15 #include <boost/test/unit_test.hpp> 22 using namespace armnn;
25 bool ConstantUsageTest(
const std::vector<BackendId>& computeDevice,
27 const std::vector<T>& inputData,
28 const std::vector<T>& constantData,
29 const std::vector<T>& expectedOutputData)
57 runtime->LoadNetwork(netId, std::move(optNet));
60 std::vector<T> outputData(inputData.size());
64 {0,
ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())}
68 {0,
Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
72 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
75 return outputData == expectedOutputData;
78 inline bool ConstantUsageFloat32Test(
const std::vector<BackendId>& backends)
82 return ConstantUsageTest(backends,
84 std::vector<float>{ 1.f, 2.f, 3.f, 4.f, 5.f, 6.f },
85 std::vector<float>{ 6.f, 5.f, 4.f, 3.f, 2.f, 1.f },
86 std::vector<float>{ 7.f, 7.f, 7.f, 7.f, 7.f, 7.f }
90 inline bool ConstantUsageUint8Test(
const std::vector<BackendId>& backends)
94 const float scale = 0.023529f;
95 const int8_t offset = -43;
100 return ConstantUsageTest(backends,
102 armnnUtils::QuantizedVector<uint8_t>({ 1.f, 2.f, 3.f, 4.f, 5.f, 6.f }, scale, offset),
103 armnnUtils::QuantizedVector<uint8_t>({ 6.f, 5.f, 4.f, 3.f, 2.f, 1.f }, scale, offset),
104 armnnUtils::QuantizedVector<uint8_t>({ 7.f, 7.f, 7.f, 7.f, 7.f, 7.f }, scale, offset)
109 template<DataType ArmnnType,
typename T = ResolveType<ArmnnType>>
110 bool Compare(T a, T b)
117 return static_cast<bool>(a) == static_cast<bool>(b);
122 constexpr
float tolerance = 0.000001f;
123 return std::fabs(static_cast<float>(a) - static_cast<float>(b)) <= tolerance;
127 int SubStringCounter(std::string&
string, std::string&& substring)
129 std::size_t found = 0;
132 while((found =
string.find(substring, found)) != std::string::npos)
136 found += substring.length();
144 const std::map<
int, std::vector<TInput>>& inputTensorData,
145 const std::map<
int, std::vector<TOutput>>& expectedOutputData,
146 std::vector<BackendId> backends)
157 runtime->LoadNetwork(netId, std::move(optNet));
160 inputTensors.reserve(inputTensorData.size());
161 for (
auto&& it : inputTensorData)
163 inputTensors.push_back({it.first,
164 ConstTensor(runtime->GetInputTensorInfo(netId, it.first), it.second.data())});
167 outputTensors.reserve(expectedOutputData.size());
168 std::map<int, std::vector<TOutput>> outputStorage;
169 for (
auto&& it : expectedOutputData)
171 std::vector<TOutput> out(it.second.size());
172 outputStorage.emplace(it.first, out);
173 outputTensors.push_back({it.first,
174 Tensor(runtime->GetOutputTensorInfo(netId, it.first),
175 outputStorage.at(it.first).data())});
179 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
182 for (
auto&& it : expectedOutputData)
184 std::vector<TOutput> out = outputStorage.at(it.first);
185 for (
unsigned int i = 0; i < out.size(); ++i)
187 BOOST_CHECK(Compare<ArmnnOType>(it.second[i], out[i]) ==
true);
192 inline void ImportNonAlignedInputPointerTest(std::vector<BackendId> backends)
194 using namespace armnn;
223 std::string ignoredErrorMessage;
226 runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
229 std::vector<float> inputData
231 1.0f, 2.0f, 3.0f, 4.0f
235 float* misalignedInputData =
reinterpret_cast<float*
>(
reinterpret_cast<char*
>(inputData.data()) + 1);
237 std::vector<float> outputData(4);
240 float* alignedOutputData = outputData.data();
248 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), alignedOutputData)}
251 runtime->GetProfiler(netId)->EnableProfiling(
true);
254 BOOST_CHECK_THROW(runtime->EnqueueWorkload(netId, inputTensors, outputTensors),
MemoryImportException);
257 inline void ExportNonAlignedOutputPointerTest(std::vector<BackendId> backends)
259 using namespace armnn;
288 std::string ignoredErrorMessage;
291 runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
294 std::vector<float> inputData
296 1.0f, 2.0f, 3.0f, 4.0f, 5.0f
300 float* alignedInputData = inputData.data();
302 std::vector<float> outputData(5);
305 float* misalignedOutputData =
reinterpret_cast<float*
>(
reinterpret_cast<char*
>(outputData.data()) + 1);
313 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), misalignedOutputData)}
