Apply TransposeFolding & GemmRewriter passes. (#41084)

cmake/external/cinn.cmake

@@ -26,7 +26,7 @@ add_definitions(-w)
 ######################################
 include(ExternalProject)
 set(CINN_PREFIX_DIR ${THIRD_PARTY_PATH}/CINN)
-set(CINN_GIT_TAG e11c5e672f9961e28cfa403d86f99808beb58817)
+set(CINN_GIT_TAG 1fd85187b6c18da4dd51f22619d093ef08d61b01)
 set(CINN_OPTIONAL_ARGS -DPY_VERSION=${PY_VERSION}
                        -DWITH_CUDA=${WITH_GPU}
                        -DWITH_CUDNN=${WITH_GPU}

paddle/fluid/framework/paddle2cinn/cinn_compiler.cc

@@ -223,9 +223,12 @@ std::unique_ptr<CinnCompiledObject> CinnCompiler::CompileGraph(
     const Target& target, std::int64_t compiled_num, void* stream) const {
   CinnGraphSymbolization symbol{compiled_num, graph, target, input_tensors};
   auto frontend_program = symbol();
-  ProgramPass::Apply(&frontend_program, target, {"Decomposer"});
   auto fetch_ids = symbol.GetFetchIds();
+  ProgramPass::Apply(&frontend_program, fetch_ids, target, {"Decomposer"});
   ::cinn::frontend::ApplyPass(&frontend_program, fetch_ids, "RemoveIdentity");
+  ::cinn::frontend::ApplyPass(&frontend_program, fetch_ids, "TransposeFolding");
+  ProgramPass::Apply(&frontend_program, fetch_ids, target, {"GemmRewriter"});
+
   auto cinn_graph = std::make_shared<::cinn::hlir::framework::Graph>(
       frontend_program, target);
   VLOG(1) << "-- The " << compiled_num << "-th compilation ("