IVGCVSW-6896 Fix pre-import when using sync execute.
* Refactor backend capability checks in LoadedNetwork.
* ImportInputs should check the number of tensors does not exceed the
number of inputs.
* In EnqueueWorkload the check for for the count of input tensors
was ignoring pre-imported inputs.
* Added checks to verify ImportInputs/ImportOutputs worked as expected
in EndToEndTestImpl.
* Improve documentation on ImportInputs/ImportOutputs in IRuntime.hpp.
* Disabled import tests in CL and Neon EndToEndTests that cannot work.
Signed-off-by: Colm Donelan <colm.donelan@arm.com>
Change-Id: Iae4b2644a1c9f01ee72bce1afb211661cc9ae2e3
diff --git a/src/backends/cl/test/ClEndToEndTests.cpp b/src/backends/cl/test/ClEndToEndTests.cpp
index fa6e027..f28679c 100644
--- a/src/backends/cl/test/ClEndToEndTests.cpp
+++ b/src/backends/cl/test/ClEndToEndTests.cpp
@@ -514,12 +514,18 @@
QLstmEndToEnd(clDefaultBackends);
}
-TEST_CASE("ClForceImportWithMisalignedInputBuffersEndToEndTest")
+TEST_CASE("ClForceImportWithMisalignedInputBuffersEndToEndTest"
+ // Currently, the CL workload for activation does not support tensor handle replacement so this test case
+ // will always fail.
+ * doctest::skip(true))
{
ForceImportWithMisalignedInputBuffersEndToEndTest(clDefaultBackends);
}
-TEST_CASE("ClForceImportWithMisalignedOutputBuffersEndToEndTest")
+TEST_CASE("ClForceImportWithMisalignedOutputBuffersEndToEndTest"
+ // Currently, the CL workload for activation does not support tensor handle replacement so this test case
+ // will always fail.
+ * doctest::skip(true))
{
ForceImportWithMisalignedOutputBuffersEndToEndTest(clDefaultBackends);
}
diff --git a/src/backends/cl/test/ClImportTensorHandleTests.cpp b/src/backends/cl/test/ClImportTensorHandleTests.cpp
index 9a075d2..1198cad 100644
--- a/src/backends/cl/test/ClImportTensorHandleTests.cpp
+++ b/src/backends/cl/test/ClImportTensorHandleTests.cpp
@@ -397,11 +397,14 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc);
-
+ // We expect the import to have succeeded.
+ CHECK(importedOutputIds.size() == 1);
// Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Retrieve the Profiler.Print() output to get the workload execution
ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
@@ -536,11 +539,15 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedOutputIds.size() == 1);
// Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Retrieve the Profiler.Print() output to get the workload execution
ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
@@ -680,11 +687,15 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedOutputIds.size() == 1);
// Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Retrieve the Profiler.Print() output to get the workload execution
ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
@@ -798,11 +809,13 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc);
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc);
+ CHECK(importedOutputIds.size() == 1);
// Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Retrieve the Profiler.Print() output to get the workload execution
ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
@@ -838,7 +851,7 @@
/*
* This is a test to check the functionality of the Forced Import functionality when using repeated inferences that
* require switching from importing to copy. For the first inference we create aligned Pointers and check they are
- * imported correctly. For the second we use similar pointers but don't use PreImporting to force fall back to copy.
+ * imported correctly. For the second we use similar pointers but don't use PreImporting.
*/
// Create runtime in which test will run
IRuntime::CreationOptions options;
@@ -959,11 +972,15 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensors, MemorySource::Malloc);
+ // We expect the import to have succeeded.
+ CHECK(importedOutputIds.size() == 1);
// Do the inference
- runtime->EnqueueWorkload(netId, inputTensors, outputTensors, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Retrieve the Profiler.AnalyzeEventsAndWriteResults() output to get the workload execution
ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
@@ -1246,11 +1263,13 @@
INFO("Run ImportInputs");
std::vector<ImportedInputId> importedInputIds =
runtime->ImportInputs(netId, inputTensorsImport, MemorySource::Malloc);
+ CHECK(importedInputIds.size() == 1);
std::vector<ImportedOutputId> importedOutputIds =
runtime->ImportOutputs(netId, outputTensorsImport, MemorySource::Malloc);
+ CHECK(importedOutputIds.size() == 1);
// Do the inference with pre-imported inputs/outputs
- runtime->EnqueueWorkload(netId, inputTensorsImport, outputTensorsImport, importedInputIds, importedOutputIds);
+ runtime->EnqueueWorkload(netId, InputTensors(), OutputTensors(), importedInputIds, importedOutputIds);
// Sync the outputs so we can read the data
arm_compute::CLScheduler::get().sync();