XCCL support (#171)

* add reduce_mean and gather

* fix format

* add kunlun allreduce and cmakefile

* add kunlun allreduce and cmakefile

* deltete cmake opt

* fix format

* fix makefile

* add DIST option in Makefile

* add xpu allgather

* delete xpu_wait()

* add xpu allgather

* delete specific compiler

* fix format

* fix gather

* add broadcast

* fix format

* fix

* fix xpu, add where operation, fix element-wise operation

* fix softmax

* fix softmax

* log internal input and output

* fix kunlun gather bugs

* update CMakeList.txt and Makefile

* fix some kunlun kernels

* fix Makefile

* fix Makefile

* set cmake version 3.12

* format

* fix where, gather and support gpt2

* "fix format"

* fix format

* copy onnx.py from master

* use KUNLUN_HOME instead of absolute path

* fix torchvision models

* support torchvison model-zoo

* fix format

* format fix, CMakeList fix

* fix review

* fix vecToString return value

* fix format

* delete  empty file

---------

Co-authored-by: wanghailu <wanghailu0717@163.com>
Co-authored-by: wanghailu <wanghailu@qiyuanlab.com>
Co-authored-by: Haojie Wang <haojie0429@gmail.com>

CMakeLists.txt

@@ -53,11 +53,13 @@ endif()
 
 set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_EXTENSIONS OFF) # -std=gnu++11 when on, -std=c++11 when off
+add_compile_options(-Wno-error=unused-variable)
 
 find_package(
   Python
   COMPONENTS Interpreter Development
   REQUIRED)
+
 # OpenMP
 find_package(OpenMP)
 if(OpenMP_C_FOUND)
@@ -282,9 +284,9 @@ if(USE_KUNLUN)
   endif()
   message(STATUS "KUNLUN_HOME: ${KUNLUN_HOME}")
 
-  include_directories("${KUNLUN_HOME}/XTDK/include/")
+  include_directories("${KUNLUN_HOME}/include/")
-  find_library(KUNLUN_RT libxpurt.so "${KUNLUN_HOME}/lib64")
+  find_library(KUNLUN_RT libxpurt.so "${KUNLUN_HOME}/so/")
-  find_library(KUNLUN_DNN libxpuapi.so "${KUNLUN_HOME}/XTDK/shlib")
+  find_library(KUNLUN_DNN libxpuapi.so "${KUNLUN_HOME}/so/")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -lstdc++ -Wall -Werror")
 
   if ((NOT DEFINED TARGET_CPU_ARCH) AND (NOT DEFINED ENV{TARGET_CPU_ARCH}))
@@ -297,6 +299,13 @@ if(USE_KUNLUN)
   endif()
   message(STATUS "TARGET_CPU_ARCH: ${TARGET_CPU_ARCH}")
 
+  if (BUILD_DIST)
+    message(STATUS "Add BUILD_DIST, use XCCL with KUNLUN XPU")
+    list(APPEND CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
+    find_package(XCCL REQUIRED)
+    add_compile_definitions(INFINI_USE_XCCL=1)
+    target_link_libraries(InfiniTensor ${XCCL_LIBRARIES})
+  endif()
   target_link_libraries(InfiniTensor ${KUNLUN_RT} ${KUNLUN_DNN} stdc++)
 endif()
 
@@ -335,6 +344,7 @@ if(BUILD_TEST)
     endif()
     if (USE_KUNLUN)
       build_test(test/kernels/kunlun/*.cc)
+      build_test(test/kunlun/*.cc)
     endif()
     if (USE_INTELCPU)
       build_test(test/kernels/intelcpu/*.cc)

Makefile

@@ -9,6 +9,7 @@ BACKTRACE ?= ON
 TEST ?= ON
 DIST ?= OFF
 NNET ?= OFF
+DIST ?= OFF
 FORMAT_ORIGIN ?=
 # Docker build options
 DOCKER_NAME ?= infinitensor

cmake/FindXCCL.cmake

@@ -0,0 +1,27 @@
+# Find the xccl libraries
+set(XCCL_INCLUDE_DIR $ENV{KUNLUN_HOME}/include CACHE PATH "Folder contains KUNLUN XCCL headers")
+set(XCCL_LIB_DIR $ENV{KUNLUN_HOME}  CACHE PATH "Folder contains KUNLUN XCCL libraries")
+
+list(APPEND CMAKE_PREFIX_PATH $ENV{KUNLUN_HOME})
+
+find_path(XCCL_INCLUDE_DIRS # ${XCCL_INCLUDE_DIR}
+  NAMES xpu/bkcl.h
+  HINTS XCCL_INCLUDE_DIR)
+
+find_library(XCCL_LIBRARIES # ${XCCL_LIB_DIR}
+  NAMES so/libbkcl.so
+  HINTS XCCL_LIB_DIR)
+
+message(STATUS "XCCL_INCLUDE_DIRS: ${XCCL_INCLUDE_DIRS}")
+message(STATUS "XCCL_LIBRARIES: ${XCCL_LIBRARIES}")
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(XCCL DEFAULT_MSG XCCL_INCLUDE_DIRS XCCL_LIBRARIES)
+
+if (XCCL_FOUND)
+  set (XCCL_HEADER_FILE "${XCCL_INCLUDE_DIRS}/xpu/bkcl.h")
+  message (STATUS "Determing XCCL version from ${XCCL_HEADER_FILE}...")
+  list (APPEND CMAKE_REQUIRED_INCLUDES ${XCCL_INCLUDE_DIRS})
+  message(STATUS "Found XCCL (include: ${XCCL_INCLUDE_DIRS}, library: ${XCCL_LIBRARIES})")
+  mark_as_advanced(XCCL_INCLUDE_DIRS XCCL_LIBRARIES)
+endif()

examples/NNmodel

@@ -1 +1 @@
-Subproject commit b896cec2dba5b8522b141ac4f89eb43074ee1b98
+Subproject commit 51d3105277f3774ed31c02ed4cd11fa92925af77

