Arm backend: Add devtools support to example

New flags on run.sh
--etdump
  Build in etdump and profiling, the etdump base64 coded and put in
  the log
--debug_build
  Build debug instead of release
--extra_build_flags
  Extra flags to pass to cmake this makes it for example possible to
  override the allocator pool size or other build time cmake flags.

The devtools build has been updated so FLATCC_EXECUTABLE can be used
to point out the executable.

Signed-off-by: Zingo Andersen <[email protected]>
Change-Id: Ic0fb1e48ee633c5fe91473bdc2db9e894b2fc4fa

CMakeLists.txt

@@ -673,10 +673,17 @@ if(EXECUTORCH_BUILD_XNNPACK)
 endif()
 
 if(EXECUTORCH_BUILD_DEVTOOLS)
-  set(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER
-      ON
-      CACHE BOOL "EXECUTORCH_BUILD_EXTENSION_DATA_LOADER" FORCE
-  )
+  if(NOT EXECUTORCH_BUILD_ARM_BAREMETAL)
+    set(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER
+        ON
+        CACHE BOOL "EXECUTORCH_BUILD_EXTENSION_DATA_LOADER" FORCE
+    )
+  else()
+    set(EXECUTORCH_BUILD_EXTENSION_DATA_LOADER
+        OFF
+        CACHE BOOL "EXECUTORCH_BUILD_EXTENSION_DATA_LOADER" FORCE
+    )
+  endif()
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/devtools)
 endif()
 

backends/arm/runtime/ArmBackendEthosU.cpp

@@ -15,6 +15,39 @@
 
 #include <ethosu_driver.h>
 
+#if defined(ET_EVENT_TRACER_ENABLED)
+#include <executorch/runtime/core/event_tracer.h>
+#include <executorch/runtime/core/event_tracer_hooks.h>
+using executorch::runtime::EventTracer;
+using executorch::runtime::EventTracerEntry;
+
+class EventTraceScope {
+ public:
+  EventTraceScope(EventTracer* event_tracer_, const char* name) {
+    event_tracer = event_tracer_;
+    event_tracer_entry_scope = event_tracer->start_profiling(name);
+  }
+  ~EventTraceScope() {
+    event_tracer->end_profiling(event_tracer_entry_scope);
+  }
+
+ private:
+  EventTracer* event_tracer;
+  EventTracerEntry event_tracer_entry_scope;
+};
+#define EXECUTORCH_PROF_SCOPE(EVENTTRACER, NAME) \
+  EventTraceScope event_tracer_scope = EventTraceScope(EVENTTRACER, NAME)
+#define EXECUTORCH_PROF_START(EVENTTRACER, SCOPE, NAME) \
+  SCOPE = EVENTTRACER->start_profiling(NAME)
+#define EXECUTORCH_PROF_END(EVENTTRACER, SCOPE) \
+  EVENTTRACER->end_profiling(SCOPE)
+
+#else
+#define EXECUTORCH_PROF_SCOPE(EVENTTRACER, NAME)
+#define EXECUTORCH_PROF_START(EVENTTRACER, SCOPE, NAME)
+#define EXECUTORCH_PROF_END(EVENTTRACER, SCOPE)
+#endif
+
 #include <executorch/backends/arm/runtime/VelaBinStream.h>
 #include <executorch/runtime/backend/interface.h>
 #include <executorch/runtime/core/error.h>
@@ -109,20 +142,38 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
       BackendExecutionContext& context,
       DelegateHandle* input_handle,
       EValue** args) const override {
+#if defined(ET_EVENT_TRACER_ENABLED)
+    EventTracer* event_tracer = context.event_tracer();
+    EventTracerEntry event_tracer_local_scope;
+#endif
+
+    EXECUTORCH_PROF_SCOPE(event_tracer, "ArmBackend::execute()");
+    ArmBackendExecuteCallbacks ArmBackend_execute_callbacks;
+
     ExecutionHandle* execution_handle = (ExecutionHandle*)input_handle;
     VelaHandles handles;
 
-    ArmBackendExecuteCallbacks ArmBackend_execute_callbacks;
     // Command stream - we know at this point it's aligned
+    EXECUTORCH_PROF_START(
+        event_tracer,
+        event_tracer_local_scope,
+        "+ArmBackend::execute()processed_data");
     char* data = (char*)execution_handle->processed->data();
+    EXECUTORCH_PROF_END(event_tracer, event_tracer_local_scope);
+
     ET_LOG(Debug, "ArmBackend::execute %p", data);
 
