Add: show conv2gemm derivation

Fix: update of derivator.h is missing in NNET_OpSearch

.gitignore

@@ -31,8 +31,7 @@
 *.out
 *.app
 
-build/
-build_debug/
+*build*/
 
 .vscode/
 

3rd-party/backward-cpp

@@ -1 +1 @@
-Subproject commit 3bb9240cb15459768adb3e7d963a20e1523a6294
+Subproject commit f30744bcf726ea3735df7ecf9e9de9ddac540283

3rd-party/googletest

@@ -1 +1 @@
-Subproject commit b796f7d44681514f58a683a3a71ff17c94edb0c1
+Subproject commit e2239ee6043f73722e7aa812a459f54a28552929

3rd-party/nlohmann_json_cmake_fetchcontent

@@ -1 +1 @@
-Subproject commit 13132dd361c8c5b5753983d5186cf54f689d90f9
+Subproject commit 6aebf09233951e4ce30a63919186a70b2b195756

3rd-party/pybind11

@@ -1 +1 @@
-Subproject commit 0bd8896a4010f2d91b2340570c24fa08606ec406
+Subproject commit 1e3400b6742288429f2069aaf5febf92d0662dae

CMakeLists.txt

@@ -129,7 +129,7 @@ if(BUILD_TEST_EINNET)
 endif()
 
 # Python bindings
-file(GLOB_RECURSE FFIS src/ffi/ffi_infinitensor.cc)
+file(GLOB_RECURSE FFIS src/ffi/ffi_callback.cc src/ffi/ffi_infinitensor.cc)
 pybind11_add_module(backend MODULE ${FFIS})
 target_link_libraries(backend PRIVATE InfiniTensor)
 
@@ -168,6 +168,7 @@ endif()
 
 if(USE_CUDA)
   add_compile_definitions(USE_CUDA=1)
+  add_compile_definitions(CUDA_API_PER_THREAD_DEFAULT_STREAM=1) # Support CUDA graph stream caputre
   # Since enable_language only executes once, rerun cmake is required if CMAKE_CUDA_HOST_COMPILER is wrong
   set(CMAKE_CUDA_HOST_COMPILER
       ${CMAKE_CXX_COMPILER}

README_CN.md

@@ -81,7 +81,7 @@ import onnx
 from pyinfinitensor.onnx import OnnxStub
 from pyinfinitensor import backend
 