examples/distributed/launch_kunlun.py

@@ -0,0 +1,213 @@
+import argparse
+import os
+import time
+import multiprocessing as mp
+from pyinfinitensor.onnx import OnnxStub, backend
+import onnx
+from onnx.external_data_helper import convert_model_to_external_data
+from onnx.shape_inference import infer_shapes_path
+import numpy as np
+from parallel_opt import parallel_model
+
+st_input_dir = "standard/inputs/"
+st_output_dir = "standard/outputs/"
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="launch distributed infinitensor")
+    parser.add_argument("--num_nodes", type=int, default=1, help="number of nodes")
+    parser.add_argument(
+        "--nproc_per_node", type=int, default=2, help="number of processes per node"
+    )
+    parser.add_argument(
+        "--name", type=str, default="test", help="name of this instance."
+    )
+    parser.add_argument(
+        "--model", type=str, default="/data1/shared/panzezhong/llama/fp32/my_llama_fp32.sim.onnx", help="path to the ONNX model file."
+    )
+    parser.add_argument("--batch_size", type=int, default=1, help="batch size.")
+    parser.add_argument("--length", type=int, default=1, help="sequence length.")
+    parser.add_argument(
+        "--gen_std",
+        default=False,
+        action="store_true",
+        help="whether to generate the standard results.",
+    )
+    parser.add_argument(
+        "--run_single",
+        default=False,
+        action="store_true",
+        help="whether run model with single process with standard inputs"
+    )
+    args = parser.parse_args()
+    print("arg setting: ", args)
+    return (
+        args.num_nodes,
+        args.nproc_per_node,
+        args.name,
+        args.model,
+        args.batch_size,
+        args.length,
+        args.gen_std,
+        args.run_single
+    )
+
+
+def run_model(model, runtime, world_size=1, rank=0, n=10):
+    stub = OnnxStub(model, runtime)
+    load_inputs(stub, world_size, rank)
+    # stub.tune()
+    stub.run()
+    # get outputs
+    time.sleep(0.01)
+    outputs = next(stub.outputs.values().__iter__()).copyout_numpy()
+
+    # bench
+    begin = time.time()
+    for _ in range(n):
+        stub.run()
+    end = time.time()
+    avg_time = (end - begin) / n
+    print(f"average time: {avg_time}")
+    return outputs
+
+
+
+def run_and_compare(name, model, runtime, world_size=1, rank = 0):
+    results = np.load(os.path.join(st_output_dir,f"output.npy"))
+    outputs = run_model(model, runtime, world_size, rank)
+    print(outputs[:100])
+    if np.isnan(outputs).any():
+        print("Nan in output")
+    print("answer argmax:", np.argmax(results))
+    print("output argmax:", np.argmax(outputs))
+    #np.testing.assert_allclose(outputs, results, rtol=1e-3, atol=1e-3)
+    getDiff(results, outputs)
+
+
+def start_worker(
+    name: str, world_size: int, rank: int, local_rank: int, model: onnx.ModelProto
+):
+    dist_name = name + "_dist"
+    model = parallel_model(model, world_size, rank)
+    extern_path = f"./{dist_name}_rank{rank}.pb"
+    if os.path.exists(extern_path):
+        os.remove(extern_path)
+    onnx.save_model(
+        model,
+        f"./{dist_name}_rank{rank}.onnx",
+        save_as_external_data=True,
+        location=extern_path,
+    )
+    infer_shapes_path(f"./{dist_name}_rank{rank}.onnx")
+    runtime = backend.KUNLUNRuntime(local_rank)
+    # print("init comm")
+    runtime.init_comm(
+        dist_name,
+        world_size,
+        rank,
+    )
+    run_and_compare(name, model, runtime, world_size, rank)
+
+
+def start_single(name, model):
+    runtime = backend.KUNLUNRuntime(0)
+    run_and_compare(name, model, runtime)
+
+
+def generate_input_output(model):
+    runtime = backend.KUNLUNRuntime(0)
+    stub = OnnxStub(model, runtime)
+    position_id = 0
+    for i, (name, tensor) in enumerate(stub.inputs.items()):
+        input = tensor.copyout_numpy()
+        if np.issubdtype(input.dtype, np.integer):
+            if input.size == 1:
+                # input = np.array([position_id])
+                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
+            else:
+                input = np.random.randint(0,2,size=input.shape, dtype=input.dtype)
+        elif input.dtype == np.bool_:
+            input = np.random.randint(0,2,size=input.shape) > 0
+        else:
+            if i == 0:
+                input = np.ones(input.shape).astype(input.dtype)
+                position_id = input.shape[-1] - 1
+            else:
+                input = np.random.rand(*input.shape).astype(input.dtype)
+        tensor.copyin_numpy(input)
+        np.save(os.path.join(st_input_dir, f"input_{i}"), input)
+    stub.run()
+    # print(stub.outputs)
+    time.sleep(0.01)
+    output = next(stub.outputs.values().__iter__()).copyout_numpy()
+    print(output[:100])
+    if np.isnan(output).any():
+        print("Nan in output")
+    np.save(os.path.join(st_output_dir, f"output"), output)
+
+
+def load_inputs(stub, world_size=1, rank=0):
+    for i, (name, tensor) in enumerate(stub.inputs.items()):
+        input = np.load(os.path.join(st_input_dir, f"input_{i}.npy"))
+        if all(x == y for x,y in zip(input.shape,tensor.shape())):
+            tensor.copyin_numpy(input)
+        else:
+            tensor.copyin_numpy(np.hsplit(input, world_size)[rank])
+
+
+def getDiff(base, test):
+    absolute_diff = np.abs(np.subtract(base, test))
+    max_absolute_diff = np.max(absolute_diff)
+
+    baseCopy = base.astype(np.float64).ravel()
+    testCopy = test.astype(np.float64).ravel()
+    upValue = np.sum(np.abs(baseCopy - testCopy))
+    downValue = np.sum(np.abs(baseCopy)) + np.float64(1e-9)
+    max_relative_diff = upValue / downValue
+    print(f"Max absolute difference: {max_absolute_diff}\nMax relative difference: {max_relative_diff}")
+
+    return max_absolute_diff, max_relative_diff
+
+
+def main():
+    nnodes, nproc_per_node, name, model_path, bs, length, gen_std, run_single = parse_args()
+
+    model = onnx.load(model_path)
+
+    # generate standart output
+    if gen_std:
+        print("Generate inputs and outputs.")
+        p = mp.Process(target=generate_input_output, args=[model])
+        p.start()
+        p.join()
+        return
+
+    # # run single process.
+    # # use standalone process to isolate cuda.
+    if run_single:
+        print("run model by single GPU.")
+        p = mp.Process(target=start_single, args=(name, model))
+        p.start()
+        p.join()
+        return 
+
+    # run distributed parallel.
+    world_size = nnodes * nproc_per_node
+    print(f"run model by {world_size} GPU in parallel.")
+    workers = [
+        mp.Process(
+            target=start_worker,
+            args=(name, world_size, rank, rank % nproc_per_node, model),
+        )
+        for rank in range(world_size)
+    ]
+
+    for w in workers:
+        w.start()
+
+    for w in workers:
+        w.join()
+
+
+if __name__ == "__main__":
+    main()

include/core/common.h

@@ -61,16 +61,30 @@ template <typename T> auto enum_to_underlying(T e) {
 }
 
 template <typename T> std::string vecToString(const std::vector<T> &vec) {
-    std::string ret;
+    std::stringstream ss;
-    ret.append("[");
+    ss << "[";
-    for (auto d : vec) {
+    for (size_t i = 0; i < vec.size(); ++i) {
-        ret.append(std::to_string(d));
+        ss << vec.at(i);
-        ret.append(",");
+        if (i < vec.size() - 1) {
+            ss << ",";
         }
-    if (!vec.empty())
+    }
-        ret.pop_back();
+    ss << "]";
-    ret.append("]");
+    return ss.str();
-    return ret;
+}
+
+template <typename T> std::string vecToString(const T *st, size_t length) {
+    std::stringstream ss;
+    ss << "[";
+    size_t i = 0;
+    for (i = 0; i < length; i++) {
+        ss << *(st + i);
+        if (i < length - 1) {
+            ss << ",";
+        }
+    }
+    ss << "]";
+    return ss.str();
 }
 
 double timeit(

include/core/runtime.h

@@ -15,6 +15,7 @@ class GraphObj;
 class GraphHandlerObj;
 class RuntimeObj;
 class BlobObj;
+template <typename T> class WorkspaceObj;
 
 using TensorBase = Ref<TensorBaseObj>;
 using Tensor = Ref<TensorObj>;
@@ -23,6 +24,7 @@ using Graph = Ref<GraphObj>;
 using GraphHandler = Ref<GraphHandlerObj>;
 using Runtime = Ref<RuntimeObj>;
 using Blob = Ref<BlobObj>;
+template <typename T> using Workspace = Ref<WorkspaceObj<T>>;
 
 using TensorVec = vector<Tensor>;
 using OpVec = vector<Operator>;

include/core/tensor.h

@@ -4,6 +4,7 @@
 #include "utils/data_convert.h"
 #include <cmath>
 #include <cstring>
+#include <fstream>
 