+    EXECUTORCH_PROF_START(
+        event_tracer,
+        event_tracer_local_scope,
+        "+ArmBackend::execute()vela_bin_read()");
     // Read key sections from the vela_bin_stream
     if (vela_bin_read(data, &handles, execution_handle->processed->size()) ==
         false) {
       ET_LOG(Error, "ArmBackend::vela_read: error, invalid binary layout");
       return Error::InvalidProgram;
     }
+    EXECUTORCH_PROF_END(event_tracer, event_tracer_local_scope);
 
     ET_LOG(
         Debug,
@@ -186,6 +237,9 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
 
       // Select a compatible copy routine
       if (both_char and permuted_input_shape) {
+        EXECUTORCH_PROF_SCOPE(
+            event_tracer,
+            "+ArmBackend::execute()handles.input.permute_CHW_to_HWC()");
         // permuted byte copy CHW to HWC
         permute_CHW_to_HWC(
             tensor_in.mutable_data_ptr<char>(),
@@ -194,6 +248,8 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
             tensor_in.size(2),
             tensor_in.size(3));
       } else if (both_char or both_int) {
+        EXECUTORCH_PROF_SCOPE(
+            event_tracer, "+ArmBackend::execute()handles.input.memcpy()");
         // Sizes match and elt size matches so memcpy
         memcpy(
             scratch_addr,
@@ -234,14 +290,18 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
         (uint64_t)handles.weight_data, (uint64_t)handles.scratch_data};
     size_t bases_size[2] = {
         handles.weight_data_size, handles.scratch_data_size};
-    int result = ethosu_invoke_v3(
+    int result = 0;
+    EXECUTORCH_PROF_START(
+        event_tracer, event_tracer_local_scope, "+ArmBackend::execute()NPU");
+    result = ethosu_invoke_v3(
         driver.get(),
         (void*)handles.cmd_data,
         handles.cmd_data_size,
         bases,
         bases_size,
         2, /* fixed array of pointers to binary interface*/
         nullptr);
+    EXECUTORCH_PROF_END(event_tracer, event_tracer_local_scope);
 
     if (result != 0) {
       ET_LOG(
@@ -277,6 +337,10 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
           &permuted_output_shape));
       if (tensor_out.scalar_type() == ScalarType::Char and
           permuted_output_shape) {
+        EXECUTORCH_PROF_SCOPE(
+            event_tracer,
+            "+ArmBackend::execute()handles.output.permute_HWC_to_CHW()");
+
         char* output_address = (char*)output_addr;
         permute_HWC_to_CHW(
             output_address,
@@ -285,6 +349,8 @@ class ArmBackend final : public ::executorch::runtime::BackendInterface {
             tensor_out.size(2),
             tensor_out.size(3));
       } else {
+        EXECUTORCH_PROF_SCOPE(
+            event_tracer, "+ArmBackend::execute()handles.output.move()");
         for (int j = 0; j < tensor_out.numel(); j++) {
           if (tensor_out.scalar_type() == ScalarType::Char) {
             char* output_address = (char*)output_addr;

devtools/CMakeLists.txt

@@ -13,12 +13,14 @@ cmake_minimum_required(VERSION 3.19)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
+set(_flatcc_source_dir ${CMAKE_CURRENT_SOURCE_DIR}/../third-party/flatcc)
+
 if(NOT CMAKE_CXX_STANDARD)
   set(CMAKE_CXX_STANDARD 17)
 endif()
 
 if(NOT FLATCC_EXECUTABLE)
-  set(FLATCC_EXECUTABLE flatcc)
+  set(FLATCC_EXECUTABLE ${_flatcc_source_dir}/bin/flatcc)
 endif()
 
 # Source root directory for executorch.
@@ -66,7 +68,7 @@ set(FLATCC_DEBUG_CLANG_SANITIZE
     OFF
     CACHE BOOL ""
 )
-set(_flatcc_source_dir ${CMAKE_CURRENT_SOURCE_DIR}/../third-party/flatcc)
+
 add_subdirectory(${_flatcc_source_dir} ${CMAKE_BINARY_DIR}/third-party/flatcc)
 
 # Fix for "relocation R_X86_64_32 against `.rodata' can not be used when making
@@ -163,7 +165,7 @@ add_custom_command(
     # Note that the flatcc project actually writes its outputs into the source
     # tree instead of under the binary directory, and there's no way to change
     # that behavior.
-    ${_flatcc_source_dir}/bin/flatcc -cwr -o
+    ${FLATCC_EXECUTABLE} -cwr -o
     ${_program_schema__include_dir}/executorch/devtools/etdump
     ${_etdump_schema__srcs}
   COMMAND rm -rf ${_etdump_schema_cleanup_paths}