-stub = OnnxStub(onnx.load("model_file"), backend.cpu_runtime())
+stub = OnnxStub.from_model(onnx.load("model_file"), backend.cpu_runtime())
 ```
 
 [`onnx.load`](https://onnx.ai/onnx/api/serialization.html#load-a-model) 是 onnx 提供的加载函数，将 onnx 文件读取为保存在内存中的 onnx 模型。
@@ -201,7 +201,7 @@ def infer(model: ModelProto, input) -> dict:
 
 
 model0 = onnx.load(sys.argv[1])
-model1 = OnnxStub(model0, backend.cpu_runtime()).to_onnx("new")
+model1 = OnnxStub.from_model(model0, backend.cpu_runtime()).to_onnx("new")
 
 input_shape = [x.dim_value for x in model1.graph.input[0].type.tensor_type.shape.dim]
 input = numpy.random.random(input_shape).astype(numpy.float32)

include/bang/bang_runtime.h

@@ -36,7 +36,7 @@ class BangRuntimeObj : public RuntimeObj {
              bool profiling = false) const;
     // double runEvaluation(const Graph &graph, int nWarmups,
     //                      int nEvaluations) const;
-    void sync() const;
+    void sync() const override;
     BangPtr alloc(size_t size) override {
         void *ptr;
         checkBangError(cnrtMalloc(&ptr, size));

include/core/common.h

@@ -75,7 +75,8 @@ template <typename T> std::string vecToString(const std::vector<T> &vec) {
 
 double timeit(
     const std::function<void()> &func,
-    const std::function<void(void)> &sync = []() {}, int warmupRounds = 200,
-    int timingRounds = 200);
+    // HACK: set timeit rounds to 10 for fast debug
+    const std::function<void(void)> &sync = []() {}, int warmupRounds = 10,
+    int timingRounds = 100);
 
 } // namespace infini

include/core/graph.h

@@ -16,7 +16,8 @@ class GraphObj : public Object {
     string toString() const override;
     Runtime getRuntime() const { return runtime; }
 
-    Tensor addTensor(Shape dim, DataType dtype = DataType::Float32);
+    Tensor addTensor(Shape dim, DataType dtype = DataType::Float32,
+                     TensorType tensorType = TensorType::Other);
     Tensor addTensor(const Tensor &tensor);
     TensorVec addTensor(const TensorVec &tensors);
     /**
@@ -47,6 +48,22 @@ class GraphObj : public Object {
         return opClone;
     }
 
+    Operator cloneOpAndCreateOutputs(Operator op, TensorVec inputs) {
+        auto shapes = *op->inferShape(inputs);
+        vector<Tensor> outputs;
+        for (auto shape : shapes)
+            outputs.emplace_back(addTensor(shape));
+        return cloneOperator(op, inputs, outputs);
+    }
+
+    Operator cloneOpAndCreateInputsOutputs(Operator op) {
+        vector<Tensor> inputs;
+        for (auto t : op->getInputs()) {
+            inputs.emplace_back(cloneTensor(t));
+        }
+        return cloneOpAndCreateOutputs(op, inputs);
+    }
+
     const TensorVec &getTensors() const { return tensors; }
     const OpVec &getOperators() const { return ops; }
     OpVec getComputeOps() const;
@@ -62,6 +79,7 @@ class GraphObj : public Object {
     void optimize();
 
     void dataMalloc();
+    void dataFree();
 
     /**
      * @brief Add an operator and create its outputs. Output tensor arguments
@@ -107,6 +125,11 @@ class GraphObj : public Object {
 
     bool checkValid() const;
 
+    /// @brief If a tensor has no source and garget, it is independent and
+    /// removed from the graph.
+    /// @return The number of removed tensors.
+    int removeIndependentTensors();
+
   private:
     /**
      * @brief Add reverse connections and Op relationship in ctor.

include/core/graph_handler.h

@@ -35,20 +35,33 @@ class GraphHandlerObj {
     Graph g;
 
   public:
-    GraphHandlerObj(Runtime runtime)
+    explicit GraphHandlerObj(Runtime runtime)
         : g(make_ref<GraphObj>(std::move(runtime))) {}
 
-    Tensor tensor(Shape dims, int dtype);
+    explicit GraphHandlerObj(Graph g) : g(std::move(g)) {}
+
+    //------ tensors
+
+    vector<Tensor> inputs() { return g->getInputs(); }
+
+    vector<Tensor> outputs() { return g->getOutputs(); }
+
+    Tensor tensor(Shape dims, int dtype, TensorType ttype);
 
     //------ operators
 
-    inline OpVec operators() { return g->getOperators(); }
+    OpVec operators() { return g->getOperators(); }
 
     Tensor conv(Tensor input, Tensor weight, Tensor output, int ph, int pw,
                 int sh, int sw, int dh, int dw);
     Tensor convTransposed2d(Tensor input, Tensor weight, Tensor output, int ph,
                             int pw, int sh, int sw, int dh, int dw, int oph,
                             int opw);
+    Tensor convNHWC(Tensor input, Tensor weight, Tensor output, int ph, int pw,
+                    int sh, int sw, int dh, int dw);
+    Tensor convTransposed2dNHWC(Tensor input, Tensor weight, Tensor output,
+                                int ph, int pw, int sh, int sw, int dh, int dw,
+                                int oph, int opw);
     Tensor matmul(Tensor a, Tensor b, Tensor y, bool transA, bool transB,
                   Tensor bias, ActType act);
     Tensor batchNorm(Tensor input, Tensor output, Tensor mean, Tensor var,
@@ -90,18 +103,23 @@ class GraphHandlerObj {
                  const optional<vector<int>> &steps);
     Tensor pad(Tensor input, Tensor output, const vector<int> &pads,
                const optional<vector<int>> &axes);
+    /// @brief Import memBound operator from a json
+    TensorVec memBound(const TensorVec &inputs, const Tensor &outputs,
+                       const string &jsonString);
 
     //------ modifiers
 
-    inline bool topo_sort() { return g->topo_sort(); }
+    bool topo_sort() { return g->topo_sort(); }
 
-    inline void optimize() { g->optimize(); }
+    void optimize() { g->optimize(); }
 
     //------ runtime
 
-    inline void data_malloc() { g->dataMalloc(); }
+    void data_malloc() { g->dataMalloc(); }
 
-    inline void run() { g->getRuntime()->run(g); }
+    void run() { g->getRuntime()->run(g); }
+
+    Graph getGraph() const;
 };
 
 } // namespace infini

include/core/mutator.h

@@ -16,6 +16,7 @@ class Mutator {
             Runtime runtime = NativeCpuRuntimeObj::getInstance())
         : candidatesLimit(candidatesLimit), runtime(runtime){};
     virtual ~Mutator(){};
+    bool hasTunedKernel = false;
 
     virtual vector<Graph> run(const Graph &in_graph) = 0;
     /**
@@ -30,6 +31,14 @@ class Mutator {
     virtual bool isMultiBranchMergable(const Graph &in_graph) {
         IT_TODO_HALT();
     }
+
+    /// @brief Fuse memory bound operators.
+    /// @return The graph after fusion. Return `nullptr` if fails.
+    virtual Graph fuseVertically(const Graph &inputGraph) { IT_TODO_HALT(); }
+
+    /// @brief Eliminate transpose and reshape.
+    /// @return The graph after elimination. Return `nullptr` if fails.
+    virtual Graph eliminateVertically(const Graph &in_graph) { IT_TODO_HALT(); }
 };
 
 } // namespace infini

include/core/operator.h

@@ -11,6 +11,7 @@ enum class OpType {
     Matmul,
     ConvTrans,
     ConvTransNHWC,
+    ConvNHWC,
     G2BMM,
     GBMM,
     Pad,
@@ -102,6 +103,10 @@ enum class OpType {
     Dropout,
     //
     MemBound = 300,
+    //
+    Conv2dReduce = 400,
+    Conv2dReduceTranspose,
+    Any
 };
 
 using KernelAttrs = std::tuple<Device, OpType, DataType>;
@@ -121,6 +126,8 @@ class OpRegistry {
             FOP(ConvBackwardData);
             FOP(Matmul);
             FOP(ConvTrans);
+            FOP(ConvTransNHWC);
+            FOP(ConvNHWC);
             FOP(G2BMM);
             FOP(GBMM);
             FOP(Pad);
@@ -141,6 +148,7 @@ class OpRegistry {
             FOP(Reshape);
             FOP(Identity);
             FOP(Shape);
+            FOP(Flatten);
             // element wise
             FOP(BatchNorm);
             FOP(Softmax);
@@ -208,8 +216,13 @@ class OpRegistry {
             FOP(BitRightShift);
             //
             FOP(MemBound);
+            //
+            FOP(Conv2dReduce);
+            FOP(Conv2dReduceTranspose);
+            FOP(Any);
         default:
-            IT_ASSERT(false);
+            IT_ASSERT(false, "Unknown OpType " +
+                                 std::to_string(enum_to_underlying(opType)));
             break;
         }
 #undef FOP

include/core/runtime.h

@@ -1,5 +1,6 @@
 #pragma once
 #include "core/common.h"
+#include "core/object.h"
 #include "core/ref.h"
 #include <memory>
 
@@ -59,10 +60,12 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
      * execution happens.
      *
      * @param graph
-     * @param profiling Whether to print breakdown of time
+     * @param printProfiling Whether to print breakdown of time
      * @return double Return the sum of perf time for each operator
      */
-    double getPerfTime(const Graph &graph, bool profiling = false) const;
+    double getPerfTime(const Graph &graph, bool printProfiling = false,
+                       bool allowEstimation = false,
+                       bool ignoreMemboundOp = false) const;
     Blob allocBlob(size_t size);
     bool isCpu() const {
         return device == Device::CPU || device == Device::INTELCPU;
@@ -76,11 +79,19 @@ class RuntimeObj : public std::enable_shared_from_this<RuntimeObj> {
     virtual void copyBlobToCPU(void *dst, const void *src,
                                size_t bytes) const = 0;
     virtual string toString() const = 0;
+    virtual void sync() const {}
+
+    map<UidBaseType, bool>
+    getCompileTimeComputableAttribute(const Graph &graph) const;
+
+    double timeNonCtcOperators(const Graph &graph, int warmup = 1000,
+                               int repeat = 1000) const;
 
   protected:
-    void printProfilingData(double totTime,
+    void printProfilingData(double totalTime,
                             const std::map<OpType, double> &opTime,
-                            const std::map<OpType, int> &opCnt) const;
+                            const std::map<OpType, int> &opCnt,
+                            const std::map<OpType, int> &opNonCtcCnt) const;
     virtual void copyBlobInsideRuntime(void *dst, const void *src,
                                        size_t bytes) const = 0;
 };

include/core/search_engine.h

@@ -4,44 +4,35 @@
 #include "graph.h"
 #include "mutator.h"
 
-#include <unordered_map>
-
 namespace infini {
 class SearchEngine {
   private:
     Runtime runtimeExec;
     Ref<Mutator> mutator;
+    std::function<bool(const Graph &, const Graph &)> graphTimeComparer;
 
   public:
-    SearchEngine(Runtime _runtime, Ref<Mutator> _mutator) {
-        runtimeExec = _runtime;
-        mutator = _mutator;
-    }
+    SearchEngine(Runtime runtime, Ref<Mutator> mutator);
     ~SearchEngine() {}
+    int searchFilter = 0;
+    bool chooseBestMutation = true;
 
   private: // Configurations
     size_t partitionThreshold =
         3;                  // cut nodes whose #in + #out >= partitionThreshold
     size_t GRAPH_SIZE = 16; // num of best graphs.
 
-  private: // Composed objects
-    std::shared_ptr<Mutator> mutationEngine;
-
   public:
-    std::shared_ptr<Mutator> getMutationEngine() { return mutationEngine; };
     struct GroupEdge {
         int v, next;
         GroupEdge() = delete;
     };
 
-    struct Candidate { // a graph with perf
-        std::shared_ptr<Graph> graph;
-        double perf = INFINITY;
-    };
-    class MetaGraph { // a graph of subgraphs, for searching.
-      public:
-        MetaGraph() {}
-        ~MetaGraph() {}
+    // struct Candidate { // a graph with perf
+    //     Graph graph;
+    //     double perf = INFINITY;
+    // };
+    struct MetaGraphObj { // a graph of subgraphs, for searching.
         struct Node {
             Graph graph;
             std::vector<int> suc;
@@ -50,31 +41,33 @@ class SearchEngine {
         };
         std::vector<Node> nodes;
     };
+    using MetaGraph = Ref<MetaGraphObj>;
 
-    Graph run(const Graph graph);                  // entrance of search engine.
+    Graph run(const Graph graph);                  // entrance to search engine.
     std::vector<Graph> search(const Graph &graph); // search for a partition.
 
   private:
     std::vector<Graph> partitionGraph(const Graph graph);
-    std::shared_ptr<MetaGraph> buildMetaGraphWithGraph(const Graph graph);
-    std::shared_ptr<MetaGraph>
-    buildMetaGraphWithPlan(const std::shared_ptr<MetaGraph> metaGraph,
-                           const std::vector<int> &plan);
+    MetaGraph buildMetaGraphWithGraph(const Graph graph);
+    MetaGraph buildMetaGraphWithPlan(const MetaGraph metaGraph,
+                                     const std::vector<int> &plan);
     // search horizontal merges
-    std::vector<std::shared_ptr<MetaGraph>>
-    searchMerge(std::shared_ptr<MetaGraph> &metaGraph);
-    void searchMergeDfs(std::shared_ptr<MetaGraph> &metaGraph,
-                        std::vector<int> &plan, std::vector<int> &frontier,
+    std::vector<MetaGraph> searchMerge(MetaGraph &metaGraph);
+    void searchMergeDfs(MetaGraph &metaGraph, std::vector<int> &plan,
+                        std::vector<int> &frontier,
                         std::vector<std::vector<int>> &plans,
                         std::unordered_set<uint64_t> &planSet);
-    std::vector<Graph>
-    searchMutation(const std::shared_ptr<MetaGraph> &metaGraph);
+    std::vector<Graph> searchMutation(const MetaGraph &metaGraph);
 
-    void printMetaGraph(Ref<SearchEngine::MetaGraph> metaGraph);
+    void printMetaGraph(MetaGraph metaGraph);
     /**
      * @brief Check whether a multi-brach graph can be merged into a single
      * branch.
      */
     bool isMultiBranchMergable(const Graph graph);
+    Graph fuseVertically(const Graph &graph);
+
+    double getEstimatedGraphPerf(Graph graph);
 };
+
 } // namespace infini

include/core/tensor.h

@@ -12,13 +12,14 @@ namespace infini {
 // TODO: how to deal with this
 using ShapeElem = int;
 using Shape = vector<ShapeElem>;
+enum class TensorType { Error = 0, Input = 1, Initialized = 2, Other = 3 };
 class TensorObj : public TensorBaseObj {
   private:
     Shape shape;
     size_t _size; // Cache of Π(shape).
     Fuid fuid;    // Cloned tensors share the same id. Tensors constructed from
                   // scratch have a new id.
-
+    TensorType tensorType;
     void copyin(const void *ptr, size_t size) {
         runtime->copyBlobFromCPU(getRawDataPtr<void *>(), ptr, size);
     }
@@ -27,7 +28,8 @@ class TensorObj : public TensorBaseObj {
     }
 
   public:
-    TensorObj(Shape shape, DataType dtype, Runtime runtime);
+    TensorObj(Shape shape, DataType dtype, Runtime runtime,
+              TensorType tensorType = TensorType::Other);
     virtual ~TensorObj() {}
     string toString() const override;
 
@@ -39,6 +41,7 @@ class TensorObj : public TensorBaseObj {
     size_t getOffset(const vector<int> &ds) const;
     void dataMalloc();
     UidBaseType getFuid() const { return fuid; }
+    TensorType getTensorType() const { return tensorType; }
 
     void load(std::string file_path);
     void save(std::string file_path);
@@ -74,25 +77,9 @@ class TensorObj : public TensorBaseObj {
     // Thus the internal state of generator cannot be updated.
     void setData(
         std::function<void(void *, size_t, DataType)> const &generator) const;
-    Tensor clone() const {
-        auto obj = make_ref<TensorObj>(*this);
-        obj->freeData();
-        obj->targets.clear();
-        obj->source.reset();
-        return obj;
-    }
-    Tensor clone(Runtime runtime) const {
-        auto obj = make_ref<TensorObj>(*this);
-        obj->runtime = runtime;
-        obj->freeData();
-        obj->targets.clear();
-        obj->source.reset();
-        if (hasData()) {
-            obj->dataMalloc();
-            obj->copyData(this);
-        }
-        return obj;
-    }
+    void setData(const Blob &_blob) { data = _blob; }
+    Tensor clone() const;
+    Tensor clone(Runtime runtime) const;
 
     void printData() const;
     bool equalData(const Tensor &rhs, double relativeError = 1e-6) const;
@@ -106,13 +93,13 @@ class TensorObj : public TensorBaseObj {
     size_t getOffsetByBroadcastOffset(size_t bcOffset, Shape bcShape) const;
 
   private:
-    template <class T> string dataToString() const {
+    template <class T> string dataToString(void *rawPtr) const {
         std::stringstream builder;
         builder << "Tensor: " << guid << std::endl;
 
         auto numDims = shape.size();
         auto dimSzVec = vector<int>(numDims, 1);
-        auto ptr = data->getPtr<T *>();
+        T *ptr = (T *)rawPtr;
         dimSzVec[numDims - 1] = shape[numDims - 1];
 
         for (int i = numDims - 1; i != 0; --i)
@@ -123,6 +110,12 @@ class TensorObj : public TensorBaseObj {
                 if (i % dimSzVec[j] == 0)
                     builder << "[";
 
+            if (iEnd > 1000 && i > 20 && i < iEnd - 20) {
+                printf("... , ");
+                i = iEnd - 20;
+                continue;
+            }
+
             builder << ptr[i];
             for (size_t j = 0; j < numDims; ++j)
                 if ((int)i % dimSzVec[j] == dimSzVec[j] - 1)

include/cuda/cuda_any.h

@@ -0,0 +1,10 @@
+#pragma once
+
+#include "operators/any.h"
+
+namespace infini {
+
+void any_kernel_mapping(vector<float *> input, vector<float *> output,
+                        const string &kernel_name, const vector<int> &attr);
+
+} // namespace infini

include/cuda/cuda_common.h

@@ -13,6 +13,7 @@
         if (cudaSuccess != err) {                                              \
             fprintf(stderr, "Cuda error in %s:%i : %s.\n", __FILE__, __LINE__, \
                     cudaGetErrorString(err));                                  \
+            IT_ASSERT(false);                                                  \
             exit(EXIT_FAILURE);                                                \
         }                                                                      \
     }

include/cuda/cuda_conv2dreduce.h

@@ -0,0 +1,31 @@
+#pragma once
+
+namespace infini {
+
+void conv2dreduce_kernel(float *input, float *bias, float *output, bool PReLU,
+                         int n, int h, int w, int f, int r, int s, int oh,
+                         int ow, int ph, int pw, int sh, int sw, int dh,
+                         int dw);
+
+void convTranspose2dreduce_kernel(float *input, float *bias, float *output,
+                                  int act, int n, int h, int w, int f, int r,
+                                  int s, int oh, int ow, int ph, int pw, int sh,
+                                  int sw, int dh, int dw);
+
+void reduceConvRxSToNCHW(float *input, float *bias, float *output, int act,
+                         int n, int h, int w, int f, int r, int s, int oh,
+                         int ow, int ph, int pw, int sh, int sw, int dh,
+                         int dw);
+
+void convTranspose2dreduce_kernel(float *input, float *bias, float *output,
+                                  int act, int n, int h, int w, int f, int r,
+                                  int s, int oh, int ow, int ph, int pw, int sh,
+                                  int sw, int dh, int dw);
+
+void conv5x5ToConv3x3Reduce(int n, int f, int h, int w, float *input,
+                            float *output, float *bias);
+
+void conv3x3ToReduce(int n, int h, int w, int f, float *input, float *output,
+                     float *bias);
+
+} // namespace infini

include/cuda/cuda_runtime.h

@@ -6,44 +6,52 @@ namespace infini {
 
 class CudaRuntimeObj : public RuntimeObj {
   private:
+    cudaStream_t stream;
     cudnnHandle_t cudnn;
     cublasHandle_t cublas;
     CudaPtr workspace;
     size_t workspaceSize;
 
-  public:
-    CudaRuntimeObj() : RuntimeObj(Device::CUDA) {
+    // Memory information
+    size_t allocatedGPUMemorySize = 0;
+    map<void *, size_t> allocationMap;
 
-        checkCudnnError(cudnnCreate(&cudnn));
-        checkCublasError(cublasCreate(&cublas));
-        // 10GB for Longformer
-        // size_t longformerNum = 3lu * (1 << 30);
-        workspaceSize = 7ll << 30; // 7 GB
-        workspace = alloc(workspaceSize);
-    }
-    virtual ~CudaRuntimeObj() {
-        try {
-            dealloc(workspace);
-            checkCudnnError(cudnnDestroy(cudnn));
-            checkCublasError(cublasDestroy(cublas));
-        } catch (const std::exception &e) {
-            std::cerr << "Error in ~CudaRuntimeObj: " << e.what() << std::endl;
-        }
-    }
+    bool cudaGraphStatus; // Whether CUDA graph stream capture is enabled
+
+    // CUDA device properties
+    cudaDeviceProp deviceProperties;
+
+    bool enableTF32 = false;
+
+  public:
+    CudaRuntimeObj();
+    virtual ~CudaRuntimeObj();
     string toString() const override;
 
     void run(const Graph &graph, bool tune = false,
              bool profiling = false) const;
     // double runEvaluation(const Graph &graph, int nWarmups,
     //                      int nEvaluations) const;
-    void sync() const;
+    void sync() const override;
     CudaPtr alloc(size_t size) override {
         void *ptr;
+        // printf("Try to cudaMalloc: %lu bytes\n", size);
         checkCudaError(cudaMalloc(&ptr, size));
-        // printf("cuda malloc: %p %lu bytes\n", ptr, size);
+        allocatedGPUMemorySize += size;
+        allocationMap[ptr] = size;
+        // printf("cuda malloc: %p %lu bytes, total %lu bytes (%.2lf GB)\n",
+        // ptr,
+        //        size, allocatedGPUMemorySize,
+        //        double(allocatedGPUMemorySize) / 1024 / 1024 / 1024);
         return ptr;
     }
-    void dealloc(void *ptr) override { checkCudaError(cudaFree(ptr)); }
+    void dealloc(void *ptr) override {
+        checkCudaError(cudaFree(ptr));
+        allocatedGPUMemorySize -= allocationMap.at(ptr);
+        allocationMap.erase(ptr);
+        // printf("cuda dealloc: %p %lu bytes, total %lu\n", ptr,
+        //        allocationMap.at(ptr), allocatedGPUMemorySize);
+    }
     cudnnHandle_t cudnnHandle() const { return cudnn; }
     cublasHandle_t cublasHandle() const { return cublas; }
     size_t getWorkspaceSize() const { return workspaceSize; }
@@ -51,6 +59,10 @@ class CudaRuntimeObj : public RuntimeObj {
         IT_ASSERT(size <= workspaceSize);
         return workspace;
     }
+    pair<int, int> getComputeCapacitiy() const {
+        return {deviceProperties.major, deviceProperties.minor};
+    }
+    int getNumSMs() const { return deviceProperties.multiProcessorCount; }
 
     void copyBlobFromCPU(void *dst, const void *src,
                          size_t bytes) const override {
@@ -69,7 +81,19 @@ class CudaRuntimeObj : public RuntimeObj {
 
     void runWithoutSync(const Graph &graph) const;
 
+    bool isInCudaGraph() const { return cudaGraphStatus; }
+    cudaStream_t getStream() const { return stream; }
+
+    double timeWithCudaGraph(Graph graph, int rounds = 50);
+    double timeWithCudaGraph(vector<std::function<void(void)>> funcs,
+                             int rounds = 50);
+    void setEnableTF32(bool state);
+    bool getEnableTF32() const { return enableTF32; }
+
   private:
     void tune(const Graph &graph, bool profiling) const;
+
+    void beginCudaGraphStreamCapture();
+    tuple<cudaGraphExec_t, size_t> endCudaGraphStreamCapture();
 };
 } // namespace infini

include/cuda/cuda_transpose.h

@@ -0,0 +1,16 @@
+#pragma once
+
+#include "operators/transpose.h"
+#include "utils/small_array.h"
+
+namespace infini {
+
+void transpose_kernel(float *input, float *output, int nDims, int size,
+                      SmallArray strides, SmallArray outputShape,
+                      vector<int> _dims_in, vector<int> _dims_out,
+                      vector<int> _perms);
+
+void invoke_transpose_last_two_dim(float *ptrA, float *ptrB, int dim0, int dim1,
+                                   int dim2, int numSMs);
+
+} // namespace infini

include/ffi/ffi_callback.h

@@ -0,0 +1,9 @@
+#include "core/graph_handler.h"
+#include "core/mutator.h"
+#include "core/search_engine.h"
+
+namespace infini {
+namespace callback {
+void exportONNX(const Graph &graph, const string &path);
+}
+} // namespace infini

include/intelcpu/mkl_runtime.h

@@ -29,7 +29,7 @@ class MklRuntimeObj : public CpuRuntimeObj {
     string toString() const override { return "INTELCPU Runtime"; };
     dnnl::engine getEngine() const { return dnnl::engine(engine, true); }
     dnnl::stream getStream() const { return dnnl::stream(stream, true); }
-    void sync() const;
+    void sync() const override;
 };
 
 } // namespace infini

include/nnet/Visitor/Serializer.h

@@ -20,6 +20,7 @@ class Serializer : public Functor<string()> {
     string visit_(const Subscript &c) override;
     string visit_(const Var &c) override;
     string visit_(const Tensor &c) override;
+    string visit_(const Func &c) override;
     string dispatchRoutine(const Routine &c);
 
     Expr buildExprTree(string key);
@@ -29,16 +30,44 @@ class Serializer : public Functor<string()> {
     Serializer(int _verobse = 0);
     virtual ~Serializer();
 
+    /**
+     * @brief Serialize the given expression to string
+     *
+     * @param expr The expression to be serialized
+     * @param msg Message of derivation
+     * @param inputs membound operator attributes
+     * @param exec_time membound operator attributes
+     * @param hint membound operator attributes
+     * @return bool Whether the serialization succeed
+     */
+    std::optional<std::string> toString(Expr const &expr,
+                                        const string &msg = "",
+                                        vector<Tensor> inputs = {},
+                                        double exec_time = -1e9,
+                                        string hint = "");
+
     /**
      * @brief Serialize the given expression to json file
      *
      * @param expr The expression to be serialized
      * @param filePath The path of json file to be output
      * @param msg Message of derivation
+     * @param inputs membound operator attributes
+     * @param exec_time membound operator attributes
+     * @param hint membound operator attributes
      * @return bool Whether the serialization succeed
      */
-    bool serialize(const Expr &expr, const string &filePath,
-                   const string &msg = "");
+    bool toFile(const Expr &expr, const string &filePath,
+                const string &msg = "", vector<Tensor> inputs = {},
+                double exec_time = -1e9, string hint = "");
+
+    /**
+     * @brief Deserialize the given json file to expression
+     *
+     * @param text The text of the expr to be deserialized
+     * @return Expression deserialized from the given json file
+     */
+    Expr fromString(const string &text);
 
     /**
      * @brief Deserialize the given json file to expression
@@ -46,7 +75,15 @@ class Serializer : public Functor<string()> {
      * @param filePath The path to file to be deserialized
      * @return Expression deserialized from the given json file
      */
-    Expr deserialize(const string &filePath);
+    Expr fromFile(const string &filePath);
+
+    tuple<Expr, vector<Tensor>, double, string>
+    deserializeAsMemobundOp(const string &filePath);
+
+    // FIXME: the order of elements in tuple is not consistent with memboundObj
+    // constructor
+    tuple<Expr, vector<Tensor>, double, string>
+    membundOpFromString(const string &data);
 };
 
-} // namespace nnet
+} // namespace nnet

include/nnet/common.h

@@ -69,7 +69,8 @@ static inline HashType genhash(string s) {
     { IT_TODO_HALT(); }
 
 #define nnet_unimplemented_continue()                                          \
-    { dbg("Unimplemented"); }
+    {}
+// { dbg("Unimplemented"); }
 
 #define nnet_assert(expr, msg) assert(((void)(msg), (expr)))
 

include/nnet/derivator.h

@@ -67,11 +67,13 @@ class Derivator {
     vector<string> ruleStates, ruleMsgs;
     int cntStates = 0;   // the number of intermediate states
     int searchState = 0; // search state in guided search
+    bool printAndExit;
+    void printDerivationRules();
 
   public:
     Derivator(int maxDepth = 8, bool enableHashPruning = true,
               LogMode mode = LogMode::NoLog,
-              PassMode passMode = PassMode::Debug);
+              PassMode passMode = PassMode::Debug, bool printAndExit = false);
     void search(Formula &origin, int depth);
     void ruleBasedDFS(Formula &origin, int depth, vector<int> _rules,
                       map<int, vector<Var>> _substituteRules = {},

include/nnet/expr.h

@@ -104,7 +104,7 @@ enum class NodeType {
     FuncNodeType
 };
 
-enum class FuncType { Relu, Tanh, PRelu };
+enum class FuncType { Relu = 1000, Tanh, PRelu };
 
 #define DEFINE_GETTYPE(CLASS, isScalar_v)                                      \
     NodeType getType() const override { return NodeType::CLASS##Type; }        \
@@ -206,7 +206,8 @@ struct IterationType {
     enum { Loop, Sum };
     constexpr static int NumIterationType = 2;
 };
-class RangeOpNode : public OperatorNode {
+class RangeOpNode : public OperatorNode,
+                    public std::enable_shared_from_this<RangeOpNode> {
   public:
     enum { Summand, END_POS };
     constexpr static int Loop = IterationType::Loop;
@@ -230,6 +231,7 @@ class RangeOpNode : public OperatorNode {
         return 0;
     };
     string toReadable() const override;
+    string getFullExpression();
     const Expr &getSummand() const { return subExprs[Summand]; }
     const vector<VarRangePair> &getVarRanges(int _index) const {
         return vars[_index];
@@ -384,13 +386,16 @@ class FuncNode : public ExprNode {
 };
 
 // Wrappers for type deduction
-Subscript makeSubscript(const Expr &tensor, const VecExpr &subscripts);
-RangeOp makeRangeOperator(const vector<VarRangePair> &_loopIters,
-                          const vector<VarRangePair> &_sumIters, Expr _summand,
-                          const vector<int> &paddings = {});
-Tensor makeTensor(const string &name, const vector<int> &shape,
-                  const vector<int> &paddings = {},
-                  const Routine &source = nullptr);
+
+// make subscript
+Subscript mSub(const Expr &tensor, const VecExpr &subscripts);
+// make range operator
+RangeOp mL(const vector<VarRangePair> &_loopIters,
+           const vector<VarRangePair> &_sumIters, Expr _summand,
+           const vector<int> &paddings = {});
+// make tensor
+Tensor mT(const string &name, const vector<int> &shape,
+          const vector<int> &paddings = {}, const Routine &source = nullptr);
 
 // Pretty output for dbg with shared_ptr
 template <typename T, typename std::enable_if_t<std::is_base_of_v<ExprNode, T>>

include/nnet/nmutator.h

@@ -7,32 +7,53 @@ namespace infini {
 class NMutator : public Mutator {
   public:
     enum class Mode { Normal, ToNaiveMembound, RuleBased };
+    using NameNToTensorT = map<string, Tensor>;
 
   private:
     // Suffix -N: NNet objects.
     // Suffix -T: tpm objects.
     // Map: NNet tensors -> tpm tensor.
-    std::map<std::string, Tensor> inputsNameNToTensorT;
+    NameNToTensorT inputsNameNToTensorT;
     Mode mode;
     const double bandwidth = double(200) * 1024 * 1024 * 1024;
     // If in RuleBased mode, use derivationRules in derivator
     const std::vector<int> derivationRules;
+    bool searchFilter = false;
+    bool enableRules = false; // Enable operator-level transformation rules
 
   public:
-    NMutator(Mode mode = Mode::Normal);
-    NMutator(Mode mode, const std::vector<int> &derivationRules);
+    NMutator(Mode mode = Mode::Normal,
+             Runtime runtime = NativeCpuRuntimeObj::getInstance(),
+             bool enableRules = false);
+    NMutator(Mode mode, const std::vector<int> &derivationRules,
+             Runtime runtime = NativeCpuRuntimeObj::getInstance(),
+             bool enableRules = false);
     ~NMutator();
 
     vector<Graph> run(const Graph &in_graph) override;
-    void setToNaiveMembound();
+    Graph fuseVertically(const Graph &in_graph) override;
+    Graph eliminateVertically(const Graph &in_graph) override;
+    bool isMultiBranchMergable(const Graph &in_graph) override;
 
-    void setMaxDepth(int _maxDepth) { maxDepth = _maxDepth; }
+    void setToNaiveMembound();
+    void setMaxDepth(int _maxDepth) {
+        maxDepth = _maxDepth;
+        searchFilter = true;
+    }
     long long cntStates = 0;
     long long cntCandidates = 0;
 
   private:
     int maxDepth = 8;
     nnet::Expr opToExpression(Operator op);
+    /// @brief
+    /// @param op
+    /// @return pair<Expr, map from NNet tensor names to InfiniTensor tensors>
+    static pair<nnet::Expr, NameNToTensorT> extractOp(Operator op);
+    static pair<nnet::Expr, NMutator::NameNToTensorT>
+    generateUnaryExpr(const Operator &op);
+    static pair<nnet::Expr, vector<nnet::Tensor>> generateRevert(Tensor in);
+
     void runSingleOp(Graph in_graph, std::vector<Graph> &out_graphs);
 
     /**
@@ -47,12 +68,28 @@ class NMutator : public Mutator {
     double memboundTime(const Shape &dims);
 
     // TODO: recover these rules
-    // Graph fuseHetConv(nnet::Expr expr, Graph in_graph);
-    // Graph transformTConv1x1(Operator op);
-    // Graph transformTConv3x3(Operator op);
-    // Graph transformDialtedConv(Operator op);
-    // Graph transformConv1x1(Operator op);
+    Graph transformConvtransposed1x1(Operator _op);
+    // Graph transformConvtransposed(Operator op);
+    vector<Graph> transformConv1x1(Operator op);
+    vector<Graph> transformConv3x3ONNX(Operator op);
+    Graph transformG2bmm(Operator op);
+    Graph transformGbmm(Operator op);
+    Graph transformDialtedConv(Operator _op);
+    vector<Graph> transformConv1xk(Operator op);
     // Graph transformConv1xk(Operator op);
+    Graph transformConvToGEMMReduce(Operator _op);
+    Graph transformConvTranposeToGEMMReduce(Operator _op);
+
+    Tensor splitTransposeMerge(Graph g, Tensor A, int dim, int chunkSize,
+                               Tensor output = nullptr);
+
+    /// @brief Construct a new graph with a chain of operators. Use the output
+    /// from the previous operator as the input of the next operator. While
+    /// constructing, the input and output tensors from inputGraph are used as
+    /// new constructed graph.
+    /// @param op The operator chain. It can have wrong input/output shapes.
+    /// @return
+    Graph constructGraphByOperatorChain(vector<Operator> ops, Graph inputGraph);
 };
 
 } // namespace infini

include/nnet/test_models.h

@@ -0,0 +1,23 @@
+#ifdef USE_CUDA
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "core/search_engine.h"
+
+namespace infini {
+
+Graph getGANGraph(int batch, Runtime runtime, int nLayers, int modelId);
+Graph getFSRCNNGraph(int batch, Runtime runtime);
+Graph getLongformer(Runtime runtime, int bs);
+vector<Tensor> runInfoGAN(int nLayers);
+Graph getConvtransposedNHWC(Runtime runtime, Shape shape, int layerId);
+Graph optimizeGraph(Graph g, Runtime _runtime, bool tuning, NMutator::Mode mode,
+                    vector<int> rules);
+void initializeGraphTensors(Graph g, double l, double r, bool useInt);
+Graph convertNCHWtoNHWCModel(Runtime runtime, Graph inG);
+Graph optimizeWithDepthConstraint(Graph g, Runtime _runtime, int maxDepth);
+Graph optimizeModel(Graph g, Runtime _runtime, string name);
+Graph optimizeModelWithRules(Graph g, Runtime _runtime, vector<int> rules);
+
+} // namespace infini
+
+#endif

include/nnet/visitor.h

@@ -49,7 +49,7 @@ template <typename R, typename... Args> class Functor<R(Args...)> {
     virtual R visit_(const Tensor &c, Args... args) FUNCTOR_DEFAULT;
     virtual R visit_(const Func &c, Args... args) FUNCTOR_DEFAULT;
     virtual R visitDefault(const Expr &c, [[maybe_unused]] Args... args) {
-        dbg(*c);
+        dbg(*c, c->getType());
         nnet_assert(0, "Reach unimplemented visit function.");
         return R();
     };

include/operators/any.h

@@ -0,0 +1,30 @@
+#pragma once
+#include "core/operator.h"
+
+namespace infini {
+
+class AnyObj : public OperatorObj {
+  private:
+    string kernelName;
+    vector<int> attr;
+
+  public:
+    AnyObj(GraphObj *graph, const TensorVec &inputs, const TensorVec &outputs,
+           const string &kernelName, const vector<int> &attr);
+
+    OP_CLONE(AnyObj);
+
+    string toString() const override;
+
+    optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
+
+    int numInputs() const override { return inputs.size(); }
+    int numOutputs() const override { return outputs.size(); }
+
+    const string getKernelName() const;
+    void setAttr(int i, int v) { attr[i] = v; }
+    vector<int> getOpAttrVector() const override;
+    vector<int> getWorkloadVector() const override;
+};
+
+} // namespace infini

include/operators/conv.h

@@ -98,7 +98,7 @@ class ConvBaseObj : public OperatorObj {
     int numInputs() const override { return 2; }
     int numOutputs() const override { return 1; }
 
-    Tensor getBias() const { return inputs[2]; }
+    Tensor getBias() const { return inputs.size() > 2 ? inputs[2] : nullptr; }
     PaddingMode getPaddingMode() const { return padding; }
     pair<int, int> inferPaddingSize() const;
 
@@ -111,7 +111,7 @@ class ConvBaseObj : public OperatorObj {
     auto getNCHWFRS() const { return tuple(n, c, h, w, f, r, s); }
     auto getPadStrideDilation() const { return tuple(ph, pw, sh, sw, dh, dw); }
     int getChannelPerGroup() const {
-        if (type == OpType::ConvTransNHWC) {
+        if (type == OpType::ConvTransNHWC || type == OpType::ConvNHWC) {
             return inputs[1]->getDims()[3];
         } else {
             return inputs[1]->getDims()[1];
@@ -149,6 +149,25 @@ class ConvObj : public ConvBaseObj {
     void setAuxilaryAttributes(PaddingMode mode) override;
 };
 
+class ConvNHWCObj : public ConvBaseObj {
+  public:
+    ConvNHWCObj(GraphObj *graph, Tensor input, Tensor weight, Tensor output,
+                int ph, int pw, int sh = 1, int sw = 1, int dh = 1, int dw = 1,
+                Tensor bias = nullptr, ActType act = ActType::None);
+    // Constructors for setting padding mode
+    ConvNHWCObj(GraphObj *graph, Tensor input, Tensor weight, Tensor output,
+                PaddingMode mode = PaddingMode::Same, int sh = 1, int sw = 1,
+                int dh = 1, int dw = 1, Tensor bias = nullptr,
+                ActType act = ActType::None);
+    OP_CLONE(ConvNHWCObj);
+
+    optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
+    int getNumGroups() const override { return c / getChannelPerGroup(); }
+
+  private:
+    void setAuxilaryAttributes(PaddingMode mode) override;
+};
+
 class ConvBackwardFilterObj : public ConvBaseObj {
   private:
     ActType act;
@@ -220,6 +239,7 @@ class ConvTransposed2dNHWCObj : public ConvBaseObj {
 
     optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
     int getNumGroups() const override { return group; }
+    std::pair<int, int> getOutputPadding() const { return {oph, opw}; }
 
   private:
     void setAuxilaryAttributes(PaddingMode mode) override;

include/operators/conv2dreduce.h

@@ -0,0 +1,62 @@
+#pragma once
+#include "core/operator.h"
+
+namespace infini {
+
+class Conv2dReduceBase : public OperatorObj {
+  protected:
+    Tensor bias;
+    int ph, pw;
+    int sh, sw;
+    int dh, dw;
+    int n, h, w, f, r, s; // c has been reduced
+    bool PReLU;
+    float paramReLU;
+
+  public:
+    Conv2dReduceBase(OpType opType, Tensor input, Tensor bias, Tensor output,
+                     bool PReLU_, float paramReLU_, int ph_, int pw_,
+                     int sh_ = 1, int sw_ = 1, int dh_ = 1, int dw_ = 1);
+
+    std::string toString() const override;
+    int numInputs() const override { return 2; }
+    int numOutputs() const override { return 1; }
+
+    int getDh() const { return dh; }
+    int getDw() const { return dw; }
+    int getPh() const { return ph; }
+    int getPw() const { return pw; }
+    int getSh() const { return sh; }
+    int getSw() const { return sw; }
+    bool getPReLU() const { return PReLU; }
+    float getParamReLU() const { return paramReLU; }
+
+    Tensor getBias() const { return bias; }
+
+    // optional<vector<Shape>> inferShape(const TensorVec &inputs) const
+    // override;
+
+  private:
+    vector<int> getWorkloadVector() const override;
+    vector<int> getOpAttrVector() const override;
+};
+
+class Conv2dReduce : public Conv2dReduceBase {
+  public:
+    Conv2dReduce(GraphObj *graph, Tensor input, Tensor bias, Tensor output,
+                 bool PReLU_, float paramReLU_, int ph_, int pw_, int sh_ = 1,
+                 int sw_ = 1, int dh_ = 1, int dw_ = 1);
+    OP_CLONE(Conv2dReduce);
+    optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
+};
+
+class Conv2dReduceTranspose : public Conv2dReduceBase {
+  public:
+    Conv2dReduceTranspose(GraphObj *graph, Tensor input, Tensor bias,
+                          Tensor output, bool PReLU_, float paramReLU_, int ph_,
+                          int pw_, int sh_ = 1, int sw_ = 1, int dh_ = 1,
+                          int dw_ = 1);
+    OP_CLONE(Conv2dReduceTranspose);
+    optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
+};
+} // namespace infini

include/operators/membound.h

@@ -6,12 +6,17 @@ namespace infini {
 
 class MemBoundObj : public OperatorObj {
   private:
-    std::vector<nnet::Tensor> nnetInputs;
-    nnet::Expr expr, simplifiedExpr;
+    nnet::Expr expr;
+    std::vector<nnet::Tensor>
+        nnetInputs; // The order of inputs in nnetInputs should be consistant
+                    // with inputs in infinitensor
     double exec_time;
     std::string hint;
-    HashType hash, simplifiedHash;
-    int n, f, h, w;
+
+    // Generated attributes
+    HashType hash;
+    nnet::Expr simplifiedExpr;
+    HashType simplifiedHash;
 
   public:
     MemBoundObj(GraphObj *graph, const TensorVec &input,
@@ -27,9 +32,12 @@ class MemBoundObj : public OperatorObj {
     int numOutputs() const override { return outputs.size(); }
     const vector<nnet::Tensor> &getNnetInputs() const { return nnetInputs; }
     const nnet::Expr getNnetExpr() const { return expr; }
+    HashType getHash() const { return hash; }
     pair<const nnet::Expr, HashType> getSimplifiedNnetExpr() const {
         return {expr, hash};
     }
+    double getEstimatedTime() const { return exec_time; }
+    string toJson() const;
 
   private:
     vector<int> getWorkloadVector() const override;

include/operators/reshape.h

@@ -19,7 +19,7 @@ class ReshapeObj : public OperatorObj {
      * @param output The output tensor.
      * @param dims The shape of the output tensor.
      */
-    ReshapeObj(GraphObj *graph, Tensor input, Tensor output, Shape dims);
+    ReshapeObj(GraphObj *graph, Tensor input, Tensor output, Shape dims = {});
     OP_CLONE(ReshapeObj);
 
     optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
@@ -60,6 +60,7 @@ class FlattenObj : public OperatorObj {
     std::string toString() const override;
     int numInputs() const override { return 1; }
     int numOutputs() const override { return 1; }
+    int getAxis() const { return axis; }
 
   private:
     vector<int> getWorkloadVector() const override;

include/operators/slice.h

@@ -7,7 +7,9 @@ namespace infini {
  *
  */
 class SliceObj : public OperatorObj {
-    template <class T> struct range_t { T start, end, step; };
+    template <class T> struct range_t {
+        T start, end, step;
+    };
     vector<range_t<int>> axes;
 
   public:

include/operators/transpose.h

@@ -3,6 +3,8 @@
 
 namespace infini {
 class TransposeObj : public OperatorObj {
+    vector<int> transposePermute;
+
   public:
     TransposeObj(GraphObj *graph, Tensor input, Tensor output,
                  vector<int> permute);
@@ -15,7 +17,6 @@ class TransposeObj : public OperatorObj {
     std::vector<int> getPermute() const { return transposePermute; }
 
   private:
-    vector<int> transposePermute = {1, 1, 1, 1};
     vector<int> getWorkloadVector() const override;
     vector<int> getOpAttrVector() const override;
 };

include/utils/data_generator.h

@@ -46,10 +46,13 @@ class RandomGenerator : public DataGenerator {
     std::mt19937 e;
     std::uniform_int_distribution<int> di;
     std::uniform_real_distribution<float> dr;
+    bool generateInteger;
 
   public:
-    RandomGenerator(double l = 0, double r = 1, unsigned int seed = 0)
-        : l(l), r(r), e(seed), di(l, r), dr(l, r) {}
+    RandomGenerator(double l = 0, double r = 1, unsigned int seed = 0,
+                    bool generateInteger = false)
+        : l(l), r(r), e(seed), di(l, r), dr(l, r),
+          generateInteger(generateInteger) {}
     virtual ~RandomGenerator() {}
 
   private:
@@ -60,7 +63,7 @@ class RandomGenerator : public DataGenerator {
     }
     void fill(float *data, size_t size) override {
         for (size_t i = 0; i < size; i++) {
-            data[i] = dr(e);
+            data[i] = (generateInteger) ? di(e) : dr(e);
         }
     }
 };

include/utils/small_array.h

@@ -0,0 +1,8 @@
+namespace infini {
+
+#define SMALL_ARRAY_SIZE 8
+struct SmallArray {
+    int data[SMALL_ARRAY_SIZE];
+};
+
+} // namespace infini

pyinfinitensor/src/pyinfinitensor/onnx.py

@@ -25,6 +25,7 @@ from onnx.shape_inference import infer_shapes
 from onnx.numpy_helper import to_array
 from typing import Dict, List, Any, Tuple, Sequence, Union, Optional
 from functools import reduce
+import numpy as np
 
 
 class OnnxStub:
@@ -37,29 +38,48 @@ class OnnxStub:
     outputs: Dict[str, backend.Tensor] = {}
     initializer: Dict[int, TensorProto] = {}
     handler: backend.GraphHandler
+    disable_check: bool
 
-    def __init__(self, model: ModelProto, runtime):
