commit e2af6f4322a1e2b8b3c391fb721a6a80c281477f
author Narumol Prangnawarat <narumol.prangnawarat@arm.com> Fri Jan 28 17:59:18 2022 +0000
committer Jim Flynn <jim.flynn@arm.com> Mon Jan 31 12:53:51 2022 +0000
tree b0dd53289e27304a6d724821459cb0f4b6343a39
parent fd313fef775ed210f8dab84452ea382a0b4164b0

IVGCVSW-6552 Add support of aligned host memory

 * Add AllocatedData functions to OutputHandler
 * Enable import aligned memory in ImportInputs
 * Enable import aligned memory in ImportOutputs
 * Allow to import input and output if the memory is aligned
 * Implement Reconfigure function on ClConvolution2dWorkload
 * End-to-end test on Ref and Cl to ensure that input and output memory
are imported when aligned

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: I9e5e4c26d1ac2f1d806803ade5f64c6479c51718

include/armnn/IRuntime.hpp

diff --git a/include/armnn/IRuntime.hpp b/include/armnn/IRuntime.hpp
index d85a3e3..042271f 100644
--- a/include/armnn/IRuntime.hpp
+++ b/include/armnn/IRuntime.hpp
@@ -216,18 +216,19 @@
     TensorInfo GetInputTensorInfo(NetworkId networkId, LayerBindingId layerId) const;
     TensorInfo GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const;
 
-
     /// ImportInputs separates the importing and mapping of InputTensors from network execution.
     /// Allowing for a set of InputTensors to be imported and mapped once, but used in execution many times.
     /// This function is not thread safe and must not be used while other threads are calling Execute().
-    /// Only compatible with AsyncEnabled networks
-    std::vector<ImportedInputId> ImportInputs(NetworkId networkId, const InputTensors& inputTensors);
+    /// Only compatible with AsyncEnabled networks and aligned memory import
+    std::vector<ImportedInputId> ImportInputs(NetworkId networkId, const InputTensors& inputTensors,
+                                              MemorySource forceImportMemorySource = MemorySource::Undefined);
 
     /// ImportOutputs separates the importing and mapping of OutputTensors from network execution.
     /// Allowing for a set of OutputTensors to be imported and mapped once, but used in execution many times.
     /// This function is not thread safe and must not be used while other threads are calling Execute().
-    /// Only compatible with AsyncEnabled networks
-    std::vector<ImportedOutputId> ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors);
+    /// Only compatible with AsyncEnabled networks and aligned memory import
+    std::vector<ImportedOutputId> ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors,
+                                                MemorySource forceImportMemorySource = MemorySource::Undefined);
 
     /// Un-import and delete the imported InputTensor/s
     /// This function is not thread safe and must not be used while other threads are calling Execute().
@@ -242,7 +243,9 @@
     /// Evaluates a network using input in inputTensors and outputs filled into outputTensors
     Status EnqueueWorkload(NetworkId networkId,
                            const InputTensors& inputTensors,
-                           const OutputTensors& outputTensors);
+                           const OutputTensors& outputTensors,
+                           std::vector<ImportedInputId> preImportedInputIds = {},
+                           std::vector<ImportedOutputId> preImportedOutputIds = {});
 
     /// This is an experimental function.
     /// Evaluates a network using input in inputTensors and outputs filled into outputTensors.