> availablePreferredBackends;
+ for (const auto& backend : backendPreferences)
{
// Check if the backend is in the available backend devices.
if (std::find(spec.m_SupportedComputeDevices.begin(),
@@ -142,7 +142,7 @@
spec.m_SupportedComputeDevices.end())
{
availablePreferredBackends.push_back(backend);
- if (armnn::Compute::CpuRef == backend) {
+ if (backend == armnn::Compute::CpuRef) {
cpuRefUsed = true;
}
}
@@ -183,7 +183,7 @@
// which haven't had a scale set and report them all back.
bErrorFound = true;
}
- for (const armnn::Compute& backend : availablePreferredBackends)
+ for (const auto& backend : availablePreferredBackends)
{
// need to set the compute device on the layer
// before we can check if it is supported
@@ -205,7 +205,7 @@
InsertConvertFp32ToFp16LayersAfter(optNetObjPtr->GetGraph(), *layer);
// Assign a supported backend to the newly introduced conversion layers
- auto AssignFirstSupportedBackend = [&](Layer* layer, Compute preferredBackend)
+ auto AssignFirstSupportedBackend = [&](Layer* layer, BackendId preferredBackend)
{
bool supportedBackendFound = false;
std::string reasonIfUnsupported;
@@ -218,7 +218,7 @@
}
else
{
- for (const Compute& backend : availablePreferredBackends)
+ for (const auto& backend : availablePreferredBackends)
{
// Skip preferred backend (we already determined that it is not supported)
if (backend == preferredBackend)
diff --git a/src/armnn/test/EndToEndTest.cpp b/src/armnn/test/EndToEndTest.cpp
index 98b1841..d34bf69 100644
--- a/src/armnn/test/EndToEndTest.cpp
+++ b/src/armnn/test/EndToEndTest.cpp
@@ -74,7 +74,7 @@
softmax->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
// optimize the network
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Loads it into the runtime.
@@ -110,7 +110,7 @@
}
template <typename T>
-void ConstantUsageTest(const std::vector<armnn::Compute>& computeDevice,
+void ConstantUsageTest(const std::vector<armnn::BackendId>& computeDevice,
const armnn::TensorInfo& commonTensorInfo,
const std::vector<T>& inputData,
const std::vector<T>& constantData,
@@ -165,7 +165,7 @@
BOOST_TEST(outputData == expectedOutputData);
}
-static void ConstantUsageFloat32Test(const std::vector<armnn::Compute>& computeDevice)
+static void ConstantUsageFloat32Test(const std::vector<armnn::BackendId>& computeDevice)
{
const armnn::TensorInfo commonTensorInfo({ 2, 3 }, armnn::DataType::Float32);
@@ -177,7 +177,7 @@
);
}
-static void ConstantUsageUint8Test(const std::vector<armnn::Compute>& computeDevice)
+static void ConstantUsageUint8Test(const std::vector<armnn::BackendId>& computeDevice)
{
armnn::TensorInfo commonTensorInfo({ 2, 3 }, armnn::DataType::QuantisedAsymm8);
@@ -197,7 +197,7 @@
BOOST_AUTO_TEST_CASE(ConstantUsage_Ref_Float32)
{
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
ConstantUsageFloat32Test(backends);
}
@@ -217,7 +217,7 @@
BOOST_AUTO_TEST_CASE(ConstantUsage_Ref_Uint8)
{
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
ConstantUsageUint8Test(backends);
}
@@ -250,7 +250,7 @@
add->GetOutputSlot(0).SetTensorInfo(tensorInfo);
// optimize the network
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Loads it into the runtime.
@@ -349,7 +349,7 @@
activation3->GetOutputSlot(0).SetTensorInfo(tensorInfo);
// optimize the network
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Loads it into the runtime.