-        model = infer_shapes(model)
-        self.handler = backend.GraphHandler(runtime)
+    @classmethod
+    def from_onnx(cls, model: ModelProto, runtime, enable_onnx_shape_infernce=True):
+        if enable_onnx_shape_infernce:
+            model = infer_shapes(model)
+        ans = OnnxStub()
+        ans.handler = backend.GraphHandler(runtime)
 
         tensors: Dict[str, backend.Tensor] = dict()
         data: Dict[str, TensorProto] = dict()
 
+        cnt_infini_inputs = 0
         for input in model.graph.input:
             dims = _take_shape_dim(input.type.tensor_type.shape)
-            tensors[input.name] = self.handler.tensor(
-                dims, input.type.tensor_type.elem_type
+            if input.name.startswith('input'):
+                tensor_type = backend.TensorType.Input
+                cnt_infini_inputs += 1
+            else:
+                tensor_type = backend.TensorType.Initialized
+            tensors[input.name] = ans.handler.tensor(
+                dims,
+                input.type.tensor_type.elem_type,
+                tensor_type,
             )
+        assert cnt_infini_inputs == 1, f'{cnt_infini_inputs} tensor names start with "input" found.'
 
         for output in model.graph.output:
             dims = _take_shape_dim(output.type.tensor_type.shape)
-            tensors[output.name] = self.handler.tensor(
-                dims, output.type.tensor_type.elem_type
+            tensors[output.name] = ans.handler.tensor(
+                dims,
+                output.type.tensor_type.elem_type,
+                backend.TensorType.Other,
             )
 
         for initializer in model.graph.initializer:
             dims = [d for d in initializer.dims]
-            tensors[initializer.name] = self.handler.tensor(dims, initializer.data_type)
+            tensors[initializer.name] = ans.handler.tensor(
+                dims,
+                initializer.data_type,
+                backend.TensorType.Initialized,
+            )
             data[initializer.name] = initializer
 
         for node in model.graph.node:
@@ -77,17 +97,18 @@ class OnnxStub:
                 )
                 if p[0] != p[2] or p[1] != p[3]:
                     adapt = "{}-adapt".format(node.output[0])
-                    tensors[adapt] = self.handler.pad(
+                    tensors[adapt] = ans.handler.pad(
                         tensors[node.input[0]], None, p, [-2, -1]
                     )
                     p = [0, 0, 0, 0]
                 else:
                     adapt = node.input[0]
 
-                if len(node.input) > 2:
+                # HACK: ignore bias
+                if len(node.input) > 3:
                     bias = "{}-bias".format(node.output[0])
                     reshape = "{}-reshape".format(node.output[0])
-                    tensors[bias] = self.handler.conv(
+                    tensors[bias] = ans.handler.conv(
                         tensors[adapt],
                         tensors[node.input[1]],
                         None,
@@ -98,7 +119,7 @@ class OnnxStub:
                         d[0],
                         d[1],
                     )
-                    tensors[reshape] = self.handler.reshape(
+                    tensors[reshape] = ans.handler.reshape(
                         tensors[node.input[2]],
                         None,
                         [
@@ -111,13 +132,13 @@ class OnnxStub:
                             1,
                         ],
                     )
-                    tensors[node.output[0]] = self.handler.add(
+                    tensors[node.output[0]] = ans.handler.add(
                         tensors[bias],
                         tensors[reshape],
                         tensors.get(node.output[0]),
                     )
                 else:
-                    tensors[node.output[0]] = self.handler.conv(
+                    tensors[node.output[0]] = ans.handler.conv(
                         tensors[adapt],
                         tensors[node.input[1]],
                         tensors.get(node.output[0]),
@@ -142,7 +163,7 @@ class OnnxStub:
                     attributes[name]
                     for name in ["dilations", "pads", "strides", "output_padding"]
                 )
-                tensors[node.output[0]] = self.handler.convTransposed2d(
+                tensors[node.output[0]] = ans.handler.convTransposed2d(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
@@ -156,7 +177,7 @@ class OnnxStub:
                     op[1],
                 )
             elif node.op_type == "MatMul":
-                tensors[node.output[0]] = self.handler.matmul(
+                tensors[node.output[0]] = ans.handler.matmul(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
@@ -175,7 +196,7 @@ class OnnxStub:
                 # FIXME unsupport attributes: `alpha` `beta`
                 assert alpha == 1.0
                 assert beta == 1.0
-                tensors[node.output[0]] = self.handler.matmul(
+                tensors[node.output[0]] = ans.handler.matmul(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
@@ -196,7 +217,7 @@ class OnnxStub:
                     attributes[name]
                     for name in ["momentum", "epsilon", "training_mode"]
                 )
-                tensors[node.output[0]] = self.handler.batchNorm(
+                tensors[node.output[0]] = ans.handler.batchNorm(
                     input, output, mean, var, scale, bias, momentum, eps, training != 0
                 )
             elif node.op_type == "MaxPool":
@@ -215,10 +236,10 @@ class OnnxStub:
                 )
                 if p[0] != p[2] or p[1] != p[3]:
                     adapt = "{}-adapt".format(node.output[0])
-                    tensors[adapt] = self.handler.pad(
+                    tensors[adapt] = ans.handler.pad(
                         tensors.get(node.input[0]), None, p, [-2, -1]
                     )
-                    tensors[node.output[0]] = self.handler.maxPool(
+                    tensors[node.output[0]] = ans.handler.maxPool(
                         tensors[adapt],
                         tensors.get(node.output[0]),
                         k[0],
@@ -231,7 +252,7 @@ class OnnxStub:
                         s[1],
                     )
                 else:
-                    tensors[node.output[0]] = self.handler.maxPool(
+                    tensors[node.output[0]] = ans.handler.maxPool(
                         tensors[node.input[0]],
                         tensors.get(node.output[0]),
                         k[0],
@@ -257,10 +278,10 @@ class OnnxStub:
                 )
                 if p[0] != p[2] or p[1] != p[3]:
                     adapt = "{}-adapt".format(node.output[0])
-                    tensors[adapt] = self.handler.pad(
+                    tensors[adapt] = ans.handler.pad(
                         tensors.get(node.input[0]), None, p, [-2, -1]
                     )
-                    tensors[node.output[0]] = self.handler.avgPool(
+                    tensors[node.output[0]] = ans.handler.avgPool(
                         tensors[adapt],
                         tensors.get(node.output[0]),
                         k[0],
@@ -273,7 +294,7 @@ class OnnxStub:
                         s[1],
                     )
                 else:
-                    tensors[node.output[0]] = self.handler.avgPool(
+                    tensors[node.output[0]] = ans.handler.avgPool(
                         tensors[node.input[0]],
                         tensors.get(node.output[0]),
                         k[0],
@@ -287,7 +308,7 @@ class OnnxStub:
                     )
             elif node.op_type == "GlobalAveragePool":
                 [_, _, h, w] = _search_shape(model, node.input[0])
-                tensors[node.output[0]] = self.handler.avgPool(
+                tensors[node.output[0]] = ans.handler.avgPool(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     h,
@@ -300,52 +321,52 @@ class OnnxStub:
                     1,
                 )
             elif node.op_type == "Add":
-                tensors[node.output[0]] = self.handler.add(
+                tensors[node.output[0]] = ans.handler.add(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Sub":
-                tensors[node.output[0]] = self.handler.sub(
+                tensors[node.output[0]] = ans.handler.sub(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Mul":
-                tensors[node.output[0]] = self.handler.mul(
+                tensors[node.output[0]] = ans.handler.mul(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Div":
-                tensors[node.output[0]] = self.handler.div(
+                tensors[node.output[0]] = ans.handler.div(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Pow":
-                tensors[node.output[0]] = self.handler.pow(
+                tensors[node.output[0]] = ans.handler.pow(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Relu":
-                tensors[node.output[0]] = self.handler.relu(
+                tensors[node.output[0]] = ans.handler.relu(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Sigmoid":
-                tensors[node.output[0]] = self.handler.sigmoid(
+                tensors[node.output[0]] = ans.handler.sigmoid(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Tanh":
-                tensors[node.output[0]] = self.handler.tanh(
+                tensors[node.output[0]] = ans.handler.tanh(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Softmax":
-                tensors[node.output[0]] = self.handler.softmax(
+                tensors[node.output[0]] = ans.handler.softmax(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     next(
@@ -353,34 +374,39 @@ class OnnxStub:
                     ),
                 )
             elif node.op_type == "Abs":
-                tensors[node.output[0]] = self.handler.abs(
+                tensors[node.output[0]] = ans.handler.abs(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Shape":
-                tensors[node.output[0]] = self.handler.shape(
+                tensors[node.output[0]] = ans.handler.shape(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Identity":
-                tensors[node.output[0]] = self.handler.identity(
+                tensors[node.output[0]] = ans.handler.identity(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                 )
             elif node.op_type == "Flatten":
-                tensors[node.output[0]] = self.handler.flatten(
+                tensors[node.output[0]] = ans.handler.flatten(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     next((attr.i for attr in node.attribute if attr.name == "axis")),
                 )
             elif node.op_type == "PRelu":
-                tensors[node.output[0]] = self.handler.pRelu(
+                # HACK: replace PRelu with Relu
+                tensors[node.output[0]] = ans.handler.relu(
                     tensors[node.input[0]],
-                    tensors[node.input[1]],
                     tensors.get(node.output[0]),
                 )
+                # tensors[node.output[0]] = ans.handler.pRelu(
+                #     tensors[node.input[0]],
+                #     tensors[node.input[1]],
+                #     tensors.get(node.output[0]),
+                # )
             elif node.op_type == "Clip":
-                tensors[node.output[0]] = self.handler.clip(
+                tensors[node.output[0]] = ans.handler.clip(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     next(_parse_data(data[node.input[1]]).__iter__(), None)
@@ -394,7 +420,7 @@ class OnnxStub:
                 perm = next(
                     (attr.ints for attr in node.attribute if attr.name == "perm"), None
                 )
-                tensors[node.output[0]] = self.handler.transpose(
+                tensors[node.output[0]] = ans.handler.transpose(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     perm,
@@ -409,7 +435,7 @@ class OnnxStub:
                 temp = reduce(lambda acc, x: acc * x, input_shape, 1)
                 if temp < 0:
                     input_shape[input_shape.index(-1)] = size // -temp
-                tensors[node.output[0]] = self.handler.reshape(
+                tensors[node.output[0]] = ans.handler.reshape(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     input_shape,
@@ -426,7 +452,7 @@ class OnnxStub:
                 for i, x in enumerate(input_shape):
                     if i not in axes:
                         output_shape.append(x)
-                tensors[node.output[0]] = self.handler.reshape(
+                tensors[node.output[0]] = ans.handler.reshape(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     output_shape,
@@ -440,13 +466,13 @@ class OnnxStub:
                 )
                 for i in axes:
                     input_shape.insert(i, 1)
-                tensors[node.output[0]] = self.handler.reshape(
+                tensors[node.output[0]] = ans.handler.reshape(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     input_shape,
                 )
             elif node.op_type == "Concat":
-                tensors[node.output[0]] = self.handler.concat(
+                tensors[node.output[0]] = ans.handler.concat(
                     [tensors[name] for name in node.input],
                     tensors.get(node.output[0]),
                     next((attr.i for attr in node.attribute if attr.name == "axis")),
@@ -454,7 +480,7 @@ class OnnxStub:
             elif node.op_type == "Split":
                 for name, tensor in zip(
                     node.output,
-                    self.handler.split(
+                    ans.handler.split(
                         tensors[node.input[0]],
                         None,
                         next(
@@ -466,14 +492,14 @@ class OnnxStub:
                 ):
                     tensors[name] = tensor
             elif node.op_type == "Gather":
-                tensors[node.output[0]] = self.handler.gather(
+                tensors[node.output[0]] = ans.handler.gather(
                     tensors[node.input[0]],
                     tensors[node.input[1]],
                     tensors.get(node.output[0]),
                     next((attr.i for attr in node.attribute if attr.name == "axis")),
                 )
             elif node.op_type == "ReduceMean":
-                tensors[node.output[0]] = self.handler.reduce_mean(
+                tensors[node.output[0]] = ans.handler.reduce_mean(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     tensors[node.input[1]] if len(node.input) > 1 else None,
@@ -481,7 +507,7 @@ class OnnxStub:
                     != 0,
                 )
             elif node.op_type == "Slice":
-                tensors[node.output[0]] = self.handler.slice(
+                tensors[node.output[0]] = ans.handler.slice(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     _parse_data(data[node.input[1]]),
@@ -490,7 +516,7 @@ class OnnxStub:
                     _parse_data(data[node.input[4]]) if len(node.input) > 4 else None,
                 )
             elif node.op_type == "Pad":
-                tensors[node.output[0]] = self.handler.pad(
+                tensors[node.output[0]] = ans.handler.pad(
                     tensors[node.input[0]],
                     tensors.get(node.output[0]),
                     _parse_data(data[node.input[1]]),
@@ -499,7 +525,7 @@ class OnnxStub:
             elif node.op_type == "Dropout":
                 for name, tensor in zip(
                     node.output,
-                    self.handler.dropout(
+                    ans.handler.dropout(
                         tensors[node.input[0]],
                         tensors.get(node.output[0]),
                         tensors.get(node.output[1]) if len(node.output) > 1 else None,
@@ -512,18 +538,35 @@ class OnnxStub:
                     ),
                 ):
                     tensors[name] = tensor
+            elif node.op_type == "MemBound":
+                attributes = _parse_attribute(node, {"expr": None})
+                expr: str = attributes["expr"]
+                assert expr is not None
+                assert (
+                    len(node.output) == 1
+                ), """MemBound with multiple
+                    outputs requires rewrite the logic of tensor creation"""
+                outputs = ans.handler.memBound(
+                    [tensors[name] for name in node.input],
+                    tensors.get(node.output[0]),
+                    expr,
+                )
+                for name, tensor in zip(node.output, outputs):
+                    tensors[name] = tensor
             else:
                 raise Exception('Unsupported operator "{}"'.format(node.op_type))
 
-        self.handler.data_malloc()
+        # FIXME: do not load data for speed
+        return ans
+        ans.handler.data_malloc()
 
         for name, obj in tensors.items():
             tensor = data.get(name)
             if tensor == None:
                 if any(input.name == name for input in model.graph.input):
-                    self.inputs[name] = obj
+                    ans.inputs[name] = obj
             else:
-                self.initializer[obj.fuid()] = tensor
+                ans.initializer[obj.fuid()] = tensor
                 if tensor.data_type == TensorProto.INT32:
                     obj.copyin_int32(_parse_data(tensor))
                 elif tensor.data_type == TensorProto.INT64:
@@ -533,8 +576,19 @@ class OnnxStub:
                 else:
                     assert False, "Unsupported Tensor Type: {}".format(tensor.data_type)
 
-        for output in model.graph.output:
-            self.outputs[output.name] = tensors[output.name]
+        return ans
+
+    @classmethod
+    def from_graph(cls, g: backend.Graph):
+        ans = OnnxStub()
+        handler = backend.GraphHandler(g)
+        for i, tensor in enumerate(handler.inputs()):
+            ans.inputs["input{}".format(i)] = tensor
+        for i, tensor in enumerate(handler.outputs()):
+            ans.inputs["output{}".format(i)] = tensor
+        ans.handler = handler
+        ans.disable_check = True
+        return ans
 
     def to_onnx(self, name: str) -> ModelProto:
         class Context:
@@ -552,6 +606,13 @@ class OnnxStub:
             outputs: List[ValueInfoProto] = []
             # saves global input tensors
             initializers: List[TensorProto] = []
+            # saves global output tensors
+            value_info: List[ValueInfoProto] = []
+
+            enable_check = False
+
+            def __init__(self, enable_check):
+                self.enable_check = enable_check
 
             def name_op(self, op: backend.Operator) -> Tuple[backend.OpType, str]:
                 ty = op.op_type()
@@ -562,12 +623,15 @@ class OnnxStub:
 
             def push_output(self, name: str, tensor: backend.Tensor) -> str:
                 self.names[tensor] = name
-                if not tensor.has_target():
-                    shape = tensor.shape()
-                    dtype = backend.tensor_dtype(tensor)
-                    value_info = make_tensor_value_info(name, dtype, shape)
-                    check_value_info(value_info)
+
+                shape = tensor.shape()
+                dtype = backend.tensor_dtype(tensor)
+                value_info = make_tensor_value_info(name, dtype, shape)
+                check_value_info(value_info)
+                if not tensor.has_target():  # if this output is a global output
                     self.outputs.append(value_info)
+                else:  # if this output is a local output
+                    self.value_info.append(value_info)
                 return name
 
             def push_input(
@@ -577,7 +641,15 @@ class OnnxStub:
                 # means that this input is a global input
                 if name is None:
                     self.count_in += 1
-                    name = "input{}".format(self.count_in)
+                    if tensor.getTensorType() == backend.TensorType.Input:
+                        name = f"input{self.count_in}_{tensor.guid()}"
+                    else:
+                        name = f"weight{self.count_in}_{tensor.guid()}"
+                        shape = tensor.shape()
+                        data = np.random.randn(*shape)
+                        self.initializers.append(
+                            make_tensor(name, TensorProto.FLOAT, shape, data)
+                        )
                     self.names[tensor] = name
                     if init != None:
                         init.name = name
@@ -605,17 +677,25 @@ class OnnxStub:
                 return name
 
             def push_node(self, node: NodeProto) -> None:
-                check_node(node)
+                if self.enable_check:
+                    check_node(node)
                 self.nodes.append(node)
 
             def build(self, name: str) -> ModelProto:
                 graph = make_graph(
-                    self.nodes, name, self.inputs, self.outputs, self.initializers
+                    self.nodes,
+                    name,
+                    self.inputs,
+                    self.outputs,
+                    self.initializers,
+                    value_info=self.value_info,
                 )
-                check_graph(graph)
+                if self.enable_check:
+                    check_graph(graph)
 
                 model = make_model(graph)
-                check_model(model)
+                if self.enable_check:
+                    check_model(model)
 
                 return model
 
@@ -625,7 +705,7 @@ class OnnxStub:
 
         ops = self.handler.operators()  # 图中所有算子（节点）
 
-        ctx = Context()
+        ctx = Context(not self.disable_check)
 
         for op in ops:
             ty, name = ctx.name_op(op)
@@ -634,11 +714,11 @@ class OnnxStub:
                 for it in op.inputs()
             ]
             outputs = [
-                ctx.push_output("{}_{}".format(name, i), it)
+                ctx.push_output(f"{name}_{i}_{it.guid()}", it)
                 for (i, it) in enumerate(op.outputs())
             ]
-            if ty == backend.OpType.Conv:
-                ph, pw, dh, dw, sh, sw = backend.conv_attrs_of(op)
+            if ty == backend.OpType.Conv or ty == backend.OpType.ConvNHWC:
+                ph, pw, sh, sw, dh, dw  = backend.conv_attrs_of(op)
                 ctx.push_node(
                     make_node(
                         ty.name,
@@ -651,7 +731,7 @@ class OnnxStub:
                         group=op.inputs()[0].shape()[1] // op.inputs()[1].shape()[1],
                     )
                 )
-            elif ty == backend.OpType.ConvTrans:
+            elif ty == backend.OpType.ConvTrans or ty == backend.OpType.ConvTransNHWC:
                 ph, pw, sh, sw, dh, dw, oph, opw = backend.conv_trans_attrs_of(op)
                 ctx.push_node(
                     make_node(
@@ -729,7 +809,8 @@ class OnnxStub:
             ]:
                 ctx.push_node(make_node(ty.name, inputs, outputs, name))
             elif ty == backend.OpType.Flatten:
-                raise Exception("TODO")
+                ctx.push_node(make_node(ty.name, inputs,
+                              outputs, axis=backend.flatten_axis_of(op)))
             elif ty == backend.OpType.Transpose:
                 perm = backend.transpose_permute_of(op)
                 ctx.push_node(make_node(ty.name, inputs, outputs, name, perm=perm))
@@ -744,7 +825,8 @@ class OnnxStub:
                         shape,
                     )
                 )
-                ctx.push_node(make_node(ty.name, inputs, outputs, name))
+                ctx.push_node(make_node(ty.name, inputs,
+                              outputs, name, allowzero=0))
             elif ty == backend.OpType.Concat:
                 axis = backend.concat_axis_of(op)
                 ctx.push_node(make_node(ty.name, inputs, outputs, name, axis=axis))
@@ -812,6 +894,62 @@ class OnnxStub:
                         ctx.push_data_input(name, "max", TensorProto.FLOAT, [], [])
                     )
                 ctx.push_node(make_node(ty.name, inputs, outputs, name))
+            elif ty == backend.OpType.Any:
+                kernel_name = backend.any_kernelName_of(op)
+                normal_op = kernel_name != 'Reduce3x3Offset_hint'
+                ctx.push_node(
+                    make_node(
+                        ty.name if normal_op else 'Reduce3x3OffsetPlugin',
+                        inputs,
+                        outputs,
+                        name,
+                        kernelName=kernel_name,
+                        domain="nnet" if normal_op else None,
+                    )
+                )
+            elif ty in [backend.OpType.ConvTransNHWC, backend.OpType.GBMM,
+                        backend.OpType.G2BMM]:
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                    )
+                )
+            elif ty == backend.OpType.Conv2dReduce:
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                    )
+                )
+            elif ty == backend.OpType.Conv2dReduceTranspose:
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                    )
+                )
+            elif ty == backend.OpType.MemBound:
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                        expr=backend.membound_expr_of(op),
+                        hash=str(backend.membound_hash_of(op)),
+                    )
+                )
             else:
                 raise Exception("Unsupported OpType", ty)
 
@@ -828,7 +966,7 @@ class OnnxStub:
 
 
 def from_onnx(model: ModelProto, runtime):
-    stub = OnnxStub(model, runtime)
+    stub = OnnxStub.from_onnx(model, runtime)
     return stub.inputs, stub.outputs, stub.handler
 
 
@@ -889,3 +1027,9 @@ def _parse_data(tensor: TensorProto) -> List[Any]:
 
 def _take_shape_dim(shape: TensorShapeProto) -> List[int]:
     return [(d.dim_value if d.dim_value > 0 else 1) for d in shape.dim]
+
+
+def save_onnx(opt_g, filename: str):
+    stub = OnnxStub.from_graph(opt_g)
+    with open(filename, "wb") as f:
+        f.write(stub.to_onnx("optimized").SerializeToString())

pyinfinitensor/src/pyinfinitensor/tensorrt_backend.py

@@ -0,0 +1,18 @@
+import subprocess
+import re
+import os
+from .onnx import save_onnx
+
+
+def get_trt_time(g):
+    onnx_filename = '/tmp/tmp.onnx'
+    save_onnx(g, onnx_filename)
+    plugin_path = os.environ['TRT_PLUGIN']
+    # LD_LIBRARY_PATH=$TRT_PLUGIN:$LD_LIBRARY_PATH trtexec --noTF32 --onnx=/home/zly/InfiniTensor_merge/build/opt_resnet.bs16.onnx --plugins=$TRT_PLUGIN/libnvinfer_plugin.so.8.2.0
+    res = subprocess.run(
+        f'trtexec --noTF32 --onnx={onnx_filename} --plugins={plugin_path}/libnvinfer_plugin.so.8.2.0'.split(' '), capture_output=True)
+    p = re.compile('GPU Compute Time.*mean = ([0-9.]+) ms')
+    output = res.stdout.decode('utf-8')
+    # err = res.stderr.decode('utf-8')
+    # print(output, '\n'*5, err)
+    return float(p.search(output).group(1))

pyinfinitensor/tests/test_onnx.py

@@ -40,9 +40,9 @@ class TestStringMethods(unittest.TestCase):
                         file=model_file, size=os.path.getsize(model_file) / 1024 / 1024
                     )
                 )
-                model = OnnxStub(onnx.load(model_file), backend.cpu_runtime()).to_onnx(
-                    "new"
-                )
+                model = OnnxStub.from_onnx(
+                    onnx.load(model_file), backend.cpu_runtime()
+                ).to_onnx("new")
                 model = infer_shapes(model)
 
     def test_tensor(self):
@@ -304,16 +304,16 @@ class TestStringMethods(unittest.TestCase):
 
     def test_frontend(self):
         handler = backend.GraphHandler(backend.cpu_runtime())
-        a = handler.tensor([1, 2, 3], 12)
-        b = handler.tensor([1, 2, 3], 12)
-        c = handler.tensor([1, 2, 3], 12)
-        d = handler.tensor([1, 2, 3], 12)
-        e = handler.tensor([1, 2, 3], 12)
+        a = handler.tensor([1, 2, 3], 12, backend.TensorType.Input)
+        b = handler.tensor([1, 2, 3], 12, backend.TensorType.Input)
+        c = handler.tensor([1, 2, 3], 12, backend.TensorType.Input)
+        d = handler.tensor([1, 2, 3], 12, backend.TensorType.Input)
+        e = handler.tensor([1, 2, 3], 12, backend.TensorType.Input)
 
         x = handler.add(
             handler.add(handler.add(handler.add(a, b, None), c, None), d, None), e, None
         )
-        y = handler.tensor([3, 2, 1], 12)
+        y = handler.tensor([3, 2, 1], 12, backend.TensorType.Other)
         handler.reshape(x, y, [3, 2, 1])
 
 

python/cpp_plugin/gen_ansor_so.py

@@ -1,106 +1,131 @@
-import re
+import os
+import sys
+import json
+from contextlib import redirect_stdout
+import time
+import logging
 
 import numpy as np
 import tvm
 from tvm import te, tir, auto_scheduler, topi
-import os
-import json
-import logging
 
 USE_CACHE = True
+logging.basicConfig()
 logger = logging.getLogger('InfiniTensor')
-logger.setLevel(logging.DEBUG)
+logger.setLevel(logging.INFO)
 
 
 def gen_ansor_so(input_tensors, input_dtypes, output_tensor, output_dtype,
                  tvm_code, func_name, nnet_expression: str,
-                 nnet_simplified_expression: str, hash_code=None):
+                 nnet_simplified_expression: str, hash_code: str = None):
     assert len(input_tensors) == len(input_dtypes)
-    
-    logging.debug(f'Work on hash {hash_code}')
+
+    logger.debug(f'Work on hash {hash_code}')
     dir_name = os.path.join(".cache", "generated_kernels", str(hash_code))
-    
+
     if not os.path.exists(dir_name):
         os.makedirs(dir_name)
-    
+
     so_fn = os.path.join(dir_name, f"{func_name}.so")
     config_fn = os.path.join(dir_name, "config_so.json")
-    
-    print("Generating Ansor op: ")
-    print(tvm_code)
-    
-    print("Input shape: ")
-    print(input_tensors)
-    print("Output shape: ")
-    print(output_tensor)
-    
+    desc_fn = os.path.join(dir_name, "desc.txt")
+    log_fn = os.path.join(dir_name, f"ansor_{func_name}_log.json")
+    out_fn = os.path.join(dir_name, "out.txt")
+
+    logger.debug(f"Generating Ansor op: {tvm_code}")
+    logger.debug(f"Input shape: {input_tensors}")
+    logger.debug(f"Output shape: {output_tensor}")
+
     if USE_CACHE and hash_code is not None:
         if os.path.exists(dir_name) and \
-            os.path.exists(so_fn) and \
-            os.path.exists(config_fn):
+                os.path.exists(so_fn) and \
+                os.path.exists(config_fn):
             print(f"Use cache in {dir_name}")
             with open(config_fn, "r") as config_fin:
                 config = json.loads(config_fin.read().strip())
                 conv_time = config["conv_time"]
 
-            logger.debug(f'Find tuning log for {hash_code}')
+            logger.info(f'Find tuning log for {hash_code} in {so_fn}')
             return so_fn, conv_time
-    
+    logger.info(f"TVM Tuning kernel with hash {hash_code}. See {out_fn}")
+
+    time_start = time.perf_counter()
+    # Print descriptions of the task
+    if USE_CACHE and hash_code is not None:
+        with redirect_stdout(open(desc_fn, "w")):
+            print("====NNET tensor expression====")
+            print(nnet_expression+"\n")
+            print("====NNET simplified tensor expression====")
+            print(nnet_simplified_expression+"\n")
+            print("====TVM compute====")
+            print(tvm_code+"\n")
+            print("Input shape: ", input_tensors)
+            print("Output shape: ", output_tensor)
+
     @auto_scheduler.register_workload(func_name)
     def compute():
         _locals = locals()
-        exec(tvm_code, {'tvm': tvm, 'te': te, 'tir': tir, 'topi': topi}, _locals)
+        exec(tvm_code, {'tvm': tvm, 'te': te,
+             'tir': tir, 'topi': topi}, _locals)
         return _locals['ret']
-    
+
     target = tvm.target.Target("cuda")
 
     task = auto_scheduler.SearchTask(func=func_name, args=(), target=target)
 
-    # Inspect the computational graph
-    print("Computational DAG:")
-    print(task.compute_dag)
+    with redirect_stdout(open(out_fn, 'w')):
+        # Inspect the computational graph
+        print("Computational DAG:")
+        print(task.compute_dag)
 
-    log_file = f"ansor_{func_name}_log.json"
-    measure_ctx = auto_scheduler.LocalRPCMeasureContext(min_repeat_ms=300)
-    tune_option = auto_scheduler.TuningOptions(
-        num_measure_trials=10,
-        runner=measure_ctx.runner,
-        measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
-        verbose=2,
-    )
+        measure_ctx = auto_scheduler.LocalRPCMeasureContext(min_repeat_ms=300)
+        tune_option = auto_scheduler.TuningOptions(
+            num_measure_trials=10,
+            runner=measure_ctx.runner,
+            measure_callbacks=[auto_scheduler.RecordToFile(log_fn)],
+            verbose=2,
+        )
 
-    # Run auto-tuning (search)
-    task.tune(tune_option)
-    # Apply the best schedule
-    sch, args = task.apply_best(log_file)
+        # Run auto-tuning (search)
+        task.tune(tune_option)
+        # Apply the best schedule
+        sch, args = task.apply_best(log_fn)
+
+        # Kill the measurement process
+        del measure_ctx
+
+        func = tvm.build(sch, args, target, name=func_name)
+        func.export_library(so_fn)
+
+        ctx = tvm.cuda(0)
+        input_a = []
+        for i, (shape, dtype) in enumerate(zip(input_tensors, input_dtypes)):
+            a_np = np.random.uniform(size=shape).astype(dtype)
+            input_a.append(tvm.nd.array(a_np, ctx))
+        a_out = tvm.nd.array(np.zeros(output_tensor, dtype=output_dtype), ctx)
+        func(a_out, *input_a)
+        evaluator = func.time_evaluator(func.entry_name, ctx, number=100)
+        conv_time = evaluator(a_out, *input_a).mean * 1e3
+
+    time_end = time.perf_counter()
 
-    # Kill the measurement process
-    del measure_ctx
-    
-    func = tvm.build(sch, args, target, name=func_name)
-    func.export_library(so_fn)
-    
-    ctx = tvm.cuda(0)
-    input_a = []
-    for i, (shape, dtype) in enumerate(zip(input_tensors, input_dtypes)):
-        a_np = np.random.uniform(size=shape).astype(dtype)
-        input_a.append(tvm.nd.array(a_np, ctx))
-    a_out = tvm.nd.array(np.zeros(output_tensor, dtype=output_dtype), ctx)
-    func(a_out, *input_a)
-    evaluator = func.time_evaluator(func.entry_name, ctx, number=100)
-    conv_time = evaluator(a_out, *input_a).mean * 1e3
-    
-    print("====NNET tensor expression====")
-    print(nnet_expression+"\n")
-    print("====NNET simplified tensor expression====")
-    print(nnet_simplified_expression+"\n")
-    print("====Time====")
-    print(conv_time)
-    
     if USE_CACHE and hash_code is not None:
         with open(config_fn, "w") as config_fout:
             config_fout.write(json.dumps({
                 "conv_time": conv_time,
+                "tuning_time": time_end - time_start,
+                "timestamp": time.strftime("%Y-%m-%d %H:%M:%S", time.gmtime()),
             }, ensure_ascii=False, indent=2))
-    
+
     return so_fn, conv_time
+
+# Read arguments from pipe, which is redirected to stdin.
+# Write generated library path to pipe.
+
+
+def pipe_gen(fd: int):
+    args = json.load(sys.stdin)  # read from pipe
+    # print(args, f'fd={fd}')
+    ret = gen_ansor_so(**args)
+    with os.fdopen(fd, 'w') as f:
+        print(ret[0], file=f, end='')  # write to pipe

python/infinitensor/__init__.py

@@ -0,0 +1,7 @@
+import backend
+from backend import *
+import sys
+
+sys.path.extend(__path__)
+
+print("import backend: {}".format(backend))

python/infinitensor/if_onnx.py

@@ -0,0 +1,941 @@
+import backend
+import onnx
+from onnx import (
+    ModelProto,
+    TensorProto,
+    NodeProto,
+    AttributeProto,
+    TensorShapeProto,
+    ValueInfoProto,
+)
+from onnx.helper import (
+    make_node,
+    make_tensor_value_info,
+    make_tensor,
+    make_graph,
+    make_model,
+)
+from onnx.checker import (
+    check_graph,
+    check_model,
+    check_node,
+    check_value_info,
+    check_tensor,
+)
+from onnx.shape_inference import infer_shapes
+from onnx.numpy_helper import to_array
+from typing import Dict, List, Any, Tuple, Sequence, Union, Optional
+from functools import reduce
+
+
+class OnnxStub:
+    """
+    The Onnx model imported into infinitensor.
+    It can be generated from an Onnx model object.
+    """
+
+    # inputs: Dict[str, backend.Tensor] = {}
+    # outputs: Dict[str, backend.Tensor] = {}
+    initializer: Dict[int, TensorProto] = {}
+    # handler: backend.GraphHandler
+
+    # def __init__(self, model: ModelProto, runtime):
+    #     model = infer_shapes(model)
+    #     self.handler = backend.GraphHandler(runtime)
+
+    #     tensors: Dict[str, backend.Tensor] = dict()
+    #     data: Dict[str, TensorProto] = dict()
+
+    #     for input in model.graph.input:
+    #         dims = _take_shape_dim(input.type.tensor_type.shape)
+    #         tensors[input.name] = self.handler.tensor(
+    #             dims, input.type.tensor_type.elem_type
+    #         )
+
+    #     for output in model.graph.output:
+    #         dims = _take_shape_dim(output.type.tensor_type.shape)
+    #         tensors[output.name] = self.handler.tensor(
+    #             dims, output.type.tensor_type.elem_type
+    #         )
+
+    #     for initializer in model.graph.initializer:
+    #         dims = [d for d in initializer.dims]
+    #         tensors[initializer.name] = self.handler.tensor(dims, initializer.data_type)
+    #         data[initializer.name] = initializer
+
+    #     for node in model.graph.node:
+    #         if node.op_type == "Conv":
+    #             attributes = _parse_attribute(
+    #                 node,
+    #                 {
+    #                     "dilations": [1, 1],
+    #                     "pads": [0, 0, 0, 0],
+    #                     "strides": [1, 1],
+    #                 },
+    #             )
+    #             (d, p, s) = (
+    #                 attributes[name] for name in ["dilations", "pads", "strides"]
+    #             )
+    #             if p[0] != p[2] or p[1] != p[3]:
+    #                 adapt = "{}-adapt".format(node.output[0])
+    #                 tensors[adapt] = self.handler.pad(
+    #                     tensors[node.input[0]], None, p, [-2, -1]
+    #                 )
+    #                 p = [0, 0, 0, 0]
+    #             else:
+    #                 adapt = node.input[0]
+
+    #             if len(node.input) > 2:
+    #                 bias = "{}-bias".format(node.output[0])
+    #                 reshape = "{}-reshape".format(node.output[0])
+    #                 tensors[bias] = self.handler.conv(
+    #                     tensors[adapt],
+    #                     tensors[node.input[1]],
+    #                     None,
+    #                     p[0],
+    #                     p[1],
+    #                     s[0],
+    #                     s[1],
+    #                     d[0],
+    #                     d[1],
+    #                 )
+    #                 tensors[reshape] = self.handler.reshape(
+    #                     tensors[node.input[2]],
+    #                     None,
+    #                     [
+    #                         1,
+    #                         reduce(
+    #                             lambda acc, x: acc * x,
+    #                             _search_shape(model, node.input[2]),
+    #                         ),
+    #                         1,
+    #                         1,
+    #                     ],
+    #                 )
+    #                 tensors[node.output[0]] = self.handler.add(
+    #                     tensors[bias],
+    #                     tensors[reshape],
+    #                     tensors.get(node.output[0]),
+    #                 )
+    #             else:
+    #                 tensors[node.output[0]] = self.handler.conv(
+    #                     tensors[adapt],
+    #                     tensors[node.input[1]],
+    #                     tensors.get(node.output[0]),
+    #                     p[0],
+    #                     p[1],
+    #                     s[0],
+    #                     s[1],
+    #                     d[0],
+    #                     d[1],
+    #                 )
+    #         elif node.op_type == "ConvTranspose":
+    #             attributes = _parse_attribute(
+    #                 node,
+    #                 {
+    #                     "dilations": [1, 1],
+    #                     "pads": [0, 0],
+    #                     "strides": [1, 1],
+    #                     "output_padding": [0, 0],
+    #                 },
+    #             )
+    #             (d, p, s, op) = (
+    #                 attributes[name]
+    #                 for name in ["dilations", "pads", "strides", "output_padding"]
+    #             )
+    #             tensors[node.output[0]] = self.handler.convTransposed2d(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #                 p[0],
+    #                 p[1],
+    #                 s[0],
+    #                 s[1],
+    #                 d[0],
+    #                 d[1],
+    #                 op[0],
+    #                 op[1],
+    #             )
+    #         elif node.op_type == "MatMul":
+    #             tensors[node.output[0]] = self.handler.matmul(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #                 False,
+    #                 False,
+    #                 None,
+    #                 backend.ActType.Linear,
+    #             )
+    #         elif node.op_type == "Gemm":
+    #             attributes = _parse_attribute(
+    #                 node, {"alpha": 1.0, "beta": 1.0, "transA": 0, "transB": 0}
+    #             )
+    #             (alpha, beta, transA, transB) = (
+    #                 attributes[name] for name in ["alpha", "beta", "transA", "transB"]
+    #             )
+    #             # FIXME unsupport attributes: `alpha` `beta`
+    #             assert alpha == 1.0
+    #             assert beta == 1.0
+    #             tensors[node.output[0]] = self.handler.matmul(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #                 transA == 1,
+    #                 transB == 1,
+    #                 tensors[node.input[2]] if len(node.input) > 2 else None,
+    #                 backend.ActType.Linear,
+    #             )
+    #         elif node.op_type == "BatchNormalization":
+    #             (input, mean, var, scale, bias) = (
+    #                 tensors[node.input[i]] for i in [0, 3, 4, 1, 2]
+    #             )
+    #             output = tensors.get(node.output[0])
+    #             attributes = _parse_attribute(
+    #                 node, {"momentum": 0.9, "epsilon": 1e-05, "training_mode": 0}
+    #             )
+    #             (momentum, eps, training) = (
+    #                 attributes[name]
+    #                 for name in ["momentum", "epsilon", "training_mode"]
+    #             )
+    #             tensors[node.output[0]] = self.handler.batchNorm(
+    #                 input, output, mean, var, scale, bias, momentum, eps, training != 0
+    #             )
+    #         elif node.op_type == "MaxPool":
+    #             attributes = _parse_attribute(
+    #                 node,
+    #                 {
+    #                     "kernel_shape": None,
+    #                     "dilations": [1, 1],
+    #                     "pads": [0, 0, 0, 0],
+    #                     "strides": [1, 1],
+    #                 },
+    #             )
+    #             (k, d, p, s) = (
+    #                 attributes[name]
+    #                 for name in ["kernel_shape", "dilations", "pads", "strides"]
+    #             )
+    #             if p[0] != p[2] or p[1] != p[3]:
+    #                 adapt = "{}-adapt".format(node.output[0])
+    #                 tensors[adapt] = self.handler.pad(
+    #                     tensors.get(node.input[0]), None, p, [-2, -1]
+    #                 )
+    #                 tensors[node.output[0]] = self.handler.maxPool(
+    #                     tensors[adapt],
+    #                     tensors.get(node.output[0]),
+    #                     k[0],
+    #                     k[1],
+    #                     d[0],
+    #                     d[1],
+    #                     0,
+    #                     0,
+    #                     s[0],
+    #                     s[1],
+    #                 )
+    #             else:
+    #                 tensors[node.output[0]] = self.handler.maxPool(
+    #                     tensors[node.input[0]],
+    #                     tensors.get(node.output[0]),
+    #                     k[0],
+    #                     k[1],
+    #                     d[0],
+    #                     d[1],
+    #                     p[0],
+    #                     p[1],
+    #                     s[0],
+    #                     s[1],
+    #                 )
+    #         elif node.op_type == "AveragePool":
+    #             attributes = _parse_attribute(
+    #                 node,
+    #                 {
+    #                     "kernel_shape": None,
+    #                     "pads": [0, 0, 0, 0],
