[llvm] [MLGO] Model selection for models lowered through EmitC (PR #212650)

[ioana ghiban via llvm-commits]

================
@@ -0,0 +1,141 @@
+# Lower MLGO models to C++ headers via the MLIR-based pipeline.
+#
+# Each entry in ${models} has the form "cli-flag,path/to/model.mlir,type".
+# For entries whose type matches ${target_type}, this function:
+#   1. Lowers model MLIR with ${mlir_opt} and translates it with ${mlir_translate}
+#      to a C++ header defining Model<N> in ${LLVM_INCLUDE_DIR}/${include_subdir}.
+#   2. Generates ${generated_file_basename}.def in ${LLVM_INCLUDE_DIR}/${include_subdir}
+#      with one MLGO_MODEL(ClassName, "cli-flag") entry per model.
+#   3. Generates ${generated_file_basename}.h in ${LLVM_INCLUDE_DIR}/${include_subdir}
+#      including all generated model headers.
+#   4. Appends the target driving generation to ${out_var} in the caller's scope.
+#   5. Defines LLVM_HAVE_MLIR_LOWERING_<TARGET_TYPE>.
+function(mlgo_lower_models models mlir_opt mlir_translate target_type
+         include_subdir generated_file_basename out_var)
+  set(def_file "${generated_file_basename}.def")
+  set(umbrella_header "${generated_file_basename}.h")
+
+  if(NOT IS_ABSOLUTE "${mlir_opt}")
+    cmake_path(ABSOLUTE_PATH mlir_opt BASE_DIRECTORY "${CMAKE_BINARY_DIR}")
+  endif()
+
+  if(NOT IS_ABSOLUTE "${mlir_translate}")
+    cmake_path(ABSOLUTE_PATH mlir_translate BASE_DIRECTORY "${CMAKE_BINARY_DIR}")
+  endif()
+
+  string(APPEND def_content
+    "/* Auto-generated by CMake */\n"
+    "#ifndef MLGO_MODEL\n"
+    "#  define MLGO_MODEL(CLASS_NAME, CLI_FLAG)\n"
+    "#endif\n\n"
+  )
+
+  set(headers_content "/* Auto-generated by CMake */\nnamespace llvm {\n")
+  set(mlgo_gen_targets "")
+  set(model_index 1)
+
+  foreach(model_info IN LISTS models)
+    # Parse comma-separated fields (from the LLVM_MLGO_MODELS flag)
+    # cli-flag,path/to/model.mlir,type
+    string(REPLACE "," ";" model_fields "${model_info}")
+    list(GET model_fields 0 cli_flag)
+    list(GET model_fields 1 model_path)
+    list(GET model_fields 2 model_type)
+
+    # Only process models matching target_type
+    if (NOT model_type STREQUAL target_type)
+      continue()
+    endif()
+
+    set(class_name "${target_type}_Model${model_index}")
+    math(EXPR model_index "${model_index} + 1")
+
+    if (NOT IS_ABSOLUTE "${model_path}")
+      get_filename_component(model_path "${model_path}" ABSOLUTE BASE_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
+    endif()
+
+    set(emitc_mlir "${CMAKE_CURRENT_BINARY_DIR}/${class_name}_emitc.mlir")
+    set(header_file "${LLVM_INCLUDE_DIR}/${include_subdir}/${class_name}.h")
+
+    # Pass pipeline to lower MLIR models to EmitC dialect
+    # TODO: Simplify with builtin pipeline for translation.
+    # TODO: It isn't ideal to hardcode this pass pipeline here. It would be better to
+    # use the transform dialect.
+    set(mlir_passes
+      "func.func(tosa-to-linalg-named,tosa-to-linalg,tosa-to-arith,tosa-to-tensor)"
+      "symbol-privatize"
+      "scalarize-single-element-tensor-return"
+      "one-shot-bufferize{bufferize-function-boundaries=true function-boundary-type-conversion=identity-layout-map buffer-alignment=0}"
+      "buffer-results-to-out-params{hoist-static-allocs=true}"
+      "func.func(promote-buffers-to-stack)"
+      "buffer-deallocation-pipeline"
+      "func.func(convert-linalg-to-loops)"
+      "expand-strided-metadata"
+      "canonicalize"
+      "memref-elide-reinterpret-cast"
+      "convert-to-emitc"
+      "wrap-emitc-func-in-class{class-name-format=${class_name}}"
----------------
ioghiban wrote:

[nit]: Could you please keep the EmitC rewrite passes separate? Now they are interleaved with the conversion passes. Perhaps place them afterwards, after the full EmitC conversion?

https://github.com/llvm/llvm-project/pull/212650