320 BOOST_CHECK_THROW(runtime->EnqueueWorkload(netId, inputTensors, outputTensors),
MemoryImportException);
324 BOOST_CHECK_THROW(runtime->EnqueueWorkload(netId, inputTensors, outputTensors),
MemoryExportException);
328 inline void ImportAlignedPointerTest(std::vector<BackendId> backends)
330 using namespace armnn;
359 std::string ignoredErrorMessage;
362 runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
365 std::vector<float> inputData
367 1.0f, 2.0f, 3.0f, 4.0f
370 std::vector<float> outputData(4);
372 std::vector<float> expectedOutput
374 1.0f, 4.0f, 9.0f, 16.0f
383 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
386 runtime->GetProfiler(netId)->EnableProfiling(
true);
389 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
393 std::stringstream ss;
395 std::string dump = ss.str();
398 std::size_t found = dump.find(
"ActivationWorkload");
399 BOOST_TEST(found != std::string::npos);
402 found = dump.find(
"SyncMemGeneric");
403 BOOST_TEST(found != std::string::npos);
406 found = dump.find(
"CopyMemGeneric");
407 BOOST_TEST(found == std::string::npos);
410 BOOST_TEST(outputData == expectedOutput);
413 inline void ImportOnlyWorkload(std::vector<BackendId> backends)
415 using namespace armnn;
440 BOOST_TEST_CHECKPOINT(
"Load Network");
443 std::string ignoredErrorMessage;
445 BOOST_TEST(runtime->LoadNetwork(netId, std::move(optNet),ignoredErrorMessage, networkProperties)
448 BOOST_TEST_CHECKPOINT(
"Generate Data");
450 std::vector<float> inputData
452 1.0f, 2.0f, 3.0f, 4.0f
455 std::vector<float> outputData(4);
457 std::vector<float> expectedOutput
459 1.0f, 4.0f, 9.0f, 16.0f
462 BOOST_TEST_CHECKPOINT(
"Create Network");
469 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
472 BOOST_TEST_CHECKPOINT(
"Get Profiler");
474 runtime->GetProfiler(netId)->EnableProfiling(
true);
476 BOOST_TEST_CHECKPOINT(
"Run Inference");
478 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
480 BOOST_TEST_CHECKPOINT(
"Print Profiler");
483 std::stringstream ss;
485 std::string dump = ss.str();
488 BOOST_TEST_CHECKPOINT(
"Find SyncMemGeneric");
489 int count = SubStringCounter(dump,
"SyncMemGeneric");
490 BOOST_TEST(count == 0);
493 BOOST_TEST_CHECKPOINT(
"Find CopyMemGeneric");
494 count = SubStringCounter(dump,
"CopyMemGeneric");
495 BOOST_TEST(count == 1);
498 BOOST_CHECK_EQUAL_COLLECTIONS(outputData.begin(), outputData.end(), expectedOutput.begin(), expectedOutput.end());
501 inline void ExportOnlyWorkload(std::vector<BackendId> backends)
503 using namespace armnn;
528 BOOST_TEST_CHECKPOINT(
"Load Network");
531 std::string ignoredErrorMessage;
533 BOOST_TEST(runtime->LoadNetwork(netId, std::move(optNet),ignoredErrorMessage, networkProperties)
536 BOOST_TEST_CHECKPOINT(
"Generate Data");
538 std::vector<float> inputData
540 1.0f, 2.0f, 3.0f, 4.0f
543 std::vector<float> outputData(4);
545 std::vector<float> expectedOutput
547 1.0f, 4.0f, 9.0f, 16.0f
550 BOOST_TEST_CHECKPOINT(
"Create Network");
557 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
560 BOOST_TEST_CHECKPOINT(
"Get Profiler");
562 runtime->GetProfiler(netId)->EnableProfiling(
true);
564 BOOST_TEST_CHECKPOINT(
"Run Inference");
566 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
568 BOOST_TEST_CHECKPOINT(
"Print Profiler");
571 std::stringstream ss;
573 std::string dump = ss.str();
576 BOOST_TEST_CHECKPOINT(
"Find SyncMemGeneric");
577 int count = SubStringCounter(dump,
"SyncMemGeneric");
578 BOOST_TEST(count == 1);
581 BOOST_TEST_CHECKPOINT(
"Find CopyMemGeneric");
582 count = SubStringCounter(dump,
"CopyMemGeneric");
583 BOOST_TEST(count == 1);
586 BOOST_CHECK_EQUAL_COLLECTIONS(outputData.begin(), outputData.end(), expectedOutput.begin(), expectedOutput.end());
589 inline void ImportAndExportWorkload(std::vector<BackendId> backends)
591 using namespace armnn;
615 BOOST_TEST_CHECKPOINT(
"Load Network");
618 std::string ignoredErrorMessage;
620 BOOST_TEST(runtime->LoadNetwork(netId, std::move(optNet),ignoredErrorMessage, networkProperties)
623 BOOST_TEST_CHECKPOINT(
"Generate Data");
625 std::vector<float> inputData
