MLECO-2345: Adding dynamic load support for FVPs
With this patch, the generic inference runner use-case can be
configured to accept the model tflite file at run-time via
the FVP's command line parameters. Same is true for the IFM
and the inference results can be dumped out too.
NOTE: this change is only for supporting the FVP, the FPGA
implementation will not allow additional loading for the
changes in this patch to be useful.
Change-Id: I1318bd5b0cfb7bb635ced6fe58d22c3e401d2547
diff --git a/source/use_case/inference_runner/src/TestModel.cc b/source/use_case/inference_runner/src/TestModel.cc
index 4512a9b..274790f 100644
--- a/source/use_case/inference_runner/src/TestModel.cc
+++ b/source/use_case/inference_runner/src/TestModel.cc
@@ -23,14 +23,34 @@
return this->m_opResolver;
}
-extern uint8_t* GetModelPointer();
-const uint8_t* arm::app::TestModel::ModelPointer()
-{
- return GetModelPointer();
-}
+#if defined(DYNAMIC_MODEL_BASE) && defined(DYNAMIC_MODEL_SIZE)
-extern size_t GetModelLen();
-size_t arm::app::TestModel::ModelSize()
-{
- return GetModelLen();
-}
\ No newline at end of file
+ const uint8_t* arm::app::TestModel::ModelPointer()
+ {
+ info("Model pointer: 0x%08x\n", DYNAMIC_MODEL_BASE);
+ return reinterpret_cast<uint8_t *>(DYNAMIC_MODEL_BASE);
+ }
+
+ size_t arm::app::TestModel::ModelSize()
+ {
+ /* TODO: Can we get the actual model size here somehow?
+ * Currently we return the reserved space. It is possible to do
+ * so by reading the memory pattern but it will not be reliable. */
+ return static_cast<size_t>(DYNAMIC_MODEL_SIZE);
+ }
+
+#else /* defined(DYNAMIC_MODEL_BASE) && defined(DYNAMIC_MODEL_SIZE) */
+
+ extern uint8_t* GetModelPointer();
+ const uint8_t* arm::app::TestModel::ModelPointer()
+ {
+ return GetModelPointer();
+ }
+
+ extern size_t GetModelLen();
+ size_t arm::app::TestModel::ModelSize()
+ {
+ return GetModelLen();
+ }
+
+#endif /* defined(DYNAMIC_MODEL_BASE) && defined(DYNAMIC_MODEL_SIZE) */
diff --git a/source/use_case/inference_runner/src/UseCaseHandler.cc b/source/use_case/inference_runner/src/UseCaseHandler.cc
index b98b1c5..66b7042 100644
--- a/source/use_case/inference_runner/src/UseCaseHandler.cc
+++ b/source/use_case/inference_runner/src/UseCaseHandler.cc
@@ -25,81 +25,150 @@
namespace arm {
namespace app {
- bool RunInferenceHandler(ApplicationContext& ctx)
- {
- auto& platform = ctx.Get<hal_platform&>("platform");
- auto& profiler = ctx.Get<Profiler&>("profiler");
- auto& model = ctx.Get<Model&>("model");
+static void PopulateInputTensor(const Model& model)
+{
+ const size_t numInputs = model.GetNumInputs();
- constexpr uint32_t dataPsnTxtInfStartX = 150;
- constexpr uint32_t dataPsnTxtInfStartY = 40;
+#if defined(DYNAMIC_IFM_BASE) && defined(DYNAMIC_IFM_SIZE)
+ size_t curInputIdx = 0;
+#endif /* defined(DYNAMIC_IFM_BASE) && defined(DYNAMIC_IFM_SIZE) */
- if (!model.IsInited()) {
- printf_err("Model is not initialised! Terminating processing.\n");
- return false;
- }
+ /* Populate each input tensor with random data. */
+ for (size_t inputIndex = 0; inputIndex < numInputs; inputIndex++) {
