Implemented tiling and fusion path for GPU

Author: AndreyPavlenko

include/gc/Transforms/Passes.td

@@ -112,6 +112,22 @@ def GpuToGpuOcl : Pass<"gpu-to-gpuocl", "ModuleOp"> {
            "Call finish() after each kernel launch.">
     ];
 }
+
+def GpuTilingAndFusion : Pass<"gpu-tiling", "func::FuncOp"> {
+  let summary = "GPU tiling and fusion path.";
+  let description = [{
+    This path creates 2 nested loops for linalg operations, that are converted to gpu.launch.
+    The outer loop is mapped to to the grid sizes and the inner loop is mapped to the block sizes.
+  }];
+  let options = [
+    Option<"euMem", "eu-mem", "size_t",
+           /*default=*/"131072",
+           "Execution Unit cache size.">,
+    Option<"euThreads", "eu-threads", "size_t",
+           /*default=*/"8",
+           "Number of threads per EU.">
+    ];
+}
 #endif // GC_USE_IMEX
 
 def IterativeTilingAndFusion : Pass<"iterative-tiling-and-fusion",

lib/gc/Transforms/GPU/CMakeLists.txt

@@ -12,6 +12,7 @@ set_property(GLOBAL APPEND PROPERTY IMEX_LIBS ${IMEX_LIBS})
 
 gc_add_mlir_library(GcGpuPasses
   AddContextArg.cpp
+  GpuTilingAndFusion.cpp
   GpuToGpuOcl.cpp
   LinalgToXeGPU.cpp
   Pipeline.cpp