 #if USE_CUDA
 #include "cuda/cuda_runtime.h"
@@ -143,6 +144,7 @@ class TensorObj : public TensorBaseObj {
     }
 
     void printData() const;
+    void dumpData(std::ofstream &ofs) const;
     bool equalData(const Tensor &rhs, double relativeError = 1e-6) const;
 
     template <typename T> bool equalData(const vector<T> &dataVector) {
@@ -198,14 +200,21 @@ class TensorObj : public TensorBaseObj {
                 if (a[i] != b[i])
                     return false;
             } else if constexpr (std::is_floating_point_v<T>) {
-                if (fabs(a[i] - b[i]) / std::max(fabs(a[i]), fabs(b[i])) >
+                if (std::min(fabs(a[i]), fabs(b[i])) == 0. &&
+                    fabs(a[i] - b[i]) > relativeError) {
+                    printf("Error on %lu: %f %f\n", i, a[i], b[i]);
+                    return false;
+                } else if (std::min(fabs(a[i]), fabs(b[i])) != 0. &&
+                           fabs(a[i] - b[i]) /
+                                   std::max(fabs(a[i]), fabs(b[i])) >
                                relativeError) {
                     printf("Error on %lu: %f %f\n", i, a[i], b[i]);
                     return false;
                 }
-            } else
+            } else {
                 static_assert(!sizeof(T), "Unsupported data type");
             }
+        }
         return true;
     }
 
@@ -239,8 +248,8 @@ class TensorObj : public TensorBaseObj {
     //         // std::cerr << "Init beginned " << std::endl;
     // #pragma omp parallel for
     //         for (size_t i = 0; i < iEnd; ++i)
-    //             data[i] = fastrand(random_seed[omp_get_thread_num() * 16]) %
+    //             data[i] = fastrand(random_seed[omp_get_thread_num() *
-    //             10000;
+    //             16]) % 10000;
     //         // std::cerr << "Init finished" << std::endl;
     //         computed = ComputedFull;
     //         return true;
@@ -285,8 +294,8 @@ class TensorObj : public TensorBaseObj {
     //         auto nDim = dims.size();
     //         auto nBroadcastDim = ds.size() - nDim;
     //         for (size_t i = 0; i < nDim; ++i)
-    //             if (ds[nBroadcastDim + i] < 0 || ds[nBroadcastDim + i] >=
+    //             if (ds[nBroadcastDim + i] < 0 || ds[nBroadcastDim +
-    //             dims[i])
+    //             i] >= dims[i])
     //                 return (size_t)-1;
     //         size_t idx = 0;
     //         for (size_t i = 0; i < nDim; ++i)
@@ -345,12 +354,14 @@ class TensorObj : public TensorBaseObj {
     //         return (g_seed >> 16) & 0x7FFF;
     //     }
 
-    //     std::vector<std::vector<int>> const *getSplittingPoints() const {
+    //     std::vector<std::vector<int>> const *getSplittingPoints()
+    //     const {
     //         assert(!splittingPoints.empty());
     //         return &splittingPoints;
     //     }
 
-    //     bool setSplittingPoints(std::vector<std::vector<int>> value) {
+    //     bool setSplittingPoints(std::vector<std::vector<int>> value)
+    //     {
     //         assert(!value.empty());
     //         splittingPoints = value;
     //         return true;

include/core/workspace.h

@@ -0,0 +1,42 @@
+#pragma once
+#include "core/runtime.h"
+
+namespace infini {
+
+template <class T> class WorkspaceObj {
+  private:
+    T workspace;           // workspace pointer
+    size_t workspaceSize;  // Size of workspace
+    size_t workspaceAlloc; // currently use workspace size
+
+  public:
+    WorkspaceObj(T workspace_, size_t workspaceSize_)
+        : workspace(workspace_), workspaceSize(workspaceSize_) {
+        workspaceAlloc = 0;
+    }
+    virtual ~WorkspaceObj() {
+        // Dealloc workspace in RuntimeObj
+        // Set workspace = nullptr here
+        workspace = nullptr;
+    }
+    size_t getWorkspaceSize() const { return workspaceSize; }
+
+    T getWorkspace(size_t size) {
+        // Get unused workspace
+        IT_ASSERT(size + workspaceAlloc <= workspaceSize);
+        auto ret = (T)(static_cast<uint8_t *>(workspace) + workspaceAlloc);
+        workspaceAlloc += size;
+        return ret;
+    }
+    T getWorkspace() {
+        // Override getWorkspace in order to dealloc in runtime
+        return workspace;
+    }
+    void resetWorkspace() {
+        // Reset workspaceAlloc every time end kernel
+        workspaceAlloc = 0;
+    }
+    size_t getWorkspaceAlloc() const { return workspaceAlloc; }
+};
+
+} // namespace infini

include/kunlun/kunlun_act_type.h

@@ -0,0 +1,23 @@
+#include "core/op_type.h"
+#include "kunlun/kunlun_common.h"
+
+namespace infini {
+using KunlunActType = xdnn::Activation_t;
+KunlunActType parseActType(ActType act) {
+    switch (act) {
+    case ActType::None:
+        return KunlunActType::LINEAR;
+    case ActType::Tanh:
+        return KunlunActType::TANH;
+    case ActType::Sigmoid:
+        return KunlunActType::SIGMOID;
+    case ActType::Relu:
+        return KunlunActType::RELU6;
+    default:
+        fprintf(stderr, "Activation Type not support yet!\n");
+        break;
+    }
+    return KunlunActType::LINEAR;
+}
+
+}; // namespace infini

include/kunlun/kunlun_common.h

@@ -3,6 +3,8 @@
 #include "xpu/runtime_ex.h"
 #include "xpu/xdnn.h"
 
+namespace xdnn = baidu::xpu::api;
+
 #define checkKUNLUNError(call)                                                 \
     {                                                                          \
         auto err = call;                                                       \

include/kunlun/kunlun_runtime.h

@@ -1,28 +1,35 @@
 #pragma once
 #include "core/runtime.h"
+#include "core/workspace.h"
 #include "kunlun/kunlun_common.h"
-
+#ifdef INFINI_USE_XCCL
+#include "kunlun/xccl_communicator.h"
+#endif
 namespace infini {
 
 class KUNLUNRuntimeObj : public RuntimeObj {
   private:
-    baidu::xpu::api::Context *xdnn;
+    xdnn::Context *ctx;
-    KUNLUNPtr workspace;
+    std::unique_ptr<CommunicatorObj> comm;
-    size_t workspaceSize;
+    // KUNLUNPtr workspace;
+    // size_t workspaceSize;
+    Workspace<KUNLUNPtr> workspace;
 
   public:
-    KUNLUNRuntimeObj() : RuntimeObj(Device::KUNLUN) {
+    KUNLUNRuntimeObj(int deviceId = 0) : RuntimeObj(Device::KUNLUN) {
-        xdnn = baidu::xpu::api::create_context();
+        xpu_set_device(deviceId);
+        ctx = xdnn::create_context();
         // 10GB for Longformer
         // size_t longformerNum = 3lu * (1 << 30);
-        workspaceSize = 3ll << 30; // 3 GB
+        size_t workspaceSize = 3llu << 30; // 3 GB
-        // std::cout<<workspaceSize/1024/1024/1024<< std::endl;
+        KUNLUNPtr wkspacePtr = alloc(workspaceSize);
-        // std::cout<<std::bitset<64>(workspaceSize)<< std::endl;
+        workspace =
-        workspace = alloc(workspaceSize);
+            make_ref<WorkspaceObj<KUNLUNPtr>>(wkspacePtr, workspaceSize);
     }
     virtual ~KUNLUNRuntimeObj() {
-        dealloc(workspace);
+        KUNLUNPtr wkspacePtr = workspace->getWorkspace();
-        baidu::xpu::api::destroy_context(xdnn);
+        dealloc(wkspacePtr);
+        xdnn::destroy_context(ctx);
     }
     string toString() const override;
 