+    #                     "strides": [1, 1],
+    #                 },
+    #             )
+    #             (k, p, s) = (
+    #                 attributes[name] for name in ["kernel_shape", "pads", "strides"]
+    #             )
+    #             if p[0] != p[2] or p[1] != p[3]:
+    #                 adapt = "{}-adapt".format(node.output[0])
+    #                 tensors[adapt] = self.handler.pad(
+    #                     tensors.get(node.input[0]), None, p, [-2, -1]
+    #                 )
+    #                 tensors[node.output[0]] = self.handler.avgPool(
+    #                     tensors[adapt],
+    #                     tensors.get(node.output[0]),
+    #                     k[0],
+    #                     k[1],
+    #                     1,
+    #                     1,
+    #                     0,
+    #                     0,
+    #                     s[0],
+    #                     s[1],
+    #                 )
+    #             else:
+    #                 tensors[node.output[0]] = self.handler.avgPool(
+    #                     tensors[node.input[0]],
+    #                     tensors.get(node.output[0]),
+    #                     k[0],
+    #                     k[1],
+    #                     1,
+    #                     1,
+    #                     p[0],
+    #                     p[1],
+    #                     s[0],
+    #                     s[1],
+    #                 )
+    #         elif node.op_type == "GlobalAveragePool":
+    #             [_, _, h, w] = _search_shape(model, node.input[0])
+    #             tensors[node.output[0]] = self.handler.avgPool(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 h,
+    #                 w,
+    #                 1,
+    #                 1,
+    #                 0,
+    #                 0,
+    #                 1,
+    #                 1,
+    #             )
+    #         elif node.op_type == "Add":
+    #             tensors[node.output[0]] = self.handler.add(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Sub":
+    #             tensors[node.output[0]] = self.handler.sub(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Mul":
+    #             tensors[node.output[0]] = self.handler.mul(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Div":
+    #             tensors[node.output[0]] = self.handler.div(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Pow":
+    #             tensors[node.output[0]] = self.handler.pow(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Relu":
+    #             tensors[node.output[0]] = self.handler.relu(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Sigmoid":
+    #             tensors[node.output[0]] = self.handler.sigmoid(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Tanh":
+    #             tensors[node.output[0]] = self.handler.tanh(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Softmax":
+    #             tensors[node.output[0]] = self.handler.softmax(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Abs":
+    #             tensors[node.output[0]] = self.handler.abs(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Shape":
+    #             tensors[node.output[0]] = self.handler.shape(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Identity":
+    #             tensors[node.output[0]] = self.handler.identity(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Flatten":
+    #             # FIXME axis must be 1
+    #             axis = next(
+    #                 (attr.i for attr in node.attribute if attr.name == "axis"), None
+    #             )
+    #             assert axis == None or axis == 1
+    #             tensors[node.output[0]] = self.handler.flatten(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "PRelu":
+    #             tensors[node.output[0]] = self.handler.pRelu(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #             )
+    #         elif node.op_type == "Clip":
+    #             tensors[node.output[0]] = self.handler.clip(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 next(_parse_data(data[node.input[1]]).__iter__(), None)
+    #                 if len(node.input) > 1
+    #                 else None,
+    #                 next(_parse_data(data[node.input[2]]).__iter__(), None)
+    #                 if len(node.input) > 2
+    #                 else None,
+    #             )
+    #         elif node.op_type == "Transpose":
+    #             perm = next(
+    #                 (attr.ints for attr in node.attribute if attr.name == "perm"), None
+    #             )
+    #             tensors[node.output[0]] = self.handler.transpose(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 perm,
+    #             )
+    #         elif node.op_type == "Reshape":
+    #             dims = _search_shape(model, node.input[0])
+    #             size = reduce(lambda acc, x: acc * x, dims)
+    #             input_shape = _parse_data(data[node.input[1]])
+    #             for i, x in enumerate(input_shape):
+    #                 if x == 0:
+    #                     input_shape[i] = dims[i]
+    #             temp = reduce(lambda acc, x: acc * x, input_shape, 1)
+    #             if temp < 0:
+    #                 input_shape[input_shape.index(-1)] = size // -temp
+    #             tensors[node.output[0]] = self.handler.reshape(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 input_shape,
+    #             )
+    #         elif node.op_type == "Squeeze":
+    #             input_shape = _search_shape(model, node.input[0])
+    #             axes = set(
+    #                 [int(i) for i in data[node.input[1]].int64_data]
+    #                 if len(node.input) > 1
+    #                 else _parse_attribute(node, {"axes": None})["axes"]
+    #             )
+    #             assert all(input_shape[d] == 1 for d in axes)
+    #             output_shape = []
+    #             for i, x in enumerate(input_shape):
+    #                 if i not in axes:
+    #                     output_shape.append(x)
+    #             tensors[node.output[0]] = self.handler.reshape(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 output_shape,
+    #             )
+    #         elif node.op_type == "Unsqueeze":
+    #             input_shape = _search_shape(model, node.input[0])
+    #             axes = (
+    #                 [int(i) for i in data[node.input[1]].int64_data]
+    #                 if len(node.input) > 1
+    #                 else _parse_attribute(node, {"axes": None})["axes"]
+    #             )
+    #             for i in axes:
+    #                 input_shape.insert(i, 1)
+    #             tensors[node.output[0]] = self.handler.reshape(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 input_shape,
+    #             )
+    #         elif node.op_type == "Concat":
+    #             tensors[node.output[0]] = self.handler.concat(
+    #                 [tensors[name] for name in node.input],
+    #                 tensors.get(node.output[0]),
+    #                 next((attr.i for attr in node.attribute if attr.name == "axis")),
+    #             )
+    #         elif node.op_type == "Split":
+    #             for name, tensor in zip(
+    #                 node.output,
+    #                 self.handler.split(
+    #                     tensors[node.input[0]],
+    #                     None,
+    #                     next(
+    #                         (attr.i for attr in node.attribute if attr.name == "axis"),
+    #                         0,
+    #                     ),
+    #                     len(node.output),
+    #                 ),
+    #             ):
+    #                 tensors[name] = tensor
+    #         elif node.op_type == "Gather":
+    #             tensors[node.output[0]] = self.handler.gather(
+    #                 tensors[node.input[0]],
+    #                 tensors[node.input[1]],
+    #                 tensors.get(node.output[0]),
+    #                 next((attr.i for attr in node.attribute if attr.name == "axis")),
+    #             )
+    #         elif node.op_type == "ReduceMean":
+    #             tensors[node.output[0]] = self.handler.reduce_mean(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 next(
+    #                     (attr.ints for attr in node.attribute if attr.name == "axes"),
+    #                     None,
+    #                 ),
+    #                 next((attr.i for attr in node.attribute if attr.name == "keepdims"))
+    #                 != 0,
+    #             )
+    #         elif node.op_type == "Slice":
+    #             tensors[node.output[0]] = self.handler.slice(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 _parse_data(data[node.input[1]]),
+    #                 _parse_data(data[node.input[2]]),
+    #                 _parse_data(data[node.input[3]]) if len(node.input) > 3 else None,
+    #                 _parse_data(data[node.input[4]]) if len(node.input) > 4 else None,
+    #             )
+    #         elif node.op_type == "Pad":
+    #             tensors[node.output[0]] = self.handler.pad(
+    #                 tensors[node.input[0]],
+    #                 tensors.get(node.output[0]),
+    #                 _parse_data(data[node.input[1]]),
+    #                 _parse_data(data[node.input[3]]) if len(node.input) > 3 else None,
+    #             )
+    #         elif node.op_type == "Dropout":
+    #             for name, tensor in zip(
+    #                 node.output,
+    #                 self.handler.dropout(
+    #                     tensors[node.input[0]],
+    #                     tensors.get(node.output[0]),
+    #                     tensors.get(node.output[1]) if len(node.output) > 1 else None,
+    #                     _parse_data(data[node.input[1]])[0]
+    #                     if len(node.input) > 1
+    #                     else 0.5,
+    #                     _parse_data(data[node.input[2]])[0]
+    #                     if len(node.input) > 2
+    #                     else False,
+    #                 ),
+    #             ):
+    #                 tensors[name] = tensor
+    #         else:
+    #             raise Exception('Unsupported operator "{}"'.format(node.op_type))
+
+    #     self.handler.data_malloc()
+
+    #     for name, obj in tensors.items():
+    #         tensor = data.get(name)
+    #         if tensor == None:
+    #             if any(input.name == name for input in model.graph.input):
+    #                 self.inputs[name] = obj
+    #         else:
+    #             self.initializer[obj.fuid()] = tensor
+    #             if tensor.data_type == TensorProto.INT32:
+    #                 obj.copyin_int32(_parse_data(tensor))
+    #             elif tensor.data_type == TensorProto.INT64:
+    #                 obj.copyin_int64(_parse_data(tensor))
+    #             elif tensor.data_type == TensorProto.FLOAT:
+    #                 obj.copyin_float(_parse_data(tensor))
+    #             else:
+    #                 assert False, "Unsupported Tensor Type: {}".format(tensor.data_type)
+
+    #     for output in model.graph.output:
+    #         self.outputs[output.name] = tensors[output.name]
+
+    def to_onnx(self, g: backend.Graph, path: str, name: str = 'my_onnx') -> ModelProto:
+        class Context:
+            # saves object names, including tensors and operators
+            names: Dict[Union[backend.Tensor, backend.Operator], str] = dict()
+            # counts the occurrence times of each operator for naming
+            count_op: Dict[backend.OpType, int] = dict()
+            # counts input and output tensors for naming
+            count_in, count_out = 0, 0
+            # saves nodes (operators)
+            nodes: List[NodeProto] = []
+            # saves global input tensors
+            inputs: List[ValueInfoProto] = []
+            # saves global output tensors
+            outputs: List[ValueInfoProto] = []
+            # saves global input tensors
+            initializers: List[TensorProto] = []
+
+            def name_op(self, op: backend.Operator) -> Tuple[backend.OpType, str]:
+                ty = op.op_type()
+                name = "{}_{}".format(ty.name, op.guid())
+                self.names[op] = name
+                self.count_op[ty] = self.count_op.get(ty, 0) + 1
+                return ty, name
+
+            def push_output(self, name: str, tensor: backend.Tensor) -> str:
+                self.names[tensor] = name
+                if not tensor.has_target():
+                    shape = tensor.shape()
+                    dtype = backend.tensor_dtype(tensor)
+                    value_info = make_tensor_value_info(name, dtype, shape)
+                    check_value_info(value_info)
+                    self.outputs.append(value_info)
+                return name
+
+            def push_input(
+                self, tensor: backend.Tensor, init: Optional[TensorProto]
+            ) -> str:
+                name = self.names.get(tensor)
+                # means that this input is a global input
+                if name is None:
+                    self.count_in += 1
+                    name = "input_{}".format(tensor.guid())
+                    self.names[tensor] = name
+                    if init != None:
+                        init.name = name
+                        self.initializers.append(init)
+                    else:
+                        shape = tensor.shape()
+                        dtype = backend.tensor_dtype(tensor)
+                        value_info = make_tensor_value_info(name, dtype, shape)
+                        check_value_info(value_info)
+                        self.inputs.append(value_info)
+                return name
+
+            def push_data_input(
+                self,
+                node_name: str,
+                attr_name: str,
+                elem_type: int,
+                shape: Sequence[int],
+                vals: Any,
+            ) -> str:
+                name = "{}_{}".format(node_name, attr_name)
+                tensor = make_tensor(name, elem_type, shape, vals)
+                check_tensor(tensor)
+                self.initializers.append(tensor)
+                return name
+
+            def push_node(self, node: NodeProto) -> None:
+                # check_node(node)
+                self.nodes.append(node)
+
+            def build(self, name: str) -> ModelProto:
+                graph = make_graph(
+                    self.nodes, name, self.inputs, self.outputs, self.initializers
+                )
+                # check_graph(graph)
+
+                model = make_model(graph)
+                # check_model(model)
+
+                return model
+
+        # 拓扑排序
+        if not g.topo_sort():
+            raise Exception("Sorting fails")
+
+        ops = g.operators()  # 图中所有算子（节点）
+
+        ctx = Context()
+
+        for op in ops:
+            ty, name = ctx.name_op(op)
+            inputs = [
+                ctx.push_input(it, self.initializer.get(it.fuid()))
+                for it in op.inputs()
+            ]
+            outputs = [
+                ctx.push_output("{}_{}_{}".format(
+                    name, i, tensor.guid()), tensor)
+                for (i, tensor) in enumerate(op.outputs())
+            ]
+            if ty == backend.OpType.Conv:
+                ph, pw, dh, dw, sh, sw = backend.conv_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        pads=[ph, pw, ph, pw],
+                        strides=[sh, sw],
+                        dilations=[dh, dw],
+                        group=op.inputs()[0].shape()[
+                            1] // op.inputs()[1].shape()[1],
+                    )
+                )
+            elif ty == backend.OpType.ConvTrans:
+                ph, pw, sh, sw, dh, dw, oph, opw = backend.conv_trans_attrs_of(
+                    op)
+                ctx.push_node(
+                    make_node(
+                        "ConvTranspose",
+                        inputs,
+                        outputs,
+                        name,
+                        pads=[ph, pw],
+                        strides=[sh, sw],
+                        dilations=[dh, dw],
+                        output_padding=[oph, opw],
+                    )
+                )
+            elif ty == backend.OpType.ConvTransNHWC:
+                # ph, pw, sh, sw, dh, dw, oph, opw = backend.conv_trans_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        "ConvTranspose",
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                        # pads=[ph, pw],
+                        # strides=[sh, sw],
+                        # dilations=[dh, dw],
+                        # output_padding=[oph, opw],
+                    )
+                )
+            elif ty == backend.OpType.MemBound:
+                # ph, pw, sh, sw, dh, dw, oph, opw = backend.conv_trans_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        "Membound",
+                        inputs,
+                        outputs,
+                        name,
+                        domain="nnet",
+                        # pads=[ph, pw],
+                        # strides=[sh, sw],
+                        # dilations=[dh, dw],
+                        # output_padding=[oph, opw],
+                    )
+                )
+            elif ty == backend.OpType.Matmul:
+                # transA, transB = backend.matmul_attrs_of(op)
+                # HACK: recover this
+                transA, transB = False, False
+                ctx.push_node(
+                    make_node(
+                        "Gemm", inputs, outputs, name, transA=transA, transB=transB
+                    )
+                )
+            elif ty == backend.OpType.BatchNorm:
+                inputs = [inputs[i] for i in [0, 3, 4, 1, 2]]
+                momentum, eps, training = backend.batch_norm_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        "BatchNormalization",
+                        inputs,
+                        outputs,
+                        name,
+                        epsilon=eps,
+                        momentum=momentum,
+                        training_mode=training,
+                    )
+                )
+            elif ty == backend.OpType.MaxPool:
+                kh, kw, dh, dw, ph, pw, sh, sw = backend.pool_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        kernel_shape=[kh, kw],
+                        pads=[ph, pw, ph, pw],
+                        dilations=[dh, dw],
+                        strides=[sh, sw],
+                    )
+                )
+            elif ty == backend.OpType.AvgPool:
+                kh, kw, dh, dw, ph, pw, sh, sw = backend.pool_attrs_of(op)
+                ctx.push_node(
+                    make_node(
+                        "AveragePool",
+                        inputs,
+                        outputs,
+                        name,
+                        kernel_shape=[kh, kw],
+                        pads=[ph, pw, ph, pw],
+                        strides=[sh, sw],
+                    )
+                )
+            elif ty in [
+                backend.OpType.Add,
+                backend.OpType.Sub,
+                backend.OpType.Mul,
+                backend.OpType.Div,
+                backend.OpType.Pow,
+                backend.OpType.Relu,
+                backend.OpType.Sigmoid,
+                backend.OpType.Tanh,
+                backend.OpType.Softmax,
+                backend.OpType.Abs,
+                backend.OpType.Identity,
+                backend.OpType.PRelu,
+            ]:
+                ctx.push_node(make_node(ty.name, inputs, outputs, name))
+            elif ty == backend.OpType.Flatten:
+                raise Exception("TODO")
+            elif ty == backend.OpType.Transpose:
+                perm = backend.transpose_permute_of(op)
+                ctx.push_node(make_node(ty.name, inputs,
+                              outputs, name, perm=perm))
+            elif ty == backend.OpType.Reshape:
+                shape = backend.reshape_shape_of(op)
+                inputs.append(
+                    ctx.push_data_input(
+                        name,
+                        "shape",
+                        TensorProto.INT64,
+                        [len(shape)],
+                        shape,
+                    )
+                )
+                ctx.push_node(make_node(ty.name, inputs, outputs, name))
+            elif ty == backend.OpType.Concat:
+                axis = backend.concat_axis_of(op)
+                ctx.push_node(make_node(ty.name, inputs,
+                              outputs, name, axis=axis))
+            elif ty == backend.OpType.Split:
+                axis = backend.split_axis_of(op)
+                num_outputs = len(outputs)
+                split = op.inputs()[0].shape()[axis] // num_outputs
+                inputs.append(
+                    ctx.push_data_input(
+                        name,
+                        "split",
+                        TensorProto.INT64,
+                        [len(outputs)],
+                        [split for _ in range(0, num_outputs)],
+                    )
+                )
+                ctx.push_node(
+                    make_node(
+                        ty.name,
+                        inputs,
+                        outputs,
+                        name,
+                        axis=axis,
+                    )
+                )
+            elif ty == backend.OpType.Gather:
+                axis = backend.gather_axis_of(op)
+                ctx.push_node(make_node(ty.name, inputs,
+                              outputs, name, axis=axis))
+            elif ty == backend.OpType.ReduceMean:
+                axes, keepdims = backend.reduce_mean_attrs_of(op)
+                inputs.append(
+                    ctx.push_data_input(
+                        name, "axes", TensorProto.INT64, [len(axes)], axes
+                    )
+                )
+                ctx.push_node(
+                    make_node(ty.name, inputs, outputs,
+                              name, keepdims=keepdims)
+                )
+            elif ty == backend.OpType.Slice:
+                raise Exception("TODO")
+            elif ty == backend.OpType.Pad:
+                pads = backend.pad_pads_of(op)
+                inputs.append(
+                    ctx.push_data_input(
+                        name, "pads", TensorProto.INT64, [len(pads)], pads
+                    )
+                )
+                ctx.push_node(make_node(ty.name, inputs, outputs, name))
+            # elif ty == backend.OpType.Clip:
+            #     min, max = backend.clip_attrs_of(op)
+            #     if min != None:
+            #         inputs.append(
+            #             ctx.push_data_input(name, "min", TensorProto.FLOAT, [], [min])
+            #         )
+            #     else:
+            #         inputs.append(
+            #             ctx.push_data_input(name, "min", TensorProto.FLOAT, [], [])
+            #         )
+            #     if max != None:
+            #         inputs.append(
+            #             ctx.push_data_input(name, "max", TensorProto.FLOAT, [], [max])
+            #         )
+            #     else:
+            #         inputs.append(
+            #             ctx.push_data_input(name, "max", TensorProto.FLOAT, [], [])
+            #         )
+            #     ctx.push_node(make_node(ty.name, inputs, outputs, name))
+            else:
+                raise Exception("Unsupported OpType", ty)
+
+        model = ctx.build(name)
+        onnx.save(model, path)
+        return model
+
+    # def init(self) -> None:
+    #     self.handler.data_malloc()
+
+    # def optimize(self) -> None:
+    #     self.handler.optimize()
+
+    # def run(self) -> None:
+    #     self.handler.run()
+
+
+# def from_onnx(model: ModelProto, runtime):
+#     stub = OnnxStub(model, runtime)
+#     return stub.inputs, stub.outputs, stub.handler
+
+
+# def _search_shape(model: ModelProto, name: str) -> List[int]:
+#     ans = (
+#         next(
+#             (
+#                 [
+#                     (d.dim_value if d.dim_value > 0 else 1)
+#                     for d in tensor.type.tensor_type.shape.dim
+#                 ]
+#                 for tensor in model.graph.value_info
+#                 if tensor.name == name
+#             ),
+#             None,
+#         )
+#         or next(
+#             (
+#                 [
+#                     (d.dim_value if d.dim_value > 0 else 1)
+#                     for d in tensor.type.tensor_type.shape.dim
+#                 ]
+#                 for tensor in model.graph.input
+#                 if tensor.name == name
+#             ),
+#             None,
+#         )
+#         or next(
+#             [int(d) for d in tensor.dims]
+#             for tensor in model.graph.initializer
+#             if tensor.name == name
+#         )
+#     )
+#     return ans
+
+
+# def _parse_attribute(node: NodeProto, attrs: Dict[str, Any] = dict()) -> Dict[str, Any]:
+#     for attr in node.attribute:
+#         if attr.name in attrs:
+#             if attr.type == AttributeProto.INT:
+#                 attrs[attr.name] = attr.i
+#             elif attr.type == AttributeProto.INTS:
+#                 attrs[attr.name] = attr.ints
+#             elif attr.type == AttributeProto.FLOAT:
+#                 attrs[attr.name] = attr.f
+#             elif attr.type == AttributeProto.STRING:
+#                 attrs[attr.name] = attr.s
+#             elif attr.type == AttributeProto.TENSOR:
+#                 attrs[attr.name] = attr.t
+#             else:
+#                 assert False, "Unsupported Attribute Type: {}".format(attr.type)
+#     return attrs
+
+
+# def _parse_data(tensor: TensorProto) -> List[Any]:
+#     return to_array(tensor).flatten().tolist()
+
+
+# def _take_shape_dim(shape: TensorShapeProto) -> List[int]:
+#     return [(d.dim_value if d.dim_value > 0 else 1) for d in shape.dim]
+
+def export_onnx(g: backend.Graph, path: str) -> None:
+    stub = OnnxStub()
+    stub.to_onnx(g, path)

run.sh

@@ -0,0 +1,2 @@
+./test_mutator --gtest_filter="NMutator.Conv3x3_to_gemm"
+./test_conv2gemm --gtest_filter="*.NCHW_FCRS_ruleBased"

src/bang/bang_runtime.cc

@@ -29,6 +29,7 @@ void BangRuntimeObj::runWithoutSync(const Graph &graph, bool tune = false,
             perfEngine.setPerfData(perfKey, record);
         } else
             record = perfData;
+        std::cout << 5 << std::endl;
 
         double t = record->time;
         totalTime += t;

src/core/graph.cc

@@ -125,12 +125,30 @@ void GraphObj::optimize() {
 
 void GraphObj::dataMalloc() {
     for (auto &tensor : tensors) {
-        tensor->dataMalloc();
+        if (tensor->getSource() && tensor->getTargets().size() > 0 &&
+            tensor->getSource()->getOpType() == OpType::Reshape) {
+            continue;
+        } else
+            tensor->dataMalloc();
+    }
+    // Fill reshape output for avoiding nullptr
+    for (auto &tensor : tensors) {
+        if (tensor->getSource() &&
+            tensor->getSource()->getOpType() == OpType::Reshape) {
+            tensor->setData(tensor->getSource()->getInputs(0)->getDataBlob());
+        }
     }
 }
 
-Tensor GraphObj::addTensor(Shape dim, DataType dtype) {
-    return tensors.emplace_back(make_ref<TensorObj>(dim, dtype, runtime));
+void GraphObj::dataFree() {
+    for (auto &tensor : tensors) {
+        tensor->freeData();
+    }
+}
+
+Tensor GraphObj::addTensor(Shape dim, DataType dtype, TensorType tensorType) {
+    return tensors.emplace_back(
+        make_ref<TensorObj>(dim, dtype, runtime, tensorType));
 }
 
 Tensor GraphObj::addTensor(const Tensor &tensor) {
@@ -228,4 +246,14 @@ bool GraphObj::checkValid() const {
     return true;
 }
 
+int GraphObj::removeIndependentTensors() {
+    TensorVec newTensors;
+    for (const auto &t : tensors)
+        if (!t->getTargets().empty() || t->getSource())
+            newTensors.emplace_back(t);
+    auto ret = tensors.size() - newTensors.size();
+    tensors = newTensors;
+    return ret;
+}
+
 } // namespace infini

src/core/graph_handler.cc

@@ -1,10 +1,12 @@
 #include "core/graph_handler.h"
+#include "nnet/Visitor/Serializer.h"
 #include "operators/batch_norm.h"
 #include "operators/concat.h"
 #include "operators/conv.h"
 #include "operators/element_wise.h"
 #include "operators/gather.h"
 #include "operators/matmul.h"
+#include "operators/membound.h"
 #include "operators/pad.h"
 #include "operators/pooling.h"
 #include "operators/reduce_mean.h"
@@ -19,8 +21,8 @@ namespace infini {
 
 static DataType dtype_repr_convert(int);
 
-Tensor GraphHandlerObj::tensor(Shape dims, int dtype) {
-    return g->addTensor(std::move(dims), dtype_repr_convert(dtype));
+Tensor GraphHandlerObj::tensor(Shape dims, int dtype, TensorType ttype) {
+    return g->addTensor(std::move(dims), dtype_repr_convert(dtype), ttype);
 }
 
 Tensor GraphHandlerObj::conv(Tensor input, Tensor weight, Tensor output, int ph,
@@ -55,6 +57,39 @@ Tensor GraphHandlerObj::convTransposed2d(Tensor input, Tensor weight,
     }
 }
 
+Tensor GraphHandlerObj::convNHWC(Tensor input, Tensor weight, Tensor output,
+                                 int ph, int pw, int sh, int sw, int dh,
+                                 int dw) {
+    if (output) {
+        g->addOpWithOutputs<ConvNHWCObj>(std::move(input), std::move(weight),
+                                         output, ph, pw, sh, sw, dh, dw);
+        return output;
+    } else {
+        return g
+            ->addOp<ConvNHWCObj>(std::move(input), std::move(weight), output,
+                                 ph, pw, sh, sw, dh, dw)
+            ->getOutput();
+    }
+}
+
+Tensor GraphHandlerObj::convTransposed2dNHWC(Tensor input, Tensor weight,
+                                             Tensor output, int ph, int pw,
+                                             int sh, int sw, int dh, int dw,
+                                             int oph, int opw) {
+    if (output) {
+        g->addOpWithOutputs<ConvTransposed2dNHWCObj>(
+            std::move(input), std::move(weight), output, ph, pw, sh, sw, dh, dw,
+            oph, opw);
+        return output;
+    } else {
+        return g
+            ->addOp<ConvTransposed2dNHWCObj>(std::move(input),
+                                             std::move(weight), output, ph, pw,
+                                             sh, sw, dh, dw, oph, opw)
+            ->getOutput();
+    }
+}
+
 Tensor GraphHandlerObj::matmul(Tensor a, Tensor b, Tensor y, bool transA,
                                bool transB, Tensor bias, ActType act) {
     if (y) {
@@ -291,6 +326,22 @@ Tensor GraphHandlerObj::pad(Tensor input, Tensor output,
     }
 }
 
+TensorVec GraphHandlerObj::memBound(const TensorVec &inputs,
+                                    const Tensor &output,
+                                    const string &jsonString) {
+    const auto &[expr, nnetInputs, execTime, hint] =
+        nnet::Serializer().membundOpFromString(jsonString);
+    if (output) {
+        g->addOpWithOutputs<MemBoundObj>(std::move(inputs), TensorVec{output},
+                                         nnetInputs, expr, execTime, hint);
+        return {output};
+    } else
+        return g
+            ->addOp<MemBoundObj>(std::move(inputs), TensorVec{nullptr},
+                                 nnetInputs, expr, execTime, hint)
+            ->getOutputs();
+}
+
 static DataType dtype_repr_convert(int dtype) {
     switch ((OnnxDType)dtype) {
     case OnnxDType::FLOAT:
@@ -314,4 +365,12 @@ static DataType dtype_repr_convert(int dtype) {
     }
 }
 
+Graph GraphHandlerObj::getGraph() const {
+    int nRemoved = g->removeIndependentTensors();
+    if (nRemoved > 0)
+        std::cout << "Removed " << nRemoved << " independent tensors"
+                  << std::endl;
+    return g;
+}
+
 } // namespace infini

src/core/operator.cc

@@ -1,6 +1,7 @@
 #include "core/operator.h"
 #include "core/graph.h"
 #include "core/hash.h"
+#include "nnet/dbg.h"
 
 namespace infini {
 
@@ -25,7 +26,8 @@ bool OperatorObj::isConcatOp() const { return type == OpType::Concat; }
 bool OperatorObj::isComputeOp() const {
     return type == OpType::Conv || type == OpType::Matmul ||
            type == OpType::ConvTrans || type == OpType::ConvTransNHWC ||
-           type == OpType::G2BMM || type == OpType::GBMM;
+           type == OpType::G2BMM || type == OpType::GBMM ||
+           type == OpType::ConvNHWC;
 }
 
 bool OperatorObj::isTransposeOp() const { return type == OpType::Transpose; }
@@ -33,8 +35,12 @@ bool OperatorObj::isTransposeOp() const { return type == OpType::Transpose; }
 bool OperatorObj::isReshapeOp() const { return type == OpType::Reshape; }
 
 bool OperatorObj::isMemBoundOp() const {
-    return type == OpType::MemBound || type == OpType::Activation ||
-           type == OpType::Transpose;
+    if (type == OpType::Any)
+        return true; // TODO: check operator attributes
+    return type == OpType::MemBound || type == OpType::Reshape ||
+           type == OpType::Activation || type == OpType::Transpose ||
+           type == OpType::Relu || type == OpType::Tanh ||
+           type == OpType::Softmax;
 }
 
 void OperatorObj::removePredecessors(const Operator &op) {
@@ -83,22 +89,31 @@ HashType OperatorObj::hash() const {
 
 bool OperatorObj::checkValid(GraphObj *graph) {
     auto optShapes = inferShape();
+    IT_ASSERT(optShapes);
     if (!optShapes) // shape inference failed
         return false;
 
     const vector<Shape> &shapes = *optShapes;
+    IT_ASSERT(shapes.size() == outputs.size());
     if (shapes.size() != outputs.size())
         return false;
     if (graph) { // if graph != nullptr, outputs should be created
         auto dataTypes = inferDataType();
         for (size_t i = 0; i < outputs.size(); i++) {
             IT_ASSERT(!outputs[i], "Find empty output while operator creation");
-            outputs[i] = graph->addTensor(shapes[i], dataTypes[i]);
+            outputs[i] =
+                graph->addTensor(shapes[i], dataTypes[i], TensorType::Other);
         }
     } else { // if outputs have been created, check their shapes
         for (size_t i = 0; i < shapes.size(); ++i) {
-            if (shapes[i] != outputs[i]->getDims())
+            IT_ASSERT(shapes[i] == outputs[i]->getDims(),
+                      (vecToString(shapes[i]) +
+                       " != " + vecToString(outputs[i]->getDims())));
+            if (shapes[i] != outputs[i]->getDims()) {
+                dbg(shapes[i], outputs[i]->getDims());
                 return false;
+            }
+            IT_ASSERT(outputs[i]->getTensorType() == TensorType::Other);
         }
     }
     return true;

src/core/runtime.cc

@@ -2,9 +2,15 @@
 #include "core/blob.h"
 #include "core/kernel.h"
 #include "core/perf_engine.h"
+#include "operators/membound.h"
 #include "utils/data_generator.h"
 #include <chrono>
 #include <cstring>
+
+#ifdef USE_CUDA
+#include "cuda_profiler_api.h"
+#endif
+
 namespace infini {
 void CpuRuntimeObj::run(const Graph &graph, bool tune, bool profiling) const {
     if (!tune && profiling)
@@ -52,17 +58,40 @@ void CpuRuntimeObj::run(const Graph &graph, bool tune, bool profiling) const {
             opCnt[op->getOpType()]++;
         }
     }
-    if (profiling)
-        printProfilingData(totalTime, opTime, opCnt);
+    // if (profiling)
+    //     printProfilingData(totalTime, opTime, opCnt);
 }
 
-double RuntimeObj::getPerfTime(const Graph &graph, bool profiling) const {
+map<UidBaseType, bool>
+RuntimeObj::getCompileTimeComputableAttribute(const Graph &graph) const {
+    map<UidBaseType, bool> ctcMap; // compile-time computable
+    // Skip static computation
+    bool status = graph->topo_sort();
+    IT_ASSERT(status, "Topological sort failed");
+    for (auto &op : graph->getOperators()) {
+        bool compileTimeComputable = true;
+        for (auto input : op->getInputs()) {
+            // FIXME: propogate the tensor type. Current only the first operator
+            // after weights are compile-time computable.
+            if (input->getTensorType() != TensorType::Initialized)
+                compileTimeComputable = false;
+        }
+        ctcMap[op->getGuid()] = compileTimeComputable;
+    }
+    return ctcMap;
+}
+
+double RuntimeObj::getPerfTime(const Graph &graph, bool profiling,
+                               bool allowEstimation,
+                               bool ignoreMemboundOp) const {
     const auto &kernelRegistry = KernelRegistry::getInstance();
     auto &perfEngine = PerfEngine::getInstance();
     // Statistics
     double totalTime = 0;
     std::map<OpType, double> opTime;
-    std::map<OpType, int> opCnt;
+    std::map<OpType, int> opCnt, opNonCtcCnt;
+    // compile-time computable
+    map<UidBaseType, bool> ctcMap = getCompileTimeComputableAttribute(graph);
 
     for (auto &op : graph->getOperators()) {
         auto kernelAttrs = KernelAttrs{device, op->getOpType(), op->getDType()};
@@ -70,11 +99,19 @@ double RuntimeObj::getPerfTime(const Graph &graph, bool profiling) const {
         auto perfKey = PerfEngine::Key{kernelAttrs, op->getOpPerfKey()};
         auto perfData = perfEngine.getPerfData(perfKey);
 
-        PerfRecord record;
-        // Tune the kernel if there is no record
-        if (!perfData) {
+        double time = -1e9;
+        if (ctcMap[op->getGuid()]) { // Compile-time computable operators
+            time = 0;
+        } else if (op->getOpType() == OpType::MemBound && ignoreMemboundOp) {
+            time = 0;
+        } else if (op->getOpType() == OpType::MemBound && allowEstimation) {
+            time = as<MemBoundObj>(op)->getEstimatedTime();
+        } else if (perfData) { // Tune the kernel if there is no record
+            time = perfData->time;
+        } else {
             // TODO: should tenosrs automatically allocate when access data?
-            // allocate memory for empty tensors and release it after profiling
+            // allocate memory for empty tensors and release it after
+            // profiling
             TensorVec allocatedTensors;
             for (auto t : op->getInputs())
                 if (!t->hasData())
@@ -88,37 +125,47 @@ double RuntimeObj::getPerfTime(const Graph &graph, bool profiling) const {
             }
 
             // Profile operators and record the results
-            record = kernel->tune(op, this);
+            PerfRecord record = kernel->tune(op, this);
+            time = record->time;
             perfEngine.setPerfData(perfKey, record);
 
             // Free allocated memory
             for (auto t : allocatedTensors)
                 t->freeData();
-        } else
-            record = perfData;
+        }
 
-        double t = record->time;
-        totalTime += t;
+        // FIXME: ignore trnapose when necessary
+        //     op->getOpType() != OpType::Transpose &&
+        //     op->getOpType() != OpType::ReduceMean
+        if (op->getOpType() != OpType::Reshape)
+            totalTime += time;
         if (profiling) {
             op->print();
-            printf(" op_time %lf\n", t);
-            opTime[op->getOpType()] += t;
+            printf("  op_time %lf\n", time);
+            opTime[op->getOpType()] += time;
             opCnt[op->getOpType()]++;
+            if (!ctcMap[op->getGuid()])
+                opNonCtcCnt[op->getOpType()]++;
+            else
+                opNonCtcCnt[op->getOpType()]; // Create a new entry
         }
     }
     if (profiling)
-        printProfilingData(totalTime, opTime, opCnt);
+        printProfilingData(totalTime, opTime, opCnt, opNonCtcCnt);
     return totalTime;
 }
 
-void RuntimeObj::printProfilingData(double totalTime,
-                                    const std::map<OpType, double> &opTime,
-                                    const std::map<OpType, int> &opCnt) const {
-    printf("%11s %3s %7s %7s %7s\n", "Op", "Cnt", "T_tot", "Percent", "T_mean");
+void RuntimeObj::printProfilingData(
+    double totalTime, const std::map<OpType, double> &opTime,
+    const std::map<OpType, int> &opCnt,
+    const std::map<OpType, int> &opNonCtcCnt) const {
+    printf("%11s %3s %5s %7s %7s %7s\n", "Op", "Cnt", "#NCtc", "T_tot",
+           "Percent", "T_mean");
     for (const auto &[type, t] : opTime) {
-        printf("%11s %3d %7.3f %7.1f %7.3f\n",
-               OpRegistry::getOpName(type).data(), opCnt.at(type), t,
-               t / totalTime * 100, t / opCnt.at(type));
+        printf("%11s %3d %5d %7.3f %7.1f %7.3f\n",
+               OpRegistry::getOpName(type).data(), opCnt.at(type),
+               opNonCtcCnt.at(type), t, t / totalTime * 100,
+               t / opCnt.at(type));
     }
 }
 
@@ -160,4 +207,44 @@ void CpuRuntimeObj::copyBlobInsideRuntime(void *dst, const void *src,
 
 string NativeCpuRuntimeObj::toString() const { return "CPU Runtime"; }
 
+double RuntimeObj::timeNonCtcOperators(const Graph &graph, int warmup,
+                                       int repeat) const {
+    const auto &kernelRegistry = KernelRegistry::getInstance();
+    auto &perfEngine = PerfEngine::getInstance();
+    // compile-time computable
+    map<UidBaseType, bool> ctcMap = getCompileTimeComputableAttribute(graph);
+    vector<tuple<Operator, Kernel *, PerfRecord>> kernels;
+    bool status = graph->topo_sort();
+    IT_ASSERT(status, "Topological sort failed");
+
+    for (auto &op : graph->getOperators()) {
+        // HACK: set correct data type
+        auto kernelAttrs =
+            KernelAttrs{device, op->getOpType(), DataType::Float32};
+        Kernel *kernel = kernelRegistry.getKernel(kernelAttrs);
+        auto perfKey = PerfEngine::Key{kernelAttrs, op->getOpPerfKey()};
+        auto perfData = perfEngine.getPerfData(perfKey);
+        if (perfData)
+            kernel->compute(op, perfData, this);
+        else
+            kernel->compute(op, this);
+        if (!ctcMap.at(op->getGuid()) && op->getOpType() != OpType::Reshape)
+            kernels.emplace_back(op, kernel, perfData);
+    }
+    for (auto &[op, kernel, perfData] : kernels) {
+        dbg(op);
+    }
+    double ret = timeit(
+        [&]() {
+            for (auto &[op, kernel, perfData] : kernels) {
+                if (perfData)
+                    kernel->compute(op, perfData, this);
+                else
+                    kernel->compute(op, this);
+            }
+        },
+        [&]() { this->sync(); }, warmup, repeat);
+    return ret;
+}
+
 } // namespace infini

src/core/search_engine.cc

@@ -1,6 +1,9 @@
 #include "core/search_engine.h"
 #include "core/hash.h"
 #include "core/runtime.h"
+#include "ffi/ffi_callback.h"
+#include "nnet/dbg.h"
+#include "operators/reshape.h"
 
 #include <algorithm>
 #include <iostream>
@@ -8,7 +11,17 @@
 
 namespace infini {
 
-void SearchEngine::printMetaGraph(Ref<SearchEngine::MetaGraph> metaGraph) {
+using MetaGraph = SearchEngine::MetaGraph;
+
+SearchEngine::SearchEngine(Runtime runtime, Ref<Mutator> mutator)
+    : runtimeExec(runtime), mutator(mutator) {
+    // Compare graph with estimated time
+    graphTimeComparer = [this](const Graph &a, const Graph &b) -> bool {
+        return getEstimatedGraphPerf(a) < getEstimatedGraphPerf(b);
+    };
+}
+
+void SearchEngine::printMetaGraph(MetaGraph metaGraph) {
     for (size_t i = 0; i < metaGraph->nodes.size(); i++) {
         auto &node = metaGraph->nodes[i];
         std::cout << "id: " << i << std::endl;
@@ -32,8 +45,7 @@ Graph SearchEngine::run(const Graph graph) {
     IT_ASSERT(runtimeExec == graph->getRuntime());
     std::cout << "[INFO] original graph: " << std::endl;
     std::cout << graph->toString();
-    std::cout << "[INFO] perf: " << runtimeExec->getPerfTime(graph)
-              << std::endl;
+    std::cout << "[INFO] perf: " << getEstimatedGraphPerf(graph) << std::endl;
 
     std::vector<Graph> partitions = partitionGraph(graph);
 
@@ -45,7 +57,6 @@ Graph SearchEngine::run(const Graph graph) {
         std::vector<Graph> candidates = search(subGraph);
         std::cout << "[INFO] size: " << candidates.size() << std::endl;
         IT_ASSERT(candidates.size() > 0);
-        std::cout << subGraph->toString() << std::endl;
         std::vector<Graph> nextGraphs;
         for (auto lastGraph : bestGraphs) {
             for (auto thisGraph : candidates) {
@@ -61,13 +72,12 @@ Graph SearchEngine::run(const Graph graph) {
                     }
                 }
                 auto tmp = make_ref<GraphObj>(runtimeExec, ops);
-                tmp->dataMalloc();
                 nextGraphs.emplace_back(tmp);
             }
         }
-        std::sort(nextGraphs.begin(), nextGraphs.end(), [&](Graph x, Graph y) {
-            return runtimeExec->getPerfTime(x) < runtimeExec->getPerfTime(y);
-        });
+        dbg("===Num" + std::to_string(nextGraphs.size()));
+        std::sort(nextGraphs.begin(), nextGraphs.end(), graphTimeComparer);
+
         if (nextGraphs.size() > GRAPH_SIZE) {
             nextGraphs.resize(GRAPH_SIZE);
         }
@@ -81,10 +91,30 @@ Graph SearchEngine::run(const Graph graph) {
     for (size_t i = 0; i < bestGraphs.size(); i++) {
         std::cout << "bestGraph " << i << ":" << std::endl;
         std::cout << bestGraphs[i]->toString();
