rpm/armnn
SHA256
1
0
Fork 0
forked from pool/armnn
Code Pull Requests Activity
98cd3ecc18
armnn/armnn-fix_arm32_dep.patch

285 lines
9.6 KiB
Diff
Raw Blame History

From dcaa6109c95034aa3b945acd50a2882e40f13370 Mon Sep 17 00:00:00 2001
From: Ferran Balaguer <ferran.balaguer@arm.com>
Date: Wed, 21 Aug 2019 13:28:38 +0100
Subject: [PATCH] IVGCVSW-3175 Add Regression Tests for Zero Copy
Signed-off-by: Ferran Balaguer <ferran.balaguer@arm.com>
Change-Id: I6f16ea0dca359283a3b187e2f046f82a7dc2ff7c
---
.../backendsCommon/test/EndToEndTestImpl.hpp | 153 ++++++++++++++++++
.../reference/test/RefEndToEndTests.cpp | 86 ++--------
2 files changed, 167 insertions(+), 72 deletions(-)
diff --git a/src/backends/backendsCommon/test/EndToEndTestImpl.hpp b/src/backends/backendsCommon/test/EndToEndTestImpl.hpp
index f8673d69..8a3e44fc 100644
--- a/src/backends/backendsCommon/test/EndToEndTestImpl.hpp
+++ b/src/backends/backendsCommon/test/EndToEndTestImpl.hpp
@@ -8,6 +8,7 @@
#include <armnn/ArmNN.hpp>
#include <armnn/INetwork.hpp>
+#include <Profiling.hpp>
#include <backendsCommon/test/QuantizeHelper.hpp>
@@ -171,4 +172,156 @@ void EndToEndLayerTestImpl(INetworkPtr network,
}
}
+inline void ImportNonAlignedPointerTest(std::vector<BackendId> backends)
+{
+ using namespace armnn;
+
+ // Create runtime in which test will run
+ IRuntime::CreationOptions options;
+ IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+ // build up the structure of the network
+ INetworkPtr net(INetwork::Create());
+
+ IConnectableLayer* input = net->AddInputLayer(0);
+
+ NormalizationDescriptor descriptor;
+ IConnectableLayer* norm = net->AddNormalizationLayer(descriptor);
+
+ IConnectableLayer* output = net->AddOutputLayer(0);
+
+ input->GetOutputSlot(0).Connect(norm->GetInputSlot(0));
+ norm->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
+ norm->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
+
+ // Optimize the network
+ IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
+
+ // Loads it into the runtime.
+ NetworkId netId;
+ runtime->LoadNetwork(netId, std::move(optNet));
+
+ // Creates structures for input & output
+ std::vector<float> inputData
+ {
+ 1.0f, 2.0f, 3.0f, 4.0f, 5.0f
+ };
+
+ // Misaligned input
+ float * misalignedInputData = inputData.data();
+ misalignedInputData++;
+
+ std::vector<float> outputData(5);
+
+ // Misaligned output
+ float * misalignedOutputData = outputData.data();
+ misalignedOutputData++;
+
+ InputTensors inputTensors
+ {
+ {0,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), misalignedInputData)},
+ };
+ OutputTensors outputTensors
+ {
+ {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), misalignedOutputData)}
+ };
+
+ // The result of the inference is not important, just the fact that there
+ // should not be CopyMemGeneric workloads.
+ runtime->GetProfiler(netId)->EnableProfiling(true);
+
+ // Do the inference
+ runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
+
+ // Retrieve the Profiler.Print() output to get the workload execution
+ ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
+ std::stringstream ss;
+ profilerManager.GetProfiler()->Print(ss);;
+ std::string dump = ss.str();
+
+ // Contains RefNormalizationWorkload
+ std::size_t found = dump.find("RefNormalizationWorkload");
+ BOOST_TEST(found != std::string::npos);
+ // No Contains SyncMemGeneric (Created when importing the output tensor handle)
+ found = dump.find("SyncMemGeneric");
+ BOOST_TEST(found == std::string::npos);
+ // Contains CopyMemGeneric
+ found = dump.find("CopyMemGeneric");
+ BOOST_TEST(found != std::string::npos);
+}
+
+inline void ImportAlignedPointerTest(std::vector<BackendId> backends)
+{
+ using namespace armnn;
+
+ // Create runtime in which test will run
+ IRuntime::CreationOptions options;
+ IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+ // build up the structure of the network
+ INetworkPtr net(INetwork::Create());
+
+ IConnectableLayer* input = net->AddInputLayer(0);
+
+ NormalizationDescriptor descriptor;
+ IConnectableLayer* norm = net->AddNormalizationLayer(descriptor);
+
+ IConnectableLayer* output = net->AddOutputLayer(0);
+
+ input->GetOutputSlot(0).Connect(norm->GetInputSlot(0));
+ norm->GetOutputSlot(0).Connect(output->GetInputSlot(0));
+
+ input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
+ norm->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
+
+ // Optimize the network
+ IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
+
+ // Loads it into the runtime.