@@ -31,6 +38,7 @@ class KUNLUNRuntimeObj : public RuntimeObj {
     // double runEvaluation(const Graph &graph, int nWarmups,
     //                      int nEvaluations) const;
     void sync() const;
+
     KUNLUNPtr alloc(size_t size) override {
         void *ptr;
         checkKUNLUNError(
@@ -38,33 +46,33 @@ class KUNLUNRuntimeObj : public RuntimeObj {
         return ptr;
     }
     void dealloc(void *ptr) override { xpu_free(ptr); }
-    baidu::xpu::api::Context *KUNLUNHandle() const { return xdnn; }
+
+    xdnn::Context *KUNLUNHandle() const { return ctx; }
+    // Get $size workspace by bytes
     KUNLUNPtr getWorkspace(size_t size) const {
-        IT_ASSERT(size <= workspaceSize);
+        auto ret = workspace->getWorkspace(size);
-        return workspace;
+        return ret;
     }
+    Workspace<KUNLUNPtr> getWorkspaceObj() const { return workspace; }
 
     void copyBlobFromCPU(void *dst, const void *src,
                          size_t bytes) const override {
         xpu_memcpy(dst, const_cast<void *>(src), bytes,
                    XPUMemcpyKind::XPU_HOST_TO_DEVICE);
     }
-
     void copyBlobToCPU(void *dst, const void *src,
                        size_t bytes) const override {
         xpu_memcpy(dst, const_cast<void *>(src), bytes,
                    XPUMemcpyKind::XPU_DEVICE_TO_HOST);
     }
-
     void copyBlobInsideRuntime(void *dst, const void *src,
                                size_t bytes) const override {
         xpu_memcpy(dst, const_cast<void *>(src), bytes,
                    XPUMemcpyKind::XPU_DEVICE_TO_DEVICE);
     }
+    void initComm(const string &name, int worldSize, int rank) final;
 
-    void initComm(const string &, int, int) override { IT_TODO_HALT(); }
+    CommunicatorObj &getCommunicator() const final { return *comm; }
-
-    CommunicatorObj &getCommunicator() const override { IT_TODO_HALT(); }
 
   private:
     void runWithoutSync(const Graph &graph, bool tune, bool profiling) const;

include/kunlun/xccl_communicator.h

@@ -0,0 +1,60 @@
+#pragma once
+#include "core/communicator.h"
+#include "xpu/bkcl.h"
+#include <chrono>
+#include <filesystem>
+#include <fstream>
+#include <thread>
+
+#define checkXcclError(call)                                                   \
+    {                                                                          \
+        auto err = call;                                                       \
+        if (BKCL_SUCCESS != err) {                                             \
+            fprintf(stderr, "XCCL error in %s:%i.\n", __FILE__, __LINE__);     \
+            exit(EXIT_FAILURE);                                                \
+        }                                                                      \
+    }
+
+namespace infini {
+
+class XcclCommunicatorObj final : public CommunicatorObj {
+  private:
+    BKCLContext_t comm;
+
+  public:
+    XcclCommunicatorObj(const string &name, int worldSize, int rank)
+        : CommunicatorObj(worldSize, rank) {
+        const std::string filePath("./" + name + "_xccl_id.bin");
+        BKCLUniqueId commId;
+        if (rank == 0) {
+            checkXcclError(bkcl_get_unique_id(&commId));
+            std::ofstream ofs(filePath, std::ios::binary);
+            ofs.write((char *)&commId, sizeof(BKCLUniqueId));
+        } else {
+            auto begin = std::chrono::steady_clock::now();
+            while (!std::filesystem::exists(filePath)) {
+                auto now = std::chrono::steady_clock::now();
+                _IT_ASSERT_2(now < begin + std::chrono::seconds(10),
+                             "time limit (10s) exceeded.");
+                std::this_thread::sleep_for(std::chrono::milliseconds(100));
+            }
+            std::ifstream ifs(filePath, std::ios::binary);
+            ifs.read((char *)&commId, sizeof(BKCLUniqueId));
+        }
+        checkXcclError(bkcl_init_rank(&comm, rank, worldSize, &commId));
+        if (rank == 0) {
+            std::filesystem::remove(filePath);
+        }
+    }
+
+    BKCLContext_t getXcclComm() { return comm; }
+
+    ~XcclCommunicatorObj() final { checkXcclError(bkcl_destroy_context(comm)); }
+    virtual string toString() const final {
+        std::ostringstream oss;
+        oss << "XCCL communicator";
+        return oss.str();
+    }
+};
+
+} // namespace infini

include/operators/unary.h

@@ -159,6 +159,7 @@ enum class CastType {
     Uint322Int64,
     Float162Float,
     BFloat162Float,
+    Float2Float,
 };
 
 class CastObj : public OperatorObj {

include/utils/broadcast_shape.h

@@ -1,14 +0,0 @@
-#pragma once
-
-namespace infini {
-void broadcastShape(const Shape &originShape, SmallArray &modifyShape,
-                    int nDims, int size) {
-    for (int i = nDims - size - 1; i >= 0; --i) {
-        modifyShape.data[i] = 1;
-    }
-    for (int i = nDims - 1; i >= nDims - size; --i) {
-        modifyShape.data[i] = originShape[i - nDims + size];
-    }
-}
-
-} // namespace infini

include/utils/operator_utils.h

@@ -5,6 +5,9 @@
 #include "core/operator.h"
 #include "core/tensor.h"
 
+#include "utils/small_array.h"
+#include <numeric>
+
 namespace infini {
 
 // Launch a broadcast shape based on the shape of input A and B
@@ -20,6 +23,12 @@ size_t delocate_index(const Shape &shapeIndex, const Shape &shape,
                       const Shape &stride);
 // Convert KernelAttrs to a string representation
 std::string get_kernel_attrs_str(const KernelAttrs &kernelAttrs);
+// VectorProd
+int shapeProd(std::vector<int>::iterator start, std::vector<int>::iterator end);
+void broadcastShape(const Shape &originShape, SmallArray &modifyShape,
+                    int nDims, int size);
+void broadcastShape(const Shape &tempShape, Shape &modifyShape);
+
 } // namespace infini
 
 #endif

include/utils/small_array.h

@@ -4,6 +4,14 @@ namespace infini {
 #define SMALL_ARRAY_SIZE 8
 struct SmallArray {
     int data[SMALL_ARRAY_SIZE];
+
+    int prod(int start, int end) {
+        int result = 1;
+        for (int i = start; i < end; ++i) {
+            result *= data[i];
+        }
+        return result;
+    }
 };
 
