Merge remote-tracking branch 'origin/zhicong/deep_tile_matmul' into layout_propagation

Author: yifeizh2

.github/workflows/build-llvm.yml

@@ -2,13 +2,17 @@ name: LLVM Build
 
 on:
   workflow_dispatch:
+  push:
+    paths:
+      - cmake/llvm-version.txt
+      - .github/workflows/build-llvm.yml
 
 permissions: read-all
 
 jobs:
   build:
     name: Build
-    runs-on: [self-hosted, 0.0.1]
+    runs-on: [self-hosted]
 
     steps:
       - uses: actions/checkout@v4
@@ -27,7 +31,7 @@ jobs:
           python3 -m pip install -r mlir/python/requirements.txt
           mkdir llvm-install
           cmake -G Ninja llvm -B build -DCMAKE_INSTALL_PREFIX=llvm-install -DMLIR_ENABLE_BINDINGS_PYTHON=ON -DPython3_EXECUTABLE=$(which python3) \
-            -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=true -DLLVM_ENABLE_PROJECTS="mlir" -DLLVM_TARGETS_TO_BUILD="X86" -DLLVM_INSTALL_UTILS=true -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DLLVM_INSTALL_GTEST=ON
+            -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=true -DLLVM_ENABLE_PROJECTS="mlir" -DLLVM_EXPERIMENTAL_TARGETS_TO_BUILD="SPIRV" -DLLVM_TARGETS_TO_BUILD="X86" -DLLVM_INSTALL_UTILS=true -DCMAKE_EXPORT_COMPILE_COMMANDS=ON -DLLVM_INSTALL_GTEST=ON
           cmake --build build --target install
           cd llvm-install
           tar -zcf ../llvm.tgz .

README.md

@@ -58,7 +58,15 @@ cmake --build . --target gc-check
 Notes:
  * `/PATH/TO/llvm-project/llvm-install` should be the install path of LLVM. If you installed LLVM elsewhere by `-DCMAKE_INSTALL_PREFIX` option when building LLVM, you need to change the path in `-DMLIR_DIR` accordingly.
  *  The cmake option `-DLLVM_EXTERNAL_LIT` is for the tests of this project. It requires the `lit` tool to be installed in the system. You can install it via `pip install lit`. If you don't need to run the tests of this repo, you can omit this option in the command line.
- * If GPU components are on (`-DGC_USE_GPU=ON`), make sure the Level-zero runtime is installed in your system. Either install Level-zero runtime via system package managers (e.g. `apt`), or follow the instructions of [IMEX](https://github.com/intel/mlir-extensions).
+
+More notes if GPU components are on (`-DGC_USE_GPU=ON`):
+ * make sure the OpenCL runtime is installed in your system. You can either
+  install using OS-provided package (Ubuntu 22.04)
+```sh
+sudo apt install -y intel-opencl-icd opencl-c-headers
+```
+  Or, download and install package from: https://github.com/intel/compute-runtime/releases
+ * the LLVM codebase needs to be patched to support XeGPU lowering (from IMEX). Please follow instructions of [IMEX](https://github.com/intel/mlir-extensions) on patching LLVM.
 
 Graph Compiler supports the following build-time options.
 

cmake/imex.cmake

@@ -4,11 +4,11 @@ get_property(IMEX_INCLUDES GLOBAL PROPERTY IMEX_INCLUDES)
 if (NOT DEFINED IMEX_INCLUDES)
     include(functions)
     set(IMEX_CHECK_LLVM_VERSION ON)
-    set(IMEX_ENABLE_L0_RUNTIME 1)
+    set(IMEX_ENABLE_L0_RUNTIME 0)
     # TODO: Change to main https://github.com/oneapi-src/oneDNN.git when all the
     # required functionality is merged.
     gc_fetch_content(imex 496b240093b5e132b60c5ee69878300fe69be300 https://github.com/Menooker/mlir-extensions
-            CMAKE_ARGS "-DMLIR_DIR=${MLIR_DIR};-DIMEX_CHECK_LLVM_VERSION=ON;-DIMEX_ENABLE_L0_RUNTIME=1"
+            CMAKE_ARGS "-DMLIR_DIR=${MLIR_DIR};-DIMEX_CHECK_LLVM_VERSION=ON;-DIMEX_ENABLE_L0_RUNTIME=0"
     )
 
     set(IMEX_INCLUDES

cmake/llvm-version.txt

@@ -1 +1 @@
-37661a17e26d9002ae9ade8c0de3932c22f16360
+89946bda5e1c7ceaf6d26634cc8c8c9498d9f7be