-        std::cout << "[INFO] perf: " << runtimeExec->getPerfTime(bestGraphs[i])
+        std::cout << "[INFO] perf: " << getEstimatedGraphPerf(bestGraphs[i])
                   << std::endl;
     }
 
+    // Fuse vertically and sort according to performance
+    for (size_t i = 0; i < bestGraphs.size(); ++i) {
+        bestGraphs[i] = fuseVertically(bestGraphs[i]);
+    }
+    std::sort(bestGraphs.begin(), bestGraphs.end(), graphTimeComparer);
+
+    // Check optimized graphs are legal
+    for (auto g : bestGraphs) {
+        g->checkValid();
+        IT_ASSERT(graph->getInputs().size() == g->getInputs().size(),
+                  graph->toString() + string("\n") + g->toString());
+        IT_ASSERT(graph->getOutputs().size() == g->getOutputs().size(),
+                  graph->toString() + string("\n") + g->toString());
+    }
+
+    std::cout << "[INFO] best fused graph: " << std::endl;
+    std::cout << "[INFO] perf: " << getEstimatedGraphPerf(bestGraphs[0])
+              << std::endl;
+    std::cout << bestGraphs[0] << std::endl;
+
     return bestGraphs[0];
 }
 
@@ -102,9 +132,9 @@ std::vector<Graph> SearchEngine::search(const Graph &graph) {
         }
     }
 
-    sort(results.begin(), results.end(), [&](Graph x, Graph y) {
-        return runtimeExec->getPerfTime(x) < runtimeExec->getPerfTime(y);
-    }); // compare with perf time
+    // compare with perf time
+    dbg("===Num" + std::to_string(results.size()));
+    std::sort(results.begin(), results.end(), graphTimeComparer);
     if (results.size() > GRAPH_SIZE) {
         results.resize(GRAPH_SIZE);
     }
@@ -112,9 +142,8 @@ std::vector<Graph> SearchEngine::search(const Graph &graph) {
 }
 
 // Build metagraph with a graph, each operator is a node.
-std::shared_ptr<SearchEngine::MetaGraph>
-SearchEngine::buildMetaGraphWithGraph(const Graph graph) {
-    auto metaGraph = std::make_shared<MetaGraph>();
+MetaGraph SearchEngine::buildMetaGraphWithGraph(const Graph graph) {
+    auto metaGraph = make_ref<MetaGraphObj>();
 
     int numOps = graph->getOperators().size();
     std::vector<int> cnt(numOps, 0);
@@ -123,7 +152,7 @@ SearchEngine::buildMetaGraphWithGraph(const Graph graph) {
     std::vector<int> q(0);
     for (size_t i = 0; i < graph->getOperators().size(); i++) {
         auto &op = graph->getOperators()[i];
-        MetaGraph::Node node;
+        MetaGraphObj::Node node;
         std::vector<Operator> ops;
         ops.emplace_back(op);
         node.graph = make_ref<GraphObj>(runtimeExec, ops);
@@ -157,9 +186,8 @@ SearchEngine::buildMetaGraphWithGraph(const Graph graph) {
 
 // Build a metagraph with graph and a plan, a plan is which ops should be a
 // node.
-std::shared_ptr<SearchEngine::MetaGraph> SearchEngine::buildMetaGraphWithPlan(
-    const std::shared_ptr<SearchEngine::MetaGraph> metaGraph,
-    const std::vector<int> &plan) {
+MetaGraph SearchEngine::buildMetaGraphWithPlan(const MetaGraph metaGraph,
+                                               const std::vector<int> &plan) {
     int numGroups = 0;
     for (auto i : plan) {
         if (i > numGroups) {
@@ -172,12 +200,12 @@ std::shared_ptr<SearchEngine::MetaGraph> SearchEngine::buildMetaGraphWithPlan(
         groups[plan[i]].emplace_back(i);
     }
 
-    auto resultMetaGraph = make_ref<MetaGraph>();
+    auto resultMetaGraph = make_ref<MetaGraphObj>();
     for (auto &group : groups) {
         std::vector<Operator> ops;
         std::unordered_set<int> preSet, sucSet;
         for (auto id : group) {
-            MetaGraph::Node node;
+            MetaGraphObj::Node node;
             for (auto op : metaGraph->nodes[id].graph->getOperators()) {
                 ops.emplace_back(op);
             }
@@ -204,9 +232,10 @@ std::shared_ptr<SearchEngine::MetaGraph> SearchEngine::buildMetaGraphWithPlan(
 }
 
 // Search how to merge multiple ops.
-std::vector<std::shared_ptr<SearchEngine::MetaGraph>>
-SearchEngine::searchMerge(std::shared_ptr<SearchEngine::MetaGraph> &metaGraph) {
+vector<MetaGraph> SearchEngine::searchMerge(MetaGraph &metaGraph) {
     IT_ASSERT(metaGraph != nullptr);
+    // HACK: disable multiple op search
+    return {metaGraph};
     std::vector<int> plan(metaGraph->nodes.size());
     for (size_t i = 0; i < plan.size(); i++) {
         plan[i] = i;
@@ -222,7 +251,7 @@ SearchEngine::searchMerge(std::shared_ptr<SearchEngine::MetaGraph> &metaGraph) {
     std::unordered_set<HashType> planSet;
     searchMergeDfs(metaGraph, plan, frontier, plans, planSet);
 
-    std::vector<std::shared_ptr<SearchEngine::MetaGraph>> metaGraphs;
+    vector<MetaGraph> metaGraphs;
     for (auto &curPlan : plans) {
         metaGraphs.emplace_back(buildMetaGraphWithPlan(metaGraph, curPlan));
     }
@@ -230,8 +259,7 @@ SearchEngine::searchMerge(std::shared_ptr<SearchEngine::MetaGraph> &metaGraph) {
 }
 
 // DFS impl for search merge.
-void SearchEngine::searchMergeDfs(std::shared_ptr<MetaGraph> &metaGraph,
-                                  std::vector<int> &plan,
+void SearchEngine::searchMergeDfs(MetaGraph &metaGraph, std::vector<int> &plan,
                                   std::vector<int> &frontier,
                                   std::vector<std::vector<int>> &plans,
                                   std::unordered_set<uint64_t> &planSet) {
@@ -320,14 +348,40 @@ void SearchEngine::searchMergeDfs(std::shared_ptr<MetaGraph> &metaGraph,
 }
 
 // Search mutation for each compute op.
-std::vector<Graph> SearchEngine::searchMutation(
-    const std::shared_ptr<SearchEngine::MetaGraph> &metaGraph) {
+std::vector<Graph> SearchEngine::searchMutation(const MetaGraph &metaGraph) {
     std::vector<Graph> graphs = {nullptr};
     // Append a node to all existing candidates
     for (auto &node : metaGraph->nodes) {
         std::vector<Graph> nextGraphs;
         if (node.type == 1) { // If it has computing OPs
             auto mutatedGraphs = mutator->run(node.graph);
+            if (mutator->hasTunedKernel)
+                chooseBestMutation = false;
+            std::sort(mutatedGraphs.begin(), mutatedGraphs.end(),
+                      graphTimeComparer);
+            if (mutatedGraphs.size() >= 10)
+                mutatedGraphs.resize(10);
+            mutatedGraphs = {mutatedGraphs[0]};
+            // if (searchFilter == 1) {
+            //     std::sort(mutatedGraphs.begin(), mutatedGraphs.end(),
+            //               graphTimeComparer);
+            //     if (mutatedGraphs.size() >= 10)
+            //         mutatedGraphs.resize(10);
+            //     mutatedGraphs = {mutatedGraphs[0]};
+            // } else if (chooseBestMutation && mutatedGraphs.size() >= 2) {
+            //     std::sort(mutatedGraphs.begin(), mutatedGraphs.end(),
+            //               graphTimeComparer);
+            //     if (mutatedGraphs.size() >= 10)
+            //         mutatedGraphs.resize(10);
+            //     mutatedGraphs = {mutatedGraphs[0]};
+            // } else { // avoid repeated kernel genreation
+            //     if (mutatedGraphs.size() >= 2) // INFOGAN
+            //         mutatedGraphs = {mutatedGraphs[1]};
+            //     // if (mutatedGraphs.size() > 2) {
+            //     //     mutatedGraphs.resize(2);
+            //     // }
+            // }
+
             for (auto graph : graphs) {
                 for (auto mutatedGraph : mutatedGraphs) {
                     std::vector<Operator> ops;
@@ -357,12 +411,8 @@ std::vector<Graph> SearchEngine::searchMutation(
                 nextGraphs.emplace_back(make_ref<GraphObj>(runtimeExec, ops));
             }
         }
-        for (auto g : nextGraphs) {
-            g->dataMalloc();
-        }
-        std::sort(nextGraphs.begin(), nextGraphs.end(), [&](Graph x, Graph y) {
-            return runtimeExec->getPerfTime(x) < runtimeExec->getPerfTime(y);
-        });
+        dbg("===Num" + std::to_string(nextGraphs.size()));
+        std::sort(nextGraphs.begin(), nextGraphs.end(), graphTimeComparer);
         if (nextGraphs.size() > GRAPH_SIZE) {
             nextGraphs.resize(GRAPH_SIZE);
         }
@@ -372,7 +422,7 @@ std::vector<Graph> SearchEngine::searchMutation(
 }
 
 bool SearchEngine::isMultiBranchMergable(const Graph graph) {
-    return mutationEngine->isMultiBranchMergable(graph);
+    return mutator->isMultiBranchMergable(graph);
 }
 
 // Split a graph into multiple independt graphs. Search engine will search for
@@ -423,7 +473,6 @@ std::vector<Graph> SearchEngine::partitionGraph(const Graph graph) {
                 std::cout << op->toString() << std::endl;
             }
             auto tmp = make_ref<GraphObj>(runtimeExec, headOps);
-            tmp->dataMalloc();
             partitions.emplace_back(tmp);
             headOps.clear();
         }
@@ -431,11 +480,100 @@ std::vector<Graph> SearchEngine::partitionGraph(const Graph graph) {
     }
     if (!headOps.empty()) {
         auto tmp = make_ref<GraphObj>(runtimeExec, headOps);
-        tmp->dataMalloc();
         partitions.emplace_back(tmp);
     }
     std::reverse(partitions.begin(), partitions.end());
     return partitions;
 }
 
+double SearchEngine::getEstimatedGraphPerf(Graph graph) {
+    // dbg(graph);
+    // // hkz
+    // callback::exportONNX(graph, "a.onnx");
+    return runtimeExec->getPerfTime(graph, false, true, true);
+}
+
+Graph SearchEngine::fuseVertically(const Graph &graph) {
+    std::unordered_map<UidBaseType, int> visitTime;
+    std::vector<Operator> ops;
+
+    graph->topo_sort();
+    int cnt = 0;
+    for (auto op : graph->getOperators()) {
+        // Skip visited OP
+        if (visitTime.find(op->getGuid()) != visitTime.end()) {
+            continue;
+        }
+        // Skip compute OP and multi-input/output OP
+        if (!op->isMemBoundOp() || (op->getPredecessors().size() != 1 &&
+                                    op->getSuccessors().size() != 1)) {
+            visitTime.emplace(op->getGuid(), ++cnt);
+            ops.emplace_back(op);
+            continue;
+        }
+        // FIXME: fuse and modify attributes of computing operators
+        if (op->getOpType() == OpType::Relu ||
+            op->getOpType() == OpType::PRelu) {
+            if (auto p = op->getInputs()[0])
+                if (auto sop = p->getSource())
+                    if (sop->getOpType() == OpType::Conv ||
+                        sop->getOpType() == OpType::Matmul) {
+                        visitTime.emplace(op->getGuid(), ++cnt);
+                        ops.emplace_back(make_ref<ReshapeObj>(
+                            nullptr, op->getInputs()[0], op->getOutputs()[0]));
+                        continue;
+                    }
+        }
+        vector<Operator> chainOps;
+        visitTime.emplace(op->getGuid(), ++cnt);
+
+        vector<Operator> tmp;
+        auto cur = op;
+        while (cur->getPredecessors().size() == 1 &&
+               cur->getPredecessors()[0]->isMemBoundOp()) {
+            cur = cur->getPredecessors()[0];
+            if (visitTime.count(cur->getGuid()))
+                break;
+            tmp.emplace_back(cur);
+            visitTime.emplace(cur->getGuid(), cnt);
+        }
+        for (int i = tmp.size() - 1; i >= 0; i--) {
+            chainOps.emplace_back(tmp[i]);
+        }
+        chainOps.emplace_back(op);
+        cur = op;
+        while (cur->getSuccessors().size() == 1 &&
+               cur->getSuccessors()[0]->isMemBoundOp()) {
+            cur = cur->getSuccessors()[0];
+            if (visitTime.count(cur->getGuid()))
+                break;
+            chainOps.emplace_back(cur);
+            visitTime.emplace(cur->getGuid(), cnt);
+        }
+        make_ref<GraphObj>(runtimeExec, chainOps)->print();
+
+        auto bestGraph = make_ref<GraphObj>(runtimeExec, chainOps);
+        // Eliminate transpose and reshape operators
+        if (auto eliminatedGraph = mutator->eliminateVertically(
+                make_ref<GraphObj>(runtimeExec, chainOps)))
+            bestGraph = eliminatedGraph;
+        // Fuse membound operators
+        if (auto optGraph = mutator->fuseVertically(bestGraph))
+            bestGraph = optGraph;
+        for (auto op : bestGraph->getOperators()) {
+            ops.emplace_back(op);
+        }
+    }
+    if (ops.empty()) {
+        IT_TODO_HALT();
+        IT_ASSERT(graph->getOutputs().size() == 1);
+        IT_ASSERT(graph->getInputs().size() == 1);
+        // auto g = make_ref<GraphObj>(runtime);
+        // TODO: add identity
+        ops.emplace_back(make_ref<ReshapeObj>(nullptr, graph->getInputs()[0],
+                                              graph->getOutputs()[0]));
+    }
+    return make_ref<GraphObj>(runtimeExec, ops);
+}
+
 } // namespace infini

src/core/tensor.cc

@@ -8,12 +8,14 @@
 
 namespace infini {
 
-TensorObj::TensorObj(Shape shape_, DataType dtype, Runtime runtime)
+TensorObj::TensorObj(Shape shape_, DataType dtype, Runtime runtime,
+                     TensorType tensorType)
     : TensorBaseObj(shape_.size(), dtype, runtime), shape(std::move(shape_)),
       _size(shape.empty()
                 ? 0
-                : std::accumulate(shape.begin(), shape.end(), 1,
-                                  [](auto acc, auto x) { return acc * x; })) {}
+                : std::accumulate(shape.begin(), shape.end(), 1lu,
+                                  [](auto acc, auto x) { return acc * x; })),
+      tensorType(tensorType) {}
 
 string TensorObj::toString() const {
     // Convert data pointer to string
@@ -24,8 +26,8 @@ string TensorObj::toString() const {
         ss << "nullptr data";
     string ret = "Tensor " + std::to_string(guid) + ", Fuid " +
                  std::to_string(fuid) + ", shape " + vecToString(shape) +
-                 ", dtype " + dtype.toString() + ", " + runtime->toString() +
-                 ", " + ss.str() + "\n";
+                 ", dtype " + dtype.toString() + ", tensorType " +
+                 std::to_string(enum_to_underlying(tensorType));
     vector<UidBaseType> targetGuids;
     for (const auto &op : targets)
         targetGuids.emplace_back(op.lock()->getGuid());
@@ -34,6 +36,7 @@ string TensorObj::toString() const {
     else
         ret += ", source None";
     ret += ", targets " + vecToString(targetGuids);
+    ret += ", " + runtime->toString() + ", " + ss.str();
     return ret;
 }
 
@@ -64,12 +67,19 @@ vector<size_t> TensorObj::getStride() const {
 
 void TensorObj::printData() const {
     IT_ASSERT(data != nullptr);
-    if (!runtime->isCpu())
-        IT_TODO_HALT();
+    void *ptr = nullptr;
+    Blob buffer;
+    if (!runtime->isCpu()) { // copy data to main memory
+        buffer = NativeCpuRuntimeObj::getInstance()->allocBlob(getBytes());
+        runtime->copyBlobToCPU(buffer->getPtr<void *>(),
+                               getRawDataPtr<void *>(), getBytes());
+        ptr = buffer->getPtr<void *>();
+    } else
+        ptr = data->getPtr<float *>();
 
 #define TRY_PRINT(N)                                                           \
     if (dtype == DataType(N))                                                  \
-        std::cout << dataToString<DT<N>::t>() << std::endl;
+        std::cout << dataToString<DT<N>::t>(ptr) << std::endl;
 
     TRY_PRINT(0)          // fmt: new line
     else TRY_PRINT(1)     //
@@ -112,8 +122,9 @@ bool TensorObj::equalData(const Tensor &rhs, double relativeError) const {
 }
 
 void TensorObj::dataMalloc() {
-    if (!data)
+    if (!data) {
         data = runtime->allocBlob(getBytes());
+    }
 }
 
 void TensorObj::copyData(const TensorObj *src) {
@@ -172,4 +183,27 @@ size_t TensorObj::getOffsetByBroadcastOffset(size_t bcOffset,
 
     return getOffsetByPos(pos, shape);
 }
+
+Tensor TensorObj::clone() const {
+    auto obj = make_ref<TensorObj>(*this);
+    obj->freeData();
+    obj->targets.clear();
+    obj->source.reset();
+    return obj;
+}
+
+Tensor TensorObj::clone(Runtime runtime) const {
+    auto obj = make_ref<TensorObj>(*this);
+    obj->runtime = runtime;
+    obj->freeData();
+    obj->targets.clear();
+    obj->source.reset();
+    // FIXME
+    // if (hasData()) {
+    //     obj->dataMalloc();
+    //     obj->copyData(this);
+    // }
+    return obj;
+}
+
 }; // namespace infini

src/cuda/cuda_runtime.cc

@@ -2,10 +2,58 @@
 #include "core/kernel.h"
 #include "core/perf_engine.h"
 #include "core/runtime.h"
+#include "cuda_profiler_api.h"
+#include "nnet/dbg.h"
+#include "operators/any.h"
 #include "operators/conv.h"
 #include "operators/matmul.h"
+#ifdef INFINI_USE_TVM
+#include "tvm/runtime/device_api.h"
+#endif
 namespace infini {
 
+CudaRuntimeObj::CudaRuntimeObj()
+    : RuntimeObj(Device::CUDA), stream(cudaStreamPerThread),
+      cudaGraphStatus(false) {
+    checkCudnnError(cudnnCreate(&cudnn));
+    checkCublasError(cublasCreate(&cublas));
+    checkCudnnError(cudnnSetStream(cudnn, stream));
+    checkCublasError(cublasSetStream(cublas, stream));
+    workspaceSize = 2ll << 30; // 2 GB
+    workspace = alloc(workspaceSize);
+    // Get CUDA device properties
+    checkCudaError(cudaGetDeviceProperties(&deviceProperties, 0));
+}
+
+CudaRuntimeObj::~CudaRuntimeObj() {
+    try {
+        dealloc(workspace);
+        checkCudnnError(cudnnDestroy(cudnn));
+        checkCublasError(cublasDestroy(cublas));
+    } catch (const std::exception &e) {
+        std::cerr << "Error in ~CudaRuntimeObj: " << e.what() << std::endl;
+    }
+}
+
+void CudaRuntimeObj::beginCudaGraphStreamCapture() {
+    enum cudaStreamCaptureStatus pCaptureStatus;
+    checkCudaError(cudaStreamIsCapturing(stream, &pCaptureStatus));
+    IT_ASSERT(pCaptureStatus == cudaStreamCaptureStatusNone);
+    cudaGraphStatus = true;
+    checkCudaError(cudaStreamBeginCapture(stream, cudaStreamCaptureModeGlobal));
+}
+
+tuple<cudaGraphExec_t, size_t> CudaRuntimeObj::endCudaGraphStreamCapture() {
+    cudaGraph_t cudaGraph;
+    cudaGraphExec_t instance;
+    checkCudaError(cudaStreamEndCapture(stream, &cudaGraph));
+    cudaGraphStatus = false;
+    size_t numCudaGraphNodes;
+    checkCudaError(cudaGraphGetNodes(cudaGraph, nullptr, &numCudaGraphNodes));
+    checkCudaError(cudaGraphInstantiate(&instance, cudaGraph, NULL, NULL, 0));
+    return {instance, numCudaGraphNodes};
+}
+
 void CudaRuntimeObj::runWithoutSync(const Graph &graph) const {
     const auto &kernelRegistry = KernelRegistry::getInstance();
     auto &perfEngine = PerfEngine::getInstance();
@@ -75,4 +123,74 @@ void CudaRuntimeObj::sync() const { checkCudaError(cudaDeviceSynchronize()); }
 
 string CudaRuntimeObj::toString() const { return "CUDA Runtime"; }
 
+double CudaRuntimeObj::timeWithCudaGraph(Graph graph, int rounds) {
+    const auto &kernelRegistry = KernelRegistry::getInstance();
+    auto &perfEngine = PerfEngine::getInstance();
+    // compile-time computable
+    map<UidBaseType, bool> ctcMap = getCompileTimeComputableAttribute(graph);
+    vector<tuple<Operator, Kernel *, PerfRecord>> kernels;
+    bool status = graph->topo_sort();
+    IT_ASSERT(status, "Topological sort failed");
+
+    for (auto &op : graph->getOperators()) {
+        // HACK: set correct data type
+        auto kernelAttrs =
+            KernelAttrs{device, op->getOpType(), DataType::Float32};
+        Kernel *kernel = kernelRegistry.getKernel(kernelAttrs);
+        auto perfKey = PerfEngine::Key{kernelAttrs, op->getOpPerfKey()};
+        auto perfData = perfEngine.getPerfData(perfKey);
+        if (perfData)
+            kernel->compute(op, perfData, this);
+        else
+            kernel->compute(op, this);
+        bool isFakeOp = (as<AnyObj>(op) &&
+                         as<AnyObj>(op)->getKernelName() == string("FakeOp"));
+        if (as<AnyObj>(op))
+            dbg(op, as<AnyObj>(op)->getKernelName() == string("FakeOp"));
+        if (!ctcMap.at(op->getGuid()) && op->getOpType() != OpType::Reshape &&
+            op->getOpType() != OpType::Flatten && !isFakeOp)
+            kernels.emplace_back(op, kernel, perfData);
+    }
+    for (auto &[op, kernel, perfData] : kernels) {
+        dbg(op);
+    }
+    vector<std::function<void(void)>> funcs;
+    for (auto &[op, kernel, perfData] : kernels) {
+        if (perfData)
+            funcs.push_back([&]() { kernel->compute(op, perfData, this); });
+        else
+            funcs.push_back([&]() { kernel->compute(op, this); });
+    }
+    return timeWithCudaGraph(funcs, rounds);
+}
+
+double
+CudaRuntimeObj::timeWithCudaGraph(std::vector<std::function<void(void)>> funcs,
+                                  int rounds) {
+// TODO: move this to kernel source?
+// Init tvm stream
+#ifdef INFINI_USE_TVM
+    DLDevice tvm_device_id = {kDLCUDA, 0};
+    auto tvm_device = tvm::runtime::DeviceAPI::Get(tvm_device_id);
+    tvm_device->SetStream(tvm_device_id, getStream());
+#endif
+    beginCudaGraphStreamCapture();
+    for (auto &f : funcs)
+        f();
+    auto [cudaGraphInstance, numCudaGraphNodes] = endCudaGraphStreamCapture();
+    // Since one TVM packed function may contaion more than one CUDA kernel, the
+    // number of captured kernels may exceed the number of operators.
+    IT_ASSERT(numCudaGraphNodes >= funcs.size(),
+              std::to_string(numCudaGraphNodes) +
+                  " != " + std::to_string(funcs.size()));
+    return timeit(
+        [&, cudaGraphInstance = cudaGraphInstance, stream = getStream()]() {
+            checkCudaError(cudaGraphLaunch(cudaGraphInstance, stream));
+        },
+        [&, stream = getStream()]() { cudaStreamSynchronize(stream); },
+        std::min(50, rounds), rounds);
+}
+
+void CudaRuntimeObj::setEnableTF32(bool state) { enableTF32 = state; }
+
 } // namespace infini

src/ffi/ffi_callback.cc

@@ -0,0 +1,22 @@
+#include "core/graph.h"
+#include <pybind11/stl.h>
+
+namespace py = pybind11;
+
+namespace infini {
+
+namespace callback {
+
+using namespace py::literals;
+
+static std::function<void(const Graph &, string)> exportONNXImpl;
+void exportONNX(const Graph &graph, const string &path) {
+    IT_ASSERT(Py_IsInitialized(), "Python interpreter is not running.");
+    static auto exportONNXImpl =
+        py::module_::import("pyinfinitensor.onnx").attr("save_onnx");
+    exportONNXImpl(graph, path);
+}
+
+} // namespace callback
+
+} // namespace infini

src/ffi/ffi_infinitensor.cc

@@ -1,9 +1,15 @@
 #include "core/graph_handler.h"
+#include "core/mutator.h"
+#include "core/search_engine.h"
+#include "nnet/nmutator.h"
+#include "nnet/test_models.h"
+#include "operators/any.h"
 #include "operators/batch_norm.h"
 #include "operators/concat.h"
 #include "operators/conv.h"
 #include "operators/gather.h"
 #include "operators/matmul.h"
+#include "operators/membound.h"
 #include "operators/pad.h"
 #include "operators/pooling.h"
 #include "operators/reduce_mean.h"
@@ -63,6 +69,8 @@ void export_values(py::module &m) {
         .VALUE(OpType, Conv)
         .VALUE(OpType, Matmul)
         .VALUE(OpType, ConvTrans)
+        .VALUE(OpType, ConvTransNHWC)
+        .VALUE(OpType, ConvNHWC)
         .VALUE(OpType, G2BMM)
         .VALUE(OpType, GBMM)
         .VALUE(OpType, Pad)
@@ -94,8 +102,16 @@ void export_values(py::module &m) {
         .VALUE(OpType, Abs)
         .VALUE(OpType, Resize)
         .VALUE(OpType, Dropout)
+        .VALUE(OpType, Conv2dReduce)
+        .VALUE(OpType, Conv2dReduceTranspose)
+        .VALUE(OpType, MemBound)
+        .VALUE(OpType, Any)
         .export_values();
 
+    py::enum_<TensorType>(m, "TensorType")
+        .VALUE(TensorType, Input)
+        .VALUE(TensorType, Initialized)
+        .VALUE(TensorType, Other);
 #undef VALUE
 }
 
@@ -132,19 +148,34 @@ static Ref<RuntimeObj> intelcpu_runtime() { return make_ref<MklRuntimeObj>(); }
 #endif
 
 static std::tuple<int, int, int, int, int, int> conv_attrs_of(Operator op) {
-    IT_ASSERT(op->getOpType() == OpType::Conv);
-    auto conv = dynamic_cast<const ConvObj *>(op.get());
-    return std::make_tuple(conv->getPh(), conv->getPw(), conv->getDh(),
-                           conv->getDw(), conv->getSh(), conv->getSw());
+    IT_ASSERT(op->getOpType() == OpType::Conv ||
+              op->getOpType() == OpType::ConvNHWC);
+    auto conv = dynamic_cast<const ConvBaseObj *>(op.get());
+    return std::make_tuple(conv->getPh(), conv->getPw(), conv->getSh(),
+                           conv->getSw(), conv->getDh(), conv->getDw());
 }
 
 static std::tuple<int, int, int, int, int, int, int, int>
 conv_trans_attrs_of(Operator op) {
-    IT_ASSERT(op->getOpType() == OpType::ConvTrans);
-    auto conv = dynamic_cast<const ConvTransposed2dObj *>(op.get());
-    auto [oph, opw] = conv->getOutputPadding();
-    return std::make_tuple(conv->getPh(), conv->getPw(), conv->getDh(),
-                           conv->getDw(), conv->getSh(), conv->getSw(), oph,
+    IT_ASSERT(op->getOpType() == OpType::ConvTrans ||
+              op->getOpType() == OpType::ConvTransNHWC);
+    auto conv = dynamic_cast<const ConvBaseObj *>(op.get());
+    int oph, opw;
+
+    if (op->getOpType() == OpType::ConvTrans) {
+        auto _conv = dynamic_cast<const ConvTransposed2dObj *>(op.get());
+        auto output_pad = _conv->getOutputPadding();
+        oph = output_pad.first;
+        opw = output_pad.second;
+    } else {
+        auto _conv = dynamic_cast<const ConvTransposed2dNHWCObj *>(op.get());
+        auto output_pad = _conv->getOutputPadding();
+        oph = output_pad.first;
+        opw = output_pad.second;
+    }
+
+    return std::make_tuple(conv->getPh(), conv->getPw(), conv->getSh(),
+                           conv->getSw(), conv->getDh(), conv->getDw(), oph,
                            opw);
 }
 
@@ -210,6 +241,11 @@ static vector<int64_t> reshape_shape_of(Operator op) {
     return ans;
 }
 
+static int flatten_axis_of(Operator op) {
+    IT_ASSERT(op->getOpType() == OpType::Flatten);
+    return as<FlattenObj>(op)->getAxis();
+}
+
 static vector<int64_t> pad_pads_of(Operator op) {
     IT_ASSERT(op->getOpType() == OpType::Pad);
     auto shape = dynamic_cast<const PadObj *>(op.get())->getPads();
@@ -219,11 +255,20 @@ static vector<int64_t> pad_pads_of(Operator op) {
     return ans;
 }
 
+static string any_kernelName_of(Operator op) {
+    IT_ASSERT(op->getOpType() == OpType::Any);
+    return as<AnyObj>(op)->getKernelName();
+}
+
 static vector<int> transpose_permute_of(Operator op) {
     IT_ASSERT(op->getOpType() == OpType::Transpose);
     return dynamic_cast<const TransposeObj *>(op.get())->getPermute();
 }
 
+static string membound_expr_of(Operator op) {
+    return as<MemBoundObj>(op)->toJson();
+}
+
 void export_functions(py::module &m) {
 #define FUNCTION(NAME) def(#NAME, &NAME)
     m.def("cpu_runtime", &NativeCpuRuntimeObj::getInstance)
@@ -248,29 +293,45 @@ void export_functions(py::module &m) {
         .FUNCTION(reduce_mean_attrs_of)
         .FUNCTION(tensor_dtype)
         .FUNCTION(reshape_shape_of)
+        .FUNCTION(flatten_axis_of)
         .FUNCTION(pad_pads_of)
         .FUNCTION(transpose_permute_of)
         .FUNCTION(concat_axis_of)
         .FUNCTION(split_axis_of)
-        .FUNCTION(gather_axis_of);
+        .FUNCTION(gather_axis_of)
+        .FUNCTION(membound_expr_of)
+        .FUNCTION(any_kernelName_of)
+        .def("membound_hash_of",
+             [](Operator op) { return as<MemBoundObj>(op)->getHash(); });
 #undef FUNCTION
 }
 
 void init_graph_builder(py::module &m) {
     using Handler = GraphHandlerObj;
 
-    py::class_<RuntimeObj, std::shared_ptr<RuntimeObj>>(m, "Runtime");
+    py::class_<Object, Ref<Object>>(m, "_Object")
+        .def("__str__", &Object::toString)
+        .def("guid", &Object::getGuid);
+    py::class_<RuntimeObj, Ref<RuntimeObj>>(m, "Runtime")
+        .def("run", &RuntimeObj::run, "graph"_a, "tune"_a = false,
+             "profiling"_a = false)
+        .def("getPerfTime", &RuntimeObj::getPerfTime, "graph"_a, "profiling"_a,
+             "allowEstimation"_a, "ignoreMemboundOp"_a)
+        .def("timeNonCtcOperators", &RuntimeObj::timeNonCtcOperators);
     py::class_<NativeCpuRuntimeObj, std::shared_ptr<NativeCpuRuntimeObj>,
                RuntimeObj>(m, "CpuRuntime");
 #ifdef USE_CUDA
-    py::class_<CudaRuntimeObj, std::shared_ptr<CudaRuntimeObj>, RuntimeObj>(
-        m, "CudaRuntime");
+    py::class_<CudaRuntimeObj, Ref<CudaRuntimeObj>, RuntimeObj>(m,
+                                                                "CudaRuntime")
+        .def("timeWithCudaGraph",
+             py::overload_cast<Graph, int>(&CudaRuntimeObj::timeWithCudaGraph))
+        .def("setEnableTF32", &CudaRuntimeObj::setEnableTF32);
 #endif
 #ifdef USE_BANG
     py::class_<BangRuntimeObj, std::shared_ptr<BangRuntimeObj>, RuntimeObj>(
         m, "BangRuntime");
 #endif
-    py::class_<TensorObj, std::shared_ptr<TensorObj>>(m, "Tensor")
+    py::class_<TensorObj, std::shared_ptr<TensorObj>, Object>(m, "Tensor")
         .def("fuid", &TensorObj::getFuid, policy::automatic)
         .def("shape", &TensorObj::getDims, policy::move)
         .def("copyin_float", &TensorObj::copyin<float>, policy::move)
@@ -281,8 +342,10 @@ void init_graph_builder(py::module &m) {
         .def("copyout_int64", &TensorObj::copyout<int64_t>, policy::move)
         .def("has_target", &TensorObj::hasTarget, policy::automatic)
         .def("src", &TensorObj::getSource, policy::move)
-        .def("printData", &TensorObj::printData, policy::automatic);
-    py::class_<OperatorObj, std::shared_ptr<OperatorObj>>(m, "Operator")
+        .def("print_data", &TensorObj::printData)
+        .def("data_malloc", &TensorObj::dataMalloc)
+        .def("getTensorType", &TensorObj::getTensorType);
+    py::class_<OperatorObj, std::shared_ptr<OperatorObj>, Object>(m, "Operator")
         .def("op_type", &OperatorObj::getOpType, policy::automatic)
         .def("inputs", py::overload_cast<>(&OperatorObj::getInputs, py::const_),
              policy::reference)
@@ -291,9 +354,16 @@ void init_graph_builder(py::module &m) {
              policy::reference);
     py::class_<Handler>(m, "GraphHandler")
         .def(py::init<Runtime>())
-        .def("tensor", &Handler::tensor, policy::move)
+        .def(py::init<Graph>())
+        .def("inputs", &Handler::inputs, policy::move)
+        .def("outputs", &Handler::outputs, policy::move)
+        .def("tensor", &Handler::tensor, policy::move, "shape"_a, "dtype"_a = 1,
+             "tensor_type"_a = TensorType::Other)
         .def("conv", &Handler::conv, policy::move)
         .def("convTransposed2d", &Handler::convTransposed2d, policy::move)
+        .def("convNHWC", &Handler::convNHWC, policy::move)
+        .def("convtransposed2dNHWC", &Handler::convTransposed2dNHWC,
+             policy::move)
         .def("matmul", &Handler::matmul, policy::move)
         .def("batchNorm", &Handler::batchNorm, policy::move)
         .def("maxPool", &Handler::maxPool, policy::move)
@@ -321,11 +391,50 @@ void init_graph_builder(py::module &m) {
         .def("reduce_mean", &Handler::reduceMean, policy::move)
         .def("slice", &Handler::slice, policy::move)
         .def("pad", &Handler::pad, policy::move)
+        .def("memBound", &Handler::memBound, policy::move)
         .def("topo_sort", &Handler::topo_sort, policy::automatic)
         .def("optimize", &Handler::optimize, policy::automatic)
         .def("operators", &Handler::operators, policy::move)
         .def("data_malloc", &Handler::data_malloc, policy::automatic)
-        .def("run", &Handler::run, policy::automatic);
+        .def("run", &Handler::run, policy::automatic)
+        .def("getGraph", &Handler::getGraph);
+    py::class_<Mutator, Ref<Mutator>>(m, "Mutator").def("run", &Mutator::run);
+    py::enum_<NMutator::Mode>(m, "NMutatorMode")
+        .value("Normal", NMutator::Mode::Normal)
+        .value("RuleBased", NMutator::Mode::RuleBased);
+    py::class_<NMutator, Ref<NMutator>, Mutator>(m, "NMutator")
+        .def(py::init<NMutator::Mode>())
+        .def(py::init<NMutator::Mode, vector<int>>())
+        .def("run", &NMutator::run);
+    py::class_<SearchEngine>(m, "SearchEngine")
+        .def(py::init<Runtime, Ref<Mutator>>())
+        .def("run", &SearchEngine::run);
+    py::class_<GraphObj, Ref<GraphObj>, Object>(m, "Graph")
+        .def("tensors", &GraphObj::getTensors)
+        .def("operators", &GraphObj::getOperators)
+        .def("inputs", &GraphObj::getInputs)
+        .def("outputs", &GraphObj::getOutputs)
+        .def("print", &GraphObj::print)
+        .def("topo_sort", &GraphObj::topo_sort);
+}
+
+void export_test_model(py::module &m) {
+#ifdef USE_CUDA
+    m.def("runInfoGAN", &runInfoGAN)
+        .def("getGANGraph", &getGANGraph)
+        .def("getFSRCNNGraph", &getFSRCNNGraph)
+        .def("getLongformer", &getLongformer)
+        .def("getConvtransposedNHWC", &getConvtransposedNHWC)
+        .def("optimizeGraph", &optimizeGraph, "graph"_a, "runtime"_a,
+             "tuning"_a = false, "mode"_a = NMutator::Mode::Normal,
+             "rules"_a = vector<int>{})
+        .def("initializeGraphTensors", &initializeGraphTensors, "g"_a,
+             "l"_a = -0.1, "r"_a = 0.1, "useInt"_a = false)
+        .def("convertNCHWtoNHWCModel", &convertNCHWtoNHWCModel)
+        .def("optimizeWithDepthConstraint", &optimizeWithDepthConstraint)
+        .def("optimizeModel", &optimizeModel)
+        .def("optimizeModelWithRules", &optimizeModelWithRules);
+#endif
 }
 
 } // namespace infini
@@ -335,4 +444,5 @@ PYBIND11_MODULE(backend, m) {
     infini::export_values(m);
     infini::export_functions(m);
     infini::init_graph_builder(m);
+    infini::export_test_model(m);
 }

src/kernels/cuda/G2BMM.cc

@@ -33,7 +33,7 @@ class G2BMMCudnn : public CudaKernelWithoutConfig {
         auto record =
             make_ref<PerfRecordObj>(std::numeric_limits<double>::max());
         const auto [warmupRounds, timingRounds] =
-            op->getB() > 100 ? tuple{1, 3} : tuple{5, 15};
+            op->getB() > 100 ? tuple{1, 1} : tuple{1, 2};
         double tmp =
             timeit([&]() { g2bmmKernel(op, context); },
                    [&]() { context->sync(); }, warmupRounds, timingRounds);

src/kernels/cuda/GBMM.cc

@@ -34,7 +34,7 @@ class GBMMCudnn : public CudaKernelWithoutConfig {
         auto record =
             make_ref<PerfRecordObj>(std::numeric_limits<double>::max());
         const auto [warmupRounds, timingRounds] =
-            op->getB() > 100 ? tuple{1, 3} : tuple{5, 15};
+            op->getB() > 100 ? tuple{1, 1} : tuple{1, 3};
         double tmp =
             timeit([&]() { gbmmKernel(op, context); },
                    [&]() { context->sync(); }, warmupRounds, timingRounds);

src/kernels/cuda/any.cc

@@ -0,0 +1,88 @@
+#include "operators/any.h"
+#include "cuda/cuda_any.h"
+#include "cuda/cuda_conv2dreduce.h"
+#include "cuda/cuda_kernel_wihtout_config.h"
+#include "cuda/cuda_runtime.h"
+
+namespace infini {
+
+class AnyCuda : public CudaKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<AnyObj>(_op);
+
+        auto inputs = op->getInputs();
+        auto outputs = op->getOutputs();
+
+        vector<float *> inputsRawPtr;
+        for (auto &input : inputs) {
+            inputsRawPtr.emplace_back(input->getRawDataPtr<float *>());
+        }
+        vector<float *> outputsRawPtr;
+        for (auto &output : outputs) {
+            outputsRawPtr.emplace_back(output->getRawDataPtr<float *>());
+        }
+
+        any_kernel_mapping(inputsRawPtr, outputsRawPtr, op->getKernelName(),
+                           op->getOpAttrVector());
+    }
+};
+
+void any_kernel_mapping(vector<float *> inputs, vector<float *> outputs,
+                        const string &kernelName, const vector<int> &attr) {
+    if (kernelName == "conv2dreduce_kernel") {
+        IT_ASSERT(attr.size() == 15);
+        IT_ASSERT(inputs.size() == 1 || inputs.size() == 2)
+        IT_ASSERT(outputs.size() == 1);
+        conv2dreduce_kernel(inputs[0], inputs.size() > 1 ? inputs[1] : nullptr,
+                            outputs[0], attr[0] != 0, attr[1], attr[2], attr[3],
+                            attr[4], attr[5], attr[6], attr[7], attr[8],
+                            attr[9], attr[10], attr[11], attr[12], attr[13],
+                            attr[14]);
+    } else if (kernelName == "reduceConvRxSToNCHW") {
+        IT_ASSERT(attr.size() == 15);
+        IT_ASSERT(inputs.size() == 1 || inputs.size() == 2)
+        IT_ASSERT(outputs.size() == 1);
+        // float *input, float *bias, float *output, int act,
+        //                          int n, int h, int w, int f, int r, int s,
+        //                          int oh, int ow, int ph, int pw, int sh, int
+        //                          sw, int dh, int dw
+        reduceConvRxSToNCHW(inputs[0], inputs.size() > 1 ? inputs[1] : nullptr,
+                            outputs[0], attr[0], attr[1], attr[2], attr[3],
+                            attr[4], attr[5], attr[6], attr[7], attr[8],
+                            attr[9], attr[10], attr[11], attr[12], attr[13],
+                            attr[14]);
+    } else if (kernelName == "convTranspose2dreduce_kernel") {
+        IT_ASSERT(attr.size() == 15);
+        IT_ASSERT(inputs.size() == 1 || inputs.size() == 2)
+        IT_ASSERT(outputs.size() == 1);
+        convTranspose2dreduce_kernel(
+            inputs[0], inputs.size() > 1 ? inputs[1] : nullptr, outputs[0],
+            attr[0] != 0, attr[1], attr[2], attr[3], attr[4], attr[5], attr[6],
+            attr[7], attr[8], attr[9], attr[10], attr[11], attr[12], attr[13],
+            attr[14]);
+    } else if (kernelName == "conv5x5ToConv3x3Reduce") {
+        IT_ASSERT(attr.size() == 4);
+        IT_ASSERT(inputs.size() == 1 || inputs.size() == 2)
+        IT_ASSERT(outputs.size() == 1);
+        conv5x5ToConv3x3Reduce(attr[0], attr[1], attr[2], attr[3], inputs[0],
+                               outputs[0],
+                               inputs.size() > 1 ? inputs[1] : nullptr);
+    } else if (kernelName == "conv3x3ToReduce") {
+        IT_ASSERT(attr.size() == 4);
+        IT_ASSERT(inputs.size() == 1 || inputs.size() == 2);
+        IT_ASSERT(outputs.size() == 1);
+        conv3x3ToReduce(attr[0], attr[1], attr[2], attr[3], inputs[0],
+                        outputs[0], inputs.size() > 1 ? inputs[1] : nullptr);
+    } else if (kernelName == "FakeOp" || kernelName == "Reduce3x3Offset_hint") {
+    } else {
+        std::cout << "Unimplemented AnyOp cuda kernel: " << kernelName
+                  << std::endl;
+        IT_TODO_HALT();
+    }
+}
+
+REGISTER_KERNEL(Device::CUDA, OpType::Any, DataType::Float32, AnyCuda,
+                "Any_CUDA_Float32");
+
+} // namespace infini

src/kernels/cuda/clip.cc

@@ -24,4 +24,4 @@ class ClipCuda : public CudaKernelWithoutConfig {
 REGISTER_KERNEL(Device::CUDA, OpType::Clip, DataType::Float32, ClipCuda,
                 "Clip_CUDA_Float32");
 
-}; // namespace infini
+} // namespace infini

src/kernels/cuda/clip.cu

@@ -1,9 +1,7 @@
 #include "core/common.h"
-#include "core/constants.h"
 #include "cuda/cuda_common.h"
 #include <math.h>
 
-using infini::E_CONSTANT;
 constexpr unsigned int num_threads() { return 32 * 4; }
 constexpr int thread_work_size() { return 4; }
 constexpr int block_work_size() { return thread_work_size() * num_threads(); }
@@ -29,4 +27,4 @@ void clip_kernel(float *input, float *output, int num, float minValue,
                                           maxValue);
 }
 
-}; // namespace infini
+} // namespace infini

src/kernels/cuda/conv.cc

@@ -1,4 +1,5 @@
 #include "operators/conv.h"
+#include "core/graph.h"
 #include "core/kernel.h"
 #include "cuda/cuda_runtime.h"
 #include <chrono>
@@ -52,7 +53,7 @@ class convCudnn : public Kernel {
                cudnnFilterDescriptor_t, cudnnTensorDescriptor_t,
                cudnnConvolutionDescriptor_t, cudnnActivationDescriptor_t,
                cudnnTensorDescriptor_t>
-    createCuDNNDescriptor(const Ref<ConvObj> &op,
+    createCuDNNDescriptor(const Ref<ConvBaseObj> &op,
                           const ConvCuDnnPerfRecord &record) const {
         void *const inData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const knData = (op->getInputs(1)->getRawDataPtr<void *>());
@@ -68,15 +69,23 @@ class convCudnn : public Kernel {
 
         int channelsPerGrp = cpg, channels = c;
 
+        // set input format
+        cudnnTensorFormat_t tensorFormat = (op->getOpType() == OpType::ConvNHWC)
+                                               ? CUDNN_TENSOR_NHWC
+                                               : CUDNN_TENSOR_NCHW;
+
         // get inputs
         cudnnTensorDescriptor_t inDesc;
         checkCudnnError(cudnnCreateTensorDescriptor(&inDesc));
         checkCudnnError(cudnnSetTensor4dDescriptor(
-            inDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, n, channels, h, w));
+            inDesc, tensorFormat, CUDNN_DATA_FLOAT, n, channels, h, w));
 
         // get kernels
         cudnnFilterDescriptor_t knDesc;
         checkCudnnError(cudnnCreateFilterDescriptor(&knDesc));
+        // FIXME: filter data layout is not changed with input data layout
+        // since FCRS shows better performance for NHWC inputs in some cases.
+        // This should be tunable.
         checkCudnnError(cudnnSetFilter4dDescriptor(knDesc, CUDNN_DATA_FLOAT,
                                                    CUDNN_TENSOR_NCHW, f,
                                                    channelsPerGrp, r, s));
@@ -84,7 +93,7 @@ class convCudnn : public Kernel {
         cudnnTensorDescriptor_t biasDesc;
         checkCudnnError(cudnnCreateTensorDescriptor(&biasDesc));
         checkCudnnError(cudnnSetTensor4dDescriptor(
-            biasDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, f, 1, 1));
+            biasDesc, tensorFormat, CUDNN_DATA_FLOAT, 1, f, 1, 1));
 
         // get convlution descriptor
         cudnnConvolutionDescriptor_t convDesc;
@@ -125,18 +134,25 @@ class convCudnn : public Kernel {
             convDesc, inDesc, knDesc, &outn, &outc, &outh, &outw));
         cudnnTensorDescriptor_t outDesc;
         checkCudnnError(cudnnCreateTensorDescriptor(&outDesc));
-        checkCudnnError(cudnnSetTensor4dDescriptor(outDesc, CUDNN_TENSOR_NCHW,
-                                                   CUDNN_DATA_FLOAT, outn, outc,
-                                                   outh, outw));
-        IT_ASSERT((vector{outn, outc, outh, outw}) ==
-                      op->getOutput()->getDims(),
-                  "cuDNN output shape mismatches with OP output shape");
+        checkCudnnError(cudnnSetTensor4dDescriptor(
+            outDesc, tensorFormat, CUDNN_DATA_FLOAT, outn, outc, outh, outw));
+
+        if (op->getOpType() == OpType::ConvNHWC) {
+            IT_ASSERT((vector{outn, outh, outw, outc}) ==
+                          op->getOutput()->getDims(),
+                      "cuDNN output shape mismatches with OP output shape");
+        } else {
+            IT_ASSERT((vector{outn, outc, outh, outw}) ==
+                          op->getOutput()->getDims(),
+                      "cuDNN output shape mismatches with OP output shape");
+        }
 
         return tuple(inData, knData, outData, inDesc, knDesc, biasDesc,
                      convDesc, actDesc, outDesc);
     }
 
-    bool cuDNNUnfused(const Ref<ConvObj> &op, const ConvCuDnnPerfRecord &record,
+    bool cuDNNUnfused(const Ref<ConvBaseObj> &op,
+                      const ConvCuDnnPerfRecord &record,
                       const CudaRuntimeObj *context) const {
         cudnnStatus_t stat;
 
@@ -219,12 +235,14 @@ class convCudnn : public Kernel {
                     const RuntimeObj *_context) const override {
         ConvCuDnnPerfRecordObj ret;
         ret.time = std::numeric_limits<double>::max();
-        auto context = dynamic_cast<const CudaRuntimeObj *>(_context);
-        auto op = as<ConvObj>(_op);
+        auto context = const_cast<CudaRuntimeObj *>(
+            dynamic_cast<const CudaRuntimeObj *>(_context));
+        auto op = as<ConvBaseObj>(_op);
+        int try_algo = op->getOpType() == OpType::ConvNHWC ? 2 : N_ALGO;
         // Both modes have the same performance. Only run cross-correlation.
         for (int mode = 1; mode < 2; mode++) {
             // Try every possible algorithm of convolution
-            for (int algo = 0; algo < N_ALGO; algo++) {
+            for (int algo = 0; algo < try_algo; algo++) {
                 auto recordRef = make_ref<ConvCuDnnPerfRecordObj>();
                 auto &record = *recordRef;
                 record.mode = mode;
@@ -251,16 +269,15 @@ class convCudnn : public Kernel {
                     record.workspaceSize, &beta, outDesc, outData);
                 if (stat != CUDNN_STATUS_SUCCESS)
                     continue;
-                record.time = timeit(
-                    [&]() {
-                        cudnnConvolutionForward(context->cudnnHandle(), &alpha,
-                                                inDesc, inData, knDesc, knData,
-                                                convDesc, ALGOS[record.algo],
-                                                wsData, record.workspaceSize,
-                                                &beta, outDesc, outData);
-                    },
-                    [&]() { context->sync(); });
-                // printf("mode:%d algo:%d :%.8lf\n", mode, algo, record.time);
+                // Time the kernel with CUDA Graph to get a precise time
+                std::function<void(void)> func = [&]() {
+                    cudnnConvolutionForward(
