Add support for Corstone-320

Enable running Executorch on Corstone-320, with Arm Ethos-U85.

Signed-off-by: Per Åstrand <[email protected]>
Change-Id: I45daa23871119ecb33cdc4abee112adc9a829076

Author: per

examples/arm/executor_runner/CMakeLists.txt

@@ -16,6 +16,8 @@ if(NOT DEFINED ET_PTE_FILE_PATH AND NOT ${SEMIHOSTING})
   )
 endif()
 
+set(TARGET_BOARD "corstone-300" CACHE STRING "Target board")
+
 # Example ExecuTorch demo for bare metal Cortex-M based systems
 set(ET_DIR_PATH
     "../../.."
@@ -55,10 +57,13 @@ endif()
 # libraries. We link against ethosu_target_init which includes all of these
 # dependencies.
 
-# For Corstone-300 FVP builds we put models into the larger DRAM area
-set(MEMORY_MODEL "dram")
-set(MEMORY_ARENA "dram")
-add_subdirectory(${ETHOS_SDK_PATH}/core_platform/targets/corstone-300 target)
+if(TARGET_BOARD STREQUAL "corstone-300")
+  add_subdirectory(${ETHOS_SDK_PATH}/core_platform/targets/corstone-300 target)
+elseif(TARGET_BOARD STREQUAL "corstone-320")
+  add_subdirectory(${ETHOS_SDK_PATH}/core_platform/targets/corstone-320 target)
+else()
+  message(FATAL_ERROR "Unsupported TARGET_BOARD: ${TARGET_BOARD}")
+endif()
 
 # Dependencies from the ExecuTorch build
 add_library(executorch STATIC IMPORTED)
@@ -171,7 +176,7 @@ endif()
 if(SEMIHOSTING)
   # Remove this when MLBEDSW-8910 is closed.
   set_source_files_properties(
-    ${ETHOS_SDK_PATH}/core_platform/targets/corstone-300/retarget.c
+    ${ETHOS_SDK_PATH}/core_platform/targets/${TARGET_BOARD}/retarget.c
     PROPERTIES HEADER_FILE_ONLY TRUE
   )
 endif()

examples/arm/executor_runner/arm_executor_runner.cpp

@@ -26,11 +26,11 @@
 // In our unit test flow, we have the capability to provide an enitre model to
 // the Corstone-3xx FVP using semi hosting. Hence, the input allocation pool
 // needs to be large enough to take an entire model. On the FVP,
-// network_model_sec is linked to the DDR, which is large (256MB on
+// input_data_sec is linked to the DDR, which is large (256MB on
 // Corstone-300).
 const size_t input_allocation_pool_size = 100 * 1024 * 1024;
 unsigned char __attribute__((
-    section("network_model_sec"),
+    section("input_data_sec"),
     aligned(16))) input_allocation_pool[input_allocation_pool_size];
 // memory for the model will be allocated from the input_allocation_pool
 char* model_pte = nullptr;
@@ -65,12 +65,12 @@ using executorch::runtime::TensorInfo;
 
 #define METHOD_ALLOCATOR_POOL_SIZE (70 * 1024 * 1024)
 unsigned char __attribute__((
-    section("network_model_sec"),
+    section("input_data_sec"),
     aligned(16))) method_allocation_pool[METHOD_ALLOCATOR_POOL_SIZE];
 
 const size_t temp_allocation_pool_size = 1 * 1024 * 1024;
 unsigned char __attribute__((
-    section("network_model_sec"),
+    section("input_data_sec"),
     aligned(16))) temp_allocation_pool[temp_allocation_pool_size];
 
 void et_pal_init(void) {}

examples/arm/executor_runner/arm_perf_monitor.cpp

@@ -38,11 +38,18 @@ void ethosu_inference_begin(struct ethosu_driver* drv, void*) {
   ETHOSU_PMU_PMCCNTR_CFG_Set_Stop_Event(drv, ETHOSU_PMU_NPU_IDLE);
   ETHOSU_PMU_PMCCNTR_CFG_Set_Start_Event(drv, ETHOSU_PMU_NPU_ACTIVE);
 
-  // Setup 4 counters
+// Setup 4 counters
+#if defined(ETHOSU55) || defined(ETHOSU65)
   ETHOSU_PMU_Set_EVTYPER(drv, 0, ETHOSU_PMU_AXI0_RD_DATA_BEAT_RECEIVED);
   ETHOSU_PMU_Set_EVTYPER(drv, 1, ETHOSU_PMU_AXI1_RD_DATA_BEAT_RECEIVED);
   ETHOSU_PMU_Set_EVTYPER(drv, 2, ETHOSU_PMU_AXI0_WR_DATA_BEAT_WRITTEN);
   ETHOSU_PMU_Set_EVTYPER(drv, 3, ETHOSU_PMU_NPU_IDLE);
+#elif defined(ETHOSU85)
+  ETHOSU_PMU_Set_EVTYPER(drv, 0, ETHOSU_PMU_EXT0_RD_DATA_BEAT_RECEIVED);
+  ETHOSU_PMU_Set_EVTYPER(drv, 1, ETHOSU_PMU_EXT1_RD_DATA_BEAT_RECEIVED);
+  ETHOSU_PMU_Set_EVTYPER(drv, 2, ETHOSU_PMU_EXT0_WR_DATA_BEAT_WRITTEN);
+  ETHOSU_PMU_Set_EVTYPER(drv, 3, ETHOSU_PMU_NPU_IDLE);
+#endif
   // Enable 4 counters
   ETHOSU_PMU_CNTR_Enable(drv, 0xf);