[OPTIMIZER] Coalesce pass no longer takes a num-warps argument (#99)

Improved design to avoid inconsistent `num-warps` value between the pass and the parent module of the operation it processes.

include/triton/Dialect/TritonGPU/IR/TritonGPUDialect.td

@@ -18,6 +18,18 @@ def TritonGPU_Dialect : Dialect {
     "triton::TritonDialect",
     "mlir::gpu::GPUDialect"
   ];
+
+  let extraClassDeclaration = [{
+    static std::string getNumWarpsAttrName() { return "triton_gpu.num-warps"; }
+    static int getNumWarps(ModuleOp mod) { 
+      if(!mod->hasAttr("triton_gpu.num-warps"))
+        llvm::report_fatal_error(
+            "TritonGPU module should contain a triton_gpu.num-warps attribute");
+      return mod->getAttr("triton_gpu.num-warps").cast<IntegerAttr>().getInt();
+    }
+  }];
+  
+
 }
 
 #endif

include/triton/Dialect/TritonGPU/Transforms/Passes.td

@@ -33,12 +33,6 @@ def TritonGPUCoalesce: Pass<"tritongpu-coalesce", "mlir::ModuleOp"> {
   let constructor = "mlir::createTritonGPUCoalescePass()";
 
   let dependentDialects = ["mlir::triton::gpu::TritonGPUDialect"];
-
-  let options = [
-    Option<"numWarps", "num-warps",
-          "int32_t", /*default*/"4",
-          "number of warps">
-  ];
 }
 
 def TritonGPUCombineOps : Pass<"tritongpu-combine", "mlir::ModuleOp"> {