Set TFLu external context

Remove PMU configuration from the InferenceJob struct and add an
external context parameter intead. The external context is passed
to the TFLu interpreter and will be returned in the
ethosu_inference_begin() and ethosu_inference_end() callbacks.

Change-Id: I6dab04c0ab5088b1325be365d77d65d1182e7441

2 files changed, 12 insertions, 32 deletions

diff --git a/applications/inference_process/include/inference_process.hpp b/applications/inference_process/include/inference_process.hpp
index 6ab453c..9635884 100644
--- a/applications/inference_process/include/inference_process.hpp
+++ b/applications/inference_process/include/inference_process.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited. All rights reserved.
+ * Copyright (c) 2019-2022 Arm Limited. All rights reserved.
  *
  * SPDX-License-Identifier: Apache-2.0
  *
@@ -53,10 +53,7 @@ struct InferenceJob {
     std::vector<DataPtr> output;
     std::vector<DataPtr> expectedOutput;
     size_t numBytesToPrint;
-    std::vector<uint8_t> pmuEventConfig;
-    bool pmuCycleCounterEnable;
-    std::vector<uint32_t> pmuEventCount;
-    uint64_t pmuCycleCounterCount;
+    void *externalContext;
 
     InferenceJob();
     InferenceJob(const std::string &name,
@@ -64,9 +61,8 @@ struct InferenceJob {
                  const std::vector<DataPtr> &input,
                  const std::vector<DataPtr> &output,
                  const std::vector<DataPtr> &expectedOutput,
-                 size_t numBytesToPrint,
-                 const std::vector<uint8_t> &pmuEventConfig,
-                 const bool pmuCycleCounterEnable);
+                 const size_t numBytesToPrint = 0,
+                 void *externalContext        = nullptr);
 
     void invalidate();
     void clean();
diff --git a/applications/inference_process/src/inference_process.cpp b/applications/inference_process/src/inference_process.cpp
index ebd9d6c..4c65005 100644
--- a/applications/inference_process/src/inference_process.cpp
+++ b/applications/inference_process/src/inference_process.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2021 Arm Limited. All rights reserved.
+ * Copyright (c) 2019-2022 Arm Limited. All rights reserved.
  *
  * SPDX-License-Identifier: Apache-2.0
  *
@@ -100,20 +100,18 @@ char *DataPtr::end() const {
     return static_cast<char *>(data) + size;
 }
 
-InferenceJob::InferenceJob() : numBytesToPrint(0) {}
+InferenceJob::InferenceJob() : numBytesToPrint(0), externalContext(nullptr) {}
 
 InferenceJob::InferenceJob(const string &_name,
                            const DataPtr &_networkModel,
                            const vector<DataPtr> &_input,
                            const vector<DataPtr> &_output,
                            const vector<DataPtr> &_expectedOutput,
-                           size_t _numBytesToPrint,
-                           const vector<uint8_t> &_pmuEventConfig,
-                           const bool _pmuCycleCounterEnable) :
+                           const size_t _numBytesToPrint,
+                           void *_externalContext) :
     name(_name),
     networkModel(_networkModel), input(_input), output(_output), expectedOutput(_expectedOutput),
-    numBytesToPrint(_numBytesToPrint), pmuEventConfig(_pmuEventConfig), pmuCycleCounterEnable(_pmuCycleCounterEnable),
-    pmuEventCount(), pmuCycleCounterCount(0) {}
+    numBytesToPrint(_numBytesToPrint), externalContext(_externalContext) {}
 
 void InferenceJob::invalidate() {
     networkModel.invalidate();
@@ -167,16 +165,14 @@ bool InferenceProcess::runJob(InferenceJob &job) {
 
     // Create the TFL micro interpreter
     tflite::AllOpsResolver resolver;
-#ifdef LAYER_BY_LAYER_PROFILER
-    tflite::LayerByLayerProfiler profiler(job.pmuEventConfig, job.pmuCycleCounterEnable);
-#else
     tflite::ArmProfiler profiler;
-#endif
-
     tflite::MicroErrorReporter errorReporter;
     tflite::MicroInterpreter interpreter(
         model, resolver, tensorArena, tensorArenaSize, &errorReporter, nullptr, &profiler);
 
+    // Set external context
+    interpreter.SetMicroExternalContext(job.externalContext);
+
     // Allocate tensors
     TfLiteStatus status = interpreter.AllocateTensors();
     if (status != kTfLiteOk) {
@@ -196,14 +192,6 @@ bool InferenceProcess::runJob(InferenceJob &job) {
         return true;
     }
 
-#ifdef LAYER_BY_LAYER_PROFILER
-    if (job.pmuCycleCounterEnable) {
-        job.pmuCycleCounterCount = profiler.GetPmuCycleCounterCount();
-    }
-
-    job.pmuEventCount.assign(profiler.GetPmuEventCount().begin(), profiler.GetPmuEventCount().end());
-#endif
-
     LOG("Inference runtime: %" PRId32 " cycles\n", profiler.GetTotalTicks());
 
     // Copy output data from TFLu arena to job descriptor
@@ -333,10 +321,6 @@ bool InferenceProcess::compareOfm(InferenceJob &job, tflite::MicroInterpreter &i
 }
 
 void InferenceProcess::printJob(InferenceJob &job, tflite::MicroInterpreter &interpreter) {
-    for (size_t i = 0; i < job.pmuEventCount.size(); i++) {
-        LOG("ethosu_pmu_cntr%zu : %" PRIu32 "\n", i, job.pmuEventCount[i]);
-    }
-
     LOG("arena_used_bytes : %zu\n", interpreter.arena_used_bytes());
 
     // Print all of the output data, or the first NUM_BYTES_TO_PRINT bytes,