Adding nnlib as subrepo

.gitmodules

@@ -68,3 +68,6 @@
 [submodule "third-party/ios-cmake"]
 	path = third-party/ios-cmake
 	url = https://github.com/leetal/ios-cmake
+[submodule "backends/cadence/hifi/third-party/nnlib/nnlib-hifi4"]
+	path = backends/cadence/hifi/third-party/nnlib/nnlib-hifi4
+	url = https://github.com/foss-xtensa/nnlib-hifi4.git

CMakeLists.txt

@@ -417,7 +417,7 @@ target_link_libraries(executorch_no_prim_ops PRIVATE program_schema)
 # Check if dl exists for this toolchain and only then link it.
 find_library(DL_LIBRARY_EXISTS NAMES dl)
 # Check if the library was found
-if(DL_LIBRARY_EXISTS)
+if(DL_LIBRARY_EXISTS AND NOT EXECUTORCH_BUILD_CADENCE)
   target_link_libraries(executorch_no_prim_ops PRIVATE dl) # For dladdr()
 endif()
 target_include_directories(
@@ -443,7 +443,9 @@ target_link_options_shared_lib(executorch)
 # Real integrations should supply their own YAML file that only lists the
 # operators necessary for the models that will run.
 #
+if(NOT EXECUTORCH_BUILD_CADENCE)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/kernels/portable)
+endif()
 
 if(EXECUTORCH_BUILD_KERNELS_CUSTOM)
   # TODO: move all custom kernels to ${CMAKE_CURRENT_SOURCE_DIR}/kernels/custom
@@ -496,6 +498,8 @@ if(EXECUTORCH_BUILD_EXECUTOR_RUNNER)
 
   if(EXECUTORCH_BUILD_KERNELS_OPTIMIZED)
     list(APPEND _executor_runner_libs optimized_native_cpu_ops_lib)
+  elseif(EXECUTORCH_BUILD_CADENCE)
+    list(APPEND _executor_runner_libs cadence_ops_lib)
   else()
     list(APPEND _executor_runner_libs portable_ops_lib)
   endif()
@@ -566,6 +570,10 @@ if(EXECUTORCH_BUILD_COREML)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/apple/coreml)
 endif()
 
+if(EXECUTORCH_BUILD_CADENCE)
+   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/backends/cadence)
+endif()
+
 if(EXECUTORCH_BUILD_PYBIND)
   add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/third-party/pybind11)
 

backends/cadence/CMakeLists.txt

@@ -12,7 +12,7 @@ if(NOT CMAKE_CXX_STANDARD)
 endif()
 
 # Set the project name.
-project(cadence_executorch_example)
+project(cadence_backend)
 
 # Source root directory for executorch.
 if(NOT EXECUTORCH_ROOT)
@@ -21,121 +21,21 @@ endif()
 
 include(${EXECUTORCH_ROOT}/build/Utils.cmake)
 
-if(NOT PYTHON_EXECUTABLE)
-  resolve_python_executable()
-endif()
-
 # Let files say "include <executorch/path/to/header.h>".
 set(_common_include_directories ${EXECUTORCH_ROOT}/..)
 