- const size_t numInputs = model.GetNumInputs();
+ TfLiteTensor* inputTensor = model.GetInputTensor(inputIndex);
-#if VERIFY_TEST_OUTPUT
- info("Initial input tensors values:\n");
- for (size_t inputIndex = 0; inputIndex < model.GetNumInputs(); inputIndex++) {
- arm::app::DumpTensor(model.GetInputTensor(inputIndex));
- }
- info("Initial output tensors values:\n");
- for (size_t outputIndex = 0; outputIndex < model.GetNumOutputs(); outputIndex++) {
- arm::app::DumpTensor(model.GetOutputTensor(outputIndex));
- }
-#endif /* VERIFY_TEST_OUTPUT */
+ debug("Populating input tensor %zu@%p\n", inputIndex, inputTensor);
+ debug("Total input size to be populated: %zu\n", inputTensor->bytes);
- /* Populate each input tensor with random data. */
- for (size_t inputIndex = 0; inputIndex < numInputs; inputIndex++) {
+ if (inputTensor->bytes > 0) {
- TfLiteTensor* inputTensor = model.GetInputTensor(inputIndex);
+ uint8_t* tData = tflite::GetTensorData<uint8_t>(inputTensor);
- debug("Populating input tensor %zu@%p\n", inputIndex, inputTensor);
- debug("Total input size to be populated: %zu\n", inputTensor->bytes);
-
- /* Create a random input. */
- if (inputTensor->bytes > 0) {
-
- uint8_t* tData = tflite::GetTensorData<uint8_t>(inputTensor);
-
- for (size_t j = 0; j < inputTensor->bytes; ++j) {
- tData[j] = static_cast<uint8_t>(std::rand() & 0xFF);
- }
+#if defined(DYNAMIC_IFM_BASE) && defined(DYNAMIC_IFM_SIZE)
+ if (curInputIdx + inputTensor->bytes > DYNAMIC_IFM_SIZE) {
+ printf_err("IFM reserved buffer size insufficient\n");
+ return;
}
+ memcpy(tData, reinterpret_cast<void *>(DYNAMIC_IFM_BASE + curInputIdx),
+ inputTensor->bytes);
+ curInputIdx += inputTensor->bytes;
+#else /* defined(DYNAMIC_IFM_BASE) */
+ /* Create a random input. */
+ for (size_t j = 0; j < inputTensor->bytes; ++j) {
+ tData[j] = static_cast<uint8_t>(std::rand() & 0xFF);
+ }
+#endif /* defined(DYNAMIC_IFM_BASE) && defined(DYNAMIC_IFM_SIZE) */
}
+ }
- /* Strings for presentation/logging. */
- std::string str_inf{"Running inference... "};
+#if defined(DYNAMIC_IFM_BASE)
+ info("%d input tensor/s populated with %d bytes with data read from 0x%08x\n",
+ numInputs, curInputIdx, DYNAMIC_IFM_BASE);
+#endif /* defined(DYNAMIC_IFM_BASE) */
+}
- /* Display message on the LCD - inference running. */
- platform.data_psn->present_data_text(
- str_inf.c_str(), str_inf.size(),
- dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
+#if defined (DYNAMIC_OFM_BASE) && defined(DYNAMIC_OFM_SIZE)
+static void PopulateDynamicOfm(const Model& model)
+{
+ /* Dump the output to a known memory location */
+ const size_t numOutputs = model.GetNumOutputs();
+ size_t curCopyIdx = 0;
+ uint8_t* const dstPtr = reinterpret_cast<uint8_t *>(DYNAMIC_OFM_BASE);
- if (!RunInference(model, profiler)) {
- return false;
+ for (size_t outputIdx = 0; outputIdx < numOutputs; ++outputIdx) {
+ TfLiteTensor* outputTensor = model.GetOutputTensor(outputIdx);
+ uint8_t* const tData = tflite::GetTensorData<uint8_t>(outputTensor);
+
+ if (tData && outputTensor->bytes > 0) {