@@ -410,7 +410,7 @@
pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
// optimize the network
- std::vector<Compute> backends = {Compute::CpuAcc, Compute::CpuRef};
+ std::vector<BackendId> backends = {Compute::CpuAcc, Compute::CpuRef};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Load it into the runtime. It should pass.
@@ -446,7 +446,7 @@
pooling->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
// optimize the network
- std::vector<Compute> backends = {Compute::CpuAcc};
+ std::vector<BackendId> backends = {Compute::CpuAcc};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(!optNet);
}
diff --git a/src/armnn/test/JsonPrinterTests.cpp b/src/armnn/test/JsonPrinterTests.cpp
index 44609ea..93f32cc 100644
--- a/src/armnn/test/JsonPrinterTests.cpp
+++ b/src/armnn/test/JsonPrinterTests.cpp
@@ -117,7 +117,7 @@
return sections;
}
-std::string SoftmaxProfilerTestSetupHelper(const std::vector<armnn::Compute>& backends)
+std::string SoftmaxProfilerTestSetupHelper(const std::vector<armnn::BackendId>& backends)
{
using namespace armnn;
@@ -239,7 +239,7 @@
}
void SetupSoftmaxProfilerWithSpecifiedBackendsAndValidateJSONPrinterResult(
- const std::vector<armnn::Compute>& backends)
+ const std::vector<armnn::BackendId>& backends)
{
// setup the test fixture and obtain JSON Printer result
std::string result = SoftmaxProfilerTestSetupHelper(backends);
@@ -250,10 +250,10 @@
std::string changeLine40;
std::string changeLine45;
- switch(backends[0]) {
- case armnn::Compute::GpuAcc: backend = "Cl";
- changeLine31 = ",\n\"OpenClKernelTimer/: softmax_layer_max_shift_exp_sum_quantized_serial GWS[,,]\": {";
- changeLine39 = R"(us"
+ if (backends[0] == armnn::Compute::GpuAcc) {
+ backend = "Cl";
+ changeLine31 = ",\n\"OpenClKernelTimer/: softmax_layer_max_shift_exp_sum_quantized_serial GWS[,,]\": {";
+ changeLine39 = R"(us"
},
"OpenClKernelTimer/: softmax_layer_norm_quantized GWS[,,]": {
"raw": [
@@ -263,7 +263,7 @@
],
"unit": "us")";
- changeLine40 = R"(
+ changeLine40 = R"(
},
"CopyMemGeneric_Execute": {
"raw": [
@@ -272,11 +272,13 @@
],
"unit": "us")";
- changeLine45 = "}\n";
- break;
- case armnn::Compute::CpuAcc: backend = "Neon";
- changeLine31 = ",\n\"NeonKernelTimer/: NEFillBorderKernel\": {";
- changeLine39 = R"(us"
+ changeLine45 = "}\n";
+ }
+ else if (backends[0] == armnn::Compute::CpuAcc)
+ {
+ backend = "Neon";
+ changeLine31 = ",\n\"NeonKernelTimer/: NEFillBorderKernel\": {";
+ changeLine39 = R"(us"
},
"NeonKernelTimer/: NELogitsDMaxKernel": {
"raw": [
@@ -293,7 +295,7 @@
],
"unit": "us")";
- changeLine40 = R"(
+ changeLine40 = R"(
},
"CopyMemGeneric_Execute": {
"raw": [
@@ -302,11 +304,9 @@
],
"unit": "us")";
- changeLine45 = "}\n";
- break;
- default:
- break;
+ changeLine45 = "}\n";
}
+
std::string testData = R"({
"ArmNN": {
"inference_measurements": {
diff --git a/src/armnn/test/NetworkTests.cpp b/src/armnn/test/NetworkTests.cpp
index 3b426fa..4f8dd7e 100644
--- a/src/armnn/test/NetworkTests.cpp
+++ b/src/armnn/test/NetworkTests.cpp
@@ -65,7 +65,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
std::ostringstream ss;
@@ -472,7 +472,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph().AllocateDynamicBuffers();
BOOST_CHECK(optNet);
@@ -503,7 +503,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(optNet);
// validate workloads
@@ -534,7 +534,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::GpuAcc };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(optNet);
// validate workloads
@@ -570,7 +570,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(!optNet);
}
@@ -597,7 +597,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_REQUIRE(optNet);