+                        context->cudnnHandle(), &alpha, inDesc, inData, knDesc,
+                        knData, convDesc, ALGOS[record.algo], wsData,
+                        record.workspaceSize, &beta, outDesc, outData);
+                };
+                record.time = context->timeWithCudaGraph({func}, 100);
+                // printf("mode:%d algo:%d :%.4lf\n", mode, algo, record.time);
 
                 // Update the tune result
                 if (ret.time > record.time)
@@ -283,7 +300,7 @@ class convCudnn : public Kernel {
 
     void compute(const Operator &_op, const PerfRecord &_record,
                  const RuntimeObj *_context) const override {
-        auto op = as<ConvObj>(_op);
+        auto op = as<ConvBaseObj>(_op);
         auto record = as<ConvCuDnnPerfRecordObj>(_record);
         auto context = dynamic_cast<const CudaRuntimeObj *>(_context);
         bool success = cuDNNUnfused(op, record, context);
@@ -294,5 +311,8 @@ class convCudnn : public Kernel {
 REGISTER_KERNEL(Device::CUDA, OpType::Conv, DataType::Float32, convCudnn,
                 "Conv_cuDNN_CUDA_Float32");
 
+REGISTER_KERNEL(Device::CUDA, OpType::ConvNHWC, DataType::Float32, convCudnn,
+                "ConvNHWC_cuDNN_CUDA_Float32");
+
 REGISTER_CONSTRUCTOR(1, ConvCuDnnPerfRecordObj::from_json);
 } // namespace infini

src/kernels/cuda/conv2dreduce.cc

@@ -0,0 +1,44 @@
+#include "operators/conv2dreduce.h"
+#include "cuda/cuda_conv2dreduce.h"
+#include "cuda/cuda_kernel_wihtout_config.h"
+#include "cuda/cuda_runtime.h"
+
+namespace infini {
+
+class Conv2dReduceCuda : public CudaKernelWithoutConfig {
+    void compute(const Operator &_op, const RuntimeObj *_context) const {
+        auto op = as<Conv2dReduceBase>(_op);
+        float *const input = (op->getInputs(0)->getRawDataPtr<float *>());
+        float *const bias =
+            op->getBias() ? (op->getBias()->getRawDataPtr<float *>()) : nullptr;
+        float *const output = (op->getOutput()->getRawDataPtr<float *>());
+
+        auto dim = op->getInputs(0)->getDims();
+        int n = dim[0], h = dim[1], w = dim[2], f = dim[3], r = dim[4],
+            s = dim[5];
+        int dh = op->getDh(), dw = op->getDw();
+        int sh = op->getSh(), sw = op->getSw();
+        int ph = op->getPh(), pw = op->getPw();
+        auto odim = op->getOutput()->getDims();
+        int oh = odim[1], ow = odim[2];
+        bool PReLU = op->getPReLU();
+        // float paramReLU = op->getParamReLU();
+
+        auto opType = op->getOpType();
+
+        if (opType == OpType::Conv2dReduce) {
+            conv2dreduce_kernel(input, bias, output, PReLU, n, h, w, f, r, s,
+                                oh, ow, ph, pw, sh, sw, dh, dw);
+        } else {
+            convTranspose2dreduce_kernel(input, bias, output, PReLU, n, h, w, f,
+                                         r, s, oh, ow, ph, pw, sh, sw, dh, dw);
+        }
+    }
+};
+
+REGISTER_KERNEL(Device::CUDA, OpType::Conv2dReduce, DataType::Float32,
+                Conv2dReduceCuda, "Conv2dReduce_CUDA_Float32");
+REGISTER_KERNEL(Device::CUDA, OpType::Conv2dReduceTranspose, DataType::Float32,
+                Conv2dReduceCuda, "Conv2dReduceTranspose_CUDA_Float32");
+
+} // namespace infini

src/kernels/cuda/conv2dreduce.cu

@@ -0,0 +1,239 @@
+#include "cuda/cuda_common.h"
+#include "nnet/dbg.h"
+
+using dtype = float;
+
+__global__ void conv2dreduce_kernel_(float *__restrict__ input,
+                                     float *__restrict__ bias,
+                                     float *__restrict__ output,
+                                     const bool PReLU, const int n, const int f,
+                                     const int h, const int w, const int oh,
+                                     const int ow, const int r, const int s,
+                                     const int ph, const int pw, const int dh,
+                                     const int dw, const int sh, const int sw) {
+    // output shape: (n, oh, ow, f)
+    // input shape: (n, h, w, f, r, s)
+    int nid = blockIdx.x, fid = blockIdx.y;
+    int hid = threadIdx.x, wid = threadIdx.y;
+    const int fchunck = r * s, wchunk = f * fchunck, hchunk = w * wchunk,
+              nchunck = h * hchunk;
+    float *nfinput = input + nid * nchunck + fid * fchunck;
+    if (nid < n && fid < f && hid < oh && wid < ow) {
+        float imm = 0.0;
+        int ihst = hid * sh - ph;
+        int iwst = wid * sw - pw;
+        for (int ri = 0; ri < r; ++ri) {
+            for (int si = 0; si < s; ++si) {
+                int ihid = ihst + ri * dh;
+                int iwid = iwst + si * dw;
+                if (ihid >= 0 && ihid < h && iwid >= 0 && iwid < w) {
+                    imm += *(nfinput + ihid * hchunk + iwid * wchunk + ri * s +
+                             si);
+                }
+            }
+        }
+        if (bias) {
+            imm += bias[fid];
+        }
+        if (PReLU) {
+            imm = imm > 0.0 ? imm : 0.0;
+        }
+        output[nid * (oh * ow * f) + hid * (ow * f) + wid * f + fid] = imm;
+    }
+}
+__global__ void convTranspose2dreduce_kernel2_(
+    float *__restrict__ input, float *__restrict__ bias,
+    float *__restrict__ output, const bool PReLU, const int n, const int f,
+    const int h, const int w, const int oh, const int ow, const int r,
+    const int s, const int ph, const int pw, const int dh, const int dw,
+    const int sh, const int sw) {
+    int warp_id = (blockDim.x / 32) * blockIdx.x + threadIdx.x / 32;
+    int lane = threadIdx.x % 32;
+    int nid = warp_id / (f * oh * ow);
+    int fid = (warp_id - nid * (f * oh * ow)) / (oh * ow);
+    int hid = (warp_id - nid * (f * oh * ow) - fid * (oh * ow)) / ow;
+    int wid = warp_id % ow;
+    if (hid >= oh || wid >= ow || nid > n || fid > f)
+        return;
+
+    const int fchunck = r * s, wchunk = f * fchunck, hchunk = w * wchunk,
+              nchunck = h * hchunk;
+    float *nfinput = input + nid * nchunck + fid * fchunck;
+    // view as conv, the true ph and pw
+    int tph = r - ph - 1, tpw = s - pw - 1;
+    int th = (h - 1) * sh + 1, tw = (w - 1) * sw + 1;
+
+    float imm = 0.0;
+    int ihst = hid - tph;
+    int iwst = wid - tpw;
+    for (int idx = lane; idx < r * s; idx += 32) {
+        int ri = idx / s;
+        int si = idx % s;
+        int ihid = ihst + r - ri - 1;
+        int iwid = iwst + s - si - 1;
+        if (ihid >= 0 && ihid < th && iwid >= 0 && iwid < tw &&
+            (ihid % sh == 0) && (iwid % sw == 0)) {
+            imm += *(nfinput + (ihid / sh) * hchunk + (iwid / sw) * wchunk +
+                     ri * s + si);
+        }
+    }
+
+    for (int k = 16; k > 0; k >>= 1) {
+        imm += __shfl_down_sync(0xffffffff, imm, k); // sum
+    }
+    if (lane == 0) {
+        if (bias) {
+            imm += bias[fid];
+        }
+        if (PReLU) {
+            imm = imm > 0.0 ? imm : 0.0;
+        }
+        output[nid * (oh * ow * f) + hid * (ow * f) + wid * f + fid] = imm;
+    }
+}
+
+__global__ void convTranspose2dreduce_kernel_(
+    float *__restrict__ input, float *__restrict__ bias,
+    float *__restrict__ output, const bool PReLU, const int n, const int f,
+    const int h, const int w, const int oh, const int ow, const int r,
+    const int s, const int ph, const int pw, const int dh, const int dw,
+    const int sh, const int sw, const int block_x_num, const int block_y_num) {
+    // assert dh = dw = 1
+    int nid = blockIdx.x / block_x_num, fid = blockIdx.y / block_y_num;
+    int hid = (blockIdx.x % block_x_num) * blockDim.x + threadIdx.x,
+        wid = (blockIdx.y % block_y_num) * blockDim.y + threadIdx.y;
+    if (hid >= oh || wid >= ow)
+        return;
+    const int fchunck = r * s, wchunk = f * fchunck, hchunk = w * wchunk,
+              nchunck = h * hchunk;
+    float *nfinput = input + nid * nchunck + fid * fchunck;
+    // view as conv, the true ph and pw
+    int tph = r - ph - 1, tpw = s - pw - 1;
+    int th = (h - 1) * sh + 1, tw = (w - 1) * sw + 1;
+    if (nid < n && fid < f && hid < oh && wid < ow) {
+        float imm = 0.0;
+        int ihst = hid - tph;
+        int iwst = wid - tpw;
+        for (int ri = 0; ri < r; ++ri) {
+            for (int si = 0; si < s; ++si) {
+                int ihid = ihst + r - ri - 1;
+                int iwid = iwst + s - si - 1;
+                if (ihid >= 0 && ihid < th && iwid >= 0 && iwid < tw &&
+                    (ihid % sh == 0) && (iwid % sw == 0)) {
+                    imm += *(nfinput + (ihid / sh) * hchunk +
+                             (iwid / sw) * wchunk + ri * s + si);
+                }
+            }
+        }
+        if (bias) {
+            imm += bias[fid];
+        }
+        if (PReLU) {
+            imm = imm > 0.0 ? imm : 0.0;
+        }
+        output[nid * (oh * ow * f) + hid * (ow * f) + wid * f + fid] = imm;
+    }
+}
+
+// nhwrsc -> nhwc
+__global__ void reduce_4x4(dtype *in, dtype *out, int act, const int N,
+                           const int F, const int H, const int W, const int IH,
+                           const int IW) {
+// #define in_index(n, h, w, r, s, f)                                             \
+//     ((((((n)*IH + h) * IW + w) * R + r) * S + s) * F + f)
+#define in_index(n, h, w, f, r, s)                                             \
+    ((((((n)*IH + h) * IW + w) * F + f) * R + r) * S + s)
+#define out_index(n, h, w, f) (((((n)*H) + (h)) * W + (w)) * F + (f))
+    const int R = 4, S = 4;
+    const int n_tasks = N * F * H * W;
+    int start = threadIdx.x + blockDim.x * blockIdx.x;
+    int stride = blockDim.x * gridDim.x;
+    for (int i = start; i < n_tasks; i += stride) {
+        int t = i, n, f, h, w;
+        f = t % F;
+        t /= F;
+        w = t % W;
+        t /= W;
+        h = t % H;
+        t /= H;
+        n = t;
+
+        // unroll this 2-iter loop
+        float sum = 0;
+        int x, y;
+        for (int r = (h + 1) & 1; r < R; r += 2) {
+            x = (h + 1 - r) / 2;
+            if (x >= 0 && x < IH) {
+                for (int s = (w + 1) & 1; s < S; s += 2) {
+                    y = (w + 1 - s) / 2;
+                    if (y >= 0 && y < IW) {
+                        sum += in[in_index(n, x, y, f, r, s)];
+                        // if (i==0)
+                        //     printf("TTT nhwf= %d,%d,%d,%d x=%d y=%d, v=%f,
+                        //     index=%d, rsf %d %d %d\n", n, h, w,
+                        //            f, x, y, in[in_index(n, x, y, r, s, f)],
+                        //            in_index(n, x, y, r, s, f), r,s,f);
+                    }
+                }
+            }
+        }
+        if (act == 0) {
+            out[out_index(n, h, w, f)] = sum;
+        } else if (act == 1) { // Relu
+            out[out_index(n, h, w, f)] = sum > 0 ? sum : 0;
+        } else if (act == 2) {
+            out[out_index(n, h, w, f)] = tanhf(sum);
+        }
+    }
+#undef in_index
+#undef out_index
+}
+
+namespace infini {
+
+void conv2dreduce_kernel(float *input, float *bias, float *output, bool PReLU,
+                         int n, int h, int w, int f, int r, int s, int oh,
+                         int ow, int ph, int pw, int sh, int sw, int dh,
+                         int dw) {
+    dim3 grid(n, f);
+    dim3 block(oh, ow);
+    // cudaStream_t stream(cudaStreamPerThread);
+    conv2dreduce_kernel_<<<grid, block, 0>>>(input, bias, output, PReLU, n, f,
+                                             h, w, oh, ow, r, s, ph, pw, dh, dw,
+                                             sh, sw);
+}
+
+void convTranspose2dreduce_kernel(float *input, float *bias, float *output,
+                                  int act, int n, int h, int w, int f, int r,
+                                  int s, int oh, int ow, int ph, int pw, int sh,
+                                  int sw, int dh, int dw) {
+    dim3 grid(n, f);
+    dim3 block(oh, ow);
+    // cudaStream_t stream(cudaStreamPerThread);
+    // puts("convTranspose2dreduce_kernel is executed");
+    if (r == 4 && s == 4 && sh == 2 && sw == 2) {
+        const int M = r * s * f, N = n * h * w;
+        reduce_4x4<<<(M * N + 127) / 128, 128>>>(input, output, act, n, f, oh,
+                                                 ow, h, w);
+    } else {
+        // puts("why use this conv2dreduce");
+        // block.x = 32;
+        // block.y = 32;
+        // int block_x_num = (oh + block.x - 1) / block.x;
+        // int block_y_num = (ow + block.y - 1) / block.y;
+        // grid.x = n * (block_x_num);
+        // grid.y = f * (block_y_num);
+        // convTranspose2dreduce_kernel_<<<grid, block, 0>>>(
+        //     input, bias, output, (bool)act, n, f, h, w, oh, ow, r, s, ph, pw,
+        //     dh, dw, sh, sw, block_x_num, block_y_num);
+
+        block.x = 128;
+        block.y = 1;
+        grid.x = (n * f * ow * oh + block.x / 32 - 1) / (block.x / 32);
+        grid.y = 1;
+        convTranspose2dreduce_kernel2_<<<grid, block, 0>>>(
+            input, bias, output, (bool)act, n, f, h, w, oh, ow, r, s, ph, pw,
+            dh, dw, sh, sw);
+    }
+}
+} // namespace infini

src/kernels/cuda/matmul.cc

@@ -6,6 +6,8 @@ namespace infini {
 
 struct MatmulCublasPerfRecordObj : public PerfRecordObj {
     int algo = CUBLAS_GEMM_DEFAULT;
+    /// @brief 0 for cublasGemmStridedBatchedEx, 1 for cublasGemmEx
+    int apiId = 0;
     void to_json(json &j) override {
         j["type"] = 2;
         j["data"] = std::make_pair(algo, time);
@@ -19,8 +21,7 @@ struct MatmulCublasPerfRecordObj : public PerfRecordObj {
     }
 };
 
-constexpr int N_ALGO = 24;
-constexpr cublasGemmAlgo_t ALGOS[N_ALGO] = {
+const vector<cublasGemmAlgo_t> Algos = {
     CUBLAS_GEMM_ALGO0,  CUBLAS_GEMM_ALGO1,  CUBLAS_GEMM_ALGO2,
     CUBLAS_GEMM_ALGO3,  CUBLAS_GEMM_ALGO4,  CUBLAS_GEMM_ALGO5,
     CUBLAS_GEMM_ALGO6,  CUBLAS_GEMM_ALGO7,  CUBLAS_GEMM_ALGO8,
@@ -30,6 +31,17 @@ constexpr cublasGemmAlgo_t ALGOS[N_ALGO] = {
     CUBLAS_GEMM_ALGO18, CUBLAS_GEMM_ALGO19, CUBLAS_GEMM_ALGO20,
     CUBLAS_GEMM_ALGO21, CUBLAS_GEMM_ALGO22, CUBLAS_GEMM_ALGO23,
 };
+const vector<cublasGemmAlgo_t> AlgosTensorOp = {
+    CUBLAS_GEMM_DFALT_TENSOR_OP,  CUBLAS_GEMM_ALGO0_TENSOR_OP,
+    CUBLAS_GEMM_ALGO1_TENSOR_OP,  CUBLAS_GEMM_ALGO2_TENSOR_OP,
+    CUBLAS_GEMM_ALGO3_TENSOR_OP,  CUBLAS_GEMM_ALGO4_TENSOR_OP,
+    CUBLAS_GEMM_ALGO5_TENSOR_OP,  CUBLAS_GEMM_ALGO6_TENSOR_OP,
+    CUBLAS_GEMM_ALGO7_TENSOR_OP,  CUBLAS_GEMM_ALGO8_TENSOR_OP,
+    CUBLAS_GEMM_ALGO9_TENSOR_OP,  CUBLAS_GEMM_ALGO10_TENSOR_OP,
+    CUBLAS_GEMM_ALGO11_TENSOR_OP, CUBLAS_GEMM_ALGO12_TENSOR_OP,
+    CUBLAS_GEMM_ALGO13_TENSOR_OP, CUBLAS_GEMM_ALGO14_TENSOR_OP,
+    CUBLAS_GEMM_ALGO15_TENSOR_OP};
+
 class matmulCublas : public Kernel {
     bool do_compute(const Operator &_op, const PerfRecord &_record,
                     const RuntimeObj *_context) const {
@@ -47,9 +59,12 @@ class matmulCublas : public Kernel {
         const int lda = op->getTransA() ? m : k, ldb = op->getTransB() ? k : n,
                   ldc = n;
         const float alpha = 1.f, beta = 0.f;
-        // TODO:use compute type
         cublasStatus_t stat;
-        if (b > 1) {
+        // Set the compute type to TF32 if enabled
+        cublasComputeType_t computeType = context->getEnableTF32()
+                                              ? CUBLAS_COMPUTE_32F_FAST_TF32
+                                              : CUBLAS_COMPUTE_32F;
+        if (record->apiId == 0) {
             // Support batch broadcast with zero stride
             int dimA = op->getInputs(0)->getDims().size();
             int dimB = op->getInputs(1)->getDims().size();
@@ -63,17 +78,23 @@ class matmulCublas : public Kernel {
                  (dimB == 3 && op->getInputs(1)->getDims()[0] == 1))
                     ? 0 // Broadcast the batch dimension if batch size is 1
                     : n * k;
+            // printf("cublasGemmStridedBatchedEx %d%d, mnk %d %d %d, alpha %f,
+            // B "
+            //        "%d %lld, A %d %lld, C %d %d, b %d %d\n",
+            //        opB, opA, n, m, k, alpha, ldb, strideB, lda, strideA, ldc,
+            //        m * n, b, record->algo);
             stat = cublasGemmStridedBatchedEx(
                 context->cublasHandle(), opB, opA, n, m, k, &alpha, inBData,
                 CUDA_R_32F, ldb, strideB, inAData, CUDA_R_32F, lda, strideA,
-                &beta, outData, CUDA_R_32F, ldc, m * n, b, CUDA_R_32F,
+                &beta, outData, CUDA_R_32F, ldc, m * n, b, computeType,
                 (cublasGemmAlgo_t)record->algo);
-        } else {
+        } else if (record->apiId == 1) {
             stat = cublasGemmEx(
                 context->cublasHandle(), opB, opA, n, m, k, &alpha, inBData,
                 CUDA_R_32F, ldb, inAData, CUDA_R_32F, lda, &beta, outData,
-                CUDA_R_32F, ldc, CUDA_R_32F, (cublasGemmAlgo_t)record->algo);
-        }
+                CUDA_R_32F, ldc, computeType, (cublasGemmAlgo_t)record->algo);
+        } else
+            IT_ASSERT(false);
         // if (stat != CUBLAS_STATUS_SUCCESS)
         //     cout << cublasGetErrorString(stat);
         return (stat == CUBLAS_STATUS_SUCCESS);
@@ -98,15 +119,29 @@ class matmulCublas : public Kernel {
         IT_ASSERT(op);
         auto ret = make_ref<MatmulCublasPerfRecordObj>();
         ret->time = std::numeric_limits<double>::max();
-        for (int i = 0; i < N_ALGO; i++) {
-            auto rcd = make_ref<MatmulCublasPerfRecordObj>();
-            rcd->algo = ALGOS[i];
-            if (!do_compute(_op, rcd, _context))
-                continue;
-            rcd->time = timeit([&]() { do_compute(_op, rcd, _context); },
-                               [&]() { context->sync(); });
-            if (rcd->time < ret->time)
-                ret = rcd;
+        vector<int> apis{0};
+        if (op->getB() == 1)
+            apis.emplace_back(1);
+
+        // Set the possible algorithm range
+        auto algos = Algos;
+        if (context->getEnableTF32()) {
+            algos.insert(algos.end(), AlgosTensorOp.begin(),
+                         AlgosTensorOp.end());
+        }
+
+        for (int api : apis) {
+            for (size_t i = 0; i < algos.size(); i++) {
+                auto rcd = make_ref<MatmulCublasPerfRecordObj>();
+                rcd->apiId = api;
+                rcd->algo = algos[i];
+                if (!do_compute(_op, rcd, _context))
+                    continue;
+                rcd->time = timeit([&]() { do_compute(_op, rcd, _context); },
+                                   [&]() { context->sync(); });
+                if (rcd->time < ret->time)
+                    ret = rcd;
+            }
         }
         IT_ASSERT(ret->time < std::numeric_limits<double>::max(),
                   "No valid algorithm found for " + op->toString());

src/kernels/cuda/membound_tvm_packed_function.cc

@@ -1,5 +1,6 @@
 #ifdef INFINI_USE_TVM
 #include "core/kernel.h"
+#include "cuda/cuda_conv2dreduce.h"
 #include "cuda/cuda_runtime.h"
 #include "dlpack/dlpack.h"
 #include "ffi/ffi_embed.h"
@@ -8,6 +9,13 @@
 #include "operators/pooling.h"
 #include "tvm/runtime/module.h"
 #include "tvm/runtime/packed_func.h"
+#include <nlohmann/json.hpp>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+using json = nlohmann::json;
 
 namespace py = pybind11;
 
@@ -22,6 +30,8 @@ class TVMRecordObj : public PerfRecordObj {
     std::string dllPath;
     std::string funcName;
     std::vector<int> inputIdx;
+    tvm::runtime::PackedFunc packedFunc;
+    bool useExistingKernel = false;
 };
 
 using TVMRecord = Ref<TVMRecordObj>;
@@ -33,9 +43,15 @@ class MemboundTVMPackedFunction : public Kernel {
         auto op = as<MemBoundObj>(_op);
         // auto context = dynamic_cast<const CudaRuntimeObj *>(_context);
         auto tvmRecord = std::dynamic_pointer_cast<TVMRecordObj>(record);
-        tvm::runtime::PackedFunc packedFunc =
-            getPackedFunction(tvmRecord->dllPath, tvmRecord->funcName);
-        IT_ASSERT(packedFunc != nullptr);
+
+        // Use user-defined kernels
+        if (tvmRecord->useExistingKernel) {
+            bool success = useExistingKernels(op);
+            IT_ASSERT(success);
+            return;
+        }
+
+        tvm::runtime::PackedFunc packedFunc = tvmRecord->packedFunc;
 
         // prepare inputs and outputs
         vector<DLTensorHolder> inputsHolder;
@@ -63,10 +79,18 @@ class MemboundTVMPackedFunction : public Kernel {
     // Premise: op is idempotent since it is called multiple times.
     PerfRecord tune(const Operator &_op,
                     const RuntimeObj *_context) const override {
-        TVMRecord ret = std::make_shared<TVMRecordObj>();
         auto op = as<MemBoundObj>(_op);
         auto context = dynamic_cast<const CudaRuntimeObj *>(_context);
 
+        // If hash matches, use user-defined kernels
+        if (useExistingKernels(op)) {
+            TVMRecord ret = std::make_shared<TVMRecordObj>();
+            ret->time = timeit([&]() { useExistingKernels(op); },
+                               [&]() { context->sync(); });
+            ret->useExistingKernel = true;
+            return ret;
+        }
+
         // invoke Ansor to tune a membound kernel
         auto [expr, hash] = op->getSimplifiedNnetExpr();
         nnet::AsTVMVisitor visitor;
@@ -93,6 +117,7 @@ class MemboundTVMPackedFunction : public Kernel {
                 if (inputName == op->getNnetInputs()[j]->getName())
                     break;
             }
+            IT_ASSERT(j < numInputs, "Cannot find input name: " + inputName);
             inputIdx.emplace_back(j);
         }
 
@@ -114,29 +139,41 @@ class MemboundTVMPackedFunction : public Kernel {
         tvm::runtime::TVMArgs args(preArgs.first.data(), preArgs.second.data(),
                                    preArgs.first.size());
 
+        TVMRecord ret = std::make_shared<TVMRecordObj>();
         ret->time = timeit([&]() { packedFunc.CallPacked(args, &rv); },
                            [&]() { context->sync(); });
         ret->kernelName = kernelName;
         ret->dllPath = dllPath;
         ret->funcName = func;
         ret->inputIdx = inputIdx;
+        ret->packedFunc = packedFunc;
 
-        return std::dynamic_pointer_cast<PerfRecordObj>(ret);
+        return ret;
+    }
+
+    std::string serializeTVMArgs(const std::vector<std::vector<int>> &inDims,
+                                 const std::vector<std::string> &inDTypes,
+                                 const std::vector<int> &outDims,
+                                 const std::string &outDType,
+                                 const std::string &lambda,
+                                 const std::string &funcName,
+                                 const std::string &nnetExprString,
+                                 const std::string &nnetSimplifiedExprString,
+                                 const HashType hashCode) const {
+        json j;
+        // Consistant with python API interface
+        j["input_tensors"] = inDims;
+        j["input_dtypes"] = inDTypes;
+        j["output_tensor"] = outDims;
+        j["output_dtype"] = outDType;
+        j["tvm_code"] = lambda;
+        j["func_name"] = funcName;
+        j["nnet_expression"] = nnetExprString;
+        j["nnet_simplified_expression"] = nnetSimplifiedExprString;
+        j["hash_code"] = std::to_string(hashCode);
+        return j.dump();
     }
 
-    /// @brief
-    /// @param inDims
-    /// @param inDTypes
-    /// @param outDims
-    /// @param outDType
-    /// @param lambda
-    /// @param funcName Generated function name
-    /// @param nnetExpressionString Save expr in string for logging.
-    /// @param nnetSimplifiedExprString Save simplified expr in string for
-    /// logging.
-    /// @param hashCode (optional) Hash code of the input expression for kernel
-    /// cache.
-    /// @return
     std::string getAnsorDLL(const std::vector<std::vector<int>> &inDims,
                             const std::vector<std::string> &inDTypes,
                             const std::vector<int> &outDims,
@@ -146,29 +183,60 @@ class MemboundTVMPackedFunction : public Kernel {
                             const std::string &nnetExprString,
                             const std::string &nnetSimplifiedExprString,
                             const HashType hashCode) const {
-        std::string dllPath;
-        try {
-            start_interpreter();
-            // Use static to avoid re-importing the module. Re-importing results
-            // in cuBLAS failure, whose root cause is not identified yet.
-            static auto func =
-                py::module::import("cpp_plugin").attr("gen_ansor_so");
-            py::tuple code =
-                func(inDims, inDTypes, outDims, outDType, lambda, funcName,
-                     nnetExprString, nnetSimplifiedExprString,
-                     std::to_string(hashCode));
-            dllPath = py::str(code[0]);
-        } catch (py::error_already_set &e) {
-            if (e.matches(PyExc_ImportError)) {
-                std::cerr << "Import Error. Don't forget to set environment "
-                             "variable PYTHONPATH to contain "
-                             "<repo-root>/python"
-                          << std::endl;
-            }
-            throw;
+        int fdP2C[2], fdC2P[2];
+        for (auto fd : {fdP2C, fdC2P}) {
+            int status = pipe(fd);
+            IT_ASSERT(status == 0, "pipe failed");
         }
+        pid_t pid = fork();
+        IT_ASSERT(pid >= 0, "fork failed");
+        if (pid == 0) { // Child process
+            close(fdP2C[1]);
+            close(fdC2P[0]);
 
-        return dllPath;
+            dup2(fdP2C[0], STDIN_FILENO);
+            close(fdP2C[0]);
+
+            string cmd =
+                "from cpp_plugin.gen_ansor_so import pipe_gen; pipe_gen(+" +
+                std::to_string(fdC2P[1]) + ")";
+            const char *const argv[] = {"python3", "-c", cmd.data(), NULL};
+            execvp("python3", const_cast<char *const *>(argv));
+        } else { // Parent process
+            close(fdP2C[0]);
+            close(fdC2P[1]);
+
+            // Write to pipe
+            string serializedArgs = serializeTVMArgs(
+                inDims, inDTypes, outDims, outDType, lambda, funcName,
+                nnetExprString, nnetSimplifiedExprString, hashCode);
+            int status = -1;
+            status =
+                write(fdP2C[1], serializedArgs.data(), serializedArgs.size());
+            IT_ASSERT((size_t)status == serializedArgs.size(),
+                      "Failed to write to pipe");
+            close(fdP2C[1]);
+
+            // Wait for TVM
+            waitpid(pid, &status, 0);
+            IT_ASSERT(WIFEXITED(status), "TVM process was terminated");
+            const int es = WEXITSTATUS(status);
+            IT_ASSERT(es == 0,
+                      "TVM process exit with code " + std::to_string(es));
+
+            // Read from pipe
+            FILE *stream;
+            stream = fdopen(fdC2P[0], "r");
+            char buf_read[257] = {0};
+            status = std::fscanf(stream, "%256c", buf_read);
+            IT_ASSERT(status == 1, "Failed to read from pipe");
+            IT_ASSERT(buf_read[256] == 0, "Pipe buffer overflow");
+            fclose(stream);
+            close(fdC2P[0]);
+            return buf_read;
+        }
+        IT_ASSERT(false, "Should not reach here");
+        return "";
     }
 
     tvm::runtime::PackedFunc getPackedFunction(string path,
@@ -214,6 +282,35 @@ class MemboundTVMPackedFunction : public Kernel {
 
         return {values, type_codes};
     }
+
+    bool useExistingKernels(Ref<MemBoundObj> op) const {
+        return false;
+        const map<HashType, tuple<int, int, int, int, int, int, int, int, int,
+                                  int, int, int, int, int, int>>
+            hashMap = {
+                // clang-format off
+{18446744073661354550ULL, {1, 1, 2, 2, 256, 4, 4, 4, 4, 1, 1, 2, 2, 1, 1}},
+{124145340ULL,            {1, 1, 4, 4, 128, 4, 4, 8, 8, 1, 1, 2, 2, 1, 1}},
+{18446744073695718019ULL, {1, 1, 8, 8, 64, 4, 4, 16, 16, 1, 1, 2, 2, 1, 1}},
+{515085072ULL,            {2, 1, 16, 16, 3, 4, 4, 32, 32, 1, 1, 2, 2, 1, 1}}
+        }; // clang-format on
+        float *input = op->getInputs(0)->getRawDataPtr<float *>();
+        float *bias = nullptr;
+        float *output = op->getOutput()->getRawDataPtr<float *>();
+        if (auto it = hashMap.find(op->getHash()); it != hashMap.end()) {
+            auto &[PReLU, n, h, w, f, r, s, oh, ow, ph, pw, sh, sw, dh, dw] =
+                it->second;
+            IT_ASSERT(op->getInputs(0)->size() ==
+                      size_t(n) * h * w * f * r * s);
+            IT_ASSERT(op->getOutput()->size() == size_t(n) * oh * ow * f);
+            convTranspose2dreduce_kernel(input, bias, output, PReLU, n, h, w, f,
+                                         r, s, oh, ow, ph, pw, sh, sw, dh, dw);
+            return true;
+        }
+        // conv2dreduce_kernel(input, bias, output, PReLU, n, h, w, f, r, s,
+        //                     oh, ow, ph, pw, sh, sw, dh, dw);
+        return false;
+    }
 };
 
 REGISTER_KERNEL(Device::CUDA, OpType::MemBound, DataType::Float32,

src/kernels/cuda/reduceConv.cu

@@ -0,0 +1,287 @@
+#include "core/common.h"
+#include <vector>
+using namespace std;
+
+template <class T>
+__global__ void reduce_merge_conv_3x3_1x1(
+    T *__restrict__ input, T *__restrict__ output, T *__restrict__ bias,
+    const int N, const int H, const int W, const int F, const int N_offset,
+    const int H_offset, const int W_offset, const int F_offset,
+    const int out_N_offset, const int out_F_offset, const int out_H_offset,
+    const int out_W_offset, const int num) {
+    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+    if (tid < num) {
+        int tmptid = tid;
+        const int n = (tmptid / out_N_offset);
+        tmptid -= n * out_N_offset;
+        const int f = tmptid / out_F_offset;
+        tmptid -= f * out_F_offset;
+        const int h = tmptid / out_H_offset;
+        tmptid -= h * out_H_offset;
+        const int w = tmptid / out_W_offset;
+
+        const int noff = n * N_offset;
+        const int hoff = h * H_offset;
+        const int woff = w * W_offset;
+        const int foff = f * F_offset;
+        input += noff + foff + woff + hoff;
+        T res = 0;
+        res += input[4];
+        res += input[9];
+
+        if (h < H - 1) {
+            res += input[H_offset + 7];
+            if (w < W - 1)
+                res += input[H_offset + W_offset + 8];
+            if (w > 0)
+                res += input[H_offset - W_offset + 6];
+        }
+        if (h > 0) {
+            res += input[1 - H_offset];
+            if (w < W - 1)
+                res += input[W_offset - H_offset + 2];
+            if (w > 0)
+                res += input[-1 * H_offset - W_offset];
+        }
+        if (w < W - 1)
+            res += input[5 + W_offset];
+        if (w > 0)
+            res += input[3 - W_offset];
+        output[tid] = max(res + bias[f], 0.f);
+    }
+}
+
+template <class T>
+__global__ void reduce_merge_conv_3x3(
+    T *__restrict__ input, T *__restrict__ output, T *__restrict__ bias,
+    const int N, const int H, const int W, const int F, const int N_offset,
+    const int H_offset, const int W_offset, const int F_offset,
+    const int out_N_offset, const int out_F_offset, const int out_H_offset,
+    const int out_W_offset, const int num, const int act) {
+    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+    if (tid < num) {
+        int tmptid = tid;
+        const int n = (tmptid / out_N_offset);
+        tmptid -= n * out_N_offset;
+        const int f = tmptid / out_F_offset;
+        tmptid -= f * out_F_offset;
+        const int h = tmptid / out_H_offset;
+        tmptid -= h * out_H_offset;
+        const int w = tmptid / out_W_offset;
+
+        const int noff = n * N_offset;
+        const int hoff = h * H_offset;
+        const int woff = w * W_offset;
+        const int foff = f * F_offset;
+        input += noff + foff + woff + hoff;
+        T res = 0;
+        res += input[4];
+
+        if (h < H - 1) {
+            res += input[H_offset + 7];
+            if (w < W - 1)
+                res += input[H_offset + W_offset + 8];
+            if (w > 0)
+                res += input[H_offset - W_offset + 6];
+        }
+        if (h > 0) {
+            res += input[1 - H_offset];
+            if (w < W - 1)
+                res += input[W_offset - H_offset + 2];
+            if (w > 0)
+                res += input[-1 * H_offset - W_offset];
+        }
+        if (w < W - 1)
+            res += input[5 + W_offset];
+        if (w > 0)
+            res += input[3 - W_offset];
+        if (act) {
+            // output[tid] = max(res + bias[f], 0.f);
+            // HACK: temperaly remove bias
+            output[tid] = max(res, 0.f);
+        } else {
+            // output[tid] = res + bias[f];
+            // HACK: temperaly remove bias
+            output[tid] = res;
+        }
+    }
+}
+
+template <class T>
+__global__ void
+reduce_2(T *__restrict__ input, T *__restrict__ output, T *__restrict__ bias,
+         const int N, const int F, const int H, const int W, const int N_offset,
+         const int F_offset, const int H_offset, const int W_offset,
+         const int out_N_offset, const int out_F_offset, const int out_H_offset,
+         const int out_W_offset, const int num) {
+    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+    if (tid < num) {
+        int tmptid = tid;
+        const int n = tmptid / out_N_offset;
+        tmptid -= n * out_N_offset;
+        const int f = tmptid / out_F_offset;
+        tmptid -= f * out_F_offset;
+        const int h = tmptid / out_H_offset;
+        tmptid -= h * out_H_offset;
+        const int w = tmptid / out_W_offset;
+
+        const int noff = n * N_offset;
+        const int foff = f * F_offset * 4;
+        const int hoff = h * H_offset;
+        const int woff = w * W_offset;
+        input += noff + foff + woff + hoff;
+        T res = input[0];
+        if (w != W - 1)
+            res += input[F_offset * 2 + 3];
+        if (h != H - 1) {
+            res += input[F_offset + 3 * H_offset];
+            if (w != W - 1)
+                res += input[F_offset * 3 + 3 * H_offset + 3];
+        }
+        // output[tid] = max(res + bias[f], 0.f);
+        // HACK: temperaly remove bias
+        output[tid] = max(res, 0.f);
+    }
+}
+
+__global__ void reduceConvRxSToNCHWKernel(
+    float *__restrict__ input, float *__restrict__ bias,
+    float *__restrict__ output, const int act, const int n, const int f,
+    const int h, const int w, const int oh, const int ow, const int r,
+    const int s, const int ph, const int pw, const int dh, const int dw) {
+    // input shape: (n, h, w, f, r, s)
+    // output shape: (n, f, oh, ow)
+    const int tid = blockIdx.x * blockDim.x + threadIdx.x;
+    const int out_N_offset = f * oh * ow, out_F_offset = oh * ow,
+              out_H_offset = ow, out_W_offset = 1;
+    const int num = out_N_offset * n;
+    if (tid < num) {
+        // output index
+        int tmptid = tid;
+        const int nid = (tmptid / out_N_offset);
+        tmptid -= nid * out_N_offset;
+        const int fid = tmptid / out_F_offset;
+        tmptid -= fid * out_F_offset;
+        const int hid = tmptid / out_H_offset;
+        tmptid -= hid * out_H_offset;
+        const int wid = tmptid / out_W_offset;
+
+        // Input index
+        const int fchunck = r * s, wchunk = f * fchunck, hchunk = w * wchunk,
+                  nchunck = h * hchunk;
+        float *__restrict__ nfinput = input + nid * nchunck + fid * fchunck;
+        float imm = 0.0;
+        const int ihst = hid, iwst = wid;
+        for (int ri = 0; ri < r; ++ri) {
+            for (int si = 0; si < s; ++si) {
+                int ihid = ihst + (ri - r / 2) * dh;
+                int iwid = iwst + (si - s / 2) * dw;
+                if (ihid >= 0 && ihid < h && iwid >= 0 && iwid < w) {
+                    imm += *(nfinput + ihid * hchunk + iwid * wchunk + ri * s +
+                             si);
+                }
+            }
+        }
+        if (bias) {
+            imm += bias[fid];
+        }
+        if (act) {
+            imm = imm > 0.0 ? imm : 0;
+        }
+        output[tid] = imm;
+    }
+}
+
+namespace infini {
+
+void hetConvToMMReduce(int n, int h, int w, int f, float *input, float *output,
+                       float *bias) {
+    const int kBlockSize = 128;
+    vector<int> in_params = {n, h, w, f}; // NHWF
+    vector<int> out_params = {n, f, h, w};
+    int in_base = 10;
+    int out_base = 1;
+    vector<int> in_offsets;
+    vector<int> out_offsets;
+    for (int i = 0; i < 4; ++i) {
+        in_offsets.push_back(in_base);
+        in_base *= in_params[3 - i];
+        out_offsets.push_back(out_base);
+        out_base *= out_params[3 - i];
+    }
+    reduce_merge_conv_3x3_1x1<float>
+        <<<(out_base + kBlockSize - 1) / kBlockSize, kBlockSize>>>(
+            input, output, bias, in_params[0], in_params[1], in_params[2],
+            in_params[3], in_offsets[3], in_offsets[2], in_offsets[1],
+            in_offsets[0], out_offsets[3], out_offsets[2], out_offsets[1],
+            out_offsets[0], out_base);
+}
+
+void conv5x5ToConv3x3Reduce(int n, int f, int h, int w, float *input,
+                            float *output, float *bias) {
+    const int kBlockSize = 128;
+    vector<int> params{n, f, h, w}; // NFHW
+    vector<int> ranges(4);
+    ranges[3] = params[3] + 2;
+    ranges[2] = params[2] + 2;
+    ranges[1] = params[1] * 4;
+    ranges[0] = params[0];
+
+    int in_base = 1;
+    int out_base = 1;
+    vector<int> in_offsets;
+    vector<int> out_offsets;
+    for (int i = 0; i < 4; ++i) {
+        in_offsets.push_back(in_base);
+        in_base *= ranges[3 - i];
+        out_offsets.push_back(out_base);
+        out_base *= params[3 - i];
+    }
+    reduce_2<float><<<(out_base + kBlockSize - 1) / kBlockSize, kBlockSize>>>(
+        input, output, bias, params[0], params[1], params[2], params[3],
+        in_offsets[3], in_offsets[2], in_offsets[1], in_offsets[0],
+        out_offsets[3], out_offsets[2], out_offsets[1], out_offsets[0],
+        out_base);
+}
+
+// [NHW,FRS] -> [NFHW]
+void conv3x3ToReduce(int n, int h, int w, int f, float *input, float *output,
+                     float *bias) {
+    const int kBlockSize = 128;
+    vector<int> in_params = {n, h, w, f}; // NHWF
+    vector<int> out_params = {n, f, h, w};
+    int in_base = 9;
+    int out_base = 1;
+    vector<int> in_offsets;
+    vector<int> out_offsets;
+    for (int i = 0; i < 4; ++i) {
+        in_offsets.push_back(in_base);
+        in_base *= in_params[3 - i];
+        out_offsets.push_back(out_base);
+        out_base *= out_params[3 - i];
+    }
+    reduce_merge_conv_3x3<float>
+        <<<(out_base + kBlockSize - 1) / kBlockSize, kBlockSize>>>(
+            input, output, bias, in_params[0], in_params[1], in_params[2],
+            in_params[3], in_offsets[3], in_offsets[2], in_offsets[1],
+            in_offsets[0], out_offsets[3], out_offsets[2], out_offsets[1],
+            out_offsets[0], out_base, 0);
+}
+
+void reduceConvRxSToNCHW(float *input, float *bias, float *output, int act,
+                         int n, int h, int w, int f, int r, int s, int oh,
+                         int ow, int ph, int pw, int sh, int sw, int dh,
+                         int dw) {
+    IT_ASSERT(sh == 1 && sw == 1,
+              "reduceConvRxSToNCHWKernel_kernel only support sh=sw=1");
+    IT_ASSERT(dh == 1 && dw == 1,
+              "reduceConvRxSToNCHWKernel_kernel only support dh=dw=1");
+    const int blocksize = 512;
+    const int gridsize = (n * f * oh * ow + blocksize - 1) / blocksize;
+
+    cudaStream_t stream(cudaStreamPerThread);
+    reduceConvRxSToNCHWKernel<<<gridsize, blocksize, 0, stream>>>(
+        input, bias, output, act, n, f, h, w, oh, ow, r, s, ph, pw, dh, dw);
+}
+
+} // namespace infini

src/kernels/cuda/reshape.cc

@@ -4,10 +4,13 @@ namespace infini {
 class CopyCuda : public CudaKernelWithoutConfig {
     void compute(const Operator &op,
                  const RuntimeObj *_context) const override {
-        auto inData = op->getInputs(0)->getRawDataPtr<void *>();
-        auto outData = op->getOutputs()[0]->getRawDataPtr<void *>();
-        cudaMemcpyAsync(outData, inData, op->getInputs(0)->getBytes(),
-                        cudaMemcpyDeviceToDevice);
+        // auto inData = op->getInputs(0)->getRawDataPtr<void *>();
+        // auto outData = op->getOutputs()[0]->getRawDataPtr<void *>();
+        // cudaMemcpyAsync(outData, inData, op->getInputs(0)->getBytes(),
+        //                 cudaMemcpyDeviceToDevice);
+
+        // HACK: optimization
+        op->getOutputs()[0]->setData(op->getInputs(0)->getDataBlob());
     }
 };
 // reshape/flatten/identity all act as copying from input to output.

src/kernels/cuda/transpose.cc

@@ -0,0 +1,70 @@
+#include "operators/transpose.h"
+#include "cuda/cuda_kernel_wihtout_config.h"
+#include "cuda/cuda_runtime.h"
+#include "cuda/cuda_transpose.h"
+
+namespace infini {
+
+class TransposeCuda : public CudaKernelWithoutConfig {
+    void generic_transpose(const Ref<TransposeObj> &op,
+                           const RuntimeObj *context) const {
+        auto input = op->getInputs(0);
+        auto output = op->getOutput();
+        void *const inputData = input->getRawDataPtr<void *>();
+        void *const outputData = output->getRawDataPtr<void *>();
+        const auto &inputShape = input->getDims();
+        const auto &outputShape = output->getDims();
+
+        const auto &perm = op->getPermute();
+        int size = input->size();
+        int nDims = input->getDims().size();
+
+        // Compute strides
+        SmallArray strides, buffer;
+        IT_ASSERT(nDims <= SMALL_ARRAY_SIZE);
+        int curStride = 1;
+        for (int i = nDims - 1; i >= 0; --i) {
+            buffer.data[i] = curStride;
+            curStride *= inputShape[i];
+        }
+        for (int i = 0; i < nDims; ++i) {
+            strides.data[i] = buffer.data[perm[i]];
+        }
+
+        SmallArray outputDims;
+        for (int i = 0; i < nDims; ++i) {
+            outputDims.data[i] = outputShape[i];
+        }
+
+        transpose_kernel((float *)inputData, (float *)outputData, nDims, size,
+                         strides, outputDims, input->getDims(),
+                         output->getDims(), perm);
+    }
+
+    void fast_transpose_last_dim(const Ref<TransposeObj> &op,
+                                 const RuntimeObj *context) const {
+        // Perm 0 2 3 1
+        auto cuda = dynamic_cast<const CudaRuntimeObj *>(context);
+        auto shape = op->getOutput()->getDims();
+        invoke_transpose_last_two_dim(
+            op->getInputs(0)->getRawDataPtr<float *>(),
+            op->getOutput()->getRawDataPtr<float *>(), shape[0],
+            shape[1] * shape[2], shape[3], cuda->getNumSMs());
+    }
+
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<TransposeObj>(_op);
+        const auto &perm = op->getPermute();
+        if (perm == vector{0, 2, 3, 1}) {
+            fast_transpose_last_dim(op, _context);
+        } else {
+            generic_transpose(op, _context);
+        }
+    }
+};
+
+REGISTER_KERNEL(Device::CUDA, OpType::Transpose, DataType::Float32,
+                TransposeCuda, "Transpose_CUDA_Float32");
+
+} // namespace infini

src/kernels/cuda/transpose.cu

@@ -0,0 +1,231 @@
+#include "core/common.h"
+#include "cuda/cuda_common.h"
+#include "utils/small_array.h"
+
+#include <cassert>
+#include <cstdint>
+#include <iostream>
+#include <numeric>
+#include <limits>
+#define CUDA_HOST_DEVICE __forceinline__ __device__ __host__
+
+// https://github.com/462630221/SampleCode
+template <typename T> struct QuotientMod {
+    T quotient;
+    T mod;
+    __host__ __device__ QuotientMod(T q, T m) : quotient(q), mod(m) {}
+};
+
+template <typename T> struct FastIntDivider {
+    FastIntDivider() {}
+    FastIntDivider(T d) { divisor_ = d; };
+    __forceinline__ __device__ __host__ T div(T n) { return n / divisor_; }
+    __forceinline__ __device__ __host__ T mod(T n) { return n % divisor_; }
+    __forceinline__ __device__ __host__ QuotientMod<T> divmod(T n) {
+        return QuotientMod<T>(n / divisor_, n % divisor_);
+    }
+    T divisor_;
+};
+
+template <> struct FastIntDivider<uint32_t> {
+    FastIntDivider(){};
+
+    FastIntDivider(uint32_t d) {
+        assert(d >= 1);
+        divisor_ = d;
+        // if put 0 to __builtin_clz, the result undefined.
+        if (d == 1) {
+            rshift_ = 0;
+        } else {
+            rshift_ = 32 - __builtin_clz(d - 1);
+        }
+        uint64_t magic_t = ((1lu << (32 + rshift_)) + d - 1) / d;
+        magic_ = uint32_t(magic_t);
+    };
+
+    __forceinline__ __device__ __host__ uint32_t div(uint32_t n) {
+#if defined(__CUDA_ARCH__)
+        uint32_t q = __umulhi(n, magic_);
+#else
+        uint32_t q = (uint64_t(n) * magic_) >> 32;
+#endif
+        // return (((n - q) >> 1) + q) >> (rshift_ - 1);
+        return (n + q) >> rshift_;
+    }
+
+    __forceinline__ __device__ __host__ QuotientMod<uint32_t>
+    divmod(uint32_t n) {
+        uint32_t q = div(n);
+        return QuotientMod<uint32_t>(q, n - divisor_ * q);
