From 6ffb9d46e0bb4750e123300bb2a12ceb167113ef Mon Sep 17 00:00:00 2001
From: Lingxiao Ma <xysmlx@gmail.com>
Date: Sun, 28 Nov 2021 16:57:14 +0900
Subject: [PATCH 1/2] Support -ftuning_list in kernel tuning pass

---
 .../engine/pass/graph/kernel_tuning.cpp       | 30 ++++++++++++++++++-
 .../engine/pass/graph/kernel_tuning.hpp       |  2 ++
 2 files changed, 31 insertions(+), 1 deletion(-)

diff --git a/src/nnfusion/engine/pass/graph/kernel_tuning.cpp b/src/nnfusion/engine/pass/graph/kernel_tuning.cpp
index b9a699c5a..6b689e005 100644
--- a/src/nnfusion/engine/pass/graph/kernel_tuning.cpp
+++ b/src/nnfusion/engine/pass/graph/kernel_tuning.cpp
@@ -20,6 +20,7 @@ DEFINE_int64(fkernel_tuning_steps, 0, "Enable automatic kernel tuning for maximu
 DEFINE_string(ftuning_blocklist,
               "",
               "List of op types that skip kernel tuning pass, e.g., \"Softmax,Add\"");
+DEFINE_string(ftuning_list, "", "List of op types for kernel tuning pass, e.g., \"Softmax,Add\"");
 DEFINE_string(fantares_perf_file, "./antares_perf.csv", "File to save Antares kernel performance.");
 DECLARE_bool(fantares_mode);
 DECLARE_string(fantares_codegen_server);
@@ -130,6 +131,7 @@ void dump_perf(std::string filename,
 
 std::pair<std::vector<std::shared_ptr<GNode>>, std::vector<std::shared_ptr<TuningStatus>>>
     get_tuning_candidates(std::shared_ptr<nnfusion::graph::Graph>& graph,
+                          const std::unordered_set<std::string> tuning_list,
                           const std::unordered_set<std::string> block_list,
                           std::unordered_map<std::string, size_t>& ir2cnt)
 {
@@ -147,6 +149,12 @@ std::pair<std::vector<std::shared_ptr<GNode>>, std::vector<std::shared_ptr<Tunin
         auto n_device_type = (*gnode)["DeviceType"].as<NNFusion_DeviceType>();
         NNFUSION_CHECK(n_device_type != UNKNOWN);
 
+        // filter ops not in TuningList
+        if (tuning_list.find(gnode->get_op_type()) == tuning_list.end())
+        {
+            continue;
+        }
+
         // filter ops in BlockList
         if (block_list.find(gnode->get_op_type()) != block_list.end())
         {
@@ -255,11 +263,31 @@ bool KernelTuning::parse_block_list()
     NNFUSION_LOG(INFO) << "Kernel Tuning BlockList: " << join(BlockList, ", ");
 }
 
+bool KernelTuning::parse_tuning_list()
+{
+    auto tuninglist_str = FLAGS_ftuning_list;
+    stringstream ss(tuninglist_str);
+    while (ss.good())
+    {
+        string substr;
+        getline(ss, substr, ',');
+        TuningList.insert(substr);
+    }
+    NNFUSION_LOG(INFO) << "Kernel Tuning List: " << join(TuningList, ", ");
+}
+
 bool KernelTuning::run_on_graph(std::shared_ptr<nnfusion::graph::Graph>& graph)
 {
     if (FLAGS_fantares_mode)
     {
+        parse_tuning_list();
         parse_block_list();
+        for (auto item : TuningList)
+        {
+            NNFUSION_CHECK(BlockList.find(item) == BlockList.end())
+                << "Kernel Tuning Pass: There are same operators in TuningList and "
+                   "TuningBlockList.";
+        }
         // register antares kernels anyway here in case kernel selection pass will use them
         register_antares_kernel();
     }
@@ -274,7 +302,7 @@ bool KernelTuning::run_on_graph(std::shared_ptr<nnfusion::graph::Graph>& graph)
     std::vector<std::shared_ptr<TuningStatus>> tuning_kernels;
     std::unordered_map<std::string, size_t> ir2cnt;
     std::vector<std::shared_ptr<GNode>> nodes;
-    std::tie(nodes, tuned_kernels) = get_tuning_candidates(graph, BlockList, ir2cnt);
+    std::tie(nodes, tuned_kernels) = get_tuning_candidates(graph, TuningList, BlockList, ir2cnt);
     for (auto gnode : nodes)
     {
         if (!(*gnode)["DeviceType"].is_valid())
diff --git a/src/nnfusion/engine/pass/graph/kernel_tuning.hpp b/src/nnfusion/engine/pass/graph/kernel_tuning.hpp
index 1d49761e2..67cf3a592 100644
--- a/src/nnfusion/engine/pass/graph/kernel_tuning.hpp
+++ b/src/nnfusion/engine/pass/graph/kernel_tuning.hpp
@@ -22,11 +22,13 @@ namespace nnfusion
 
             private:
                 bool parse_block_list();
+                bool parse_tuning_list();
                 bool insert_to_kernel_cache(
                     const std::vector<std::shared_ptr<nnfusion::graph::GNode>>& nodes);
 
             private:
                 std::unordered_set<std::string> BlockList;
+                std::unordered_set<std::string> TuningList;
             };
         }
     }

From 5332e23445537c57e3ca581a59b3adbc3dd46f19 Mon Sep 17 00:00:00 2001
From: Lingxiao Ma <xysmlx@users.noreply.github.com>
Date: Sun, 5 Dec 2021 14:00:27 +0800
Subject: [PATCH 2/2] fix bug in AntaresCudaKernelEmitter when IR is empty

---
 src/nnfusion/core/kernels/cuda_gpu/cuda_emitter.hpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/nnfusion/core/kernels/cuda_gpu/cuda_emitter.hpp b/src/nnfusion/core/kernels/cuda_gpu/cuda_emitter.hpp
index fada9792d..baf5b3551 100644
--- a/src/nnfusion/core/kernels/cuda_gpu/cuda_emitter.hpp
+++ b/src/nnfusion/core/kernels/cuda_gpu/cuda_emitter.hpp
@@ -287,6 +287,7 @@ namespace nnfusion
                                                    << ctx->gnode->get_op_type();
                                 log_cache.insert(ctx->gnode->get_op_type());
                             }
+                            return;
                         }
 
                         kernel_info =