+ if (curCopyIdx + outputTensor->bytes > DYNAMIC_OFM_SIZE) {
+ printf_err("OFM reserved buffer size insufficient\n");
+ return;
+ }
+ memcpy(dstPtr + curCopyIdx, tData, outputTensor->bytes);
+ curCopyIdx += outputTensor->bytes;
}
+ }
- /* Erase. */
- str_inf = std::string(str_inf.size(), ' ');
- platform.data_psn->present_data_text(
- str_inf.c_str(), str_inf.size(),
- dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
-
- info("Final results:\n");
- profiler.PrintProfilingResult();
+ info("%d output tensor/s worth %d bytes copied to 0x%08x\n",
+ numOutputs, curCopyIdx, DYNAMIC_OFM_BASE);
+}
+#endif /* defined (DYNAMIC_OFM_BASE) && defined(DYNAMIC_OFM_SIZE) */
#if VERIFY_TEST_OUTPUT
- for (size_t outputIndex = 0; outputIndex < model.GetNumOutputs(); outputIndex++) {
- arm::app::DumpTensor(model.GetOutputTensor(outputIndex));
- }
+static void DumpInputs(const Model& model, const char* message)
+{
+ info("%s\n", message);
+ for (size_t inputIndex = 0; inputIndex < model.GetNumInputs(); inputIndex++) {
+ arm::app::DumpTensor(model.GetInputTensor(inputIndex));
+ }
+}
+
+static void DumpOutputs(const Model& model, const char* message)
+{
+ info("%s\n", message);
+ for (size_t outputIndex = 0; outputIndex < model.GetNumOutputs(); outputIndex++) {
+ arm::app::DumpTensor(model.GetOutputTensor(outputIndex));
+ }
+}
#endif /* VERIFY_TEST_OUTPUT */
- return true;
+bool RunInferenceHandler(ApplicationContext& ctx)
+{
+ auto& platform = ctx.Get<hal_platform&>("platform");
+ auto& profiler = ctx.Get<Profiler&>("profiler");
+ auto& model = ctx.Get<Model&>("model");
+
+ constexpr uint32_t dataPsnTxtInfStartX = 150;
+ constexpr uint32_t dataPsnTxtInfStartY = 40;
+
+ if (!model.IsInited()) {
+ printf_err("Model is not initialised! Terminating processing.\n");
+ return false;
}
+#if VERIFY_TEST_OUTPUT
+ DumpInputs(model, "Initial input tensors values");
+ DumpOutputs(model, "Initial output tensors values");
+#endif /* VERIFY_TEST_OUTPUT */
+
+ PopulateInputTensor(model);
+
+#if VERIFY_TEST_OUTPUT
+ DumpInputs(model, "input tensors populated");
+#endif /* VERIFY_TEST_OUTPUT */
+
+ /* Strings for presentation/logging. */
+ std::string str_inf{"Running inference... "};
+
+ /* Display message on the LCD - inference running. */
+ platform.data_psn->present_data_text(
+ str_inf.c_str(), str_inf.size(),
+ dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
+
+ if (!RunInference(model, profiler)) {
+ return false;
+ }
+
+ /* Erase. */
+ str_inf = std::string(str_inf.size(), ' ');
+ platform.data_psn->present_data_text(
+ str_inf.c_str(), str_inf.size(),
+ dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
+
+ info("Final results:\n");
+ profiler.PrintProfilingResult();
+
+#if VERIFY_TEST_OUTPUT
+ DumpOutputs(model, "output tensors post inference");
+#endif /* VERIFY_TEST_OUTPUT */
+
+#if defined (DYNAMIC_OFM_BASE) && defined(DYNAMIC_OFM_SIZE)
+ PopulateDynamicOfm(model);
+#endif /* defined (DYNAMIC_OFM_BASE) && defined(DYNAMIC_OFM_SIZE) */
+
+ return true;
+}
+
} /* namespace app */
} /* namespace arm */