627 1.0f, 2.0f, 3.0f, 4.0f
630 std::vector<float> outputData(4);
632 std::vector<float> expectedOutput
634 1.0f, 4.0f, 9.0f, 16.0f
637 BOOST_TEST_CHECKPOINT(
"Create Network");
644 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
647 BOOST_TEST_CHECKPOINT(
"Get Profiler");
649 runtime->GetProfiler(netId)->EnableProfiling(
true);
651 BOOST_TEST_CHECKPOINT(
"Run Inference");
653 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
655 BOOST_TEST_CHECKPOINT(
"Print Profiler");
658 std::stringstream ss;
660 std::string dump = ss.str();
663 BOOST_TEST_CHECKPOINT(
"Find SyncMemGeneric");
664 int count = SubStringCounter(dump,
"SyncMemGeneric");
665 BOOST_TEST(count == 1);
668 BOOST_TEST_CHECKPOINT(
"Find CopyMemGeneric");
669 count = SubStringCounter(dump,
"CopyMemGeneric");
670 BOOST_TEST(count == 0);
673 BOOST_CHECK_EQUAL_COLLECTIONS(outputData.begin(), outputData.end(), expectedOutput.begin(), expectedOutput.end());
676 inline void ExportOutputWithSeveralOutputSlotConnectionsTest(std::vector<BackendId> backends)
678 using namespace armnn;
708 std::string ignoredErrorMessage;
711 runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
714 std::vector<float> inputData
716 1.0f, 2.0f, 3.0f, 4.0f
719 std::vector<float> outputData0(4);
720 std::vector<float> outputData1(4);
728 {0,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData0.data())},
729 {1,
armnn::Tensor(runtime->GetOutputTensorInfo(netId, 1), outputData1.data())}
734 runtime->GetProfiler(netId)->EnableProfiling(
true);
737 runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
741 std::stringstream ss;
743 std::string dump = ss.str();
745 std::size_t found = std::string::npos;
749 found = dump.find(
"RefActivationWorkload");
753 found = dump.find(
"NeonActivationWorkload");
757 found = dump.find(
"ClActivationWorkload");
760 BOOST_TEST(found != std::string::npos);
762 found = dump.find(
"SyncMemGeneric");
763 BOOST_TEST(found == std::string::npos);
765 found = dump.find(
"CopyMemGeneric");
766 BOOST_TEST(found != std::string::npos);
static IRuntimePtr Create(const CreationOptions &options)
Interface for a layer that is connectable to other layers via InputSlots and OutputSlots.
CPU Execution: Reference C++ kernels.
static ProfilerManager & GetInstance()
std::unique_ptr< IRuntime, void(*)(IRuntime *runtime)> IRuntimePtr
typename ResolveTypeImpl< DT >::Type ResolveType
std::vector< std::pair< LayerBindingId, class ConstTensor > > InputTensors
Copyright (c) 2020 ARM Limited.
BOOST_CHECK(profilingService.GetCurrentState()==ProfilingState::WaitingForAck)
virtual void SetTensorInfo(const TensorInfo &tensorInfo)=0
A tensor defined by a TensorInfo (shape and data type) and a mutable backing store.
IOptimizedNetworkPtr Optimize(const INetwork &network, const std::vector< BackendId > &backendPreferences, const IDeviceSpec &deviceSpec, const OptimizerOptions &options=OptimizerOptions(), Optional< std::vector< std::string > &> messages=EmptyOptional())
Create an optimized version of the network.
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
std::vector< std::pair< LayerBindingId, class Tensor > > OutputTensors
std::unique_ptr< IOptimizedNetwork, void(*)(IOptimizedNetwork *network)> IOptimizedNetworkPtr
void SetQuantizationScale(float scale)
GPU Execution: OpenCL: ArmCompute.
An ActivationDescriptor for the ActivationLayer.
CPU Execution: NEON: ArmCompute.
virtual const IInputSlot & GetInputSlot(unsigned int index) const =0
Get a const input slot handle by slot index.
virtual const IOutputSlot & GetOutputSlot(unsigned int index) const =0
Get the const output slot handle by slot index.
void SetQuantizationOffset(int32_t offset)
std::unique_ptr< INetwork, void(*)(INetwork *network)> INetworkPtr
virtual int Connect(IInputSlot &destination)=0
armnn::Runtime::CreationOptions::ExternalProfilingOptions options
void Print(std::ostream &outStream) const override
Print stats for events in JSON Format to the given output stream.
static INetworkPtr Create()
ActivationFunction m_Function
The activation function to use (Sigmoid, TanH, Linear, ReLu, BoundedReLu, SoftReLu, LeakyReLu, Abs, Sqrt, Square).