update

Author: yifeizh2

.github/workflows/clang-tidy.yml

@@ -14,7 +14,7 @@ jobs:
 
     steps:
     - name: Install OpenMP
-      run: "sudo apt install -y libomp-dev"
+      run: "sudo apt install -y libomp-dev opencl-c-headers"
 
     - name: Fetch sources
       uses: actions/checkout@v4

CMakeLists.txt

@@ -98,12 +98,15 @@ get_property(GC_TOOLS GLOBAL PROPERTY GC_TOOLS)
 get_property(GC_MLIR_LIBS GLOBAL PROPERTY GC_MLIR_LIBS)
 get_property(GC_PASS_LIBS GLOBAL PROPERTY GC_PASS_LIBS)
 get_property(GC_DIALECT_LIBS GLOBAL PROPERTY GC_DIALECT_LIBS)
+get_property(IMEX_LIBS GLOBAL PROPERTY IMEX_LIBS)
+
 install(TARGETS
   GcInterface
   ${GC_TOOLS}
   ${GC_MLIR_LIBS}
   ${GC_PASS_LIBS}
   ${GC_DIALECT_LIBS}
+  ${IMEX_LIBS}
   EXPORT ${PROJECT_NAME}Targets
   ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
   LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}

README.md

@@ -1,9 +1,7 @@
 [<img src="https://scan.coverity.com/projects/30281/badge.svg">](https://scan.coverity.com/projects/intel-graph-compiler)
 
 # Graph Compiler
-Graph Compiler is an end-to-end, MLIR-based compiler designed to enhance the performance of deep learning workloads. It accepts computation graphs from the frontend, applies domain-specific optimizations and transformations, generates code, and manages runtime execution.
-
-The current frontend for Graph Compiler is [oneDNN Graph API](https://oneapi-src.github.io/oneDNN/graph_extension.html).
+Graph Compiler is an end-to-end, MLIR-based compiler designed to enhance the performance of deep learning workloads. It accepts computation graphs from the frontend (primarily linalg on tensors), applies domain-specific optimizations and transformations, generates code, and manages runtime execution.
 
 ## Build instructions
 

cmake/functions.cmake

@@ -117,4 +117,36 @@ function(gc_add_mlir_dialect_library name)
     if(GcInterface IN_LIST ARGN)
         target_link_libraries(obj.${name} PUBLIC GcInterface)
     endif()
-endfunction()
+endfunction()
+
+macro(gc_add_mlir_tool name)
+    # the dependency list copied from mlir/tools/mlir-cpu-runner/CMakeLists.txt of upstream
+    if(NOT DEFINED LLVM_LINK_COMPONENTS)
+        set(LLVM_LINK_COMPONENTS
+          Core
+          Support
+          nativecodegen
+          native
+        )
+    endif()
+    if(NOT DEFINED MLIR_LINK_COMPONENTS)
+        gc_set_mlir_link_components(MLIR_LINK_COMPONENTS
+          MLIRAnalysis
+          MLIRBuiltinToLLVMIRTranslation
+          MLIRExecutionEngine
+          MLIRIR
+          MLIRJitRunner
+          MLIRLLVMDialect
+          MLIRLLVMToLLVMIRTranslation
+          MLIRToLLVMIRTranslationRegistration
+          MLIRParser
+          MLIRTargetLLVMIRExport
+          MLIRSupport
+        )
+    endif()
+    add_mlir_tool(${ARGV})
+    #LLVM_LINK_COMPONENTS is processed by LLVM cmake in add_llvm_executable
+    target_link_libraries(${name} PRIVATE GcInterface ${MLIR_LINK_COMPONENTS})
+    llvm_update_compile_flags(${name})
+    set_property(GLOBAL APPEND PROPERTY GC_TOOLS ${name})
+endmacro()

cmake/imex-version.txt

@@ -1 +1 @@
-8209807be6148d81fda6f439a01b77696986dd3e
+25123cc3692fdbcf837510f39de47ff353d482fc

cmake/imex.cmake

@@ -14,7 +14,8 @@ if (NOT DEFINED IMEX_INCLUDES)
 
     # TODO: Change to main https://github.com/intel/mlir-extensions when all the
     # required functionality is merged.
-    gc_fetch_content(imex "${IMEX_HASH}" https://github.com/intel/mlir-extensions
+    set(IMEX_URL https://github.com/intel/mlir-extensions)
+    gc_fetch_content(imex "${IMEX_HASH}" "${IMEX_URL}"
             SET IMEX_CHECK_LLVM_VERSION=ON IMEX_ENABLE_L0_RUNTIME=0
     )
 
@@ -24,4 +25,5 @@ if (NOT DEFINED IMEX_INCLUDES)
             ${imex_SOURCE_DIR}/src
     )
     set_property(GLOBAL PROPERTY IMEX_INCLUDES ${IMEX_INCLUDES})
+    target_compile_options(GcInterface INTERFACE -DGC_USE_IMEX)
 endif ()

cmake/llvm-version.txt

@@ -1 +1 @@
-3191587666aa3d1e53966bc8876614c7197fac4f
+f6a756f35a4d0719a96b4e214905369d565d87da

cmake/onednn_lite_config.cmake

@@ -28,7 +28,7 @@ if(${DNNL_CPU_RUNTIME} STREQUAL "OMP")
 endif()
 
 if(${DNNL_CPU_RUNTIME} STREQUAL "TBB")
-    include("${DNNL_PATH}/cmake/TBB.cmake")
+    find_package(TBB REQUIRED)
 endif()
 
 ########## copied from main cmake file of DNNL

doc/GPUPipeline.md renamed to docs/GPU_pipeline_overview.md

@@ -33,6 +33,9 @@ struct MatmulConfig {
   uint32_t innerMostMBlock, innerMostNBlock, innerMostKBlock;
 };
 
+bool validateConfig(const MatmulConfig &cfg, ArrayRef<uint32_t> shape,
+                    bool allowIndivisibleInnerblock, bool isVNNIMM2D);
+
 enum DimType { Batch, M, N, K };
 
 // Extract the index of the given DimType in the DimType list