@@ -676,7 +676,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::Undefined };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(!optNet);
@@ -738,7 +738,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::Undefined, armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::Undefined, armnn::Compute::CpuRef };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(optNet);
@@ -774,7 +774,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc,
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc,
armnn::Compute::GpuAcc,
armnn::Compute::CpuRef };
@@ -818,7 +818,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
@@ -851,7 +851,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
// build up the structure of the network
armnn::INetworkPtr net(armnn::INetwork::Create());
@@ -909,7 +909,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
armnn::OptimizerOptions optimizerOptions;
optimizerOptions.m_ReduceFp32ToFp16 = true;
@@ -976,7 +976,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = {armnn::Compute::GpuAcc};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
armnn::OptimizerOptions optimizerOptions;
optimizerOptions.m_ReduceFp32ToFp16 = true;
diff --git a/src/armnn/test/ProfilerTests.cpp b/src/armnn/test/ProfilerTests.cpp
index 8d76844..650c13a 100644
--- a/src/armnn/test/ProfilerTests.cpp
+++ b/src/armnn/test/ProfilerTests.cpp
@@ -163,7 +163,7 @@
armnn::NetworkId networkIdentifier = 1;
armnn::INetworkPtr mockNetwork(armnn::INetwork::Create());
mockNetwork->AddInputLayer(0, "test layer");
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
runtime->LoadNetwork(networkIdentifier, armnn::Optimize(*mockNetwork, backends, runtime->GetDeviceSpec()));
// Check that now there's a profiler registered for this thread (created and registered by the loading the network).
diff --git a/src/armnn/test/RuntimeTests.cpp b/src/armnn/test/RuntimeTests.cpp
index 0237387..76f5774 100644
--- a/src/armnn/test/RuntimeTests.cpp
+++ b/src/armnn/test/RuntimeTests.cpp
@@ -39,7 +39,7 @@
armnn::NetworkId networkIdentifier1 = 1;
armnn::INetworkPtr mockNetwork1(armnn::INetwork::Create());
mockNetwork1->AddInputLayer(0, "test layer");
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
runtime->LoadNetwork(networkIdentifier1, Optimize(*mockNetwork1, backends, runtime->GetDeviceSpec()));
// Mock network 2.
@@ -71,7 +71,7 @@
BOOST_GLOBAL_FIXTURE(DisableGlobalLeakChecking);
-void CreateAndDropDummyNetwork(const std::vector<armnn::Compute>& backends, armnn::Runtime& runtime)
+void CreateAndDropDummyNetwork(const std::vector<armnn::BackendId>& backends, armnn::Runtime& runtime)
{
armnn::NetworkId networkIdentifier;
{
@@ -129,7 +129,7 @@
armnn::Runtime runtime(options);
armnn::RuntimeLoadedNetworksReserve(&runtime);
- std::vector<armnn::Compute> backends = {armnn::Compute::GpuAcc};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
{
// Do a warmup of this so we make sure that all one-time
// initialization happens before we do the leak checking.
@@ -158,7 +158,7 @@
armnn::Runtime runtime(options);
armnn::RuntimeLoadedNetworksReserve(&runtime);
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuAcc};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuAcc};
{
// Do a warmup of this so we make sure that all one-time
// initialization happens before we do the leak checking.
@@ -187,7 +187,7 @@
armnn::Runtime runtime(options);
armnn::RuntimeLoadedNetworksReserve(&runtime);
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
{
// Do a warmup of this so we make sure that all one-time
// initialization happens before we do the leak checking.