+ NetworkId netId;
+ runtime->LoadNetwork(netId, std::move(optNet));
+
+ // Creates structures for input & output
+ std::vector<float> inputData
+ {
+ 1.0f, 2.0f, 3.0f, 4.0f
+ };
+
+ std::vector<float> outputData(4);
+
+ InputTensors inputTensors
+ {
+ {0,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())},
+ };
+ OutputTensors outputTensors
+ {
+ {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
+ };
+
+ // The result of the inference is not important, just the fact that there
+ // should not be CopyMemGeneric workloads.
+ runtime->GetProfiler(netId)->EnableProfiling(true);
+
+ // Do the inference
+ runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
+
+ // Retrieve the Profiler.Print() output to get the workload execution
+ ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
+ std::stringstream ss;
+ profilerManager.GetProfiler()->Print(ss);;
+ std::string dump = ss.str();
+
+ // Contains RefNormalizationWorkload
+ std::size_t found = dump.find("RefNormalizationWorkload");
+ BOOST_TEST(found != std::string::npos);
+ // Contains SyncMemGeneric
+ found = dump.find("SyncMemGeneric");
+ BOOST_TEST(found != std::string::npos);
+ // No contains CopyMemGeneric
+ found = dump.find("CopyMemGeneric");
+ BOOST_TEST(found == std::string::npos);
+}
+
} // anonymous namespace
diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp
index 31e9b339..ee42c9e9 100644
--- a/src/backends/reference/test/RefEndToEndTests.cpp
+++ b/src/backends/reference/test/RefEndToEndTests.cpp
@@ -322,78 +322,6 @@ BOOST_AUTO_TEST_CASE(TrivialMin)
BOOST_TEST(outputData[3] == 2);
}
-BOOST_AUTO_TEST_CASE(RefNoCopyWorkloads)
-{
- using namespace armnn;
-
- // Create runtime in which test will run
- IRuntime::CreationOptions options;
- IRuntimePtr runtime(armnn::IRuntime::Create(options));
-
- // build up the structure of the network
- INetworkPtr net(INetwork::Create());
-
- IConnectableLayer* input = net->AddInputLayer(0);
-
- NormalizationDescriptor descriptor;
- IConnectableLayer* norm = net->AddNormalizationLayer(descriptor);
-
- IConnectableLayer* output = net->AddOutputLayer(0);
-
- input->GetOutputSlot(0).Connect(norm->GetInputSlot(0));
- norm->GetOutputSlot(0).Connect(output->GetInputSlot(0));
-
- input->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
- norm->GetOutputSlot(0).SetTensorInfo(TensorInfo({ 1, 1, 4, 1 }, DataType::Float32));
-
- // Optimize the network
- IOptimizedNetworkPtr optNet = Optimize(*net, defaultBackends, runtime->GetDeviceSpec());
-
- // Loads it into the runtime.
- NetworkId netId;
- runtime->LoadNetwork(netId, std::move(optNet));
-
- // Creates structures for input & output
- std::vector<float> inputData
- {
- 1.0f, 2.0f, 3.0f, 4.0f
- };
-
- std::vector<float> outputData(4);
-
- InputTensors inputTensors
- {
- {0,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())},
- };
- OutputTensors outputTensors
- {
- {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
- };
-
- // The result of the inference is not important, just the fact that there
- // should not be CopyMemGeneric workloads.
- runtime->GetProfiler(netId)->EnableProfiling(true);
-
- // Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
-
- // Retrieve the Profiler.Print() output to get the workload execution
- ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
- std::stringstream ss;
- profilerManager.GetProfiler()->Print(ss);;
- std::string dump = ss.str();
-
- // Contains RefNormalizationWorkload
- std::size_t found = dump.find("RefNormalizationWorkload");
- BOOST_TEST(found != std::string::npos);
- // Contains SyncMemGeneric
- found = dump.find("SyncMemGeneric");
- BOOST_TEST(found != std::string::npos);
- // No contains CopyMemGeneric
- found = dump.find("CopyMemGeneric");
- BOOST_TEST(found == std::string::npos);
-}
-
BOOST_AUTO_TEST_CASE(RefEqualSimpleEndToEndTest)
{
const std::vector<uint8_t> expectedOutput({ 1, 1, 1, 1, 0, 0, 0, 0,
@@ -1023,4 +951,18 @@ BOOST_AUTO_TEST_CASE(RefResizeNearestNeighborEndToEndInt16NhwcTest)
ResizeNearestNeighborEndToEnd<armnn::DataType::QuantisedSymm16>(defaultBackends, armnn::DataLayout::NHWC);
}
+#if !defined(__ANDROID__)
+// Only run these tests on non Android platforms
+BOOST_AUTO_TEST_CASE(RefImportNonAlignedPointerTest)
+{
+ ImportNonAlignedPointerTest(defaultBackends);
+}
+
+BOOST_AUTO_TEST_CASE(RefImportAlignedPointerTest)
+{
+ ImportAlignedPointerTest(defaultBackends);
+}
+
+#endif
+
BOOST_AUTO_TEST_SUITE_END()
\ No newline at end of file
View Git Blame Copy Permalink