Arm backend: Add support for ETDump of outputs

Differential Revision: D82454608

Pull Request resolved: #14143

Author: zingo

backends/arm/scripts/build_executor_runner.sh

@@ -44,7 +44,7 @@ help() {
     echo "  --memory_mode=<CONFIG>               Vela memory mode, used for setting the Timing Adapter parameters of the Corstone platforms."
     echo "                                       Valid values are Shared_Sram(for Ethos-U55, Ethos-U65, Ethos-85), Sram_Only(for Ethos-U55, Ethos-U65, Ethos-U85) or Dedicated_Sram(for Ethos-U65, Ethos-U85)."
     echo "                                       Default: Shared_Sram for the Ethos-U55 and Sram_Only for the Ethos-U85"
-    echo "  --etdump                             Adds Devtools etdump support to track timing, etdump area will be base64 encoded in the log"
+    echo "  --etdump                             Adds Devtools etdump support to track timing and output, etdump area will be base64 encoded in the log"
     echo "  --extra_build_flags=<FLAGS>          Extra flags to pass to cmake like -DET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE=60000 Default: none "
     echo "  --output=<FOLDER>                    Output folder Default: <MODEL>/<MODEL>_<TARGET INFO>.pte"
     echo "  --et_build_root=<FOLDER>             Build output root folder to use, defaults to ${et_build_root}"
@@ -161,7 +161,7 @@ if [ "$bundleio" = true ] ; then
 fi
 
 if [ "$build_with_etdump" = true ] ; then
-    build_with_etdump_flags=" -DEXECUTORCH_ENABLE_EVENT_TRACER=ON "
+    build_with_etdump_flags=" -DEXECUTORCH_ENABLE_EVENT_TRACER=ON -DET_DUMP_INTERMEDIATE_OUTPUTS=ON "
 fi
 
 echo "Building with BundleIO/etdump/extra flags: ${build_bundleio_flags} ${build_with_etdump_flags} ${extra_build_flags}"

backends/arm/scripts/run_fvp.sh

@@ -21,6 +21,7 @@ elf_file=""
 data_file=""
 target="ethos-u55-128"
 timeout="600"
+etrecord_file=""
 
 help() {
     echo "Usage: $(basename $0) [options]"
@@ -29,6 +30,7 @@ help() {
     echo "  --data=<FILE>@<ADDRESS>  Place a file in memory at this address, useful to emulate a PTE flashed into memory instead as part of the code."
     echo "  --target=<TARGET>        Target to build and run for Default: ${target}"
     echo "  --timeout=<TIME_IN_SEC>  Maximum target runtime, used to detect hanging, might need to be higer on large models Default: ${timeout}"
+    echo "  --etrecord=<FILE>        If ETDump is used you can supply a ETRecord file matching the PTE"
     exit 0
 }
 
@@ -39,6 +41,7 @@ for arg in "$@"; do
       --data=*) data_file="--data ${arg#*=}";;
       --target=*) target="${arg#*=}";;
       --timeout=*) timeout="${arg#*=}";;
+      --etrecord=*) etrecord_file="${arg#*=}";;
       *)
       ;;
     esac
@@ -115,15 +118,23 @@ echo "Checking for a etdump in log"
 ! grep "#\[RUN THIS\]" ${log_file} >/dev/null
 if [ $? != 0 ]; then
     echo "Found ETDump in log!"
+    devtools_extra_args=""
     echo "#!/bin/sh" > etdump_script.sh
     sed -n '/^#\[RUN THIS\]$/,/^#\[END\]$/p' ${log_file} >> etdump_script.sh
     # You can run etdump_script.sh if you do
     # $ chmod a+x etdump_script.sh
     # $ ./etdump_script.sh
     # But lets not trust the script as a bad patch would run bad code on your machine
-    grep ">etdump.bin" etdump_script.sh | cut -d\" -f2- | cut -d\" -f1 >etdump.base64
-    base64 -d etdump.base64 >etdump.bin
-    python3 -m devtools.inspector.inspector_cli --etdump_path etdump.bin  --source_time_scale cycles --target_time_scale cycles
+    grep ">etdump.bin" etdump_script.sh | cut -d\" -f2- | cut -d\" -f1 | base64 -d >etdump.bin
+    ! grep ">debug_buffer.bin" etdump_script.sh >/dev/null
+    if [ $? != 0 ]; then
+        grep ">debug_buffer.bin" etdump_script.sh | cut -d\" -f2- | cut -d\" -f1 | base64 -d >debug_buffer.bin
+        devtools_extra_args="${devtools_extra_args} --debug_buffer_path debug_buffer.bin"
+    fi
+    if [[ ${etrecord_file} != "" ]]; then
+        devtools_extra_args="${devtools_extra_args} --etrecord_path ${etrecord_file}"
+    fi
+    python3 -m devtools.inspector.inspector_cli --etdump_path etdump.bin ${devtools_extra_args} --source_time_scale cycles --target_time_scale cycles
 fi
 