@@ -243,7 +243,7 @@
VALGRIND_COUNT_LEAKS(leakedBefore, dubious, reachableBefore, suppressed);
// build a mock-network and load it into the runtime
- std::vector<armnn::Compute> backends = {armnn::Compute::GpuAcc};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::GpuAcc};
{
armnn::TensorInfo inputTensorInfo(armnn::TensorShape({ 7, 7 }), armnn::DataType::Float32);
armnn::TensorInfo outputTensorInfo(armnn::TensorShape({ 7, 7 }), armnn::DataType::Float32);
@@ -331,7 +331,7 @@
mockNetwork1->AddInputLayer(0, "test layer");
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
runtime.LoadNetwork(networkIdentifier1, Optimize(*mockNetwork1, backends, runtime.GetDeviceSpec()));
}
@@ -379,7 +379,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(optNet);
@@ -406,7 +406,7 @@
armnn::IRuntime::CreationOptions options;
armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
- std::vector<armnn::Compute> backends = { armnn::Compute::GpuAcc };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::GpuAcc };
armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
BOOST_CHECK(optNet);
@@ -442,7 +442,7 @@
normalize->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
// optimize the network
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Load it into the runtime. It should success.
@@ -476,7 +476,7 @@
normalize->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 4 }, DataType::Float32));
// Allow fallback to CpuRef.
- std::vector<armnn::Compute> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
+ std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc, armnn::Compute::CpuRef };
// optimize the network
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
@@ -524,7 +524,7 @@
armnn::DataType::QuantisedAsymm8
));
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
std::vector<std::string> errMessages;
armnn::IOptimizedNetworkPtr optNet = Optimize(
*net,
@@ -533,7 +533,7 @@
OptimizerOptions(),
errMessages
);
-
+
BOOST_TEST(errMessages.size() == 1);
BOOST_TEST(errMessages[0] ==
"ERROR: output 0 of layer Softmax (softmax) is of type "
diff --git a/src/armnnCaffeParser/test/TestInputs.cpp b/src/armnnCaffeParser/test/TestInputs.cpp
index 616d75d..bd13a09 100644
--- a/src/armnnCaffeParser/test/TestInputs.cpp
+++ b/src/armnnCaffeParser/test/TestInputs.cpp
@@ -42,7 +42,7 @@
armnn::NetworkId netId;
// Check everything works normally
- std::vector<armnn::Compute> backends = {armnn::Compute::CpuRef};
+ std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
{
network = parser->CreateNetworkFromString(explicitInput.c_str(), {}, { "data" });
BOOST_TEST(network.get());
diff --git a/src/armnnTfLiteParser/test/ParserFlatbuffersFixture.hpp b/src/armnnTfLiteParser/test/ParserFlatbuffersFixture.hpp
index 404cbf0..e9cbb24 100644
--- a/src/armnnTfLiteParser/test/ParserFlatbuffersFixture.hpp
+++ b/src/armnnTfLiteParser/test/ParserFlatbuffersFixture.hpp
@@ -44,7 +44,7 @@
std::vector<uint8_t> m_GraphBinary;
std::string m_JsonString;
std::unique_ptr<ITfLiteParser, void (*)(ITfLiteParser *parser)> m_Parser;
- std::vector<std::pair<armnn::IRuntimePtr, armnn::Compute>> m_Runtimes;
+ std::vector<std::pair<armnn::IRuntimePtr, armnn::BackendId>> m_Runtimes;
armnn::NetworkId m_NetworkIdentifier;
/// If the single-input-single-output overload of Setup() is called, these will store the input and output name
diff --git a/src/armnnUtils/ParserPrototxtFixture.hpp b/src/armnnUtils/ParserPrototxtFixture.hpp
index c502ad9..1adada7 100644
--- a/src/armnnUtils/ParserPrototxtFixture.hpp
+++ b/src/armnnUtils/ParserPrototxtFixture.hpp
@@ -59,7 +59,7 @@
std::string m_Prototext;
std::unique_ptr<TParser, void(*)(TParser* parser)> m_Parser;
- std::vector<std::pair<armnn::IRuntimePtr, armnn::Compute>> m_Runtimes;
+ std::vector<std::pair<armnn::IRuntimePtr, armnn::BackendId>> m_Runtimes;
armnn::NetworkId m_NetworkIdentifier;
/// If the single-input-single-output overload of Setup() is called, these will store the input and output name
diff --git a/src/backends/cl/test/Fp16SupportTest.cpp b/src/backends/cl/test/Fp16SupportTest.cpp
index 90bef36..e8d4591 100644
--- a/src/backends/cl/test/Fp16SupportTest.cpp
+++ b/src/backends/cl/test/Fp16SupportTest.cpp
@@ -73,7 +73,7 @@
additionLayer->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
// optimize the network
- std::vector<Compute> backends = {Compute::GpuAcc};
+ std::vector<BackendId> backends = {Compute::GpuAcc};
IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
// Loads it into the runtime.