-# Find prebuilt executorch lib
-find_package(executorch CONFIG REQUIRED)
-
-add_compile_options(
-  -DSDK_DEBUGCONSOLE=1
-  -DSERIAL_PORT_TYPE_UART=1
-  -DDEBUG_CONSOLE_RX_ENABLE=0
-  -DDEBUG
-  -DCPU_MIMXRT685SFVKB_dsp
-  -DMCUXPRESSO_SDK
-  -g
-  -O0
-  -Wall
-  -fsigned-char
-  -Wno-missing-braces
-  -fmessage-length=0
-  -DPRINTF_FLOAT_ENABLE=1
-)
-
-if(NOT DEFINED NXP_SDK_ROOT_DIR)
-  message(FATAL_ERROR "NXP_SDK_ROOT_DIR is not set")
-endif()
-
-# lint_cmake: -linelength
-set(SOURCES
-    ${NXP_SDK_ROOT_DIR}/components/lists/fsl_component_generic_list.c
-    ${NXP_SDK_ROOT_DIR}/components/uart/fsl_adapter_usart.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_clock.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_common.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_common_dsp.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_flexcomm.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_gpio.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_mu.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_reset.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers/fsl_usart.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/system_MIMXRT685S_dsp.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/utilities/debug_console_lite/fsl_assert.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/utilities/debug_console_lite/fsl_debug_console.c
-    ${NXP_SDK_ROOT_DIR}/boards/evkmimxrt685/dsp_examples/mu_polling/dsp/board_hifi4.c
-    ${NXP_SDK_ROOT_DIR}/boards/evkmimxrt685/dsp_examples/mu_polling/dsp/pin_mux.c
-    ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/utilities/str/fsl_str.c
-)
-
-add_library(dsp_mu_polling_libs STATIC ${SOURCES})
-
-target_include_directories(
-  dsp_mu_polling_libs
-  PUBLIC ${NXP_SDK_ROOT_DIR}
-         ${NXP_SDK_ROOT_DIR}/components/uart
-         ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/drivers
-         ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/utilities/debug_console_lite
-         ${NXP_SDK_ROOT_DIR}/components/lists
-         ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S
-         ${NXP_SDK_ROOT_DIR}/CMSIS/Core/Include
-         ${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/utilities/str
-         ${NXP_SDK_ROOT_DIR}/boards/evkmimxrt685/dsp_examples/mu_polling/dsp
-)
-
-add_library(extension_runner_util STATIC IMPORTED)
-set_property(
-  TARGET extension_runner_util
-  PROPERTY
-    IMPORTED_LOCATION
-    "${CMAKE_CURRENT_LIST_DIR}/../../cmake-out/extension/runner_util/libextension_runner_util.a"
-)
-
+add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/hifi/third-party/nnlib)  
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/hifi/operators)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/hifi/kernels)
 
-# Generate the model header file
-add_custom_command(
-  OUTPUT ${CMAKE_BINARY_DIR}/model_pte.h
-  COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/utils/gen_header.py
-          --model_path ${MODEL_PATH} --header_output_path ${CMAKE_BINARY_DIR}
-  COMMENT "Converting .pte model to header file..."
-  DEPENDS ${CMAKE_CURRENT_LIST_DIR}/utils/gen_header.py
-)
-
-add_custom_target(gen_model_header DEPENDS ${CMAKE_BINARY_DIR}/model_pte.h)
 
-add_executable(cadence_executorch_example executor_runner.cpp)
-add_dependencies(cadence_executorch_example gen_model_header)
-
-# lint_cmake: -linelength
-target_include_directories(
-  cadence_executorch_example PUBLIC ${ROOT_DIR}/.. ${CMAKE_BINARY_DIR}
-                                    ${_common_include_directories}
+install(
+  TARGETS cadence_ops_lib
+  DESTINATION lib
+  INCLUDES
+  DESTINATION ${_common_include_directories}
 )
 
-target_link_options(
-  cadence_executorch_example PRIVATE
-  -mlsp=${NXP_SDK_ROOT_DIR}/devices/MIMXRT685S/xtensa/min-rt
-)
-target_link_libraries(
-  cadence_executorch_example dsp_mu_polling_libs cadence_ops_lib
-  extension_runner_util executorch
-)
 
-add_custom_command(
-  TARGET cadence_executorch_example
-  POST_BUILD
-  COMMAND
-    ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/utils/post_compilation.py
-    ${CMAKE_CURRENT_BINARY_DIR}/${PROJECT_NAME} ${CMAKE_BINARY_DIR}
-  COMMENT
-    "Generating .bin files that can be used to flash the DSP with. Copy over
-    the dsp_text_release.bin and dsp_data_release.bin that are generated into
-    your NXP MCUXpresso IDE workspace and flash the DSP with these binaries."
-    DEPENDS
-    ${CMAKE_CURRENT_LIST_DIR}/utils/post_compilation.py
-)
+
+

backends/cadence/cadence.cmake

@@ -41,8 +41,8 @@ set(CMAKE_CROSSCOMPILING TRUE)
 set(CMAKE_C_COMPILER ${TOOLCHAIN_HOME}/bin/${CROSS_COMPILE_TARGET}-clang)
 set(CMAKE_CXX_COMPILER ${TOOLCHAIN_HOME}/bin/${CROSS_COMPILE_TARGET}-clang++)
 