 } // namespace infini

pyinfinitensor/src/pyinfinitensor/onnx.py

@@ -1,4 +1,4 @@
-import backend
+import backend
 from onnx import (
     ModelProto,
     TensorProto,
@@ -208,8 +208,8 @@ class OnnxStub:
                 )
             elif node.op_type == "MatMul":
                 tensors[node.output[0]] = self.handler.matmul(
-                    tensors[node.input[0]],
+                    tensors[node.input[0]], # input
-                    tensors[node.input[1]],
+                    tensors[node.input[1]], # weight
                     tensors.get(node.output[0]),
                     False,
                     False,

src/core/graph_handler.cc

@@ -695,6 +695,8 @@ static CastType inferCastType(Tensor input, int to) {
         return CastType::Float162Float;
     } else if (iType == DataType::BFloat16 && oType == DataType::Float32) {
         return CastType::BFloat162Float;
+    } else if (iType == DataType::Float32 && oType == DataType::Float32) {
+        return CastType::Float2Float;
     } else {
         IT_TODO_HALT_MSG("Unsupported CastType : input_type is " +
                          iType.toString() + " output_type is " +

src/core/tensor.cc

@@ -66,6 +66,36 @@ void TensorObj::setShape(Shape shape_) {
     _size = size;
 }
 
+void TensorObj::dumpData(std::ofstream &ofs) const {
+    IT_ASSERT(data != nullptr);
+    if (!runtime->isCpu())
+        IT_TODO_HALT();
+
+#define TRY_DUMP(N)                                                            \
+    if (dtype == DataType(N))                                                  \
+        ofs << dataToString<DT<N>::t>() << std::endl;
+
+    TRY_DUMP(0)           // fmt: new line
+    else TRY_DUMP(1)      //
+        else TRY_DUMP(2)  //
+        else TRY_DUMP(3)  //
+        else TRY_DUMP(4)  //
+        else TRY_DUMP(5)  //
+        else TRY_DUMP(6)  //
+        else TRY_DUMP(7)  //
+        else TRY_DUMP(8)  //
+        else TRY_DUMP(9)  //
+        else TRY_DUMP(10) //
+        else TRY_DUMP(11) //
+        else TRY_DUMP(12) //
+        else TRY_DUMP(13) //
+        else TRY_DUMP(16) //
+        else IT_TODO_HALT();
+    ofs.flush();
+
+#undef TRY_DUMP
+}
+
 void TensorObj::printData() const {
     IT_ASSERT(data != nullptr);
     if (!runtime->isCpu())

src/ffi/ffi_infinitensor.cc

@@ -429,7 +429,9 @@ void init_graph_builder(py::module &m) {
 #endif
 #ifdef USE_KUNLUN
     py::class_<KUNLUNRuntimeObj, std::shared_ptr<KUNLUNRuntimeObj>, RuntimeObj>(
-        m, "KUNLUNRuntime");
+        m, "KUNLUNRuntime")
+        .def(py::init<int>(), py::arg("device") = 0)
+        .def("init_comm", &KUNLUNRuntimeObj::initComm);
 #endif
     py::class_<TensorObj, std::shared_ptr<TensorObj>>(m, "Tensor",
                                                       py::buffer_protocol())

src/kernels/cpu/unary.cc

@@ -145,6 +145,9 @@ class NativeUnary : public CpuKernelWithoutConfig {
         case OpType::Atanh:
             _doCompute = aTanhCompute<T>;
             break;
+        case OpType::Acosh:
+            _doCompute = aCoshCompute<T>;
+            break;
         default:
             IT_TODO_HALT();
         }

src/kernels/cuda/where.cc

@@ -2,7 +2,7 @@
 #include "cuda/cuda_kernel_wihtout_config.h"
 #include "cuda/cuda_runtime.h"
 #include "cuda/cuda_where.h"
-#include "utils/broadcast_shape.h"
+#include "utils/operator_utils.h"
 
 namespace infini {
 

src/kernels/kunlun/all_gather.cc

@@ -0,0 +1,43 @@
+#ifdef INFINI_USE_XCCL
+#include "operators/all_gather.h"
+#include "kunlun/kunlun_kernel_without_config.h"
+#include "kunlun/kunlun_runtime.h"
+#include "kunlun/xccl_communicator.h"
+
+namespace infini {
+class AllGatherXCCL : public KUNLUNKernelWithoutConfig {
+  public:
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<AllGatherObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+        int world_size = op->getWorldSize();
+        IT_ASSERT(world_size == context->getCommunicator().getWorldSize());
+        void *input = op->getInputs(0)->getRawDataPtr<void *>();
+        KUNLUNPtr output_temp =
+            context->getWorkspace(op->getInputs(0)->getBytes() * world_size);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        size_t bytes = op->getInputs(0)->getBytes();
+        size_t count = bytes / op->getDType().getSize();
+
+        BKCLContext_t comm =
+            dynamic_cast<XcclCommunicatorObj &>(context->getCommunicator())
+                .getXcclComm();
+        // TODO: Using the default stream 0
+        checkXcclError(
+            bkcl_all_gather(comm, input, count, output_temp, BKCL_FLOAT, 0));
+
+        for (int i = 0; i < world_size; ++i) {
+            Tensor output = op->getOutput(i);
+            context->copyBlobInsideRuntime(
+                output->getRawDataPtr<float *>(),
+                static_cast<float *>(output_temp) + i * count, bytes);
+        }
+    }
+};
+
+REGISTER_KERNEL(Device::KUNLUN, OpType::AllGather, AllGatherXCCL,
+                "AllGatcher_XCCL_KUNLUN");
+} // namespace infini
+#endif

src/kernels/kunlun/all_reduce.cc

@@ -0,0 +1,49 @@
+#ifdef INFINI_USE_XCCL
+#include "operators/all_reduce.h"
+#include "kunlun/kunlun_kernel_without_config.h"
+#include "kunlun/kunlun_runtime.h"
+#include "kunlun/xccl_communicator.h"
+
+namespace infini {
+class AllReduceXCCL : public KUNLUNKernelWithoutConfig {
+  public:
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<AllReduceBaseObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+        void *input = op->getInputs(0)->getRawDataPtr<void *>();
+        void *output = op->getOutput(0)->getRawDataPtr<void *>();
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        size_t count = op->getInputs(0)->size();
+
+        BKCLContext_t comm =
+            dynamic_cast<XcclCommunicatorObj &>(context->getCommunicator())
+                .getXcclComm();
+        checkXcclError(bkcl_all_reduce(comm, input, output, count,
+                                       BKCLDataType::BKCL_FLOAT, getRedOp(),
+                                       0));
+    }
+    virtual BKCLOp getRedOp() const = 0;
+};
+
+class AllReduceSumXCCL : public AllReduceXCCL {
+    BKCLOp getRedOp() const override { return BKCLOp::BKCL_ADD; }
+};
+
+class AllReduceMinXCCL : public AllReduceXCCL {
+    BKCLOp getRedOp() const override { return BKCLOp::BKCL_MIN; }
+};
+
+class AllReduceMaxXCCL : public AllReduceXCCL {
+    BKCLOp getRedOp() const override { return BKCLOp::BKCL_MAX; }
+};
+
+REGISTER_KERNEL(Device::KUNLUN, OpType::AllReduceSum, AllReduceSumXCCL,
+                "AllReduce_Sum_XCCL_KUNLUN");
+REGISTER_KERNEL(Device::KUNLUN, OpType::AllReduceMax, AllReduceMaxXCCL,
+                "AllReduce_Max_XCCL_KUNLUN");
+REGISTER_KERNEL(Device::KUNLUN, OpType::AllReduceMin, AllReduceMinXCCL,
+                "AllReduce_Min_XCCL_KUNLUN");
+} // namespace infini
+#endif