diff --git a/tests/ExecuteNetwork/ExecuteNetwork.cpp b/tests/ExecuteNetwork/ExecuteNetwork.cpp
index ee20747..7f1bcd3 100644
--- a/tests/ExecuteNetwork/ExecuteNetwork.cpp
+++ b/tests/ExecuteNetwork/ExecuteNetwork.cpp
@@ -147,7 +147,7 @@
printf("\n");
}
-void RemoveDuplicateDevices(std::vector<armnn::Compute>& computeDevices)
+void RemoveDuplicateDevices(std::vector<armnn::BackendId>& computeDevices)
{
// Mark the duplicate devices as 'Undefined'.
for (auto i = computeDevices.begin(); i != computeDevices.end(); ++i)
@@ -166,11 +166,11 @@
computeDevices.end());
}
-bool CheckDevicesAreValid(const std::vector<armnn::Compute>& computeDevices)
+bool CheckDevicesAreValid(const std::vector<armnn::BackendId>& computeDevices)
{
return (!computeDevices.empty()
&& std::none_of(computeDevices.begin(), computeDevices.end(),
- [](armnn::Compute c){ return c == armnn::Compute::Undefined; }));
+ [](armnn::BackendId c){ return c == armnn::Compute::Undefined; }));
}
} // namespace
@@ -178,7 +178,7 @@
template<typename TParser, typename TDataType>
int MainImpl(const char* modelPath,
bool isModelBinary,
- const std::vector<armnn::Compute>& computeDevice,
+ const std::vector<armnn::BackendId>& computeDevice,
const char* inputName,
const armnn::TensorShape* inputTensorShape,
const char* inputTensorDataFilePath,
@@ -232,7 +232,7 @@
// This will run a test
int RunTest(const std::string& modelFormat,
const std::string& inputTensorShapeStr,
- const vector<armnn::Compute>& computeDevice,
+ const vector<armnn::BackendId>& computeDevice,
const std::string& modelPath,
const std::string& inputName,
const std::string& inputTensorDataFilePath,
@@ -360,7 +360,7 @@
"caffe-binary, caffe-text, tflite-binary, onnx-binary, onnx-text, tensorflow-binary or tensorflow-text.")
("model-path,m", po::value(&modelPath), "Path to model file, e.g. .caffemodel, .prototxt, .tflite,"
" .onnx")
- ("compute,c", po::value<std::vector<armnn::Compute>>()->multitoken(),
+ ("compute,c", po::value<std::vector<armnn::BackendId>>()->multitoken(),
"The preferred order of devices to run layers on by default. Possible choices: CpuAcc, CpuRef, GpuAcc")
("input-name,i", po::value(&inputName), "Identifier of the input tensor in the network.")