include/gc/Analysis/MatmulConfigAnalysis.h

@@ -22,63 +22,57 @@ using namespace mlir;
 struct SystemDesc {
   // get runtime OMP_NUM_THREADS
   uint32_t getNumThreads() {
-    DataLayout layout = DataLayout(module);
-    MLIRContext *ctx = module->getContext();
     std::optional<Attribute> numThreads = layout.getDevicePropertyValue(
         Builder(ctx).getStringAttr("CPU" /* device ID*/),
         Builder(ctx).getStringAttr("num_threads"));
-    if (numThreads && dyn_cast<IntegerAttr>(*numThreads)) {
-      return cast<IntegerAttr>(*numThreads).getInt();
+    if (numThreads && isa<IntegerAttr>(*numThreads)) {
+      return dyn_cast<IntegerAttr>(*numThreads).getInt();
     }
     return 1;
   }
   // get cache size by cacheLevel
   size_t getCacheSize(uint8_t cacheLevel) {
-    DataLayout layout = DataLayout(module);
-    MLIRContext *ctx = module->getContext();
-
     if (cacheLevel == 1) {
       std::optional<Attribute> cacheSize = layout.getDevicePropertyValue(
           Builder(ctx).getStringAttr("CPU" /* device ID*/),
           Builder(ctx).getStringAttr("L1_cache_size_in_bytes"));
-      if (cacheSize && dyn_cast<IntegerAttr>(*cacheSize)) {
-        return cast<IntegerAttr>(*cacheSize).getInt();
+      if (cacheSize && isa<IntegerAttr>(*cacheSize)) {
+        return dyn_cast<IntegerAttr>(*cacheSize).getInt();
       }
     } else if (cacheLevel == 2) {
       std::optional<Attribute> cacheSize = layout.getDevicePropertyValue(
           Builder(ctx).getStringAttr("CPU" /* device ID*/),
           Builder(ctx).getStringAttr("L2_cache_size_in_bytes"));
-      if (cacheSize && dyn_cast<IntegerAttr>(*cacheSize)) {
-        return cast<IntegerAttr>(*cacheSize).getInt();
+      if (cacheSize && isa<IntegerAttr>(*cacheSize)) {
+        return dyn_cast<IntegerAttr>(*cacheSize).getInt();
       }
     } else if (cacheLevel == 3) {
       std::optional<Attribute> cacheSize = layout.getDevicePropertyValue(
           Builder(ctx).getStringAttr("CPU" /* device ID*/),
           Builder(ctx).getStringAttr("L3_cache_size_in_bytes"));
-      if (cacheSize && dyn_cast<IntegerAttr>(*cacheSize)) {
-        return cast<IntegerAttr>(*cacheSize).getInt();
+      if (cacheSize && isa<IntegerAttr>(*cacheSize)) {
+        return dyn_cast<IntegerAttr>(*cacheSize).getInt();
       }
     }
     return 0;
   }
 
   // get the maximum vector length in bits
   size_t getMaxVectorLength() {
-    DataLayout layout = DataLayout(module);
-    MLIRContext *ctx = module->getContext();
     std::optional<Attribute> maxVectorLength = layout.getDevicePropertyValue(
         Builder(ctx).getStringAttr("CPU" /* device ID*/),
         Builder(ctx).getStringAttr("max_vector_width"));
-    if (maxVectorLength && dyn_cast<IntegerAttr>(*maxVectorLength)) {
-      return cast<IntegerAttr>(*maxVectorLength).getInt();
+    if (maxVectorLength && isa<IntegerAttr>(*maxVectorLength)) {
+      return dyn_cast<IntegerAttr>(*maxVectorLength).getInt();
     }
     return 512;
   }
 
-  SystemDesc(ModuleOp m) : module(m) {}
+  SystemDesc(ModuleOp m) : layout(m), ctx(m->getContext()) {}
 
 private:
-  ModuleOp module;
+  DataLayout layout;
+  MLIRContext *ctx;
 };
 
 // The configuration for matmul tiling

include/gc/CMakeLists.txt

@@ -1,2 +1,2 @@
 add_subdirectory(Dialect)
-add_subdirectory(Transforms)
+add_subdirectory(Transforms)

include/gc/Dialect/CMakeLists.txt

@@ -1,4 +1,4 @@
 add_subdirectory(CPURuntime)
 add_subdirectory(OneDNNGraph)
 add_subdirectory(Microkernel)
-add_subdirectory(Linalgx)
+add_subdirectory(Linalgx)

include/gc/Dialect/Microkernel/CMakeLists.txt

@@ -1,3 +1,8 @@
+set(LLVM_TARGET_DEFINITIONS MicrokernelEnum.td)
+mlir_tablegen(MicrokernelEnum.h.inc -gen-enum-decls)
+mlir_tablegen(MicrokernelEnum.cpp.inc -gen-enum-defs)
+add_public_tablegen_target(MLIRMicrokernelAttrDefIncGen)
+
 add_mlir_dialect(MicrokernelOps microkernel)
 add_mlir_doc(MicrokernelOps MicrokernelOps gc/Dialect/Microkernel/ -gen-op-doc)
 add_mlir_doc(MicrokernelDialect MicrokernelDialect gc/Dialect/Microkernel/ -gen-dialect-doc)

include/gc/Dialect/Microkernel/MicrokernelDialect.h

@@ -10,6 +10,7 @@
 #define GC_DIALECTS_MICROKERNELDIALECT_H
 
 #include "mlir/IR/Dialect.h"
+#include "mlir/Interfaces/DestinationStyleOpInterface.h"
 
 #include "gc/Dialect/Microkernel/MicrokernelOpsDialect.h.inc"
 

include/gc/Dialect/Microkernel/MicrokernelDialect.td

@@ -15,15 +15,20 @@ include "mlir/IR/OpBase.td"
 // Microkernel dialect definition.
 //===----------------------------------------------------------------------===//
 
-def MicrokernelDialect : Dialect {
+def Microkernel_Dialect : Dialect {
     let name = "microkernel";
     let summary = "A dialect for microkernel abstraction.";
     let description = [{
-        The dialect wraps the BRGEMM API to set up the HW context etc.
+        This dialect contains wrappers for microkernel primitives like BRGEMM.
     }];
     let cppNamespace = "::mlir::microkernel";
-
-    let useDefaultTypePrinterParser = 1;
 }
 
+//===----------------------------------------------------------------------===//
+// Base microkernel operation definition.
+//===----------------------------------------------------------------------===//
+
+class Microkernel_Op<string mnemonic, list<Trait> traits = []> :
+        Op<Microkernel_Dialect, mnemonic, traits>;
+
 #endif // MICROKERNEL_DIALECT