src/kernels/kunlun/batch_norm.cc

@@ -26,7 +26,7 @@ class BatchNormXdnn : public KUNLUNKernelWithoutConfig {
         int h = dims[2];
         int c = dims[1];
         int n = dims[0];
-        auto ret = baidu::xpu::api::batch_norm_infer<float>(
+        auto ret = xdnn::batch_norm_infer<float>(
             context->KUNLUNHandle(), (float *)input, (float *)output, n, c, h,
             w, op->getEps(), (float *)scale, (float *)bias, (float *)mean,
             (float *)var, true);

src/kernels/kunlun/broadcast.cc

@@ -0,0 +1,32 @@
+#ifdef INFINI_USE_XCCL
+#include "operators/broadcast.h"
+#include "kunlun/kunlun_kernel_without_config.h"
+#include "kunlun/kunlun_runtime.h"
+#include "kunlun/xccl_communicator.h"
+
+namespace infini {
+class BroadcastXCCL : public KUNLUNKernelWithoutConfig {
+  public:
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<BroadcastObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+        void *input = op->getInputs(0)->getRawDataPtr<void *>();
+        void *output = op->getOutput()->getRawDataPtr<void *>();
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        size_t count = op->getInputs(0)->getBytes() / op->getDType().getSize();
+
+        BKCLContext_t comm =
+            dynamic_cast<XcclCommunicatorObj &>(context->getCommunicator())
+                .getXcclComm();
+        // TODO: Using default stream 0 for now.
+        checkXcclError(bkcl_broadcast(comm, input, output, count, BKCL_FLOAT,
+                                      op->getRoot(), 0));
+    }
+};
+
+REGISTER_KERNEL(Device::KUNLUN, OpType::Broadcast, BroadcastXCCL,
+                "Broadcast_XCCL_KUNLUN");
+} // namespace infini
+#endif

src/kernels/kunlun/cast.cc

@@ -17,74 +17,78 @@ class CastXdnn : public KUNLUNKernelWithoutConfig {
         int ret = 0;
         switch (type) {
         case CastType::Float2Float16:
-            ret = baidu::xpu::api::cast<float, float16>(
+            ret = xdnn::cast<float, float16>(
                 context->KUNLUNHandle(), (float *)aData, (float16 *)cData, len);
             break;
         case CastType::Float2Int64:
-            ret = baidu::xpu::api::cast<float, int64_t>(
+            ret = xdnn::cast<float, int64_t>(
                 context->KUNLUNHandle(), (float *)aData, (int64_t *)cData, len);
             break;
         case CastType::Float2Int32:
-            ret = baidu::xpu::api::cast<float, int>(
+            ret = xdnn::cast<float, int>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (float *)aData, (int *)cData, len);
+                                         (float *)aData, (int *)cData, len);
             break;
         case CastType::Float2Int16:
-            ret = baidu::xpu::api::cast<float, int16_t>(
+            ret = xdnn::cast<float, int16_t>(
                 context->KUNLUNHandle(), (float *)aData, (int16_t *)cData, len);
             break;
         case CastType::Float2Int8:
-            ret = baidu::xpu::api::cast<float, int8_t>(
+            ret = xdnn::cast<float, int8_t>(
                 context->KUNLUNHandle(), (float *)aData, (int8_t *)cData, len);
             break;
         case CastType::Int322Float:
-            ret = baidu::xpu::api::cast<int, float>(
+            ret = xdnn::cast<int, float>(context->KUNLUNHandle(), (int *)aData,
-                context->KUNLUNHandle(), (int *)aData, (float *)cData, len);
+                                         (float *)cData, len);
             break;
         case CastType::Int322Int8:
-            ret = baidu::xpu::api::cast<int, int8_t>(
+            ret = xdnn::cast<int, int8_t>(context->KUNLUNHandle(), (int *)aData,
-                context->KUNLUNHandle(), (int *)aData, (int8_t *)cData, len);
+                                          (int8_t *)cData, len);
             break;
         case CastType::Int322Int16:
-            ret = baidu::xpu::api::cast<int, int16_t>(
+            ret = xdnn::cast<int, int16_t>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int *)aData, (int16_t *)cData, len);
+                                           (int *)aData, (int16_t *)cData, len);
             break;
         case CastType::Int162Float:
-            ret = baidu::xpu::api::cast<int16_t, float>(
+            ret = xdnn::cast<int16_t, float>(
                 context->KUNLUNHandle(), (int16_t *)aData, (float *)cData, len);
             break;
         case CastType::Int162Int32:
-            ret = baidu::xpu::api::cast<int16_t, int>(
+            ret = xdnn::cast<int16_t, int>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int16_t *)aData, (int *)cData, len);
+                                           (int16_t *)aData, (int *)cData, len);
             break;
         case CastType::Int82Float:
-            ret = baidu::xpu::api::cast<int8_t, float>(
+            ret = xdnn::cast<int8_t, float>(
                 context->KUNLUNHandle(), (int8_t *)aData, (float *)cData, len);
             break;
         case CastType::Int82Int16:
-            ret = baidu::xpu::api::cast<int8_t, int16_t>(
+            ret = xdnn::cast<int8_t, int16_t>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int8_t *)aData, (int16_t *)cData,
+                                              (int8_t *)aData, (int16_t *)cData,
                                               len);
             break;
         case CastType::Int82Int32:
-            ret = baidu::xpu::api::cast<int8_t, int>(
+            ret = xdnn::cast<int8_t, int>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int8_t *)aData, (int *)cData, len);
+                                          (int8_t *)aData, (int *)cData, len);
             break;
         case CastType::Int322Int64:
-            ret = baidu::xpu::api::cast<int, int64_t>(
+            ret = xdnn::cast<int, int64_t>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int *)aData, (int64_t *)cData, len);
+                                           (int *)aData, (int64_t *)cData, len);
             break;
         case CastType::Int642Int32:
-            ret = baidu::xpu::api::cast<int64_t, int>(
+            ret = xdnn::cast<int64_t, int>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (int64_t *)aData, (int *)cData, len);
+                                           (int64_t *)aData, (int *)cData, len);
             break;
         case CastType::Int642Float:
-            ret = baidu::xpu::api::cast<int64_t, float>(
+            ret = xdnn::cast<int64_t, float>(
                 context->KUNLUNHandle(), (int64_t *)aData, (float *)cData, len);
             break;
         case CastType::Float162Float:
-            ret = baidu::xpu::api::cast<float16, float>(
+            ret = xdnn::cast<float16, float>(
                 context->KUNLUNHandle(), (float16 *)aData, (float *)cData, len);
             break;
+        case CastType::Float2Float:
+            ret = xdnn::copy<float>(context->KUNLUNHandle(), (float *)aData,
+                                    (float *)cData, len);
+            break;
         default:
             IT_TODO_HALT();
         }

src/kernels/kunlun/concat.cc

@@ -26,8 +26,8 @@ class ConcatXdnn : public KUNLUNKernelWithoutConfig {
             }
             dims.push_back(dim);
         }
-        auto ret = baidu::xpu::api::concat<float>(
+        auto ret = xdnn::concat<float>(context->KUNLUNHandle(), inputsData,
-            context->KUNLUNHandle(), inputsData, (float *)cData, dims, axis);
+                                       (float *)cData, dims, axis);
         assert(ret == 0);
         return;
     }

src/kernels/kunlun/conv.cc

@@ -24,11 +24,17 @@ class ConvXdnn : public KUNLUNKernelWithoutConfig {
         std::vector<int> stride = {sh, sw};
         std::vector<int> dilation = {dh, dw};
 