("subgraph-number,n", po::value<size_t>(&subgraphId)->default_value(0), "Id of the subgraph to be "
@@ -414,7 +414,7 @@
boost::trim(outputName);
// Get the preferred order of compute devices.
- std::vector<armnn::Compute> computeDevices = vm["compute"].as<std::vector<armnn::Compute>>();
+ std::vector<armnn::BackendId> computeDevices = vm["compute"].as<std::vector<armnn::BackendId>>();
// Remove duplicates from the list of compute devices.
RemoveDuplicateDevices(computeDevices);
@@ -466,7 +466,7 @@
"caffe-binary, caffe-text, onnx-binary, onnx-text, tflite-binary, tensorflow-binary or tensorflow-text.")
("model-path,m", po::value(&modelPath), "Path to model file, e.g. .caffemodel, .prototxt,"
" .tflite, .onnx")
- ("compute,c", po::value<std::vector<armnn::Compute>>()->multitoken(),
+ ("compute,c", po::value<std::vector<std::string>>()->multitoken(),
"The preferred order of devices to run layers on by default. Possible choices: CpuAcc, CpuRef, GpuAcc")
("input-name,i", po::value(&inputName), "Identifier of the input tensor in the network.")
("subgraph-number,x", po::value<size_t>(&subgraphId)->default_value(0), "Id of the subgraph to be executed."
@@ -588,7 +588,7 @@
else // Run single test
{
// Get the preferred order of compute devices.
- std::vector<armnn::Compute> computeDevices = vm["compute"].as<std::vector<armnn::Compute>>();
+ std::vector<armnn::BackendId> computeDevices = vm["compute"].as<std::vector<armnn::BackendId>>();
// Remove duplicates from the list of compute devices.
RemoveDuplicateDevices(computeDevices);
diff --git a/tests/InferenceModel.hpp b/tests/InferenceModel.hpp
index 2e0aff9..8645c90 100644
--- a/tests/InferenceModel.hpp
+++ b/tests/InferenceModel.hpp
@@ -3,15 +3,15 @@
// SPDX-License-Identifier: MIT
//
#pragma once
-#include "armnn/ArmNN.hpp"
+#include <armnn/ArmNN.hpp>
#if defined(ARMNN_TF_LITE_PARSER)
-#include "armnnTfLiteParser/ITfLiteParser.hpp"
+#include <armnnTfLiteParser/ITfLiteParser.hpp>
#endif
#include <HeapProfiling.hpp>
#if defined(ARMNN_ONNX_PARSER)
-#include "armnnOnnxParser/IOnnxParser.hpp"
+#include <armnnOnnxParser/IOnnxParser.hpp>
#endif
#include <boost/exception/exception.hpp>
@@ -20,6 +20,7 @@
#include <boost/format.hpp>
#include <boost/program_options.hpp>
#include <boost/filesystem.hpp>
+#include <boost/lexical_cast.hpp>
#include <map>
#include <string>
@@ -40,7 +41,7 @@
std::string m_InputBinding;
std::string m_OutputBinding;
const armnn::TensorShape* m_InputTensorShape;
- std::vector<armnn::Compute> m_ComputeDevice;
+ std::vector<armnn::BackendId> m_ComputeDevice;
bool m_EnableProfiling;
size_t m_SubgraphId;
bool m_IsModelBinary;
@@ -195,8 +196,6 @@
return { { output.first, armnn::Tensor(output.second, outputTensorData.data()) } };
}
-
-
template <typename IParser, typename TDataType>
class InferenceModel
{
@@ -207,7 +206,7 @@
struct CommandLineOptions
{
std::string m_ModelDir;
- std::vector<armnn::Compute> m_ComputeDevice;
+ std::vector<armnn::BackendId> m_ComputeDevice;
bool m_VisualizePostOptimizationModel;
bool m_EnableFp16TurboMode;
};
@@ -216,11 +215,13 @@
{
namespace po = boost::program_options;
