Arm backend: Set REGIONCFG registers of the Ethos-U (pytorch#10388)

The REGIONCFG registers need to be set differently depending on the
memory mode for optimal performance.

Minor fix to how we create the base64 encoded binary file when using
etdump. The new command is more portable and works on Ubuntu & Mac

Author: gggekov

examples/arm/executor_runner/CMakeLists.txt

@@ -221,6 +221,7 @@ elseif(SYSTEM_CONFIG STREQUAL "Ethos_U55_Deep_Embedded")
         )
     elseif(MEMORY_MODE STREQUAL "Sram_Only")
       target_compile_definitions(ethosu_target_common INTERFACE
+      ETHOSU_MODEL=1
       # Configure NPU architecture timing adapters
       # This is just example numbers and you should make this match your hardware
       # SRAM
@@ -411,6 +412,53 @@ else()
   message(FATAL_ERROR "Unsupported SYSTEM_CONFIG: ${SYSTEM_CONFIG}")
 endif()
 
+# The REGIONCFG registers of the Ethos-U control whether the NPU
+# reads/writes data through the SRAM or the external memory.
+# By default, the Ethos-U driver provides REGIONCFG configuration for Shared Sram memory mode.
+# For Sram_Only and Dedicated_Sram memory modes, we need to change the settings for optimal performance.
+#
+# Currently, the convention used by Vela and the Ethos-U driver is that the NPU uses:
+# Region 0 for traffic of the Read-Only data(weights & biases)
+# Region 1 for traffic of of the intermediate Read/Write buffers required for the computation
+# Region 2 for traffic of of the cache in Dedicated_Sram memory mode(not applicable in Sram_Only or Shared_Sram)
+#
+# NOTE: The above convention is determined by the Vela compiler and the Ethos-U driver and can change in the future.
+#
+# Common definitions:
+# For Ethos-U55/U65/U85, region configs are set as:
+#   0 or 1 = AXI0 (Ethos-U55 or Ethos-U65) or AXI_SRAM(Ethos-U85)
+#   2 or 3 = AXI1 (Ethos-U55 or Ethos-U65) or AXI_EXT(Ethos-U85)
+#
+# When we compile a model for Sram_Only, the memory traffic for Region 0 and Region 1 should pass via the SRAM(hence regioncfg = 1)
+# When we compile a model for Dedicated_Sram, the memory traffic for Region 0 should pass via the external memory(3),
+# the memory traffic of Region 1 should pass via the external memory(3) and the traffic for Region 2 should pass via the SRAM(0)
+#
+
+if(MEMORY_MODE STREQUAL "Sram_Only")
+  target_compile_definitions(ethosu_core_driver PRIVATE
+      NPU_QCONFIG=1
+      NPU_REGIONCFG_0=1
+      NPU_REGIONCFG_1=0
+      NPU_REGIONCFG_2=0
+      NPU_REGIONCFG_3=0
+      NPU_REGIONCFG_4=0
+      NPU_REGIONCFG_5=0
+      NPU_REGIONCFG_6=0
+      NPU_REGIONCFG_7=0)
+  elseif(MEMORY_MODE STREQUAL "Dedicated_Sram")
+    target_compile_definitions(ethosu_core_driver PRIVATE
+      NPU_QCONFIG=3
+      NPU_REGIONCFG_0=3
+      NPU_REGIONCFG_1=3
+      NPU_REGIONCFG_2=0
+      NPU_REGIONCFG_3=0
+      NPU_REGIONCFG_4=0
+      NPU_REGIONCFG_5=0
+      NPU_REGIONCFG_6=0
+      NPU_REGIONCFG_7=0)
+endif()
+
+
 # Dependencies from the ExecuTorch build
 add_library(executorch STATIC IMPORTED)
 set_property(

examples/arm/executor_runner/arm_executor_runner.cpp

@@ -766,7 +766,7 @@ int main(int argc, const char* argv[]) {
       encoded_buf[encoded_len] = 0x00; // Ensure null termination
       ET_LOG(Info, "Writing etdump.bin [base64]");
       printf(
-          "#---\nbase64 -i -d <<<\"\\\n%s\\\n\" >etdump.bin\npython3 -m devtools.inspector.inspector_cli --etdump_path etdump.bin  --source_time_scale cycles --target_time_scale cycles\n#---\n",
+          "#---\necho \"%s\" | base64 -d >etdump.bin\npython3 -m devtools.inspector.inspector_cli --etdump_path etdump.bin  --source_time_scale cycles --target_time_scale cycles\n#---\n",
           encoded_buf);
     } else {
       ET_LOG(