-set(CMAKE_C_FLAGS_INIT "-stdlib=libc++")
-set(CMAKE_CXX_FLAGS_INIT "-stdlib=libc++")
+set(CMAKE_C_FLAGS_INIT "-stdlib=libc++ -mtext-section-literals -mlongcalls")
+set(CMAKE_CXX_FLAGS_INIT "-stdlib=libc++ -mtext-section-literals -mlongcalls")
 set(CMAKE_SYSROOT ${TOOLCHAIN_HOME}/${SYSROOT_TARGET})
 set(CMAKE_LINKER ${TOOLCHAIN_HOME}/bin/xt-ld)
 add_link_options(-lm -stdlib=libc++ -Wl,--no-as-needed -static)

backends/cadence/hifi/kernels/CMakeLists.txt

@@ -20,3 +20,5 @@ target_include_directories(
          ${NN_LIB_BASE_DIR}/xa_nnlib/algo/ndsp/hifi4/include/
          ${NXP_SDK_ROOT_DIR}/middleware/dsp/naturedsp/hifi4/include/
 )
+
+target_link_libraries(cadence_kernels PRIVATE xa_nnlib)

backends/cadence/hifi/kernels/kernels.h

@@ -12,6 +12,31 @@
 #include "stddef.h"
 #include "xa_type_def.h"
 
+
+extern "C" WORD32 xa_nn_elm_quantize_f32_asym8s(WORD8 * __restrict__ p_out,
+                                     const FLOAT32 * __restrict__ p_inp,
+                                     FLOAT32 out_scale,
+                                     WORD32  out_zero_bias,
+                                     WORD32  num_elm);
+
+/*extern "C" WORD32 xa_nn_elm_quantize_f32_asym8u(UWORD8 * __restrict__ p_out,
+                                     const FLOAT32 * __restrict__ p_inp,
+                                     FLOAT32 out_scale,
+                                     WORD32  out_zero_bias,
+                                     WORD32  num_elm); */
+
+extern "C" WORD32 xa_nn_elm_dequantize_asym8s_f32(FLOAT32 * __restrict__ p_out,
+                                       const WORD8 * __restrict__ p_inp,
+                                       WORD32  inp_zero_bias,
+                                       FLOAT32 inp_scale,
+                                       WORD32  num_elm);
+
+/*extern "C" WORD32 xa_nn_elm_dequantize_asym8u_f32(FLOAT32 * __restrict__ p_out,
+                                       const UWORD8 * __restrict__ p_inp,
+                                       WORD32   inp_zero_bias,
+                                       FLOAT32  inp_scale,
+                                       WORD32   num_elm);*/
+
 namespace impl {
 namespace HiFi {
 namespace kernels {