 echo "Checking for problems in log:"

backends/arm/test/test_model.py

@@ -184,7 +184,7 @@ def build_ethosu_runtime(
             "--build_type=Release",
             f"--system_config={system_config}",
             f"--memory_mode={memory_mode}",
-            f"--extra_build_flags=-DET_DUMP_OUTPUT=OFF {extra_flags}",
+            f"--extra_build_flags=-DET_LOG_DUMP_OUTPUT=OFF {extra_flags}",
             f"--output={elf_build_path}",
         ]
     )

examples/arm/executor_runner/CMakeLists.txt

@@ -6,30 +6,59 @@
 cmake_minimum_required(VERSION 3.20)
 project(arm_executor_runner)
 
-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_MODEL_PTE_ADDR
   "Place in memory that the PTE file is located/flashed, if set to OFF the PTE is built into the code as a big data area."
   OFF
 )
-option(ET_BUNDLE_IO "Set to compile in BundleIO support" OFF)
-option(ET_ATOL "Set atol to use for BundleIO testing" OFF)
-option(ET_RTOL "Set rtol to use for BundleIO testing" OFF)
-option(ET_DUMP_INPUT "Dump input in log" OFF)
-option(ET_DUMP_OUTPUT "Dump output in log" ON)
-option(FETCH_ETHOS_U_CONTENT
-       "Fetch ethos_u dependencies instead of relying on pre-downloads" ON
-)
+
 set(ET_NUM_INFERENCES
     "1"
     CACHE STRING "Number of inferences to run"
 )
 
+option(ET_LOG_DUMP_INPUT "Dump input in log" OFF)
+option(ET_LOG_DUMP_OUTPUT "Dump output in log" ON)
+
+option(ET_BUNDLE_IO "Set to compile in BundleIO support" OFF)
+set(ET_ATOL
+    "0.01"
+    CACHE STRING "Set atol to use for BundleIO testing (Requires ET_BUNDLE_IO)"
+)
+set(ET_RTOL
+    "0.01"
+    CACHE STRING "Set atol to use for BundleIO testing (Requires ET_BUNDLE_IO)"
+)
+
+option(
+  ET_DUMP_OUTPUTS
+  "Collect and print outputs as a base64 buffer in the log (Requires EXECUTORCH_ENABLE_EVENT_TRACER)"
+  OFF
+)
+option(
+  ET_DUMP_INTERMEDIATE_OUTPUTS
+  "Collect and print intermediate outputs as a base64 buffer in the log (Requires EXECUTORCH_ENABLE_EVENT_TRACER)"
+  OFF
+)
+set(ET_DEBUG_BUFFER_SIZE
+    "2097152"
+    CACHE
+      STRING
+      "Size of buffer to collect intermediate outputs/outputs buffers (Requires EXECUTORCH_ENABLE_EVENT_TRACER and ET_DUMP_OUTPUTS or ET_DUMP_INTERMEDIATE_OUTPUTS)"
+)
+
+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(FETCH_ETHOS_U_CONTENT
+       "Fetch ethos_u dependencies instead of relying on pre-downloads" ON
+)
+
 if(NOT DEFINED ET_MODEL_PTE_ADDR
    AND NOT DEFINED ET_PTE_FILE_PATH
    AND NOT DEFINED SEMIHOSTING
@@ -322,37 +351,29 @@ if(NOT ${ET_MODEL_PTE_ADDR} AND NOT SEMIHOSTING)
   add_dependencies(arm_executor_runner gen_model_header)
 endif()
 