examples/arm/CMakeLists.txt

@@ -57,3 +57,8 @@ generate_bindings_for_kernels(
 gen_operators_lib(
   LIB_NAME "arm_portable_ops_lib" KERNEL_LIBS portable_kernels DEPS executorch
 )
+
+if(EXECUTORCH_ENABLE_EVENT_TRACER)
+  target_compile_options(executorch INTERFACE -DET_EVENT_TRACER_ENABLED)
+  target_compile_options(portable_ops_lib INTERFACE -DET_EVENT_TRACER_ENABLED)
+endif()

examples/arm/aot_arm_compiler.py

@@ -273,7 +273,7 @@ def get_compile_spec(
                 target,
                 system_config="Ethos_U55_High_End_Embedded",
                 memory_mode="Shared_Sram",
-                extra_flags="--debug-force-regor --output-format=raw --verbose-operators --verbose-cycle-estimate",
+                extra_flags="--debug-force-regor --output-format=raw --verbose-operators",
             )
             .set_permute_memory_format(True)
             .set_quantize_io(True)
@@ -286,7 +286,7 @@ def get_compile_spec(
                 target,
                 system_config="Ethos_U85_SYS_DRAM_Mid",
                 memory_mode="Shared_Sram",
-                extra_flags="--output-format=raw --verbose-operators --verbose-cycle-estimate",
+                extra_flags="--output-format=raw --verbose-operators",
             )
             .set_permute_memory_format(True)
             .set_quantize_io(True)

examples/arm/ethos-u-setup/core_platform/patches/0001-Move-rodata-to-the-DDR.patch

@@ -20,11 +20,12 @@ index b458fc6..8d4bc73 100644
 
      KEEP(*(.eh_frame*))
    } > ITCM :rom_exec
-@@ -280,7 +280,7 @@ SECTIONS
+@@ -280,7 +280,8 @@ SECTIONS
  #endif
      * (expected_output_data_sec)
      * (sec_command_stream, sec_weight_data, sec_input_data)
 -
++    *(.got*)
 +    *(.rodata*)
      * (ethosu_core_in_queue)
      * (ethosu_core_out_queue)

examples/arm/executor_runner/CMakeLists.txt

@@ -10,7 +10,6 @@ option(SEMIHOSTING "Enable semihosting" OFF)
 option(ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE "Set ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE to specify memory alloction pool size" OFF)
 option(ET_ARM_BAREMETAL_TEMP_ALLOCATOR_POOL_SIZE "Set ET_ARM_BAREMETAL_TEMP_ALLOCATOR_POOL_SIZE to specify temp alloction pool size" OFF)
 
-
 if(NOT DEFINED ET_PTE_FILE_PATH AND NOT ${SEMIHOSTING})
   message(
     FATAL_ERROR
@@ -220,10 +219,8 @@ target_sources(
 # Include the target's bare-metal linker script
 ethosu_eval_link_options(arm_executor_runner)
 
-# Need whole-archive to ensure C++ ctor's are called - this may be wasteful for
-# bin size as we link in a number of other symbols
-target_link_libraries(
-  arm_executor_runner
+set(arm_executor_runner_link)
+list(APPEND arm_executor_runner_link
   extension_runner_util
   ethosu_target_init
   executorch
@@ -237,6 +234,44 @@ target_link_libraries(
   -Xlinker -Map=arm_executor_runner.map
 )
 
+if(EXECUTORCH_ENABLE_EVENT_TRACER)
+  target_compile_options(arm_executor_runner PUBLIC -DET_EVENT_TRACER_ENABLED)
+
+  add_library(etdump STATIC IMPORTED)
+  set_property(
+      TARGET etdump
+      PROPERTY IMPORTED_LOCATION
+           "${ET_BUILD_DIR_PATH}/lib/libetdump.a"
+  )
+
+  if(CMAKE_BUILD_TYPE MATCHES "Debug")
+    set(FLATCCRT_LIB flatccrt_d)
+  else()
+    set(FLATCCRT_LIB flatccrt)
+  endif()
+
+  add_library(${FLATCCRT_LIB} STATIC IMPORTED)
+  set_property(
+      TARGET ${FLATCCRT_LIB}
+      PROPERTY IMPORTED_LOCATION
+           "${ET_BUILD_DIR_PATH}/lib/lib${FLATCCRT_LIB}.a"
+  )
+
+  list(APPEND arm_executor_runner_link
+    etdump
+    ${FLATCCRT_LIB}
+  )
+endif()
+
+# Need whole-archive to ensure C++ ctor's are called - this may be wasteful for
+# bin size as we link in a number of other symbols
+target_link_libraries(
+  arm_executor_runner
+  ${arm_executor_runner_link}
+)
+
+target_link_options( arm_executor_runner PUBLIC LINKER:-Map=arm_executor_runner.map )
+
 # ET headers and generated headers includes
 target_include_directories(
   arm_executor_runner PRIVATE ${ET_INCLUDE_PATH} ${CMAKE_CURRENT_BINARY_DIR}