+    }
+
+    uint32_t magic_;
+    uint32_t rshift_;
+    uint32_t divisor_;
+};
+
+void test_fast_u32() {
+    uint32_t d = 1;
+
+    FastIntDivider<uint32_t> diver(d);
+    std::cout << "7/3= " << uint32_t(7) / uint32_t(d) << " " << diver.div(7)
+              << std::endl;
+}
+
+constexpr unsigned int num_threads() { return 32 * 4; }
+constexpr int thread_work_size() { return 4; }
+constexpr int block_work_size() { return thread_work_size() * num_threads(); }
+
+__global__ void _transpose_kernel(float *input, float *output, int nDims,
+                                  int size, infini::SmallArray strides,
+                                  infini::SmallArray outputShape) {
+    int outputIdx = blockIdx.x * blockDim.x + threadIdx.x;
+    if (outputIdx < size) {
+        int inputIdx = 0;
+        int v = outputIdx;
+        for (int i = nDims - 1; i >= 0; --i) {
+            inputIdx += v % outputShape.data[i] * strides.data[i];
+            v /= outputShape.data[i];
+        }
+#if __CUDA_ARCH__ >= 350 || defined(USE_ROCM)
+        output[outputIdx] = __ldg(input + inputIdx);
+#else
+        output[outputIdx] = input[inputIdx];
+#endif
+    }
+}
+
+template <typename T, int NUM> struct Array {
+    CUDA_HOST_DEVICE T &operator[](unsigned int index) { return data[index]; }
+    CUDA_HOST_DEVICE const T &operator[](unsigned int index) const {
+        return data[index];
+    }
+    CUDA_HOST_DEVICE constexpr int size() const { return NUM; }
+
+    CUDA_HOST_DEVICE Array() {
+#ifndef __CUDA_ARCH__
+        for (int i = 0; i < NUM; i++) {
+            data[i] = T();
+        }
+#endif
+    }
+
+    T data[NUM];
+};
+
+/**
+ * @brief Optimize : Reorganize
+ *
+ */
+template <int NUM_AXES, int UNROLL, int BLOCK_SIZE, typename T>
+__global__ void
+transpose_kernel_v3(const T *data_in, T *data_out,
+                    const Array<uint32_t, NUM_AXES> perm_strides,
+                    Array<FastIntDivider<uint32_t>, NUM_AXES> out_strides,
+                    const size_t all_cnt) {
+    uint32_t out_offset_reg[UNROLL];
+    uint32_t in_offset_reg[UNROLL];
+#pragma unroll
+    for (int i = 0; i < UNROLL; ++i) {
+        out_offset_reg[i] =
+            blockIdx.x * BLOCK_SIZE * UNROLL + threadIdx.x + i * BLOCK_SIZE;
+        in_offset_reg[i] = 0;
+    }
+
+#pragma unroll
+    for (int j = 0; j < NUM_AXES; ++j) {
+#pragma unroll
+        for (int i = 0; i < UNROLL; ++i) {
+            QuotientMod<uint32_t> d = out_strides[j].divmod(out_offset_reg[i]);
+            in_offset_reg[i] += d.quotient * perm_strides[j];
+            out_offset_reg[i] = d.mod;
+        }
+    }
+
+    T ld_reg[UNROLL];
+    uint32_t offset = blockIdx.x * BLOCK_SIZE * UNROLL + threadIdx.x;
+    if (offset + BLOCK_SIZE * UNROLL <= all_cnt) {
+#pragma unroll
+        for (int i = 0; i < UNROLL; ++i) {
+            ld_reg[i] = data_in[in_offset_reg[i]];
+        }
+#pragma unroll
+        for (int i = 0; i < UNROLL; ++i) {
+            data_out[offset + i * BLOCK_SIZE] = ld_reg[i];
+        }
+    } else {
+#pragma unroll
+        for (int i = 0; i < UNROLL; ++i) {
+            if (offset + i * BLOCK_SIZE < all_cnt) {
+                ld_reg[i] = data_in[in_offset_reg[i]];
+            }
+        }
+#pragma unroll
+        for (int i = 0; i < UNROLL; ++i) {
+            if (offset + i * BLOCK_SIZE < all_cnt) {
+                data_out[offset + i * BLOCK_SIZE] = ld_reg[i];
+            }
+        }
+    }
+}
+
+
+
+template <typename T> T AccMul(std::vector<T> vec) {
+    return std::accumulate(vec.begin(), vec.end(), T(1), std::multiplies<T>());
+}
+
+namespace infini {
+// void transpose_kernel(float *input, float *output, int nDims, int size,
+//                       SmallArray strides, SmallArray outputShape) {
+//     int blocksize = block_work_size();
+//     int gridsize = (size + block_work_size() - 1) / block_work_size();
+//     _transpose_kernel<<<gridsize, blocksize>>>(input, output, nDims, size,
+//                                                strides, outputShape);
+// }
+
+std::vector<uint32_t> GetStrides(std::vector<uint32_t> dims) {
+    std::vector<uint32_t> strides(dims.size(), 1);
+    for (int i = dims.size() - 2; i >= 0; --i) {
+        strides[i] = strides[i + 1] * dims[i + 1];
+    }
+    return strides;
+}
+
+void transpose_kernel(float *input, float *output, int nDims, int size,
+                      SmallArray _strides, SmallArray _outputShape,
+                      vector<int> _dims_in, vector<int> _dims_out,
+                      vector<int> _perms) {
+    constexpr int NUM_AXES = 4;
+    IT_ASSERT(nDims <= NUM_AXES);
+    constexpr int UNROLL = 8 / sizeof(float);
+    constexpr int BLOCK_SIZE = 128;
+
+    vector<uint32_t> dims_in, dims_out, perms;
+    for (auto v : _dims_in)
+        dims_in.push_back(v);
+    for (auto v : _dims_out)
+        dims_out.push_back(v);
+    for (auto v : _perms)
+        perms.push_back(v);
+
+    size_t all_cnt = AccMul(dims_in);
+
+    auto strides_in = GetStrides(dims_in);
+    auto strides_out = GetStrides(dims_out);
+
+    const int grid =
+        (all_cnt + BLOCK_SIZE * UNROLL - 1) / (BLOCK_SIZE * UNROLL);
+    Array<uint32_t, NUM_AXES> perm_strides;
+    Array<FastIntDivider<uint32_t>, NUM_AXES> out_strides_fast;
+    for (int i = 0; i < NUM_AXES; ++i) {
+        out_strides_fast[i] = FastIntDivider<uint32_t>(strides_out[i]);
+        perm_strides[i] = strides_in[perms[i]];
+    }
+
+    transpose_kernel_v3<NUM_AXES, UNROLL, BLOCK_SIZE, float>
+        <<<grid, BLOCK_SIZE, 0>>>(
+            input, output, perm_strides, out_strides_fast, all_cnt);
+}
+
+} // namespace infini

src/kernels/cuda/transpose_fast.cu

@@ -0,0 +1,194 @@
+#include "cuda/cuda_common.h"
+#include <assert.h>
+#include <vector>
+
+template <int numSM, int numWarp>
+__global__ void kernel_transpose_last(float *ptrA, float *ptrB, int dim0,
+                                      int dim1, int dim2) {
+    int laneId = threadIdx.x % 32;
+    int warpId = blockIdx.x * numWarp + threadIdx.x / 32;
+    int n1 = (dim1 + 31) / 32;
+    int n2 = (dim2 + 31) / 32;
+    float bufA[32];
+    for (int i = warpId; i < dim0 * n1 * n2; i += numSM * numWarp) {
+        // clock_t ck0 = clock();
+        int i0 = i / (n1 * n2);
+        int i1 = (i % (n1 * n2)) / n2;
+        int i2 = (i % (n1 * n2)) % n2;
+        int offsetA = i0 * dim1 * dim2 + i2 * 32 * dim1 + i1 * 32;
+        int offsetB = i0 * dim1 * dim2 + i1 * 32 * dim2 + i2 * 32;
+        int ld1 = min(32, dim1 - i1 * 32);
+        int ld2 = min(32, dim2 - i2 * 32);
+        // if (i == 4 && laneId == 0)
+        //     printf("%d %d\n", ld1, ld2);
+
+        if (ld2 == 32) {
+#pragma unroll
+            for (int i = 0; i < 32; i++) {
+                if ((laneId + i) % 32 < ld1) {
+                    bufA[i] = ptrA[offsetA + i * dim1 + (laneId + i) % 32];
+                }
+            }
+        } else if (ld2 == 17) {
+#pragma unroll
+            for (int i = 0; i < 17; i++) {
+                if ((laneId + i) % 32 < ld1) {
+                    bufA[i] = ptrA[offsetA + i * dim1 + (laneId + i) % 32];
+                }
+            }
+        } else if (ld2 == 4) {
+#pragma unroll
+            for (int i = 0; i < 4; i++) {
+                if ((laneId + i) % 32 < ld1) {
+                    bufA[i] = ptrA[offsetA + i * dim1 + (laneId + i) % 32];
+                }
+            }
+        } else {
+            for (int i = 0; i < ld2; i++) {
+                if ((laneId + i) % 32 < ld1) {
+                    bufA[i] = ptrA[offsetA + i * dim1 + (laneId + i) % 32];
+                }
+            }
+        };
+
+        if (ld1 == 32) {
+#pragma unroll
+            for (int i = 0; i < 32; i++) {
+                if ((i + 32 - laneId) % 32 < ld2) {
+                    ptrB[offsetB + i * dim2 + (i + 32 - laneId) % 32] =
+                        bufA[(i + 32 - laneId) % 32];
+                }
+            }
+        } else if (ld1 == 17) {
+#pragma unroll
+            for (int i = 0; i < 17; i++) {
+                if ((i + 32 - laneId) % 32 < ld2) {
+                    ptrB[offsetB + i * dim2 + (i + 32 - laneId) % 32] =
+                        bufA[(i + 32 - laneId) % 32];
+                }
+            }
+        } else if (ld1 == 4) {
+#pragma unroll
+            for (int i = 0; i < 4; i++) {
+                if ((i + 32 - laneId) % 32 < ld2) {
+                    ptrB[offsetB + i * dim2 + (i + 32 - laneId) % 32] =
+                        bufA[(i + 32 - laneId) % 32];
+                }
+            }
+        } else {
+            for (int i = 0; i < ld1; i++) {
+                if ((i + 32 - laneId) % 32 < ld2) {
+                    ptrB[offsetB + i * dim2 + (i + 32 - laneId) % 32] =
+                        bufA[(i + 32 - laneId) % 32];
+                }
+            }
+        };
+    }
+}
+
+namespace infini {
+
+/// @brief
+/// @param ptrA Input tensor of shape [dim0, dim2, dim1]
+/// @param ptrB Output tensor  of shape [dim0, dim1, dim2]
+/// @param dim0
+/// @param dim1
+/// @param dim2
+void invoke_transpose_last_two_dim(float *ptrA, float *ptrB, int dim0, int dim1,
+                                   int dim2, int numSMs) {
+    constexpr int numWarps = 4;
+    dim3 gridDim(numSMs, 1);
+    dim3 blockDim(numWarps * 32, 1);
+    if (numSMs == 80) { // V100
+        kernel_transpose_last<80, numWarps>
+            <<<gridDim, blockDim>>>(ptrA, ptrB, dim0, dim1, dim2);
+    } else if (numSMs == 108) { // A100
+        kernel_transpose_last<108, numWarps>
+            <<<gridDim, blockDim>>>(ptrA, ptrB, dim0, dim1, dim2);
+    } else {
+        IT_TODO_HALT_MSG(std::string("transpose_last_two_dim with ") +
+                         std::to_string(numSMs) + " SMs is not implemented");
+    }
+    // cudaCheckError();
+}
+
+} // namespace infini
+
+// constexpr int numWarm = 128, numEval = 128;
+//
+// void eval_transpose_last(const std::vector<int> &shape) {
+//     assert(shape.size() == 3);
+//     int size = shape[0] * shape[1] * shape[2];
+//     float *dataA, *dataB;
+//     dataA = (float *)malloc(size * sizeof(float));
+//     dataB = (float *)malloc(size * sizeof(float));
+//     for (int i0 = 0; i0 < shape[0]; i0++) {
+//         for (int i2 = 0; i2 < shape[2]; i2++) {
+//             for (int i1 = 0; i1 < shape[1]; i1++) {
+//                 dataA[i0 * shape[1] * shape[2] + i2 * shape[1] + i1] =
+//                     i0 * shape[1] * shape[2] + i2 * shape[1] + i1;
+//             }
+//         }
+//     }
+//     float *ptrA, *ptrB;
+//     checkCudaError(cudaMalloc(&ptrA, size * sizeof(float)));
+//     checkCudaError(cudaMalloc(&ptrB, size * sizeof(float)));
+//     checkCudaError(
+//         cudaMemcpy(ptrA, dataA, size * sizeof(float),
+//         cudaMemcpyHostToDevice));
+
+//     invoke_transpose_last_two_dim(ptrA, ptrB, shape[0], shape[1], shape[2]);
+//     checkCudaError(
+//         cudaMemcpy(dataB, ptrB, size * sizeof(float),
+//         cudaMemcpyDeviceToHost));
+//     for (int i0 = 0; i0 < shape[0]; i0++) {
+//         for (int i1 = 0; i1 < shape[1]; i1++) {
+//             for (int i2 = 0; i2 < shape[2]; i2++) {
+//                 if (dataA[i0 * shape[1] * shape[2] + i1 + i2 * shape[1]] !=
+//                     dataB[i0 * shape[1] * shape[2] + i1 * shape[2] + i2]) {
+//                     std::cout
+//                         << i0 << " " << i1 << " " << i2 << " "
+//                         << dataA[i0 * shape[1] * shape[2] + i1 + i2 *
+//                         shape[1]]
+//                         << " "
+//                         << dataB[i0 * shape[1] * shape[2] + i1 * shape[2] +
+//                         i2]
+//                         << std::endl;
+//                     exit(-1);
+//                 }
+//             }
+//         }
+//     }
+//     cudaEvent_t st, ed;
+//     checkCudaError(cudaEventCreate(&st));
+//     checkCudaError(cudaEventCreate(&ed));
+//     for (int i = 0; i < numWarm; i++) {
+//         invoke_transpose_last_two_dim(ptrA, ptrB, shape[0], shape[1],
+//         shape[2]);
+//     }
+//     checkCudaError(cudaEventRecord(st));
+//     for (int i = 0; i < numEval; i++) {
+//         invoke_transpose_last_two_dim(ptrA, ptrB, shape[0], shape[1],
+//         shape[2]);
+//     }
+//     checkCudaError(cudaEventRecord(ed));
+//     checkCudaError(cudaEventSynchronize(st));
+//     checkCudaError(cudaEventSynchronize(ed));
+//     float time;
+//     checkCudaError(cudaEventElapsedTime(&time, st, ed));
+//     float bandwidth = size * 2 * sizeof(float) * numEval / time / 1e6;
+//     std::cout << "transpose_last: " << shape[0] << " " << shape[1] << " "
+//               << shape[2] << " time: " << time / numEval
+//               << " ms. bandwidth: " << bandwidth << " GB/s" << std::endl;
+// }
+
+// Performance evaluation
+// int main() {
+//     eval_transpose_last({16, 1024, 256});
+//     eval_transpose_last({16, 14 * 14, 1024});
+//     eval_transpose_last({16, 7 * 7, 2048});
+//     eval_transpose_last({16, 7 * 7, 128});
+//     eval_transpose_last({1, 14 * 14, 1024});
+//     eval_transpose_last({1, 7 * 7, 2048});
+//     eval_transpose_last({1, 7 * 7, 128});
+// }

src/kernels/cuda/unary.cc

@@ -24,9 +24,11 @@ class ActivationCudnn : public CudaKernelWithoutConfig {
         void *const outputData = (op->getOutput()->getRawDataPtr<void *>());
 
         cudnnTensorDescriptor_t inputDesc, outputDesc;
-        auto dim = op->getInputs(0)->getDims();
-        if (dim.size() != 4)
-            IT_TODO_HALT();
+        auto _dim = op->getInputs(0)->getDims();
+        IT_ASSERT_TODO(_dim.size() <= 4);
+        vector<int> dim(4, 1);
+        for (int i = 0; i < (int)_dim.size(); i++) // Unsqueeze to 4D
+            dim[i + 4 - _dim.size()] = _dim[i];
         int n = dim[0], c = dim[1], h = dim[2], w = dim[3];
 
         // get inputs

src/nnet/App/test_models.cc

@@ -0,0 +1,560 @@
+#ifdef USE_CUDA
+#include "core/blob.h"
+#include "core/dummy_mutator.h"
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "core/search_engine.h"
+#include "cuda/cuda_runtime.h"
+#include "ffi/ffi_callback.h"
+#include "nnet/nmutator.h"
+#include "operators/G2BMM.h"
+#include "operators/GBMM.h"
+#include "operators/conv.h"
+#include "operators/element_wise.h"
+#include "operators/matmul.h"
+#include "operators/pooling.h"
+#include "operators/reshape.h"
+#include "operators/softmax.h"
+#include "operators/transpose.h"
+#include "operators/unary.h"
+#include "test.h"
+#include <pybind11/stl.h>
+
+namespace infini {
+
+// Channel, kernelSize, pad, stride, isTanh
+using GANConfigs = vector<tuple<int, int, int, int, bool>>;
+using DetailedConfigs =
+    vector<tuple<int, int, int, int, int, int, int, int, int, int, bool>>;
+
+static const vector<int> metaRules = {3, 2, 2, 2, 2, 5, 8, 8, 6, 91, 90};
+
+DetailedConfigs getGANConfigs(int id, int batch) {
+    // The first conv can be transformed into gemm without reduction
+    //                                       n, f,    h, w,     c, r, s, stride,
+    //                                       pad, dilation
+    GANConfigs ret;
+    const DetailedConfigs infoConfigs = {
+        {batch, 228, 1, 1, 448, 2, 2, 1, 0, 1, false},
+        {batch, 448, 2, 2, 256, 4, 4, 2, 1, 1, false},
+        {batch, 256, 4, 4, 128, 4, 4, 2, 1, 1, false},
+        {batch, 128, 8, 8, 64, 4, 4, 2, 1, 1, false},
+        {batch, 64, 16, 16, 3, 4, 4, 2, 1, 1, true}};
+    const DetailedConfigs dcganConfigs = {
+        {batch, 100, 1, 1, 512, 4, 4, 1, 0, 1, false},
+        {batch, 512, 4, 4, 256, 4, 4, 2, 1, 1, false},
+        {batch, 256, 8, 8, 128, 4, 4, 2, 1, 1, false},
+        {batch, 128, 16, 16, 64, 4, 4, 2, 1, 1, false},
+        {batch, 64, 32, 32, 3, 4, 4, 2, 1, 1, true}};
+    DetailedConfigs details;
+    if (id == 0) { // InfoGAN
+        dbg("Use InfoGAN configs");
+        details = infoConfigs;
+    } else if (id == 1) { // DCGAN
+        dbg("Use DCGAN configs");
+        details = dcganConfigs;
+    } else
+        IT_ASSERT(false);
+    return details;
+}
+
+// NHWC format
+Graph getGANGraph(int batch, Runtime runtime, int nLayers, int modelId) {
+    IT_ASSERT(1 <= nLayers && nLayers <= 5);
+    Graph g = make_ref<GraphObj>(runtime);
+    vector<Tensor> weights;
+    auto configs = getGANConfigs(modelId, batch);
+
+    Tensor input;
+    {
+        auto &[n, f, h, w, c, r, s, stride, pad, dilation, isTanh] = configs[0];
+        input = g->addTensor({batch, 1, 1, f}, DataType::Float32,
+                             TensorType::Input);
+    }
+    for (int i = 0; i < (int)configs.size() && i < nLayers; ++i) {
+        // auto [channel, kernelSize, pad, stride, tanh] = configs[i];
+        auto &[n, f, h, w, c, r, s, stride, pad, dilation, isTanh] = configs[i];
+        IT_ASSERT(input->getDims()[3] == f);
+        auto weight = g->addTensor({f, r, s, c}, DataType::Float32,
+                                   TensorType::Initialized); // f, r, s, c
+        input = g->addOp<ConvTransposed2dNHWCObj>(input, weight, nullptr, pad,
+                                                  pad, stride, stride, 1, 1)
+                    ->getOutput();
+        if (isTanh) {
+            input = g->addOp<TanhObj>(input, nullptr)->getOutput();
+        } else {
+            input = g->addOp<ReluObj>(input, nullptr)->getOutput();
+        }
+    }
+    return g;
+}
+
+// NHWC
+Graph getFSRCNNGraph(int batch, Runtime runtime) {
+    // n, c, h, w, f, r, s, stride, pad, dilation, has_pReLU
+    const DetailedConfigs fsrcnn_config = {
+        {batch, 1, 32, 32, 56, 5, 5, 1, 2, 1, true},
+        {batch, 56, 32, 32, 12, 1, 1, 1, 0, 1, true},
+        {batch, 12, 32, 32, 12, 3, 3, 1, 1, 1, false},
+        {batch, 12, 32, 32, 12, 3, 3, 1, 1, 1, false},
+        {batch, 12, 32, 32, 12, 3, 3, 1, 1, 1, false},
+        {batch, 12, 32, 32, 12, 3, 3, 1, 1, 1, true},
+        {batch, 12, 32, 32, 56, 1, 1, 1, 0, 1, true},
+        {batch, 56, 32, 32, 1, 9, 9, 1, 3, 4, false} // ConvTransNHWC
+        // n, f, h, w, c, r, s, stride, pad, dilation, has_pReLU
+    };
+
+    Graph g = make_ref<GraphObj>(runtime);
+
+    Tensor input;
+    {
+        auto &[n, c, h, w, f, r, s, stride, pad, dilation, has_pReLU] =
+            fsrcnn_config[0];
+        input = g->addTensor({batch, h, w, c}, DataType::Float32,
+                             TensorType::Input);
+    }
+
+    for (int i = 0; i < (int)fsrcnn_config.size() - 1; ++i) {
+        // auto [channel, kernelSize, pad, stride, tanh] = configs[i];
+        auto &[n, c, h, w, f, r, s, stride, pad, dilation, has_pReLU] =
+            fsrcnn_config[i];
+        IT_ASSERT(input->getDims()[3] == c);
+        auto weight = g->addTensor({f, r, s, c}, DataType::Float32,
+                                   TensorType::Initialized); // f, r, s, c
+        input = g->addOp<ConvNHWCObj>(input, weight, nullptr, pad, pad, stride,
+                                      stride, 1, 1)
+                    ->getOutput();
+        if (has_pReLU) {
+            input = g->addOp<ReluObj>(input, nullptr)->getOutput();
+        }
+    }
+
+    // last operator is a ConvTransNHWC
+    {
+        auto &[n, f, h, w, c, r, s, stride, pad, dilation, has_pReLU] =
+            fsrcnn_config[fsrcnn_config.size() - 1];
+        IT_ASSERT(input->getDims()[3] == f);
+        auto weight = g->addTensor({f, r, s, c}, DataType::Float32,
+                                   TensorType::Initialized); // f, r, s, c
+        input = g->addOp<ConvTransposed2dNHWCObj>(input, weight, nullptr, pad,
+                                                  pad, stride, stride, 1, 1)
+                    ->getOutput();
+    }
+    return g;
+}
+
+Graph getLongformer(Runtime runtime, int bs) {
+    const int seqlen = 10000, w = 1000, featlen = 512, heads = 8, d = 4;
+    const int hidden = featlen, hiddenPerHead = hidden / heads;
+    assert(hidden % heads == 0);
+    Graph g = make_ref<GraphObj>(runtime);
+
+    auto i0 = g->addTensor({bs, seqlen, featlen}, DataType::Float32,
+                           TensorType::Input);
+    auto w0 = g->addTensor({featlen, hidden}, DataType::Float32,
+                           TensorType::Initialized);
+    auto w1 =
+        g->addTensor({512, 512}, DataType::Float32, TensorType::Initialized);
+    auto w2 =
+        g->addTensor({512, 512}, DataType::Float32, TensorType::Initialized);
+    // Feed forward
+    auto w3 =
+        g->addTensor({512, 512}, DataType::Float32, TensorType::Initialized);
+    auto bias3 =
+        g->addTensor({512}, DataType::Float32, TensorType::Initialized);
+    auto w4 =
+        g->addTensor({512, 512}, DataType::Float32, TensorType::Initialized);
+    auto bias4 =
+        g->addTensor({512}, DataType::Float32, TensorType::Initialized);
+
+    auto q0 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                           TensorType::Other);
+    auto k0 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                           TensorType::Other);
+    auto v0 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                           TensorType::Other);
+
+    auto q1 = g->addTensor({bs, seqlen, heads, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto k1 = g->addTensor({bs, seqlen, heads, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto v1 = g->addTensor({bs, seqlen, heads, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+
+    auto q2 = g->addTensor({bs, heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto k2 = g->addTensor({bs, heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto v2 = g->addTensor({bs, heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+
+    auto q3 = g->addTensor({bs * heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto k3 = g->addTensor({bs * heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+    auto v3 = g->addTensor({bs * heads, seqlen, hiddenPerHead},
+                           DataType::Float32, TensorType::Other);
+
+    auto prob = g->addTensor({bs * heads, seqlen, 2 * w + 1}, DataType::Float32,
+                             TensorType::Other);
+    auto probSoftmax = g->addTensor({bs * heads, seqlen, 2 * w + 1},
+                                    DataType::Float32, TensorType::Other);
+    auto attn = g->addTensor({bs * heads, seqlen, hiddenPerHead},
+                             DataType::Float32, TensorType::Other);
+
+    auto t00 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                            TensorType::Other);
+    auto t01 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                            TensorType::Other);
+    auto t02 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                            TensorType::Other);
+    // auto t10 = g->addTensor({bs, seqlen, hidden});
+    auto t11 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                            TensorType::Other);
+    auto t12 = g->addTensor({bs, seqlen, hidden}, DataType::Float32,
+                            TensorType::Other);
+    auto output = g->addTensor({bs, seqlen, featlen}, DataType::Float32,
+                               TensorType::Other);
+
+    g->addOpWithOutputs<MatmulObj>(i0, w0, q0, false, true);
+    g->addOpWithOutputs<MatmulObj>(i0, w1, k0, false, true);
+    g->addOpWithOutputs<MatmulObj>(i0, w2, v0, false, true);
+    g->addOpWithOutputs<ReshapeObj>(q0, q1);
+    g->addOpWithOutputs<ReshapeObj>(k0, k1);
+    g->addOpWithOutputs<ReshapeObj>(v0, v1);
+    // For example, when perm=(1, 0, 2), given an input tensor of shape (1,
+    // 2, 3), the output shape will be (2, 1, 3).
+    g->addOpWithOutputs<TransposeObj>(q1, q2, vector{0, 2, 1, 3});
+    g->addOpWithOutputs<TransposeObj>(k1, k2, vector{0, 2, 1, 3});
+    g->addOpWithOutputs<TransposeObj>(v1, v2, vector{0, 2, 1, 3});
+    g->addOpWithOutputs<ReshapeObj>(q2, q3);
+    g->addOpWithOutputs<ReshapeObj>(k2, k3);
+    g->addOpWithOutputs<ReshapeObj>(v2, v3);
+    // Attention
+    g->addOpWithOutputs<G2BMMObj>(q3, k3, prob, w, d);
+    g->addOpWithOutputs<SoftmaxObj>(prob, probSoftmax, 2);
+    g->addOpWithOutputs<GBMMObj>(probSoftmax, v3, attn, d);
+    auto attn2 = g->addOp<ReshapeObj>(attn, nullptr,
+                                      vector{bs, heads, seqlen, hiddenPerHead})
+                     ->getOutput();
+    auto t000 =
+        g->addOp<TransposeObj>(attn2, nullptr, vector{0, 2, 1, 3})->getOutput();
+    g->addOpWithOutputs<ReshapeObj>(t000, t00);
+
+    // Feed forward
+    g->addOpWithOutputs<MatmulObj>(t00, w3, t01, false, true, bias3);
+    g->addOpWithOutputs<ReluObj>(t01, t02);
+    g->addOpWithOutputs<MatmulObj>(t02, w4, t11, false, true, bias4);
+    g->addOpWithOutputs<ReluObj>(t11, t12);
+    g->addOpWithOutputs<AddObj>(t12, i0, output);
+    return g;
+}
+
+Graph getConvtransposedNHWC(Runtime runtime, Shape shape, int layerId) {
+    IT_ASSERT(0 <= layerId && layerId < 5);
+    Graph g = make_ref<GraphObj>(runtime);
+    vector<Tensor> weights;
+    vector<tuple<int, int, int, int, bool>> cs{
+        // Channel, kernelSize, pad, stride, isTanh
+        {448, 2, 0, 1, false}, {256, 4, 1, 2, false}, {128, 4, 1, 2, false},
+        {64, 4, 1, 2, false},  {3, 4, 1, 2, true},
+    };
+
+    Tensor input = g->addTensor(shape, DataType::Float32, TensorType::Input);
+    for (int i = layerId; i < layerId + 1; ++i) {
+        auto [channel, kernelSize, pad, stride, tanh] = cs[i];
+        int f = input->getDims()[3]; // n, h, w, f
+        auto weight = g->addTensor({f, kernelSize, kernelSize, channel},
+                                   DataType::Float32,
+                                   TensorType::Initialized); // f, r, s, c
+        input = g->addOp<ConvTransposed2dNHWCObj>(input, weight, nullptr, pad,
+                                                  pad, stride, stride, 1, 1)
+                    ->getOutput();
+        if (tanh) {
+            input = g->addOp<TanhObj>(input, nullptr)->getOutput();
+        } else {
+            input = g->addOp<ReluObj>(input, nullptr)->getOutput();
+        }
+    }
+    return g;
+}
+
+void printGraph(Graph g) {
+    g->print();
+    puts("============ Data ============");
+    for (auto t : g->getTensors()) {
+        dbg(t);
+        t->printData();
+    }
+}
+
+void initializeGraphTensors(Graph g, double l, double r, bool useInt) {
+    g->dataMalloc();
+    auto gen = RandomGenerator(-0.1, 0.1, 0, useInt);
+    for (auto t : g->getInputs()) {
+        t->setData(gen);
+    }
+    for (auto t : g->getOutputs()) {
+        t->setData(ZeroGenerator());
+    }
+}
+
+Graph convertNCHWtoNHWCModel(Runtime runtime, Graph inG) {
+    // Construct new graph
+    // IT_ASSERT(inG->getInputs().size() == 1);
+    IT_ASSERT(inG->getOutputs().size() == 1);
+    bool status = inG->topo_sort();
+    IT_ASSERT(status);
+    auto g = make_ref<GraphObj>(runtime);
+    map<UidBaseType, Tensor> tensors;
+    for (const auto &t : inG->getTensors())
+        if (t->getDims().size() != 4)
+            return nullptr;
+    auto getTensor = [&g, &tensors](const Tensor &inTensor) {
+        auto uid = inTensor->getGuid();
+        if (auto it = tensors.find(uid); it == tensors.end()) {
+            Shape s = inTensor->getDims();
+            s = vector{s[0], s[2], s[3], s[1]};
+            tensors[uid] = g->addTensor(s, inTensor->getDType(),
+                                        inTensor->getTensorType());
+        }
+        return tensors[uid];
+    };
+    for (auto op : inG->getOperators()) {
+        TensorVec inputs, outputs;
+        for (auto &t : op->getInputs())
+            inputs.emplace_back(getTensor(t));
+        for (auto &t : op->getOutputs())
+            outputs.emplace_back(getTensor(t));
+        if (auto cOp = as<ConvObj>(op)) {
+            const auto &[ph, pw, sh, sw, dh, dw] = cOp->getPadStrideDilation();
+            auto bias =
+                cOp->getBias() ? g->cloneTensor(cOp->getBias()) : nullptr;
+            g->addOpWithOutputs<ConvNHWCObj>(inputs[0], inputs[1], outputs[0],
+                                             ph, pw, sh, sw, dh, dw, bias,
+                                             cOp->getAct());
+        } else if (const auto &cOp = as<ConvTransposed2dObj>(op)) {
+            const auto &[ph, pw, sh, sw, dh, dw] = cOp->getPadStrideDilation();
+            const auto &[oph, opw] = cOp->getOutputPadding();
+            auto group = cOp->getNumGroups();
+            auto bias =
+                cOp->getBias() ? g->cloneTensor(cOp->getBias()) : nullptr;
+            g->addOpWithOutputs<ConvTransposed2dNHWCObj>(
+                inputs[0], inputs[1], outputs[0], ph, pw, sh, sw, dh, dw, oph,
+                opw, group, bias, cOp->getAct());
+        } else if (const auto &cOp = as<MaxPoolObj>(op)) {
+            auto t = g->addOp<ReshapeObj>(inputs[0], nullptr,
+                                          cOp->getInputs(0)->getDims())
+                         ->getOutput();
+            auto tt = g->addTensor(cOp->getOutput()->getDims(),
+                                   cOp->getOutput()->getDType());
+            g->cloneOperator(op, {t}, {tt});
+            g->addOpWithOutputs<ReshapeObj>(tt, outputs[0]);
+        } else {
+            dbg(op);
+            g->cloneOperator(op, inputs, outputs);
+        }
+    }
+    return g;
+}
+
+Graph optimizeModelWithRules(Graph g, Runtime _runtime, vector<int> rules) {
+    auto runtime = as<CudaRuntimeObj>(_runtime);
+    // make_ref<NMutator>(NMutator::Mode::RuleBased, metaRules, runtime);
+    Ref<NMutator> mutator =
+        make_ref<NMutator>(NMutator::Mode::RuleBased, rules, runtime);
+    vector<Graph> bestGraphs;
+    SearchEngine searchEngine(runtime, mutator);
+    g->dataFree();
+    return searchEngine.run(g);
+}
+
+Graph optimizeModel(Graph g, Runtime _runtime, string name) {
+    auto runtime = as<CudaRuntimeObj>(_runtime);
+    Ref<NMutator> mutator = make_ref<NMutator>(NMutator::Mode::Normal, runtime);
+    vector<Graph> bestGraphs;
+    SearchEngine searchEngine(runtime, mutator);
+    g->dataFree();
+    return searchEngine.run(g);
+}
+
+Graph optimizeGraph(Graph g, Runtime _runtime, bool tuning, NMutator::Mode mode,
+                    vector<int> rules) {
+    auto runtime = as<CudaRuntimeObj>(_runtime);
+    Runtime cpu = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(cpu);
+    //    vector<int>{3, 2, 2, 5, 8, 8, 6, 90}); // Conv2gemm
+    //    vector<int>{3, 2, 2, 2, 2, 5, 8, 8, 6, 91, 90}); // TConv
+    Ref<NMutator> mutator;
+    if (mode == NMutator::Mode::Normal) {
+        dbg(mode);
+        mutator = make_ref<NMutator>(mode, runtime);
+    } else if (mode == NMutator::Mode::RuleBased) {
+        dbg(mode, rules);
+        IT_ASSERT_TODO(rules.size() > 0);
+        mutator = make_ref<NMutator>(mode, rules, runtime);
+    } else
+        IT_TODO_HALT();
+    vector<Graph> bestGraphs;
+    SearchEngine searchEngine(runtime, mutator);
+    g->dataFree();
+    return searchEngine.run(g);
+
+    bestGraphs.emplace_back(searchEngine.run(g));
+    g->topo_sort();
+    dbg(g, bestGraphs[0], bestGraphs.size());
+    g->print();
+
+    g->dataMalloc();
+    map<UidBaseType, Tensor> fuidToInputTensor;
+    for (auto t : g->getInputs()) {
+        IT_ASSERT(fuidToInputTensor.count(t->getFuid()) == 0);
+        fuidToInputTensor[t->getFuid()] = t;
+    }
+
+    auto gen = RandomGenerator(-0.1, 0.1, 0);
+    for (auto t : g->getInputs()) {
+        t->setData(gen);
+    }
+    for (auto t : g->getOutputs()) {
+        t->setData(ZeroGenerator());
+    }
+    runtime->run(g);
+    // dbg("Baseline graph");
+    // printGraph(g);
+    // dbg(runtme->getPerfTime(g, true));
+    g->dataFree();
+
+    for (size_t i = 0; i < bestGraphs.size(); i++) {
+        auto bestGraphCpu = bestGraphs[i];
+        auto bestGraph =
+            make_ref<GraphObj>(runtime, bestGraphCpu->getOperators());
+        bestGraph->topo_sort();
+
+        // bestGraph->dataMalloc();
+        // // Initialize inputs with random data
+        // for (auto t : bestGraph->getInputs()) {
+        //     t->copyData(fuidToInputTensor[t->getFuid()]);
+        // }
+
+        // // Initialize outputs with zeros
+        // for (auto t : bestGraph->getOutputs()) {
+        //     t->setData(ZeroGenerator());
+        // }
+
+        // dbg(bestGraph);
+        // dbg(bestGraph->getOutputs());
+
+        // if (tuning) {
+        //     runtime->run(bestGraph, true);  // Tune kernels
+        //     runtime->run(bestGraph, false); // Execute transfomraed graph
+
+        //     // FIXME: g is freed
+        //     auto go0 = gCpu->cloneTensor(g->getOutputs()[0]);
+        //     auto bgo0 = gCpu->cloneTensor(bestGraph->getOutputs()[0]);
+        //     // EXPECT_TRUE(go0->equalData(bgo0, 1e-3));
+        //     dbg(go0->equalData(bgo0, 1e-3));
+        //     dbg(runtime->getPerfTime(bestGraph, true));
+        //     dbg(runtime->timeNonCtcOperators(bestGraph));
+        //     // dbg(runtime->timeWithCudaGraph(bestGraph));
+        // }
+
+        // dbg("Best graph");
+        // printGraph(bestGraph);
+        return bestGraph;
+    }
+    return nullptr;
+}
+
+Graph optimizeWithDepthConstraint(Graph g, Runtime _runtime, int maxDepth) {
+    auto runtime = as<CudaRuntimeObj>(_runtime);
+    Runtime cpu = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(cpu);
+    Ref<NMutator> mutator = make_ref<NMutator>(NMutator::Mode::Normal, runtime);
+    mutator->setMaxDepth(maxDepth);
+    g->dataFree();
+    SearchEngine searchEngine(runtime, mutator);
+    searchEngine.searchFilter = 1;
+    return searchEngine.run(g);
+}
+
+vector<Tensor> runInfoGAN(int nLayers) {
+    auto cuda = make_ref<CudaRuntimeObj>();
+    Runtime cpu = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(cpu);
+
+    Graph g = getGANGraph(1, cuda, nLayers, 0);
+
+    auto mutator =
+        make_ref<NMutator>(NMutator::Mode::RuleBased,
+                           vector<int>{3, 2, 2, 2, 2, 5, 8, 8, 6, 91, 90});
+    // // Translate OP to membound without derivation
+    // mutator->setToNaiveMembound();
+
+    vector<Graph> bestGraphs;
+    SearchEngine searchEngine(cuda, mutator);
+    bestGraphs.emplace_back(searchEngine.run(g));
+    g->topo_sort();
+    dbg(g, bestGraphs[0], bestGraphs.size());
+    g->print();
+
+    g->dataMalloc();
+    map<UidBaseType, Tensor> fuidToInputTensor;
+    for (auto t : g->getInputs()) {
+        IT_ASSERT(fuidToInputTensor.count(t->getFuid()) == 0);
+        fuidToInputTensor[t->getFuid()] = t;
+    }
+
+    auto gen = RandomGenerator(-0.1, 0.1, 0);
+    // auto gen = RandomGenerator(-5, 5, 0, true);
+    for (auto t : g->getInputs()) {
+        t->setData(gen);
+    }
+    for (auto t : g->getOutputs()) {
+        t->setData(ZeroGenerator());
+    }
+    cuda->run(g);
+    dbg("Baseline graph");
+    printGraph(g);
+    dbg(cuda->getPerfTime(g, true));
+
+    for (size_t i = 0; i < bestGraphs.size(); i++) {
+        auto bestGraphCpu = bestGraphs[i];
+        auto bestGraph = make_ref<GraphObj>(cuda, bestGraphCpu->getOperators());
+        bestGraph->topo_sort();
+
+        bestGraph->dataMalloc();
+        // Initialize inputs with random data
+        for (auto t : bestGraph->getInputs()) {
+            t->copyData(fuidToInputTensor[t->getFuid()]);
+        }
+
+        // Initialize outputs with zeros
+        for (auto t : bestGraph->getOutputs()) {
+            t->setData(ZeroGenerator());
+        }
+
+        dbg(bestGraph);
+        dbg(bestGraph->getOutputs());
+
+        cuda->run(bestGraph, true);  // Tune kernels
+        cuda->run(bestGraph, false); // Execute transfomraed graph
+
+        auto go0 = gCpu->cloneTensor(g->getOutputs()[0]);
+        auto bgo0 = gCpu->cloneTensor(bestGraph->getOutputs()[0]);
+        // EXPECT_TRUE(go0->equalData(bgo0, 1e-3));
+        std::cout << go0->equalData(bgo0, 1e-3) << std::endl;
+        bgo0->printData();
+        go0->printData();
+        dbg(cuda->getPerfTime(bestGraph, true));
+
+        dbg("Best graph");
+        printGraph(bestGraph);
+        callback::exportONNX(bestGraph, "best_graph.onnx"); // Debug
+        return {g->getOutputs()[0], bestGraph->getOutputs()[0]};
+    }
+    return {};
+}
+
+} // namespace infini
+#endif

src/nnet/Pass/MatchMemBoundKernel.cc

@@ -12,12 +12,11 @@ void MatchMemBoundKernel::transform(Formula &origin, int depth, Expr &rCur) {
     const auto &inputs = InputVisitor().getInputs(rangeOp);
     auto source =
         make_ref<ElementWiseNode>(rangeOp, inputs, rangeOp->getOutputShape());
-    auto tensor =
-        makeTensor(newTensorName(), rangeOp->getOutputShape(), {}, source);
+    auto tensor = mT(newTensorName(), rangeOp->getOutputShape(), {}, source);
     // The original code directly appends candidate. But it seems should be done
     // by the search.
     // appendCanddiate(as<TensorNode>(tensor), depth);
     nextStep(origin, depth, rCur, tensor);
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Pass/Rule3StageSplit.cc

@@ -38,11 +38,10 @@ void Rule3StageSplit::transform(Formula &origin, int depth, Expr &rCur) {
 
         // if no sum iterator, the stage is redundant
         assert(!innerSumVars.empty());
-        auto inner =
-            makeRangeOperator(innerLoopVars, innerSumVars, cur->getSummand());
+        auto inner = mL(innerLoopVars, innerSumVars, cur->getSummand());
         auto subscriptedInner = make_ref<SubscriptNode>(inner, indexForInner);
-        auto outer = makeRangeOperator(cur->getLoopVarRanges(), outerSumVars,
-                                       subscriptedInner);
+        auto outer =
+            mL(cur->getLoopVarRanges(), outerSumVars, subscriptedInner);
         outer->setPaddings(cur->getPaddings());
 
         // next searching step
@@ -79,4 +78,4 @@ Rule3StageSplit::getSplitSummationIters(RangeOp rangeOp) {
     return ret;
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Pass/Rule6KenerlMatching.cc

@@ -25,8 +25,8 @@ void Rule6KenerlMatching::transform(Formula &origin, int depth, Expr &rCur) {
     }
     { // Match element-wise OP
         auto replaces = matchElementWise(cur);
-        if (!replaces.empty())
-            dbg(rCur);
+        // if (!replaces.empty())
+        //     dbg(rCur);
         for (auto newCur : replaces)
             nextStep(origin, depth, rCur, newCur);
     }
@@ -50,8 +50,8 @@ VecExpr Rule6KenerlMatching::matchElementWise(const RangeOp &rangeOp) {
     const auto &inputs = InputVisitor().getInputs(rangeOp);
     auto source =
         make_ref<ElementWiseNode>(rangeOp, inputs, rangeOp->getOutputShape());
-    auto newTensor = makeTensor(newTensorName(), newShape, {}, source);
+    auto newTensor = mT(newTensorName(), newShape, {}, source);
     return {newTensor};
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Pass/Rule8GuidedDLT.cc

@@ -265,10 +265,9 @@ Expr Rule8GuidedDLT::guidedDLTMoreVar2(const RangeOp &cur,
     const auto sourceRoutine = make_ref<ElementWiseNode>(
         sourceExpr, vector<Tensor>{originalTensor}, newShape);
     // build stage connections
-    const auto newTensor =
-        makeTensor(newTensorName(), newShape, {}, sourceRoutine);
-    const auto &newSub = makeSubscript(
-        newTensor, VecExpr(tensorDimAxes.begin(), tensorDimAxes.end()));
+    const auto newTensor = mT(newTensorName(), newShape, {}, sourceRoutine);
+    const auto &newSub =
+        mSub(newTensor, VecExpr(tensorDimAxes.begin(), tensorDimAxes.end()));
     // TODO [1124]: get variable mapping and reorder L according to it
     // dbg(cur, originalSub, newSub, newVarRanges, replace.toReadable(),
     //     tensorDimAxes, newShape);
@@ -311,7 +310,7 @@ Expr Rule8GuidedDLT::buildGuidedDLTSource(const Subscript &originalSub,
     vector<VarRangePair> loopVarRangePairs;
     for (size_t i = 0; i < tensorDimAxes.size(); ++i)
         loopVarRangePairs.emplace_back(tensorDimAxes[i], pair(0, newShape[i]));
-    return makeRangeOperator(loopVarRangePairs, {}, newSub);
+    return mL(loopVarRangePairs, {}, newSub);
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Pass/Rule90TwoStageElementWise.cc

@@ -47,8 +47,8 @@ Rule90TwoStageElementWise::matchTwoStageElementWise(const RangeOp &rangeOp) {
     const auto &inputs = InputVisitor().getInputs(rangeOp);
     auto source =
         make_ref<ElementWiseNode>(rangeOp, inputs, rangeOp->getOutputShape());
-    auto newTensor = makeTensor(newTensorName(), newShape, {}, source);
+    auto newTensor = mT(newTensorName(), newShape, {}, source);
     return {newTensor};
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Visitor/FullPrinterVisitor.cc

@@ -13,7 +13,8 @@ string FullPrinterVisitor::print(const Expr &root) {
     oss << "==> ROOT\n" << root->toReadable() << "\n";
     for (size_t i = 0; i < q.size(); ++i) {
         const auto &[name, routine, tensor] = q[i];
-        oss << "==> " << name << " : ";
+        oss << "==> " << name << " " << infini::vecToString(tensor->getShape())
+            << " : ";
         if (routine) {
             oss << routine->toReadable() << "\n";
             if (routine->getExpr()) {

src/nnet/Visitor/MatmulTransposeMutator.cc

@@ -45,27 +45,26 @@ VecExpr MatmulTransposeMutator::transpose(const Tensor &tensor) {
         auto _va = make_ref<VarNode>("transA");
         auto _vb = make_ref<VarNode>("transB");
         auto _vc = make_ref<VarNode>("swapAB");
-        auto fakeSub = makeSubscript(matmul->getExpr(), {_va, _vb});
-        auto fakeRangeWrapperForHackHash =
-            makeRangeOperator({{_va, {0, Atrans + 100}},
-                               {_vb, {0, Btrans + 100}},
-                               {_vc, {0, ABswap + 100}}},
-                              {}, fakeSub);
+        auto fakeSub = mSub(matmul->getExpr(), {_va, _vb});
+        auto fakeRangeWrapperForHackHash = mL({{_va, {0, Atrans + 100}},
+                                               {_vb, {0, Btrans + 100}},
+                                               {_vc, {0, ABswap + 100}}},
+                                              {}, fakeSub);
         Matmul newMatmul =
             make_ref<MatmulNode>(fakeRangeWrapperForHackHash, inputs[0],
                                  inputs[1], b, m, n, k, transa, transb);
-        auto newTensor = makeTensor(derivator.newTensorName(), newShape,
-                                    newPaddings, newMatmul);
+        auto newTensor =
+            mT(derivator.newTensorName(), newShape, newPaddings, newMatmul);
         // build output transpose
         if (ABswap) {
             vector<Var> vars{derivator.getNewVar(), derivator.getNewVar()};
-            auto sub = makeSubscript(newTensor, {vars[1], vars[0]});
+            auto sub = mSub(newTensor, {vars[1], vars[0]});
             vector<VarRangePair> loopVRs;
             // Sicne inputs array may be swaped, use the orignal tensor shape
             for (int i = 0; i < 2; ++i) {
                 loopVRs.emplace_back(vars[i], Range(0, tensor->getShape(i)));
             }
-            auto rangeOp = makeRangeOperator(loopVRs, {}, sub);
+            auto rangeOp = mL(loopVRs, {}, sub);
             ret.emplace_back(rangeOp);
         } else
             ret.emplace_back(newTensor);
@@ -85,8 +84,8 @@ optional<Tensor> MatmulTransposeMutator::transposeInput(const Tensor &tensor) {
         assert(!rangeOp->hasPaddings());
         // auto paddings = rangeOp->getPaddings();
         // std::swap(paddings[0], paddings[1]);
-        auto sub = makeSubscript(rangeOp, {loopVRs[1].first, loopVRs[0].first});
-        auto newRangeOp = makeRangeOperator(loopVRs, {}, sub);
+        auto sub = mSub(rangeOp, {loopVRs[1].first, loopVRs[0].first});
+        auto newRangeOp = mL(loopVRs, {}, sub);
         // ElementWise newElementWise = make_ref<ElementWiseNode>(*ew);
         auto outputShape = ew->getOutputShape();
         std::swap(outputShape[0], outputShape[1]);
@@ -97,8 +96,8 @@ optional<Tensor> MatmulTransposeMutator::transposeInput(const Tensor &tensor) {
         auto tensorPaddings = tensor->getPaddings();
         std::swap(tensorShape[0], tensorShape[1]);
         std::swap(tensorPaddings[0], tensorPaddings[1]);
-        ret = makeTensor(derivator.newTensorName(), tensorShape, tensorPaddings,
-                         newElementWise);
+        ret = mT(derivator.newTensorName(), tensorShape, tensorPaddings,
+                 newElementWise);
         // } else if (!tensor->getSource()) {
     } else {
         nnet_unimplemented_continue();
@@ -107,4 +106,4 @@ optional<Tensor> MatmulTransposeMutator::transposeInput(const Tensor &tensor) {
     return ret;
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/Visitor/PatternMatcher.cc

@@ -32,8 +32,7 @@ RangeOp PatternMatcher::getOffsetCur() {
     }
     auto newSummand = ReplaceKit::replaceMultipleExprs(
         originalCur->getSummand(), itersFromNonZero, psis);
-    return makeRangeOperator(newLoopVarRanges, originalCur->getSumVarRanges(),
-                             newSummand);
+    return mL(newLoopVarRanges, originalCur->getSumVarRanges(), newSummand);
 }
 
 VecExpr PatternMatcher::matchKernel(const Pattern &pattern,
@@ -106,9 +105,9 @@ VecExpr PatternMatcher::applyWrapper(const VecExpr &exprs) {
         }
     }
     for (auto &expr : exprs) {
-        auto newSub = makeSubscript(expr, indexes);
-        ret.emplace_back(makeRangeOperator(originalCur->getLoopVarRanges(), {},
-                                           newSub, originalCur->getPaddings()));
+        auto newSub = mSub(expr, indexes);