backends/cadence/hifi/operators/dequantize_per_tensor.cpp

@@ -31,16 +31,21 @@ void dequantize_per_tensor_out(
 
   if (input.scalar_type() == ScalarType::Byte) {
     const uint8_t* input_data = input.const_data_ptr<uint8_t>();
-    impl::HiFi::kernels::dequantize<uint8_t>(
-        out_data, input_data, scale, zero_point, numel);
+#if 0 //NNLIB_OPT (not available in nnlib)
+    xa_nn_elm_dequantize_asym8u_f32(out_data, input_data, zero_point, scale, numel);
+#else    
+    impl::HiFi::kernels::dequantize<uint8_t>(out_data, input_data, scale, zero_point, numel);
+#endif
   } else if (input.scalar_type() == ScalarType::Char) {
     const int8_t* input_data = input.const_data_ptr<int8_t>();
-    impl::HiFi::kernels::dequantize<int8_t>(
-        out_data, input_data, scale, zero_point, numel);
+#if NNLIB_OPT
+    xa_nn_elm_dequantize_asym8s_f32(out_data, input_data, zero_point, scale, numel);
+#else    
+    impl::HiFi::kernels::dequantize<int8_t>(out_data, input_data, scale, zero_point, numel);
+#endif
   } else if (input.scalar_type() == ScalarType::Int) {
     const int32_t* input_data = input.const_data_ptr<int32_t>();
-    impl::HiFi::kernels::dequantize<int32_t>(
-        out_data, input_data, scale, zero_point, numel);
+    impl::HiFi::kernels::dequantize<int32_t>(out_data, input_data, scale, zero_point, numel);
   } else {
     ET_CHECK_MSG(false, "Unhandled input dtype %hhd", input.scalar_type());
   }

backends/cadence/hifi/operators/quantize_per_tensor.cpp

@@ -33,16 +33,21 @@ void quantize_per_tensor_out(
 
   if (out.scalar_type() == ScalarType::Byte) {
     uint8_t* out_data = out.mutable_data_ptr<uint8_t>();
-    impl::HiFi::kernels::quantize<uint8_t>(
-        out_data, input_data, 1. / scale, zero_point, numel);
+#if 0 //NNLIB_OPT (not available in nnlib)     
+    xa_nn_elm_quantize_f32_asym8u(out_data, input_data, scale, zero_point, numel);
+#else    
+    impl::HiFi::kernels::quantize<uint8_t>(out_data, input_data, 1. / scale, zero_point, numel);
+#endif
   } else if (out.scalar_type() == ScalarType::Char) {
     int8_t* out_data = out.mutable_data_ptr<int8_t>();
-    impl::HiFi::kernels::quantize<int8_t>(
-        out_data, input_data, 1. / scale, zero_point, numel);
+#if NNLIB_OPT
+    xa_nn_elm_quantize_f32_asym8s(out_data, input_data, scale, zero_point, numel);
+#else    
+    impl::HiFi::kernels::quantize<int8_t>(out_data, input_data, 1. / scale, zero_point, numel);
+#endif        
   } else if (out.scalar_type() == ScalarType::Int) {
     int32_t* out_data = out.mutable_data_ptr<int32_t>();
-    impl::HiFi::kernels::quantize<int32_t>(
-        out_data, input_data, 1. / scale, zero_point, numel);
+    impl::HiFi::kernels::quantize<int32_t>(out_data, input_data, 1. / scale, zero_point, numel);
   } else {
     ET_CHECK_MSG(false, "Unhandled input dtype %hhd", out.scalar_type());
   }

backends/cadence/hifi/third-party/nnlib/CMakeLists.txt

@@ -0,0 +1,33 @@
+
+cmake_minimum_required(VERSION 3.10.0)
+project(cadence_nnlib)
+
+
+add_custom_target( nnlib_target ALL COMMAND 
+                    make install_nnlib -f makefile -C ${EXECUTORCH_ROOT}/backends/cadence/hifi/third-party/nnlib/nnlib-hifi4/xa_nnlib/build 
+                    OBJDIR=${CMAKE_CURRENT_BINARY_DIR}/obj 
+                    LIBDIR=${CMAKE_CURRENT_BINARY_DIR}/lib 
+                    -j8 )
+
+add_library(xa_nnlib STATIC IMPORTED GLOBAL)
+add_dependencies(xa_nnlib nnlib_target)
+
+message("NNLIB")
+message("${CMAKE_CURRENT_BINARY_DIR}")
+
+set_property(
+  TARGET xa_nnlib
+  PROPERTY 
+  IMPORTED_LOCATION "${CMAKE_CURRENT_BINARY_DIR}/lib/xa_nnlib.a"
+)
+
+
+
+
+
+
+
+
+
+
+