Add input tensor CRC to Generic Benchmark application.

Only use -O3 -LNO:simd with the xtensa decompress.cc target.
rkuester · Nov 20, 2024 · 0a49b2a · 0a49b2a
1 parent 83dafce
commit 0a49b2a
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 4 deletions.
diff --git a/tensorflow/lite/micro/tools/benchmarking/generic_model_benchmark.cc b/tensorflow/lite/micro/tools/benchmarking/generic_model_benchmark.cc
@@ -170,6 +170,16 @@ void ShowOutputCRC32(tflite::MicroInterpreter* interpreter) {
   }
 }
 
+void ShowInputCRC32(tflite::MicroInterpreter* interpreter) {
+  GenCRC32Table();
+  for (size_t i = 0; i < interpreter->inputs_size(); ++i) {
+    TfLiteTensor* input = interpreter->input_tensor(i);
+    uint8_t* input_values = tflite::GetTensorData<uint8_t>(input);
+    uint32_t crc32_value = ComputeCRC32(input_values, input->bytes);
+    MicroPrintf("Input CRC32: 0x%X", crc32_value);
+  }
+}
+
 int Benchmark(const uint8_t* model_data, tflite::PrettyPrintType print_type) {
   static Profiler profiler;
   static Profiler profiler2;
@@ -225,6 +235,9 @@ int Benchmark(const uint8_t* model_data, tflite::PrettyPrintType print_type) {
   uint32_t seed = kRandomSeed;
   while (true) {
     SetRandomInput(seed++, interpreter);
+    ShowInputCRC32(&interpreter);
+    MicroPrintf("");  // null MicroPrintf serves as a newline.
+
     TfLiteStatus status = interpreter.Invoke();
     if ((status != kTfLiteOk) && (static_cast<int>(status) != kTfLiteAbort)) {
       MicroPrintf("Model interpreter invocation failed: %d\n", status);

diff --git a/tensorflow/lite/micro/tools/make/ext_libs/xtensa.inc b/tensorflow/lite/micro/tools/make/ext_libs/xtensa.inc
@@ -58,10 +58,6 @@ else ifeq ($(TARGET_ARCH), hifi5)
   CCFLAGS += $(PLATFORM_FLAGS)
   CXXFLAGS += $(PLATFORM_FLAGS)
 
-  # override KERNEL_OPTIMIZATION_LEVEL to enable higher performance
-  # Xtensa intrinsics.
-  KERNEL_OPTIMIZATION_LEVEL := -O3 -LNO:simd
-
   NNLIB_PATH := $(MAKEFILE_DIR)/downloads/xa_nnlib_hifi5
   NDSPLIB_PATH := $(MAKEFILE_DIR)/downloads/ndsplib-hifi5
 

diff --git a/tensorflow/lite/micro/tools/make/targets/xtensa_makefile.inc b/tensorflow/lite/micro/tools/make/targets/xtensa_makefile.inc
@@ -97,4 +97,10 @@ ifeq ($(OPTIMIZED_KERNEL_DIR), xtensa)
   $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/xtensa/lstm_eval.cc \
   $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/xtensa/lstm_eval_hifi.cc \
   $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/xtensa/unidirectional_sequence_lstm.cc
+
+  # override KERNEL_OPTIMIZATION_LEVEL to enable higher performance
+  # Xtensa intrinsics.
+$(KERNEL_OBJDIR)$(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/xtensa/decompress.o: $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/xtensa/decompress.cc
+	@mkdir -p $(dir $@)
+	$(CXX) $(CXXFLAGS) -O3 -LNO:simd $(INCLUDES) -c $< -o $@
 endif