-if(SEMIHOSTING)
-  target_compile_definitions(arm_executor_runner PUBLIC SEMIHOSTING)
-endif()
-
-if(ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE)
+if(ET_MODEL_PTE_ADDR)
   target_compile_definitions(
-    arm_executor_runner
-    PUBLIC
-      ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE}
+    arm_executor_runner PUBLIC -DET_MODEL_PTE_ADDR=${ET_MODEL_PTE_ADDR}
   )
 endif()
 
-target_compile_definitions(
-  arm_executor_runner
-  PUBLIC
-    ET_ARM_BAREMETAL_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE}
-)
-if(DEFINED ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE)
+if(ET_NUM_INFERENCES)
   target_compile_definitions(
-    arm_executor_runner
-    PUBLIC
-      ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE}
+    arm_executor_runner PUBLIC ET_NUM_INFERENCES=${ET_NUM_INFERENCES}
   )
 endif()
 
-if(ET_MODEL_PTE_ADDR)
-  target_compile_definitions(
-    arm_executor_runner PUBLIC -DET_MODEL_PTE_ADDR=${ET_MODEL_PTE_ADDR}
-  )
+if(ET_LOG_DUMP_INPUT)
+  target_compile_definitions(arm_executor_runner PUBLIC -DET_LOG_DUMP_INPUT)
+endif()
+
+if(ET_LOG_DUMP_OUTPUT)
+  target_compile_definitions(arm_executor_runner PUBLIC -DET_LOG_DUMP_OUTPUT)
 endif()
 
+# Devtool BundleIO: Use Bundle PTE with input and reference output included to
+# check if it matches.
+
 if(ET_BUNDLE_IO)
   target_compile_definitions(arm_executor_runner PUBLIC -DET_BUNDLE_IO)
 endif()
@@ -365,17 +386,50 @@ if(ET_RTOL)
   target_compile_definitions(arm_executor_runner PUBLIC ET_RTOL=${ET_RTOL})
 endif()
 
-if(ET_DUMP_INPUT)
-  target_compile_definitions(arm_executor_runner PUBLIC -DET_DUMP_INPUT)
+# Devtools ETDump: Speed and dumping output
+
+if(ET_DUMP_OUTPUTS)
+  target_compile_definitions(arm_executor_runner PUBLIC -DET_DUMP_OUTPUTS)
 endif()
 
-if(ET_DUMP_OUTPUT)
-  target_compile_definitions(arm_executor_runner PUBLIC -DET_DUMP_OUTPUT)
+if(ET_DUMP_INTERMEDIATE_OUTPUTS)
+  target_compile_definitions(
+    arm_executor_runner PUBLIC -DET_DUMP_INTERMEDIATE_OUTPUTS
+  )
 endif()
 
-if(ET_NUM_INFERENCES)
+if(ET_DEBUG_BUFFER_SIZE)
   target_compile_definitions(
-    arm_executor_runner PUBLIC ET_NUM_INFERENCES=${ET_NUM_INFERENCES}
+    arm_executor_runner PUBLIC ET_DEBUG_BUFFER_SIZE=${ET_DEBUG_BUFFER_SIZE}
+  )
+endif()
+
+# Semihosting FVP (FVP Simulator can access host filesystem)
+
+if(SEMIHOSTING)
+  target_compile_definitions(arm_executor_runner PUBLIC SEMIHOSTING)
+endif()
+
+# Memory buffer sizes for Executorch flow
+
+if(ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE)
+  target_compile_definitions(
+    arm_executor_runner
+    PUBLIC
+      ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_METHOD_ALLOCATOR_POOL_SIZE}
+  )
+endif()
+
+target_compile_definitions(
+  arm_executor_runner
+  PUBLIC
+    ET_ARM_BAREMETAL_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE}
+)
+if(DEFINED ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE)
+  target_compile_definitions(
+    arm_executor_runner
+    PUBLIC
+      ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE=${ET_ARM_BAREMETAL_FAST_SCRATCH_TEMP_ALLOCATOR_POOL_SIZE}
   )
 endif()