+        ret.emplace_back(mL(originalCur->getLoopVarRanges(), {}, newSub,
+                            originalCur->getPaddings()));
     }
     return ret;
 }

src/nnet/Visitor/Serializer.cc

@@ -73,17 +73,52 @@ string Serializer::visit_(const Tensor &c) {
     return key;
 }
 
-bool Serializer::serialize(const Expr &expr, const string &filePath,
-                           const string &msg) {
+string Serializer::visit_(const Func &c) {
+    const string key = std::to_string(id++);
+    j[key]["type"] = c->getType();
+    j[key]["funcType"] = c->getFuncType();
+    j[key]["object"] = dispatch(c->getObject());
+    return key;
+}
+
+std::optional<std::string> Serializer::toString(const Expr &expr,
+                                                const string &msg,
+                                                vector<Tensor> inputs,
+                                                double exec_time, string hint) {
     // Metadata
     j["Version"] = VERSION;
     j["Msg"] = msg;
+    j["exec_time"] = exec_time;
+    j["hint"] = hint;
     // Expressions and routines
     id = 0;
     dispatch(expr);
-    std::ofstream fout(filePath);
-    fout << std::setw(4) << j << std::endl;
-    return true;
+
+    // Input tensors
+    vector<string> inputsIndices;
+    for (const auto &tensor : inputs) {
+        inputsIndices.emplace_back(std::to_string(id));
+        dispatch(tensor);
+    }
+    j["nnetInputs"] = inputsIndices;
+
+    // To string
+    std::stringstream ss;
+    ss << std::setw(4) << j << std::endl;
+    return {ss.str()};
+}
+
+bool Serializer::toFile(const Expr &expr, const string &filePath,
+                        const string &msg, vector<Tensor> inputs,
+                        double exec_time, string hint) {
+    if (auto s = toString(expr, msg, inputs, exec_time, hint)) {
+        // Write to file
+        std::ofstream fout(filePath);
+        fout << *s;
+        return true;
+    } else {
+        return false;
+    }
 }
 
 string Serializer::dispatchRoutine(const Routine &c) {
@@ -129,7 +164,15 @@ string Serializer::dispatchRoutine(const Routine &c) {
     return key;
 }
 
-Expr Serializer::deserialize(const string &filePath) {
+Expr Serializer::fromString(const string &text) {
+    std::stringstream str;
+    str << text;
+    str >> j;
+    assert(j["Version"] == VERSION);
+    return buildExprTree("0");
+}
+
+Expr Serializer::fromFile(const string &filePath) {
     std::ifstream fin(filePath);
     fin >> j;
     assert(j["Version"] == VERSION);
@@ -160,7 +203,7 @@ Expr Serializer::buildExprTree(string key) {
         }
         auto summand = buildExprTree(j[key]["summand"]);
         auto paddings = j[key]["paddings"].get<std::vector<int>>();
-        auto rangeOp = makeRangeOperator(loopIters, sumIters, summand);
+        auto rangeOp = mL(loopIters, sumIters, summand);
         rangeOp->setPaddings(paddings);
         return rangeOp;
     }
@@ -180,6 +223,10 @@ Expr Serializer::buildExprTree(string key) {
         return make_ref<TensorNode>(j[key]["name"], j[key]["shape"],
                                     j[key]["paddings"], source);
     }
+    case NodeType::FuncNodeType: {
+        auto object = buildExprTree(j[key]["object"]);
+        return make_ref<FuncNode>(object, j[key]["funcType"]);
+    }
     default: {
         nnet_unimplemented_halt();
         break;
@@ -242,4 +289,25 @@ Routine Serializer::buildRoutine(string key) {
     return nullptr;
 }
 
-} // namespace nnet
+tuple<Expr, vector<Tensor>, double, string>
+Serializer::deserializeAsMemobundOp(const string &filePath) {
+    std::ifstream fin(filePath);
+    fin >> j;
+    assert(j["Version"] == VERSION);
+    vector<Tensor> inputs;
+    for (const auto &input : j["nnetInputs"])
+        inputs.emplace_back(as<TensorNode>(buildExprTree(input)));
+    return {buildExprTree("0"), inputs, j["exec_time"], j["hint"]};
+}
+
+tuple<Expr, vector<Tensor>, double, string>
+Serializer::membundOpFromString(const string &data) {
+    j = json::parse(data);
+    assert(j["Version"] == VERSION);
+    vector<Tensor> inputs;
+    for (const auto &input : j["nnetInputs"])
+        inputs.emplace_back(as<TensorNode>(buildExprTree(input)));
+    return {buildExprTree("0"), inputs, j["exec_time"], j["hint"]};
+}
+
+} // namespace nnet

src/nnet/derivator.cc

@@ -136,9 +136,10 @@ void Derivator::dfs(Formula &origin, int depth) {
 }
 
 Derivator::Derivator(int maxDepth, bool enableHashPruning, LogMode logMode,
-                     PassMode passMode)
+                     PassMode passMode, bool printAndExit)
     : maxDepth(maxDepth), logMode(logMode), passMode(passMode),
-      enableHashPruning(enableHashPruning), cntAppliedRules(12) {}
+      enableHashPruning(enableHashPruning), cntAppliedRules(12),
+      printAndExit(printAndExit) {}
 
 int Derivator::getNumIntermediateStates() { return cntStates; }
 
@@ -405,12 +406,15 @@ Expr Derivator::mergeMemboundStages(VecExpr stages) {
 void Derivator::appendCanddiate(const Tensor &tensor, int depth) {
     // if (!CountRoutineVisitor().match(tensor, 1, 0, 3))
     //     return;
+    if (intermediateStates.size() > 1 && printAndExit)
+        printDerivationRules();
 
     candidates.emplace_back(tensor, depth);
     // dbg("!!!!!!!!!!!!!!!Success!!!!!!!!!!!!!!!");
     if (enableEquivalenceCheck)
         checkDerivationEquivalence();
-    // printIntermediateStates();
+    if (candidates.size() > 2)
+        printIntermediateStates();
 }
 
 bool Derivator::checkOOB(const RangeOp &rangeOp, bool halt) {
@@ -478,6 +482,7 @@ void Derivator::printStatistics() {
     printf("#Hashed intermediate states = %lu\n", visited.size());
     printf("#Iteratos = %d\n", nIteratorNames);
     printf("#Tensors = %d\n", nTensorNames);
+    printf("#Print and Exit mode = %d\n", printAndExit);
 }
 
 void Derivator::setDumpFirstSuccess(const string &_logFnPrefix) {
@@ -490,6 +495,9 @@ void Derivator::printIntermediateStates() {
     // Skip in NoLog mode
     if (logMode == LogMode::NoLog)
         return;
+    if (intermediateStates.size() > 1 && printAndExit)
+        printDerivationRules();
+
     assert(intermediateStates.size() == ruleStates.size());
     assert(intermediateStates.size() == ruleMsgs.size());
     for (size_t i = 0; i < intermediateStates.size(); ++i) {
@@ -499,16 +507,17 @@ void Derivator::printIntermediateStates() {
         std::cout << FullPrinterVisitor().print(intermediateStates[i]) << endl;
         if (logMode == LogMode::DumpFristCandiate) {
             Serializer serializer;
-            serializer.serialize(intermediateStates[i],
-                                 logFnPrefix + to_string(i) + ".expr", msg);
+            serializer.toFile(intermediateStates[i],
+                              logFnPrefix + to_string(i) + ".expr", msg);
         }
     }
-    for (size_t i = 0; i < intermediateStates.size(); ++i) {
-        if (auto cur = as<RangeOpNode>(intermediateStates[i]))
-            if (CheckOOBVisitor().checkRangeOp(cur)) {
-                printf("OOB detected depth=%lu\n", i);
-            }
-    }
+    // FIXME
+    // for (size_t i = 0; i < intermediateStates.size(); ++i) {
+    //     if (auto cur = as<RangeOpNode>(intermediateStates[i]))
+    //         if (CheckOOBVisitor().checkRangeOp(cur)) {
+    //             printf("OOB detected depth=%lu\n", i);
+    //         }
+    // }
     if (logMode == LogMode::DumpFristCandiate) {
         puts("Serializaiton finished.");
         exit(0);
@@ -528,4 +537,23 @@ Derivator::PassMode Derivator::getPassMode() { return passMode; }
 
 Derivator::LogMode Derivator::getLogMode() { return logMode; }
 
+void Derivator::printDerivationRules() {
+    int cntRules = 0, cntNonGuideRules = 0;
+    bool startGuided = false;
+    std::cout << ruleStates.size() << "rules" << std::endl;
+    for (size_t i = 1; i < ruleStates.size(); ++i) {
+        int ruleId = ruleStates[i][4] - '0';
+        if (ruleId != 4)
+            ++cntRules;
+        if (ruleId == 8)
+            startGuided = true;
+        if (!startGuided && ruleId != 4)
+            ++cntNonGuideRules;
+    }
+    printf("#Steps w/o converging derivation %d,  #Steps w/ converging "
+           "derivation %d\n",
+           cntRules, cntNonGuideRules);
+    exit(0);
+}
+
 } // namespace nnet

src/nnet/dlt.cc

@@ -60,7 +60,7 @@ optional<Expr> DLT::apply(const RangeOp &rangeOp, const Subscript &subscript,
             // Maybe there are bugs...
             // assert(index != nullptr);
             if (index == nullptr) {
-                std::cout << "Warning empty" << std::endl;
+                // std::cout << "Warning empty" << std::endl;
                 return {};
             }
         }
@@ -83,12 +83,11 @@ optional<Expr> DLT::apply(const RangeOp &rangeOp, const Subscript &subscript,
     // HACK [important] fix this fake tensor.
     auto elementRoutine = make_ref<ElementWiseNode>(
         // FIXME: implement transpose
-        // makeTensor(newTensorName + "_DLT", {}), vector<Tensor>{tensor},
+        // mT(newTensorName + "_DLT", {}), vector<Tensor>{tensor},
         // shape0);
-        makeTensor("__DLT", {}), vector<Tensor>{tensor}, shape0);
-    auto dltedTensor =
-        makeTensor(newTensorName, shape0, dltedPaddings, elementRoutine);
-    auto dltedSubscript = makeSubscript(dltedTensor, index0);
+        mT("__DLT", {}), vector<Tensor>{tensor}, shape0);
+    auto dltedTensor = mT(newTensorName, shape0, dltedPaddings, elementRoutine);
+    auto dltedSubscript = mSub(dltedTensor, index0);
     return optional<Expr>(std::in_place, dltedSubscript);
 }
 

src/nnet/expr.cc

@@ -1,4 +1,5 @@
 #include "nnet/expr.h"
+#include "nnet/Visitor/FullPrinterVisitor.h"
 #include "nnet/Visitor/GetTensorsVisitor.h"
 
 namespace nnet {
@@ -367,19 +368,19 @@ Expr operator/(const Expr &lhs, const int rhs) {
 }
 
 // Wrappers for type deduction
-Subscript makeSubscript(const Expr &tensor, const VecExpr &subscripts) {
+Subscript mSub(const Expr &tensor, const VecExpr &subscripts) {
     return make_ref<SubscriptNode>(tensor, subscripts);
 }
 
-RangeOp makeRangeOperator(const vector<VarRangePair> &_loopIters,
-                          const vector<VarRangePair> &_sumIters, Expr _summand,
-                          const vector<int> &paddings) {
+RangeOp mL(const vector<VarRangePair> &_loopIters,
+           const vector<VarRangePair> &_sumIters, Expr _summand,
+           const vector<int> &paddings) {
     return make_ref<RangeOpNode>(_loopIters, _sumIters, _summand, paddings);
 }
 
 // Wrappers for type deduction
-Tensor makeTensor(const string &name, const vector<int> &shape,
-                  const vector<int> &paddings, const Routine &source) {
+Tensor mT(const string &name, const vector<int> &shape,
+          const vector<int> &paddings, const Routine &source) {
     if (paddings.size() == 0)
         return make_ref<TensorNode>(name, shape,
                                     vector<int>((int)shape.size(), 0), source);
@@ -463,4 +464,9 @@ void FuncNode::setObject(Expr e) {
     object = e;
 }
 
+string RangeOpNode::getFullExpression() {
+    FullPrinterVisitor printer;
+    return printer.print(this->shared_from_this());
+}
+
 } // namespace nnet

src/nnet/iterator_table.cc

@@ -296,10 +296,9 @@ const Pattern &MatmulPattern::getMatmulPattern() {
         auto k = make_ref<VarNode>("_Matmul_k");
         auto A = make_ref<TensorNode>("_Matmul_A", vector<int>({M, K}));
         auto B = make_ref<TensorNode>("_Matmul_B", vector<int>({N, K}));
-        auto subA = makeSubscript(A, {m, k});
-        auto subB = makeSubscript(B, {n, k});
-        auto range = makeRangeOperator({{m, {0, M}}, {n, {0, N}}},
-                                       {{k, {0, K}}}, subA * subB);
+        auto subA = mSub(A, {m, k});
+        auto subB = mSub(B, {n, k});
+        auto range = mL({{m, {0, M}}, {n, {0, N}}}, {{k, {0, K}}}, subA * subB);
         auto success = exprIT.analyzeExpr(range);
         assert(success);
         exprIT.buildTable({0, 1});
@@ -317,11 +316,10 @@ const Pattern &ConvPattern::getPattern() {
         // auto n = make_ref<VarNode>("_Matmul_n");
         auto A = make_ref<TensorNode>("_Conv_A", vector<int>({N, C, H, W}));
         auto B = make_ref<TensorNode>("_Conv_K", vector<int>({F, C, R, S}));
-        auto subA = makeSubscript(A, {n, c, h + r, w + s});
-        auto subB = makeSubscript(B, {f, c, r, s});
-        auto range = makeRangeOperator(
-            {{n, {0, 0}}, {f, {0, 0}}, {h, {0, 0}}, {w, {0, 0}}},
-            {{c, {0, 0}}, {r, {0, 0}}, {s, {0, 0}}}, subA * subB);
+        auto subA = mSub(A, {n, c, h + r, w + s});
+        auto subB = mSub(B, {f, c, r, s});
+        auto range = mL({{n, {0, 0}}, {f, {0, 0}}, {h, {0, 0}}, {w, {0, 0}}},
+                        {{c, {0, 0}}, {r, {0, 0}}, {s, {0, 0}}}, subA * subB);
         auto success = exprIT.analyzeExpr(range);
         assert(success);
         exprIT.buildTable({0, 1});
@@ -350,7 +348,7 @@ Expr ConvPattern::buildExpr(
     auto shape = conv->getShape();
     auto rangeOpShape = as<RangeOpNode>(expr)->getOutputShape();
     assert(shape.size() == rangeOpShape.size());
-    dbg(shape, rangeOpShape);
+    // dbg(shape, rangeOpShape);
     for (size_t i = 0; i < shape.size(); ++i) {
         if (shape[i] != rangeOpShape[i]) {
             dbg("Warning: unmatched Conv output", shape, rangeOpShape);
@@ -404,11 +402,10 @@ const Pattern &Sg2bmmPattern::getPattern() {
         // auto n = make_ref<VarNode>("_Matmul_n");
         auto A = make_ref<TensorNode>("_Sg2bmm_A", vector<int>{Batch, M, K});
         auto B = make_ref<TensorNode>("_Sg2bmm_B", vector<int>{Batch, M, K});
-        auto subA = makeSubscript(A, {b, m, k});
-        auto subB = makeSubscript(B, {b, m + w, k});
-        auto range =
-            makeRangeOperator({{b, {0, Batch}}, {m, {0, M}}, {w, {-W, W + 1}}},
-                              {{k, {0, K}}}, subA * subB);
+        auto subA = mSub(A, {b, m, k});
+        auto subB = mSub(B, {b, m + w, k});
+        auto range = mL({{b, {0, Batch}}, {m, {0, M}}, {w, {-W, W + 1}}},
+                        {{k, {0, K}}}, subA * subB);
         auto success = exprIT.analyzeExpr(range);
         assert(success);
         exprIT.buildTableWithDefaultMap();
@@ -458,11 +455,10 @@ const Pattern &LongformerGBMMPattern::getPattern() {
         auto A =
             make_ref<TensorNode>("_lo_A", vector<int>{Batch, M, 2 * W + 1});
         auto B = make_ref<TensorNode>("_lo_B", vector<int>{Batch, M, N});
-        auto subA = makeSubscript(A, {b, m, w});
-        auto subB = makeSubscript(B, {b, m + w, n});
-        auto range =
-            makeRangeOperator({{b, {0, Batch}}, {m, {0, M}}, {n, {0, M}}},
-                              {{w, {-W, W + 1}}}, subA * subB);
+        auto subA = mSub(A, {b, m, w});
+        auto subB = mSub(B, {b, m + w, n});
+        auto range = mL({{b, {0, Batch}}, {m, {0, M}}, {n, {0, M}}},
+                        {{w, {-W, W + 1}}}, subA * subB);
         auto success = exprIT.analyzeExpr(range);
         assert(success);
         exprIT.buildTableWithDefaultMap();
@@ -536,11 +532,10 @@ Expr ConvPattern::getExpr(Tensor A, Tensor K, int N, int C, int H, int W, int F,
     DEFINE_VAR(f);
     DEFINE_VAR(r);
     DEFINE_VAR(s);
-    auto subA = makeSubscript(A, {n, c, h + r - R / 2, w + s - S / 2});
-    auto subB = makeSubscript(K, {f, c, r, s});
-    auto range =
-        makeRangeOperator({{n, {0, N}}, {f, {0, F}}, {h, {0, H}}, {w, {0, W}}},
-                          {{c, {0, C}}, {r, {0, R}}, {s, {0, S}}}, subA * subB);
+    auto subA = mSub(A, {n, c, h + r - R / 2, w + s - S / 2});
+    auto subB = mSub(K, {f, c, r, s});
+    auto range = mL({{n, {0, N}}, {f, {0, F}}, {h, {0, H}}, {w, {0, W}}},
+                    {{c, {0, C}}, {r, {0, R}}, {s, {0, S}}}, subA * subB);
     return range;
 }
 
@@ -572,13 +567,13 @@ Expr ConvTransPattern::getExpr(Tensor A, Tensor K, int N, int C, int H, int W,
     //                               vector<int>{0, padding, padding, 0});
     // auto K = make_ref<TensorNode>("K", vector<int>({R, S, F, C}));
 
-    auto subA = makeSubscript(A, {n, x1 + r - 1, y1 + s - 1, f});
+    auto subA = mSub(A, {n, x1 + r - 1, y1 + s - 1, f});
     auto subK =
-        // makeSubscript(K, {(R - 2) - 2 * r + x2, (S - 2) - 2 * s + y2, f, c});
-        makeSubscript(K, {f, (R - 2) - 2 * r + x2, (S - 2) - 2 * s + y2, c});
+        // mSub(K, {(R - 2) - 2 * r + x2, (S - 2) - 2 * s + y2, f, c});
+        mSub(K, {f, (R - 2) - 2 * r + x2, (S - 2) - 2 * s + y2, c});
     // x1=(h+1)//2, x2=(h+1)%2, y1=(w+1)//2
 
-    auto range1 = makeRangeOperator(
+    auto range1 = mL(
         {
             {n, {0, N}},
             {c, {0, C}},
@@ -588,10 +583,10 @@ Expr ConvTransPattern::getExpr(Tensor A, Tensor K, int N, int C, int H, int W,
             {y2, {0, 2}},
         },
         {{f, {0, F}}, {r, {0, R / 2}}, {s, {0, S / 2}}}, subA * subK);
-    auto sub0 = makeSubscript(
+    auto sub0 = mSub(
         range1, {n, c, (h + 1) / 2, (h + 1) % 2, (w + 1) / 2, (w + 1) % 2});
-    auto range0 = makeRangeOperator(
-        {{n, {0, N}}, {h, {0, OH}}, {w, {0, OW}}, {c, {0, C}}}, {}, sub0);
+    auto range0 =
+        mL({{n, {0, N}}, {h, {0, OH}}, {w, {0, OW}}, {c, {0, C}}}, {}, sub0);
     return range0;
 }
 
@@ -606,11 +601,10 @@ pair<Expr, pair<Tensor, Tensor>> Sg2bmmPattern::getExpr(int Batch, int M, int K,
     auto B = make_ref<TensorNode>("B", vector<int>({Batch, M, K}),
                                   vector<int>{0, D * W, 0});
 
-    auto subA = makeSubscript(A, {b, m, k});
-    auto subB = makeSubscript(B, {b, m + D * (w - W), k});
-    auto range =
-        makeRangeOperator({{b, {0, Batch}}, {m, {0, M}}, {w, {0, 2 * W + 1}}},
-                          {{k, {0, K}}}, subA * subB);
+    auto subA = mSub(A, {b, m, k});
+    auto subB = mSub(B, {b, m + D * (w - W), k});
+    auto range = mL({{b, {0, Batch}}, {m, {0, M}}, {w, {0, 2 * W + 1}}},
+                    {{k, {0, K}}}, subA * subB);
     return {range, {A, B}};
 }
 
@@ -624,10 +618,10 @@ LongformerGBMMPattern::getExpr(int Batch, int M, int W, int K, int dilation) {
                                   vector<int>{0, 0, 0});
     auto B = make_ref<TensorNode>("B", vector<int>({Batch, M, K}),
                                   vector<int>{0, dilation * W, 0});
-    auto subA = makeSubscript(A, {b, m, w});
-    auto subB = makeSubscript(B, {b, m + dilation * w - dilation * W, n});
-    auto range = makeRangeOperator({{b, {0, Batch}}, {m, {0, M}}, {n, {0, K}}},
-                                   {{w, {0, 2 * W + 1}}}, subA * subB);
+    auto subA = mSub(A, {b, m, w});
+    auto subB = mSub(B, {b, m + dilation * w - dilation * W, n});
+    auto range = mL({{b, {0, Batch}}, {m, {0, M}}, {n, {0, K}}},
+                    {{w, {0, 2 * W + 1}}}, subA * subB);
     return {range, {A, B}};
 }
 
@@ -642,10 +636,10 @@ pair<Expr, pair<Tensor, Tensor>> MatmulPattern::getExpr(bool transA,
                                   vector<int>{0, 0, 0});
     auto B = make_ref<TensorNode>("B", vector<int>({Batch, K, N}),
                                   vector<int>{0, 0, 0});
-    auto subA = makeSubscript(A, {b, m, k});
-    auto subB = makeSubscript(B, {b, k, n});
-    auto range = makeRangeOperator({{b, {0, Batch}}, {m, {0, M}}, {n, {0, N}}},
-                                   {{k, {0, K}}}, subA * subB);
+    auto subA = mSub(A, {b, m, k});
+    auto subB = mSub(B, {b, k, n});
+    auto range = mL({{b, {0, Batch}}, {m, {0, M}}, {n, {0, N}}}, {{k, {0, K}}},
+                    subA * subB);
     return {range, {A, B}};
 }
 

src/nnet/replace_kit.cc

@@ -17,8 +17,8 @@ RangeOp ReplaceKit::replaceRangeOpIterator(const RangeOp &rangeOp,
                                               replace.oldIters.size() +
                                               replace.newIters.size());
         // Check the number of loop iterators
-        return makeRangeOperator(newVarRangePairs, rangeOp->getSumVarRanges(),
-                                 replacedSummand);
+        return mL(newVarRangePairs, rangeOp->getSumVarRanges(),
+                  replacedSummand);
     } else if (replace.iteratorType == IterationType::Sum) {
         for (const auto &[var, range] : rangeOp->getSumVarRanges()) {
             if (!replace.isReplaced(var))
@@ -27,8 +27,8 @@ RangeOp ReplaceKit::replaceRangeOpIterator(const RangeOp &rangeOp,
         assert(newVarRangePairs.size() == rangeOp->getSumVarRanges().size() -
                                               replace.oldIters.size() +
                                               replace.newIters.size());
-        return makeRangeOperator(rangeOp->getLoopVarRanges(), newVarRangePairs,
-                                 replacedSummand, rangeOp->getPaddings());
+        return mL(rangeOp->getLoopVarRanges(), newVarRangePairs,
+                  replacedSummand, rangeOp->getPaddings());
     }
     assert(false);
     return nullptr;
@@ -55,7 +55,7 @@ Subscript ReplaceKit::buildSubscirptForLoopVarReplace(const RangeOp &inner,
     //     } else
     //         subs.emplace_back(inner->getLoopVar(i));
     // }
-    return makeSubscript(inner, subs);
+    return mSub(inner, subs);
 }
 
 RangeOp
@@ -89,4 +89,4 @@ Expr ReplaceKit::replaceExpr(const Expr &cur, const Expr &pattern,
     return ret;
 }
 
-} // namespace nnet
+} // namespace nnet

src/nnet/test.cc

@@ -6,7 +6,7 @@
 namespace nnet {
 
 int matchExprResult(Derivator &derivator, string fn) {
-    auto ans = Serializer().deserialize(fn);
+    auto ans = Serializer().fromFile(fn);
     auto hashAns = HashVisitor()(ans);
     int match = 0;
     for (const auto &candidate : derivator.getCandidates()) {
@@ -19,14 +19,14 @@ int matchExprResult(Derivator &derivator, string fn) {
 bool checkExprLogSame(string fnPrefix, int start, int end) {
     Serializer serializer;
     string fn0 = fnPrefix + to_string(start) + ".expr";
-    Expr expr0 = serializer.deserialize(fn0);
+    Expr expr0 = serializer.fromFile(fn0);
     RangeOp range0 = as<RangeOpNode>(expr0);
     Interpreter interpreter(range0);
     auto ans0 = interpreter.interpretUniformSample(range0);
     dbg(expr0, ans0);
     for (int i = start + 1; i < end; ++i) {
         string fn1 = fnPrefix + to_string(i) + ".expr";
-        Expr expr1 = serializer.deserialize(fn1);
+        Expr expr1 = serializer.fromFile(fn1);
         RangeOp range1 = as<RangeOpNode>(expr1);
         dbg(fn1, expr1);
         auto ans1 = interpreter.interpretUniformSample(range1);
@@ -67,4 +67,4 @@ bool checkExprsEquvivalence(VecExpr exprs) {
     return true;
 }
 
-} // namespace nnet
+} // namespace nnet

src/operators/any.cc

@@ -0,0 +1,74 @@
+#include "operators/any.h"
+
+namespace infini {
+
+AnyObj::AnyObj(GraphObj *graph, const TensorVec &inputs,
+               const TensorVec &outputs, const string &kernelName,
+               const vector<int> &attr)
+    : OperatorObj(OpType::Any, inputs, outputs), kernelName(kernelName),
+      attr(attr) {
+    IT_ASSERT(checkValid(graph));
+    // Outputs must assigned when constructing AnyObj
+    IT_ASSERT(!outputs.empty());
+    for (auto &output : outputs)
+        IT_ASSERT(output != nullptr && output->size() > 0);
+}
+
+string AnyObj::toString() const {
+    std::ostringstream os;
+    os << "Any[" << getGuid() << "](";
+    for (size_t i = 0; i < inputs.size(); ++i) {
+        os << "i" << i << "=" << inputs[i]->getGuid();
+        if (i != inputs.size() - 1)
+            os << " ";
+    }
+    os << ", ";
+    for (size_t i = 0; i < outputs.size(); ++i) {
+        os << "o" << i << "=" << outputs[i]->getGuid();
+        if (i != outputs.size() - 1)
+            os << " ";
+    }
+    os << ", ";
+    os << "kernel name: " << kernelName << ", ";
+    os << "attr = [";
+    for (size_t i = 0; i < attr.size(); ++i) {
+        os << attr[i];
+        if (i != attr.size() - 1)
+            os << ", ";
+    }
+    os << "])\n";
+    return os.str();
+}
+
+optional<vector<Shape>> AnyObj::inferShape(const TensorVec &inputs) const {
+    vector<Shape> ret;
+    for (auto output : outputs) {
+        ret.emplace_back(output->getDims());
+    }
+    return ret;
+}
+
+const string AnyObj::getKernelName() const { return kernelName; }
+
+vector<int> AnyObj::getOpAttrVector() const { return attr; };
+
+vector<int> AnyObj::getWorkloadVector() const {
+    vector<int> ret = {};
+    for (auto &input : inputs) {
+        auto inputDims = input->getDims();
+        ret.insert(ret.end(), inputDims.begin(), inputDims.end());
+    }
+    for (auto &output : outputs) {
+        auto outputDims = output->getDims();
+        ret.insert(ret.end(), outputDims.begin(), outputDims.end());
+    }
+    for (auto c : kernelName) {
+        ret.emplace_back(c);
+    }
+    for (auto at : attr) {
+        ret.emplace_back(at);
+    }
+    return ret;
+}
+
+} // namespace infini

src/operators/conv.cc

@@ -21,9 +21,8 @@ string ConvBaseObj::toString() const {
     std::ostringstream os;
     os << OpRegistry::getOpName(getOpType()) << "[" << getGuid() << "]";
     os << "(";
-    if (inputs.size() == 2) {
-        os << vecToString(inputs[0]->getDims()) << ",";
-        os << vecToString(inputs[1]->getDims()) << ",";
+    for (auto &input : inputs) {
+        os << vecToString(input->getDims()) << ",";
     }
     os << "p=[" << ph << "," << pw << "],";
     os << "s=[" << sh << "," << sw << "],";
@@ -114,6 +113,75 @@ optional<vector<Shape>> ConvObj::inferShape(const TensorVec &inputs) const {
     return {{{on, oc, oh, ow}}};
 }
 
+void ConvNHWCObj::setAuxilaryAttributes(PaddingMode mode) {
+    const Tensor &input = inputs[0];
+    const Tensor &weight = inputs[1];
+    n = input->getDims()[0], c = input->getDims()[3], h = input->getDims()[1],
+    w = input->getDims()[2], f = weight->getDims()[0], r = weight->getDims()[1],
+    s = weight->getDims()[2];
+    if (mode == PaddingMode::Same) {
+        int oh = h / sh;
+        int ow = w / sw;
+        ph = (h - oh * sh + (r - sh) * dh) / 2;
+        pw = (w - ow * sw + (s - sw) * dw) / 2;
+    } else if (mode == PaddingMode::Valid) {
+        ph = pw = 0;
+    }
+}
+
+ConvNHWCObj::ConvNHWCObj(GraphObj *graph, Tensor input, Tensor weight,
+                         Tensor output, int ph, int pw, int sh, int sw, int dh,
+                         int dw, Tensor bias, ActType act)
+    : ConvBaseObj(OpType::ConvNHWC, {input, weight}, output, ph, pw, sh, sw, dh,
+                  dw, input, weight, act) {
+    if (bias)
+        IT_TODO_HALT();
+    setAuxilaryAttributes(PaddingMode::Other);
+    IT_ASSERT(checkValid(graph));
+}
+
+ConvNHWCObj::ConvNHWCObj(GraphObj *graph, Tensor input, Tensor weight,
+                         Tensor output, PaddingMode mode, int sh, int sw,
+                         int dh, int dw, Tensor bias, ActType act)
+    : ConvBaseObj(OpType::ConvNHWC, {input, weight}, output, mode, sh, sw, dh,
+                  dw, input, weight, act) {
+    if (bias)
+        IT_TODO_HALT();
+    setAuxilaryAttributes(mode);
+    IT_ASSERT(checkValid(graph));
+}
+
+optional<vector<Shape>> ConvNHWCObj::inferShape(const TensorVec &inputs) const {
+    const auto &input = inputs[0], &weight = inputs[1];
+    auto n = input->getDims()[0];
+    auto h = input->getDims()[1];
+    auto w = input->getDims()[2];
+    auto f = weight->getDims()[0];
+    auto r = weight->getDims()[1];
+    auto s = weight->getDims()[2];
+    int on = n, oc = f;
+    int oh = 0, ow = 0;
+    // For NCHW+FCRS layout, C of input is divisable by C of weight
+    if (input->getDims()[3] % weight->getDims()[3] != 0)
+        return {};
+    // Set padding size
+    if (padding == PaddingMode::Other) {
+        oh = (h - (r - sh) * dh + ph * 2) / sh;
+        ow = (w - (s - sw) * dw + pw * 2) / sw;
+    } else if (padding == PaddingMode::Same) {
+        oh = h / sh;
+        ow = w / sw;
+        // ph = (h - oh * sh + (r - sh) * dh) / 2;
+        // pw = (w - ow * sw + (s - sw) * dw) / 2;
+    } else if (padding == PaddingMode::Valid) {
+        int ph = 0;
+        int pw = 0;
+        oh = (h - (r - sh) * dh + ph * 2) / sh;
+        ow = (w - (s - sw) * dw + pw * 2) / sw;
+    }
+    return {{{on, oh, ow, oc}}};
+}
+
 ConvTransposed2dObj::ConvTransposed2dObj(GraphObj *graph, Tensor input,
                                          Tensor weight, Tensor output, int ph,
                                          int pw, int sh, int sw, int dh, int dw,

src/operators/conv2dreduce.cc

@@ -0,0 +1,98 @@
+#include "operators/conv2dreduce.h"
+
+namespace infini {
+
+Conv2dReduceBase::Conv2dReduceBase(OpType opType, Tensor input, Tensor bias_,
+                                   Tensor output, bool PReLU_, float paramReLU_,
+                                   int ph_, int pw_, int sh_, int sw_, int dh_,
+                                   int dw_)
+    : OperatorObj(opType, {input}, {output}), bias(bias_), ph(ph_), pw(pw_),
+      sh(sh_), sw(sw_), dh(dh_), dw(dw_), PReLU(PReLU_), paramReLU(paramReLU_) {
+    // expect input shape is (n, h, w, f, r, s)
+    auto inputShape = input->getDims();
+    IT_ASSERT(inputShape.size() == 6);
+    n = inputShape[0];
+    h = inputShape[1];
+    w = inputShape[2];
+    f = inputShape[3];
+    r = inputShape[4];
+    s = inputShape[5];
+
+    if (bias) {
+        auto biasShape = bias->getDims();
+        IT_ASSERT(biasShape.size() == 1);
+        IT_ASSERT(biasShape[0] == f);
+    }
+}
+
+std::string Conv2dReduceBase::toString() const {
+    std::ostringstream os;
+    os << OpRegistry::getOpName(getOpType()) << "[" << getGuid() << "]";
+    os << "(";
+    if (inputs.size() == 2) {
+        os << vecToString(inputs[0]->getDims()) << ",";
+        os << vecToString(inputs[1]->getDims()) << ",";
+    } else {
+        os << vecToString(inputs[0]->getDims()) << ",";
+    }
+    os << "p=[" << ph << "," << pw << "],";
+    os << "s=[" << sh << "," << sw << "],";
+    os << "d=[" << dh << "," << dw << "],";
+    os << "PReLU=" << (PReLU ? "true" : "false") << ",";
+    // os << "act=" << enum_to_underlying(act) << ",";
+    os << "input=" << inputs[0]->getGuid() << ",";
+    if (bias != nullptr) {
+        os << "bias=" << bias->getGuid() << ",";
+    }
+    os << "output=" << outputs[0]->getGuid() << ")";
+    return os.str();
+}
+
+std::vector<int> Conv2dReduceBase::getWorkloadVector() const {
+    return {enum_to_underlying(type), n, h, w, f, r, s, ph, pw, sh, sw, dh, dw};
+}
+
+std::vector<int> Conv2dReduceBase::getOpAttrVector() const {
+    return {enum_to_underlying(type), ph, pw, sh, sw, dh, dw};
+}
+
+Conv2dReduce::Conv2dReduce(GraphObj *graph, Tensor input, Tensor bias,
+                           Tensor output, bool PReLU_, float paramReLU_,
+                           int ph_, int pw_, int sh_, int sw_, int dh_, int dw_)
+    : Conv2dReduceBase(OpType::Conv2dReduce, input, bias, output, PReLU_,
+                       paramReLU_, ph_, pw_, sh_, sw_, dh_, dw_) {
+    IT_ASSERT(checkValid(graph));
+}
+
+optional<vector<Shape>>
+Conv2dReduce::inferShape(const TensorVec &inputs) const {
+    // const auto &input = inputs[0], &bias = inputs[1];
+    int on = n, of = f;
+    int oh = (h + ph * 2 - dh * (r - 1) - 1) / sh + 1;
+    int ow = (w + pw * 2 - dw * (s - 1) - 1) / sw + 1;
+
+    return {{{on, oh, ow, of}}};
+}
+
+Conv2dReduceTranspose::Conv2dReduceTranspose(GraphObj *graph, Tensor input,
+                                             Tensor bias, Tensor output,
+                                             bool PReLU_, float paramReLU_,
+                                             int ph_, int pw_, int sh_, int sw_,
+                                             int dh_, int dw_)
+    : Conv2dReduceBase(OpType::Conv2dReduceTranspose, input, bias, output,
+                       PReLU_, paramReLU_, ph_, pw_, sh_, sw_, dh_, dw_) {
+    IT_ASSERT(dh_ == 1);
+    IT_ASSERT(dw_ == 1);
+    IT_ASSERT(checkValid(graph));
+}
+
+optional<vector<Shape>>
+Conv2dReduceTranspose::inferShape(const TensorVec &inputs) const {
+    // const auto &input = inputs[0], &bias = inputs[1];
+    int on = n, of = f;
+    int oh = (h - 1) * sh - 2 * ph + dh * (r - 1) + 1;
+    int ow = (w - 1) * sw - 2 * pw + dw * (s - 1) + 1;
+
+    return {{{on, oh, ow, of}}};
+}
+} // namespace infini

src/operators/membound.cc

@@ -2,6 +2,7 @@
 #include "nnet/Visitor/CheckOOBVisitor.h"
 #include "nnet/Visitor/HashVisitor.h"
 #include "nnet/Visitor/MergeMemboundMutator.h"
+#include "nnet/Visitor/Serializer.h"
 
 namespace infini {
 
@@ -9,8 +10,8 @@ MemBoundObj::MemBoundObj(GraphObj *graph, const TensorVec &input,
                          const TensorVec &output,
                          const std::vector<nnet::Tensor> &nnetInputs,
                          nnet::Expr expr, double exec_time, std::string hint)
-    : OperatorObj(OpType::MemBound, input, output), nnetInputs(nnetInputs),
-      expr(expr), exec_time(exec_time), hint(hint) {
+    : OperatorObj(OpType::MemBound, input, output), expr(expr),
+      nnetInputs(nnetInputs), exec_time(exec_time), hint(hint) {
     IT_ASSERT(checkValid(graph));
     IT_ASSERT(!checkOOB(expr));
     hash = calcHash(expr);
@@ -45,7 +46,7 @@ string MemBoundObj::toString() const {
     os << "exec_time=" << exec_time << ", ";
     os << "NNet Inputs=[";
     for (const auto &tensor : nnetInputs)
-        os << tensor->toReadable() << ",";
+        os << tensor->toReadable() << vecToString(tensor->getShape()) << ",";
     os << "]";
     os << ", ExprHash=" << hash;
     os << ", SimplifiedExprHash=" << simplifiedHash;
@@ -60,11 +61,18 @@ string MemBoundObj::toString() const {
 
 optional<vector<Shape>> MemBoundObj::inferShape(const TensorVec &inputs) const {
     // inputs have to match nnetInputs excatly
-    if (inputs.size() != nnetInputs.size())
+    if (inputs.size() != nnetInputs.size()) {
+        std::cout << "Num mismatch" << inputs.size() << " "
+                  << nnetInputs.size();
         return {};
+    }
     for (size_t i = 0; i < inputs.size(); ++i)
-        if (inputs[i]->getDims() != nnetInputs[i]->getShape())
+        if (inputs[i]->getDims() != nnetInputs[i]->getShape()) {
+            std::cout << "Shape mismatch " << inputs[i]
+                      << vecToString(inputs[i]->getDims()) << " "
+                      << vecToString(nnetInputs[i]->getShape());
             return {};
+        }
     return {{nnet::as<nnet::RangeOpNode>(expr)->getOutputShape()}};
 }
 
@@ -83,4 +91,9 @@ bool MemBoundObj::checkOOB(nnet::Expr expr) {
         nnet::as<nnet::RangeOpNode>(expr));
 }
 
+string MemBoundObj::toJson() const {
+    return *nnet::Serializer().toString(expr, "MemBoundObj::toJson", nnetInputs,
+                                        exec_time, hint);
+}
+
 } // namespace infini

src/operators/reshape.cc

@@ -2,7 +2,8 @@
 
 namespace infini {
 ReshapeObj::ReshapeObj(GraphObj *graph, Tensor input, Tensor output, Shape dims)
-    : OperatorObj(OpType::Reshape, {input}, {output}), dims(std::move(dims)) {
+    : OperatorObj(OpType::Reshape, {input}, {output}),
+      dims(dims.size() == 0 ? output->getDims() : dims) {
     IT_ASSERT(checkValid(graph));
 }
 
@@ -19,9 +20,9 @@ optional<vector<Shape>> ReshapeObj::inferShape(const TensorVec &inputs) const {
 std::string ReshapeObj::toString() const {
     std::ostringstream os;
     os << "Reshape[" << getGuid() << "]";
-    os << "(";
+    os << "(input dim=";
     os << vecToString(inputs[0]->getDims()) << ",";
-    os << "dims=" << vecToString(dims) << ",";
+    os << "output dims=" << vecToString(dims) << ",";
     os << "input=" << inputs[0]->getGuid() << ",";
     os << "output=" << outputs[0]->getGuid() << ")";
     return os.str();

src/operators/transpose.cc

@@ -4,13 +4,7 @@ namespace infini {
 TransposeObj::TransposeObj(GraphObj *graph, Tensor input, Tensor output,
                            vector<int> permute)
     : OperatorObj(OpType::Transpose, {input}, {output}) {
-    if (permute.size() != 4) {
-        IT_TODO_HALT();
-    }
-    transposePermute[0] = permute[0];
-    transposePermute[1] = permute[1];
-    transposePermute[2] = permute[2];
-    transposePermute[3] = permute[3];
+    transposePermute = permute;
     IT_ASSERT(checkValid(graph));
 }
 
@@ -20,7 +14,8 @@ TransposeObj::inferShape(const TensorVec &inputs) const {
     auto input = A->getDims();
     auto output = input;
 
-    for (int i = 0; i < 4; ++i) {
+    auto nDims = input.size();
+    for (size_t i = 0; i < nDims; ++i) {
         output[i] = input[transposePermute[i]];
     }
     return {{output}};
@@ -32,7 +27,8 @@ std::string TransposeObj::toString() const {
     os << "(";
     os << vecToString(inputs[0]->getDims()) << ",";
     os << "input=" << inputs[0]->getGuid() << ",";
-    os << "output=" << outputs[0]->getGuid() << ")";
+    os << "output=" << outputs[0]->getGuid() << ",";
+    os << "perm=" << vecToString(transposePermute) << ")";
     return os.str();
 }
 

test/core/test_graph_handler.cc

@@ -7,9 +7,10 @@ namespace infini {
 TEST(Handler, matmul) {
     auto runtime = NativeCpuRuntimeObj::getInstance();
     auto handler = make_ref<GraphHandlerObj>(runtime);
-    auto i = handler->tensor({1, 2, 3}, OnnxDType::UINT32);
-    auto w = handler->tensor({1, 3, 4}, OnnxDType::UINT32);
-    auto o = handler->tensor({1, 2, 4}, OnnxDType::UINT32);
+    auto i = handler->tensor({1, 2, 3}, OnnxDType::UINT32, TensorType::Input);
+    auto w =
+        handler->tensor({1, 3, 4}, OnnxDType::UINT32, TensorType::Initialized);
+    auto o = handler->tensor({1, 2, 4}, OnnxDType::UINT32, TensorType::Other);
     handler->matmul(i, w, o, false, false, nullptr, ActType::None);
 }
 

test/core/test_graph_replace.cc

@@ -48,7 +48,7 @@ TEST(SubGraphRewriter, subGraphMatch1) {
     SubGraphRewriter v(g);
     vector<MatchGraph> subgs = v.findMatch(subG);
 
-    EXPECT_TRUE(subgs.size() == 2);
+    EXPECT_TRUE(subgs.size() == 2u);
 }
 
 TEST(MatchGraph, single_input) {
@@ -116,12 +116,12 @@ TEST(MatchGraph, single_input) {
 
     auto o4 = v.addSubGraph(subG, TensorVec{add1->getOutput(0)});
 
-    EXPECT_EQ(g->getOperators().size(), 52);
+    EXPECT_EQ(g->getOperators().size(), 52u);
     vector<MatchGraph> subgs = v.findMatch(subG);
-    EXPECT_TRUE(subgs.size() == 5);
+    EXPECT_TRUE(subgs.size() == 5u);
 
     vector<MatchGraph> subgs1 = v.findMatch(subG1);
-    EXPECT_TRUE(subgs1.size() == 4);
+    EXPECT_TRUE(subgs1.size() == 4u);
 
     // test replace
     Tensor sii0 =
@@ -135,7 +135,7 @@ TEST(MatchGraph, single_input) {
     }
 
     v.replaceSubGraph(subG, subG2);
-    EXPECT_EQ(g->getOperators().size(), 37);
+    EXPECT_EQ(g->getOperators().size(), 37u);
 }
 
 TEST(MatchGraph, multi_input) {
@@ -186,17 +186,17 @@ TEST(MatchGraph, multi_input) {
                                      nullptr);
 
         auto matches = v.findMatch(subG);
-        EXPECT_EQ(2, matches.size());
+        EXPECT_EQ(2u, matches.size());
 
         auto div0 = g->addOp<DivObj>(reduce1->getOutput(0), i2, nullptr);
         auto add1 =
             g->addOp<AddObj>(sub0->getOutput(), div0->getOutput(), nullptr);
         matches = v.findMatch(subG);
-        EXPECT_EQ(1, matches.size());
+        EXPECT_EQ(1u, matches.size());
 
         // two matched subgraphs overlaped,so only replaced one sub graph
         v.replaceSubGraph(subG, replaceG);
-        EXPECT_EQ(1, v.findMatch(replaceG).size());
+        EXPECT_EQ(1u, v.findMatch(replaceG).size());
     }
 }
 
@@ -240,7 +240,7 @@ TEST(MatchGraph, multi_output) {
     {
         auto input = g->cloneTensor(i);
         auto outs = v.addSubGraph(subg0, {input});
-        EXPECT_EQ(2, outs.size());
+        EXPECT_EQ(2u, outs.size());
         Tensor w0 = g->addTensor(Shape{96, 64, 3, 3}, DataType::UInt32);
         auto conv0 = g->addOp<ConvObj>(outs[0], w0, nullptr, 1, 1);
         auto relu0 = g->addOp<ReluObj>(conv0->getOutput(0), nullptr);
@@ -263,11 +263,11 @@ TEST(MatchGraph, multi_output) {
     }
 
     auto matches = v.findMatch(subg0);
-    EXPECT_EQ(1, matches.size());
+    EXPECT_EQ(1u, matches.size());
 
     v.replaceSubGraph(subg0, subg1);
     auto matches2 = v.findMatch(subg1);
-    EXPECT_EQ(1, matches2.size());
+    EXPECT_EQ(1u, matches2.size());
 }
 
 // gcn
@@ -354,16 +354,16 @@ TEST(MatchGraph, multi_input_output) {
             v.addSubGraph(subg0, {relu->getOutput(0), maxPool->getOutput(0)});
         auto out1 =
             v.addSubGraph(subg1, {maxPool->getOutput(0), relu->getOutput(0)});
-        EXPECT_EQ(2, out0.size());
-        EXPECT_EQ(2, out1.size());
+        EXPECT_EQ(2u, out0.size());
+        EXPECT_EQ(2u, out1.size());
         auto div = g->addOp<DivObj>(out0[0], out1[1], nullptr);
         auto sub = g->addOp<SubObj>(out0[1], out1[0], nullptr);
     }
 
-    EXPECT_EQ(2, v.findMatch(subg0).size());
-    EXPECT_EQ(2, v.findMatch(subg1).size());
+    EXPECT_EQ(2u, v.findMatch(subg0).size());
+    EXPECT_EQ(2u, v.findMatch(subg1).size());
     v.replaceSubGraph(subg0, subg2);
-    EXPECT_EQ(v.findMatch(subg2).size(), 2);
+    EXPECT_EQ(v.findMatch(subg2).size(), 2u);
 }
 
 /* One Node having two or more successors is not supported yet.

test/kernels/cuda/test_cuda_any.cc

@@ -0,0 +1,57 @@
+#include "core/graph.h"
+#include "core/kernel.h"
+#include "core/runtime.h"
+#include "cuda/cuda_runtime.h"
+#include "operators/any.h"
+
+#include "test.h"
+
+namespace infini {
+TEST(cuda_Any, anyKernel) {
+    // conv2dreduce
+    {
+        // Construct Runtime and graph for CPU and CUDA
+        Runtime cpu =
+            NativeCpuRuntimeObj::getInstance(); // CPUruntime is singleton
+        Graph gCpu = make_ref<GraphObj>(cpu);
+        Runtime cuda = make_ref<CudaRuntimeObj>();
+        Graph gCuda = make_ref<GraphObj>(cuda);
+
+        auto generator = IncrementalGenerator();
+
+        int PRelu = 0, n = 1, h = 4, w = 4, f = 2, r = 3, s = 3, oh = 4, ow = 4,
+            ph = 1, pw = 1, sh = 1, sw = 1, dh = 1, dw = 1;
+        string kernelName = "conv2dreduce_kernel";
+        vector<int> attr{PRelu, n,  h,  w,  f,  r,  s, oh,
+                         ow,    ph, pw, sh, sw, dh, dw};
+
+        // Build input data on CPu
+        Tensor i0Cpu = gCpu->addTensor({n, 1, h, w}, DataType::Float32);
+        Tensor w0Cpu = gCpu->addTensor({f, 1, r, s}, DataType::Float32);
+        // Malloc data for all tensors in a graph. Do we need implicit
+        // allocation?
+        gCpu->dataMalloc();
+        i0Cpu->setData(generator);
+        w0Cpu->setData(generator);
+        // Copy input tensors from CPU to CUDA
+        Tensor i0Cuda = gCuda->cloneTensor(i0Cpu);
+        Tensor w0Cuda = gCuda->cloneTensor(w0Cpu);
+        Tensor o0Cuda = gCuda->addTensor({n, f, oh, ow});
+        auto anyOp = gCuda->addOpWithOutputs<AnyObj>(
+            TensorVec{i0Cuda, w0Cuda}, TensorVec{o0Cuda}, kernelName, attr);
+        anyOp->print();
+        // allocate CUDA memory
+        gCuda->dataMalloc();
+        std::cout << "data malloc success..." << std::endl;
+        // Execute on CUDA
+        cuda->run(gCuda);
+        std::cout << "cuda run success..." << std::endl;
+        // copy output from CUDA to CPU
+        auto o0Cpu = gCpu->cloneTensor(anyOp->getOutput());
+        // check results on CPU
+        EXPECT_TRUE(1);
+        // print a tensor/operator/graph by print()
+        gCuda->print();
+    }
+}
+} // namespace infini

test/kernels/cuda/test_cuda_conv.cc

@@ -43,6 +43,42 @@ void testConvCudnn(
     gCuda->print();
 }
 
+void testConvNHWCCudnn(
+    const std::function<void(void *, size_t, DataType)> &generator,
+    vector<float> ansVec) {
+    // Construct Runtime and graph for CPU and CUDA
+    Runtime cpu = NativeCpuRuntimeObj::getInstance(); // CPUruntime is singleton
+    Graph gCpu = make_ref<GraphObj>(cpu);
+    Runtime cuda = make_ref<CudaRuntimeObj>();
+    Graph gCuda = make_ref<GraphObj>(cuda);
+    // Set input data on CPU in a CPU Graph
+    Tensor i0Cpu = gCpu->addTensor({1, 4, 4, 3}, DataType::Float32);
+    Tensor w0Cpu = gCpu->addTensor({2, 3, 3, 3}, DataType::Float32);
+    // Malloc data for all tensors in a graph. Do we need implicit allocation?
+    gCpu->dataMalloc();
+    i0Cpu->setData(generator);
+    w0Cpu->setData(generator);
+
+    // Copy input tensors from CPU to CUDA
+    Tensor i0Cuda = gCuda->cloneTensor(i0Cpu);
+    Tensor w0Cuda = gCuda->cloneTensor(w0Cpu);
+    // Build CUDA graph
+    auto conv =
+        gCuda->addOp<ConvNHWCObj>(i0Cuda, w0Cuda, nullptr, 1, 1, 2, 1, 1, 2);
+    // allocate CUDA memory
+    gCuda->dataMalloc();
+    // Execute on CUDA
+    cuda->run(gCuda);
+    // copy output from CUDA to CPU
+    auto o0Cpu = gCpu->cloneTensor(conv->getOutput());
+    o0Cpu->print();
+    o0Cpu->printData();
+    // check results on CPU
+    EXPECT_TRUE(o0Cpu->equalData(ansVec));
+    // print a tensor/operator/graph by print()
+    gCuda->print();
+}
+
 TEST(cuDNN_Conv, run) {
     testConvCudnn(OneGenerator(),
                   vector<float>{12, 12, 18, 18, 12, 12, 18, 18});
@@ -51,6 +87,14 @@ TEST(cuDNN_Conv, run) {
         vector<float>{4794, 4386, 8199, 7506, 11274, 10542, 20835, 19656});
 }
 
+TEST(cuDNN_Conv, runNHWC) {
+    testConvNHWCCudnn(OneGenerator(),
+                      vector<float>{12., 12., 12., 12., 18., 18., 18., 18.});
+    testConvNHWCCudnn(
+        IncrementalGenerator(),
+        vector<float>{3350, 7562, 2306, 5546, 9480, 24546, 7185, 20793});
+}
+
 TEST(cuDNN_Conv, tune) {
     Runtime cpu = NativeCpuRuntimeObj::getInstance(); // CPUruntime is singleton
     Graph gCpu = make_ref<GraphObj>(cpu);