-        auto ret = baidu::xpu::api::conv2d<float, float, float, float>(
+        // TODO: Convolution operators still have some accuracy problems
+        checkKUNLUNError((xdnn::conv2d<float, float, float, float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
             (float *)cData, n, c, h, w, f, ksize, stride, pads, dilation, g,
-            nullptr, nullptr, nullptr, true);
+            nullptr, nullptr, nullptr, true)));
-        assert(ret == 0);
+
+        // checkKUNLUNError((xdnn::conv2d_fusion<float, float, float, float>(
+        //     context->KUNLUNHandle(), (float *const)aData, (float
+        //     *const)bData, (float *)cData, n, c, h, w, f, ksize, stride, pads,
+        //     dilation, g, nullptr, nullptr, nullptr, true, nullptr, nullptr,
+        //     xdnn::Activation_t::LINEAR)));
         return;
     }
 };

src/kernels/kunlun/conv_trans.cc

@@ -37,8 +37,7 @@ class ConvTransXdnn : public KUNLUNKernelWithoutConfig {
         if (dimOutput.size() != 4)
             IT_TODO_HALT();
 
-        auto ret =
+        auto ret = xdnn::conv2d_transpose<float, float, float, float>(
-            baidu::xpu::api::conv2d_transpose<float, float, float, float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
             (float *)cData, n, c, h, w, f, ksize, stride, pads, dilation, g,
             nullptr, nullptr, nullptr, isNCHW);

src/kernels/kunlun/element_wise.cc

@@ -1,6 +1,7 @@
 #include "operators/element_wise.h"
 #include "kunlun/kunlun_kernel_without_config.h"
 #include "kunlun/kunlun_runtime.h"
+#include "utils/operator_utils.h"
 
 namespace infini {
 class AddXdnn : public KUNLUNKernelWithoutConfig {
@@ -22,10 +23,9 @@ class AddXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_add<float>(
+        checkKUNLUNError(xdnn::broadcast_add<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -49,10 +49,9 @@ class SubXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_sub<float>(
+        checkKUNLUNError(xdnn::broadcast_sub<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -76,10 +75,9 @@ class MulXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_mul<float>(
+        checkKUNLUNError(xdnn::broadcast_mul<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -95,18 +93,40 @@ class DivXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
 
+        auto aSize = op->getInputs(0)->size();
         auto aDim = op->getInputs(0)->getDims();
+        auto bSize = op->getInputs(1)->size();
         auto bDim = op->getInputs(1)->getDims();
-        if (aDim.size() == 0) {
+        auto dtype = op->getDType();
-            aDim.push_back(1);
+
-        }
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_div<float>(
+
+        if (aSize == bSize) {
+            // Do ElementWise Sub with no broadcast
+            checkKUNLUNError(xdnn::div<float>(context->KUNLUNHandle(),
+                                              (float *)aData, (float *)bData,
+                                              (float *)cData, aSize));
+        } else {
+            // Do broadcast div
+            Shape aligned = infer_broadcast(aDim, bDim);
+            if (aligned == aDim) {
+                // BData need to be broadcasted
+                checkKUNLUNError(xdnn::broadcast_div<float>(
                     context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+                    (float *)cData, aDim, bDim));
-        assert(ret == 0);
+            } else {
+                // Use workspace to broadcast aData
+                KUNLUNPtr wks = context->getWorkspace(bSize * dtype.getSize());
+                checkKUNLUNError(xdnn::broadcast<float>(
+                    context->KUNLUNHandle(), (float *)aData, (float *)wks, aDim,
+                    bDim));
+                checkKUNLUNError(xdnn::div<float>(context->KUNLUNHandle(),
+                                                  (float *)wks, (float *)bData,
+                                                  (float *)cData, bSize));
+            }
+        }
         return;
     }
 };
@@ -131,10 +151,9 @@ class PowXdnn : public KUNLUNKernelWithoutConfig {
             bDim.push_back(1);
         }
 
-        auto ret = baidu::xpu::api::broadcast_pow<float>(
+        checkKUNLUNError(xdnn::broadcast_pow<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -158,10 +177,9 @@ class MaxXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_max<float>(
+        checkKUNLUNError(xdnn::broadcast_max<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -185,10 +203,9 @@ class MinXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_min<float>(
+        checkKUNLUNError(xdnn::broadcast_min<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -204,7 +221,9 @@ class EqualXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -214,12 +233,11 @@ class EqualXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_equal<float>(
+        checkKUNLUNError(xdnn::broadcast_equal<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (bool *)wsData, aDim, bDim);
+            (bool *)wsData, aDim, bDim));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -235,7 +253,8 @@ class GreaterEqualXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -245,12 +264,11 @@ class GreaterEqualXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_greater_equal<float>(
+        checkKUNLUNError(xdnn::broadcast_greater_equal<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (bool *)wsData, aDim, bDim);
+            (bool *)wsData, aDim, bDim));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -266,7 +284,8 @@ class GreaterThanXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        KUNLUNPtr wsData =
+            context->getWorkspace(len * (op->getDType()).getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -276,12 +295,11 @@ class GreaterThanXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_greater_than<float>(
+        checkKUNLUNError(xdnn::broadcast_greater_than<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (bool *)wsData, aDim, bDim);
+            (bool *)wsData, aDim, bDim));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -297,7 +315,8 @@ class LessEqualXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -307,12 +326,11 @@ class LessEqualXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_less_equal<float>(
+        checkKUNLUNError(xdnn::broadcast_less_equal<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (bool *)wsData, aDim, bDim);
+            (bool *)wsData, aDim, bDim));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -328,7 +346,8 @@ class LessThanXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -338,12 +357,11 @@ class LessThanXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_less_than<float>(
+        checkKUNLUNError(xdnn::broadcast_less_than<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (bool *)wsData, aDim, bDim);
+            (bool *)wsData, aDim, bDim));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -367,10 +385,9 @@ class FloorDivXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::broadcast_floordiv<float>(
+        checkKUNLUNError(xdnn::broadcast_floordiv<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)bData,
-            (float *)cData, aDim, bDim);
+            (float *)cData, aDim, bDim));
-        assert(ret == 0);
         return;
     }
 };
@@ -388,10 +405,9 @@ class MSELossXdnn : public KUNLUNKernelWithoutConfig {
         size_t len = op->getOutput()->size();
 
