Summary: Add Statefull FC Cortex-m linearOps

Integrate with CMSIS-NN with per-channel quantization support

Test Plan:
Run e2e test on FVP simulator
./examples/arm/run_mcu_models_fvp.sh --target=cortex-m55 --models=qlinear

Reviewers:

Subscribers:

Tasks:

Tags:

Author: Github Executorch

backends/cortex_m/CMakeLists.txt

@@ -21,31 +21,59 @@ include(${EXECUTORCH_ROOT}/tools/cmake/Utils.cmake)
 include(${EXECUTORCH_ROOT}/tools/cmake/Codegen.cmake)
 include(FetchContent)
 
-# CMSIS-NN version to download
-set(CMSIS_NN_VERSION
-    "v4.1.0"
-    CACHE STRING "CMSIS-NN version to download"
-)
-
-# Declare CMSIS-NN as a FetchContent project
-FetchContent_Declare(
-  cmsis_nn
-  GIT_REPOSITORY https://github.com/ARM-software/CMSIS-NN.git
-  GIT_TAG ${CMSIS_NN_VERSION}
-)
-
-# Download and make CMSIS-NN available
-FetchContent_MakeAvailable(cmsis_nn)
-
-# Print paths for debugging
-message(STATUS "CMSIS-NN source dir: ${cmsis_nn_SOURCE_DIR}")
-message(STATUS "CMSIS-NN binary dir: ${cmsis_nn_BINARY_DIR}")
+set(USE_LOCAL_CMSIS_NN OFF)
+
+if(USE_LOCAL_CMSIS_NN)
+  if(NOT EXISTS ${CMSIS_NN_LOCAL_PATH})
+    message(
+      FATAL_ERROR "CMSIS-NN local path does not exist: ${CMSIS_NN_LOCAL_PATH}"
+    )
+  endif()
+  if(NOT EXISTS ${CMSIS_NN_LOCAL_LIB})
+    message(
+      FATAL_ERROR "CMSIS-NN local lib does not exist: ${CMSIS_NN_LOCAL_LIB}"
+    )
+  endif()
+  message(STATUS "Using CMSIS-NN from: ${CMSIS_NN_LOCAL_PATH}")
+  add_subdirectory(${CMSIS_NN_LOCAL_PATH} cmsis_nn_build)
+  # Add MVEI define to cmsis-nn target
+  target_compile_definitions(cmsis-nn PUBLIC ARM_MATH_MVEI=1)
+
+  set(CMSIS_NN_INCLUDE_DIR "${CMSIS_NN_LOCAL_PATH}/Include")
+  set(CMSIS_NN_LIB "${CMSIS_NN_LOCAL_LIB}/libcmsis-nn.a")
+
+else()
+  message(STATUS "Using  CMSIS-NN from via : FetchContent")
+  set(CMSIS_NN_VERSION
+      "v7.0.0"
+      CACHE STRING "CMSIS-NN version to download"
+  )
+  FetchContent_Declare(
+    cmsis_nn
+    GIT_REPOSITORY https://github.com/ARM-software/CMSIS-NN.git
+    GIT_TAG ${CMSIS_NN_VERSION}
+  )
+  FetchContent_MakeAvailable(cmsis_nn)
+  # Add MVEI define to cmsis-nn target
+  target_compile_definitions(cmsis-nn PUBLIC ARM_MATH_MVEI=1)
+  # Get the correct source and binary dirs
+  FetchContent_GetProperties(
+    cmsis_nn
+    SOURCE_DIR CMSIS_NN_SOURCE_DIR
+    BINARY_DIR CMSIS_NN_BINARY_DIR
+  )
+  set(CMSIS_NN_INCLUDE_DIR "${CMSIS_NN_SOURCE_DIR}/Include")
+  set(CMSIS_NN_LIB "${CMSIS_NN_BINARY_DIR}/libcmsis-nn.a")
+  message(STATUS " ${CMSIS_NN_SOURCE_DIR}")
+  message(STATUS "CMSIS-NN binary dir: ${CMSIS_NN_BINARY_DIR}")
+endif()
 
 # Cortex-M ops kernel sources
 set(_cortex_m_kernels__srcs
     ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantize_per_tensor.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_dequantize_per_tensor.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantized_add.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/ops/op_quantized_linear.cpp
 )
 
 # Generate C++ bindings to register kernels into Executorch (for runtime)
@@ -66,13 +94,11 @@ target_include_directories(
   cortex_m_kernels
   PRIVATE ${EXECUTORCH_ROOT}/..
           ${EXECUTORCH_ROOT}/runtime/core/portable_type/c10
-          ${cmsis_nn_SOURCE_DIR}/Include
+          ${CMSIS_NN_INCLUDE_DIR}
 )
 
 # Link directly to the CMSIS-NN static library file
-target_link_libraries(
-  cortex_m_kernels PUBLIC ${cmsis_nn_BINARY_DIR}/libcmsis-nn.a executorch
-)
+target_link_libraries(cortex_m_kernels PUBLIC ${CMSIS_NN_LIB} executorch)
 
 # Add dependency to ensure CMSIS-NN builds before we try to link. Use the actual
 # CMSIS-NN target name (usually 'cmsis-nn')

backends/cortex_m/ops/cortex_m_ops_common.h

@@ -32,16 +32,19 @@ inline void validate_cmsis_nn_tensor_requirements(
   // Basic dtype validation
   ET_CHECK_MSG(
       input1.scalar_type() == expected_dtype,
-      "Input1 dtype must be %hhd",
-      expected_dtype);
+      "Input1 dtype must be %hhd, got %hhd",
+      expected_dtype,
+      input1.scalar_type());
   ET_CHECK_MSG(
       input2.scalar_type() == expected_dtype,
-      "Input2 dtype must be %hhd",
-      expected_dtype);
+      "Input2 dtype must be %hhd, got %hhd",
+      expected_dtype,
+      input2.scalar_type());
   ET_CHECK_MSG(
       output.scalar_type() == expected_dtype,
-      "Output dtype must be %hhd",
-      expected_dtype);
+      "Output dtype must be %hhd, got %hhd",
+      expected_dtype,
+      output.scalar_type());
 
   // Dim order consistency
   ET_CHECK_MSG(
@@ -114,6 +117,27 @@ inline void validate_quantization_params(
       "Single quant Output");
 }
 
+inline bool validate_per_channel_quant_params(
+    const int32_t* multipliers,
+    const int32_t* shifts,
+    int num_channels) {
+  for (int i = 0; i < num_channels; ++i) {
+    if (multipliers[i] < (1LL << 30) || multipliers[i] > ((1LL << 31) - 1)) {
+      ET_LOG(
+          Error,
+          "weight_multiplier[%d] out of CMSIS-NN range: %d",
+          i,
+          multipliers[i]);
+      return false;
+    }
+    if (shifts[i] < 0 || shifts[i] > 31) {
+      ET_LOG(Error, "weight_shift[%d] out of range: %d", i, shifts[i]);
+      return false;
+    }
+  }
+  return true;
+}
+
 inline Error resize_to_broadcast_target_size(
     const Tensor& input1,
     const Tensor& input2,