+ std::vector<armnn::BackendId> defaultBackends = {armnn::Compute::CpuAcc, armnn::Compute::CpuRef};
+
desc.add_options()
("model-dir,m", po::value<std::string>(&options.m_ModelDir)->required(),
"Path to directory containing model files (.caffemodel/.prototxt/.tflite)")
- ("compute,c", po::value<std::vector<armnn::Compute>>(&options.m_ComputeDevice)->default_value
- ({armnn::Compute::CpuAcc, armnn::Compute::CpuRef}),
+ ("compute,c", po::value<std::vector<armnn::BackendId>>(&options.m_ComputeDevice)->default_value
+ (defaultBackends),
"Which device to run layers on by default. Possible choices: CpuAcc, CpuRef, GpuAcc")
("visualize-optimized-model,v",
po::value<bool>(&options.m_VisualizePostOptimizationModel)->default_value(false),
diff --git a/tests/InferenceTest.hpp b/tests/InferenceTest.hpp
index 32d828d..3ea7096 100644
--- a/tests/InferenceTest.hpp
+++ b/tests/InferenceTest.hpp
@@ -4,8 +4,8 @@
//
#pragma once
-#include "armnn/ArmNN.hpp"
-#include "armnn/TypesUtils.hpp"
+#include <armnn/ArmNN.hpp>
+#include <armnn/TypesUtils.hpp>
#include "InferenceModel.hpp"
#include <Logging.hpp>
@@ -30,6 +30,20 @@
return in;
}
+inline std::istream& operator>>(std::istream& in, armnn::BackendId& backend)
+{
+ std::string token;
+ in >> token;
+ armnn::Compute compute = armnn::ParseComputeDevice(token.c_str());
+ if (compute == armnn::Compute::Undefined)
+ {
+ in.setstate(std::ios_base::failbit);
+ throw boost::program_options::validation_error(boost::program_options::validation_error::invalid_option_value);
+ }
+ backend = compute;
+ return in;
+}
+
namespace test
{
diff --git a/tests/MultipleNetworksCifar10/MultipleNetworksCifar10.cpp b/tests/MultipleNetworksCifar10/MultipleNetworksCifar10.cpp
index 34fdbf0..f9fdf8b 100644
--- a/tests/MultipleNetworksCifar10/MultipleNetworksCifar10.cpp
+++ b/tests/MultipleNetworksCifar10/MultipleNetworksCifar10.cpp
@@ -36,7 +36,8 @@
namespace po = boost::program_options;
- std::vector<armnn::Compute> computeDevice;
+ std::vector<armnn::BackendId> computeDevice;
+ std::vector<armnn::BackendId> defaultBackends = {armnn::Compute::CpuAcc, armnn::Compute::CpuRef};
std::string modelDir;
std::string dataDir;
@@ -48,8 +49,7 @@
("help", "Display help messages")
("model-dir,m", po::value<std::string>(&modelDir)->required(),
"Path to directory containing the Cifar10 model file")
- ("compute,c", po::value<std::vector<armnn::Compute>>(&computeDevice)->default_value
- ({armnn::Compute::CpuAcc, armnn::Compute::CpuRef}),
+ ("compute,c", po::value<std::vector<armnn::BackendId>>(&computeDevice)->default_value(defaultBackends),
"Which device to run layers on by default. Possible choices: CpuAcc, CpuRef, GpuAcc")
("data-dir,d", po::value<std::string>(&dataDir)->required(),
"Path to directory containing the Cifar10 test data");
@@ -200,7 +200,7 @@
}
catch (const std::exception& e)
{
- // Coverity fix: various boost exceptions can be thrown by methods called by this test.
+ // Coverity fix: various boost exceptions can be thrown by methods called by this test.
std::cerr << "WARNING: MultipleNetworksCifar10: An error has occurred when running the "
"multiple networks inference tests: " << e.what() << std::endl;
return 1;