Fix AntaresCpuKernelEmitter and add ir_based_fusion in GENERIC_CPU backend (#351)

* Update AntaresCpuKernelEmitter for Antares v0.2.x

* Add IRBasedFusion for GENERIC_CPU backend

Co-authored-by: Lingxiao Ma <lingm@microsoft.com>

Author: xysmlx

src/nnfusion/core/kernels/cpu/cpu_kernel_emitter.cpp

@@ -151,20 +151,19 @@ LanguageUnit_p cpu::AntaresCpuKernelEmitter::emit_function_body()
     auto& lu = *_lu;
 
     // extract kernel code
-    const char* s_func_pattern = "// [thread_compute]\n";
+    const char* s_func_pattern = "// [thread_extent] ";
     const char* e_func_pattern = "\n}\n";
     const char* s_rank_pattern = "__rank__ = ";
     const char* e_rank_pattern = "\n";
     std::string::size_type s_func_pos = antares_code.find(s_func_pattern);
     std::string::size_type e_func_pos = antares_code.rfind(e_func_pattern);
 
-    if (s_func_pos != std::string::npos || e_func_pos != std::string::npos)
+    if (s_func_pos == std::string::npos || e_func_pos == std::string::npos)
         return nullptr;
 
     NNFUSION_CHECK(s_func_pos != std::string::npos && e_func_pos != std::string::npos);
 
-    std::string func_body = antares_code.substr(s_func_pos + strlen(s_func_pattern),
-                                                e_func_pos - s_func_pos - strlen(s_func_pattern));
+    std::string func_body = antares_code.substr(s_func_pos, e_func_pos - s_func_pos);
     std::string::size_type s_rank_pos = func_body.find(s_rank_pattern);
     std::string::size_type e_rank_pos = func_body.find(e_rank_pattern);
     std::string rank_str = func_body.substr(s_rank_pos + strlen(s_rank_pattern),

src/nnfusion/engine/device/cpu.cpp

@@ -4,6 +4,7 @@
 #include "cpu.hpp"
 #include "reversed_dfs_visitor.hpp"
 
+#include "nnfusion/engine/pass/extract_graph_signature.hpp"
 #include "nnfusion/engine/pass/graph/assign_async_info_pass.hpp"
 #include "nnfusion/engine/pass/graph/assign_layout_pass.hpp"
 #include "nnfusion/engine/pass/graph/autodiff_pass.hpp"
@@ -14,6 +15,7 @@
 #include "nnfusion/engine/pass/graph/gemm_fusion_pass.hpp"
 #include "nnfusion/engine/pass/graph/gnode_device_dispatcher.hpp"
 #include "nnfusion/engine/pass/graph/gradient_weight_mapping_pass.hpp"
+#include "nnfusion/engine/pass/graph/ir_based_fusion_pass.hpp"
 #include "nnfusion/engine/pass/graph/kernel_fusion_pass.hpp"
 #include "nnfusion/engine/pass/graph/kernel_profiling_pass.hpp"
 #include "nnfusion/engine/pass/graph/kernel_selection.hpp"
@@ -23,8 +25,6 @@
 #include "nnfusion/engine/pass/graph/pattern_substitution.hpp"
 #include "nnfusion/engine/pass/graph/runtime_const_folding_pass.hpp"
 #include "nnfusion/engine/pass/graph/vector_dot_transpose_pass.hpp"
-
-#include "nnfusion/engine/pass/extract_graph_signature.hpp"
 #include "nnfusion/engine/pass/tensor/inplace_tensor_analysis.hpp"
 #include "nnfusion/engine/pass/tensor/liveness_analysis.hpp"
 #include "nnfusion/engine/pass/tensor/tensor_device_dispatcher.hpp"
@@ -51,6 +51,7 @@ CpuEngine::CpuEngine()
     g_passes->push_back(make_shared<AssignLayoutPass>());
     g_passes->push_back(make_shared<OpInplacePass>());
 
+    g_passes->push_back(make_shared<IRBasedFusionPass>());
     g_passes->push_back(make_shared<PatternSubstitutionPass>());
 
     // Kernel selection