[ARCH] Add UnifiedHardware for backend configs

include/CMakeLists.txt

@@ -1 +1 @@
-add_subdirectory(triton)
+add_subdirectory(triton)

include/flagtree/Common/UnifiedHardwareBase.h

@@ -0,0 +1,27 @@
+#ifndef UNIFIED_HARDWARE_BASE_H
+#define UNIFIED_HARDWARE_BASE_H
+
+#include <optional> 
+
+namespace mlir {
+namespace flagtree {
+//this is the unified hardware abstraction for hardware
+//to determined if these abstraction is specified, using std::optional is needed
+//using in passes: if(uh_flagtree->xxx()){...}
+
+class UnifiedHardware{
+
+public:
+    virtual ~UnifiedHardware() = default;
+
+    //DMA
+    virtual std::optional<int> getAllocSpaceForDMATag() const {
+        return std::nullopt;
+    }
+
+};
+
+} // namespace flagtree
+} // namespace mlir
+
+#endif // UNIFIED_HARDWARE_BASE_H

include/triton/Dialect/Triton/IR/TritonAttrDefs.td

@@ -4,6 +4,7 @@
 include "mlir/IR/EnumAttr.td"
 include "mlir/IR/AttrTypeBase.td"
 
+
 // Attributes for LoadOp and StoreOp
 def TT_CacheModifierAttr : I32EnumAttr<
     "CacheModifier", "",

third_party/aipu/backend/compiler.py

@@ -7,13 +7,16 @@
 from mlir.passmanager import PassManager
 from mlir.ir import Context, Module
 
+from triton._C.libtriton.aipu import UnifiedHardwareAIPU
+
 from dataclasses import dataclass
 import functools
 import hashlib
 from typing import Any, Dict, Tuple
 from types import ModuleType
 
 
+
 def min_dot_size(target: GPUTarget):
     return lambda lhsType, rhsType: (1, 1, 1)
 
@@ -80,6 +83,7 @@ def load_dialects(self, ctx):
 
     @staticmethod
     def make_ttir(mod, metadata, opt):
+
         pm = ir.pass_manager(mod.context)
         pm.enable_debug()
         passes.common.add_inliner(pm)
@@ -96,6 +100,19 @@ def make_ttir(mod, metadata, opt):
 
     @staticmethod
     def make_linalg(mod, metadata, opt):
+
+        import ctypes
+        uh_aipu = UnifiedHardwareAIPU()
+        obj_ptr = id(uh_aipu)
+        uh_aipu_ptr = ctypes.cast(obj_ptr, ctypes.c_void_p)
+
+
+        from mlir.ir import IntegerAttr, IntegerType
+        ctx = Context()
+        ptr_type = IntegerType.get_unsigned(64, ctx)
+        ptr_attr = IntegerAttr.get(ptr_type, uh_aipu_ptr.value) 
+        # mod.set_attr("uh_aipu_ptr", ptr_attr) #报错
+
         pm = ir.pass_manager(mod.context)
         pm.enable_debug()
         # add pass here

third_party/aipu/triton_aipu.cc

@@ -22,6 +22,8 @@
 #include "triton-shared/Conversion/TritonToLinalg/TritonToLinalg.h"
 #include "triton-shared/Conversion/TritonToLinalgExperimental/TritonToLinalgExperimental.h"
 
+#include "unified_hardware_aipu.h"
+
 #include <pybind11/functional.h>
 #include <pybind11/pybind11.h>
 #include <pybind11/stl.h>
@@ -69,6 +71,7 @@ void init_triton_aipu(py::module &&m) {
   init_triton_aipu_common(m.def_submodule("common"));
   auto passes = m.def_submodule("passes");
   init_triton_aipu_passes_convert(passes.def_submodule("convert"));
+
   // load dialects
   m.def("load_dialects", [](mlir::MLIRContext &context) {
     using namespace mlir;
@@ -88,4 +91,9 @@ void init_triton_aipu(py::module &&m) {
     context.loadAllAvailableDialects();
   });
   // register passes here
+
+  //flagtree
+  py::class_<mlir::aipu::UnifiedHardwareAIPU>(m, "UnifiedHardwareAIPU")
+        .def(py::init<>());
+
 }

third_party/aipu/unified_hardware_aipu.h

@@ -0,0 +1,22 @@
+#ifndef UNIFIED_HARDWARE_AIPU_H
+#define UNIFIED_HARDWARE_AIPU_H
+
+#include <optional> 
+
+#include "flagtree/Common/UnifiedHardwareBase.h"
+
+namespace mlir {
+namespace aipu {
+
+class UnifiedHardwareAIPU final : public mlir::flagtree::UnifiedHardware {
+
+    //DMA
+    std::optional<int> getAllocSpaceForDMATag() const override{
+        return std::optional<int>(11);
+    }
+};
+
+} // namespace aipu
+} // namespace mlir
+
+#endif // UNIFIED_HARDWARE_AIPU_H