         auto dim = op->getInputs(0)->getDims();
-        auto ret = baidu::xpu::api::mse_loss<float>(
+        checkKUNLUNError(xdnn::mse_loss<float>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (float *)aData, (float *)bData,
+                                               (float *)aData, (float *)bData,
-            (float *)cData, len);
+                                               (float *)cData, len));
-        assert(ret == 0);
         return;
     }
 };
@@ -407,7 +423,8 @@ class AndXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -417,12 +434,11 @@ class AndXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::logical_and<bool>(
+        checkKUNLUNError(xdnn::logical_and<bool>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (bool *)aData, (bool *)bData,
+                                                 (bool *)aData, (bool *)bData,
-            (bool *)wsData, len);
+                                                 (bool *)wsData, len));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -438,7 +454,8 @@ class OrXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -448,12 +465,11 @@ class OrXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::logical_or<bool>(
+        checkKUNLUNError(xdnn::logical_or<bool>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (bool *)aData, (bool *)bData,
+                                                (bool *)aData, (bool *)bData,
-            (bool *)wsData, len);
+                                                (bool *)wsData, len));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -469,7 +485,8 @@ class XorXdnn : public KUNLUNKernelWithoutConfig {
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
         auto bDim = op->getInputs(1)->getDims();
@@ -479,12 +496,11 @@ class XorXdnn : public KUNLUNKernelWithoutConfig {
         if (bDim.size() == 0) {
             bDim.push_back(1);
         }
-        auto ret = baidu::xpu::api::logical_xor<bool>(
+        checkKUNLUNError(xdnn::logical_xor<bool>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (bool *)aData, (bool *)bData,
+                                                 (bool *)aData, (bool *)bData,
-            (bool *)wsData, len);
+                                                 (bool *)wsData, len));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };
@@ -499,14 +515,14 @@ class NotXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         size_t len = op->getOutput()->size();
-        KUNLUNPtr wsData = context->getWorkspace(len);
+        auto dtype = op->getDType();
+        KUNLUNPtr wsData = context->getWorkspace(len * dtype.getSize());
 
         auto aDim = op->getInputs(0)->getDims();
-        auto ret = baidu::xpu::api::logical_not<bool>(
+        checkKUNLUNError(xdnn::logical_not<bool>(
-            context->KUNLUNHandle(), (bool *)aData, (bool *)wsData, len);
+            context->KUNLUNHandle(), (bool *)aData, (bool *)wsData, len));
-        ret = baidu::xpu::api::cast<bool, float>(
+        checkKUNLUNError((xdnn::cast<bool, float>(
-            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len);
+            context->KUNLUNHandle(), (bool *)wsData, (float *)cData, len)));
-        assert(ret == 0);
         return;
     }
 };

src/kernels/kunlun/gather.cc

@@ -1,4 +1,5 @@
 #include "operators/gather.h"
+#include "core/common.h"
 #include "kunlun/kunlun_kernel_without_config.h"
 #include "kunlun/kunlun_runtime.h"
 
@@ -10,17 +11,18 @@ class GatherXdnn : public KUNLUNKernelWithoutConfig {
         IT_ASSERT(op->getDType() == DataType::Float32);
         auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
 
-        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
+        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>()); // data
-        void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
+        void *const bData =
+            (op->getInputs(1)->getRawDataPtr<void *>()); // indice
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
 
-        auto shape = op->getInputs(0)->getDims();
+        Shape aShape = op->getInputs(0)->getDims();
-        auto index = op->getInputs(1)->getDims();
+        Tensor bTensor = op->getInputs(1);
-        auto axis = op->getAxis();
+        int axis = op->getAxis();
-        auto ret = baidu::xpu::api::gather<float, int>(
+        checkKUNLUNError((baidu::xpu::api::gather<float, int>(
             context->KUNLUNHandle(), (float *)aData, (int *)bData,
-            (float *)cData, shape, index.size(), axis);
+            (float *)cData, aShape, bTensor->size(), axis)));
-        assert(ret == 0);
+
         return;
     }
 };

src/kernels/kunlun/matmul.cc

@@ -1,30 +1,123 @@
 #include "operators/matmul.h"
+#include "kunlun/kunlun_act_type.h"
+#include "kunlun/kunlun_common.h"
 #include "kunlun/kunlun_kernel_without_config.h"
 #include "kunlun/kunlun_runtime.h"
+#include "utils/operator_utils.h"
 
 namespace infini {
 class MatmulXdnn : public KUNLUNKernelWithoutConfig {
+
     void compute(const Operator &_op,
                  const RuntimeObj *_context) const override {
+        // This kernel do C = act(alpha * x * w + beta * bias)
         auto op = as<MatmulObj>(_op);
         IT_ASSERT(op->getDType() == DataType::Float32);
         auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
-        void *const cData = (op->getOutput()->getRawDataPtr<void *>());
+        void *const outData = (op->getOutput()->getRawDataPtr<void *>());
+
+        Shape aDims = op->getInputs(0)->getDims();
+        Shape bDims = op->getInputs(1)->getDims();
+        Shape cDims = op->getOutput()->getDims();
+
+        const auto [b, m, n, k] = op->getBMNK();
         bool transA = op->getTransA();
         bool transB = op->getTransB();
+        int rankA = op->getInputs(0)->getRank();
+        int rankB = op->getInputs(1)->getRank();
+        int rankAligned = std::max(rankA, rankB);
+        IT_ASSERT(rankAligned <= SMALL_ARRAY_SIZE);
 
-        auto b = op->getB();
+        float alpha = 1.f, beta = 0.f;
-        auto m = op->getM();
+        Tensor biasTensor = op->getBias();
-        auto n = op->getN();
+        DataType dtype = op->getDType();
-        auto k = op->getK();
 
-        auto ret = baidu::xpu::api::fc_batched<float, float, float, float>(
+        if (b > 1) {
-            context->KUNLUNHandle(), b, transA, transB, m, n, k, 1.0,
+            SmallArray alignedAShape;
-            (float *)aData, m * k, (float *)bData, n * k, 0.0, (float *)cData,
+            SmallArray alignedBShape;
-            m * n, nullptr, nullptr);
+            // Padding 1 in aShape and bShape in order to align rank
-        assert(ret == 0);
+            broadcastShape(aDims, alignedAShape, rankAligned, rankA);
+            broadcastShape(bDims, alignedBShape, rankAligned, rankB);
+            // Calculate batch dim
+            int batchA = alignedAShape.prod(0, rankAligned - 2);
+            int batchB = alignedBShape.prod(0, rankAligned - 2);
+            // View aShape bShape to 3 dim
+            Shape aDimsMatmul = {batchA, aDims[rankA - 2], aDims[rankA - 1]};
+            Shape bDimsMatmul = {batchB, bDims[rankB - 2], bDims[rankB - 1]};
+            auto numOutput = op->getOutput()->size();
+            KUNLUNPtr wkspace = nullptr;
+            void *AData = nullptr;
+            void *BData = nullptr;
+            void *CData = nullptr;
+            if (batchA != batchB) {
+                // If bs not equal, then broadcast
+                IT_ASSERT(batchA == 1 || batchB == 1);
+                if (batchA == 1) {
+                    // Broadcast aShapeMatmul in batch dimension
+                    Shape aDimsTarget = {b, aDimsMatmul[1], aDimsMatmul[2]};
+                    auto numInput =
+                        shapeProd(aDimsTarget.begin(), aDimsTarget.end());
+                    wkspace = context->getWorkspace(numInput * dtype.getSize());
+                    checkKUNLUNError(xdnn::broadcast<float>(
+                        context->KUNLUNHandle(), (float *)aData,
+                        (float *)wkspace, aDimsMatmul, aDimsTarget));
+                    AData = wkspace;
+                    BData = bData;
+                    CData =
+                        biasTensor
+                            ? context->getWorkspace(numOutput * dtype.getSize())
+                            : outData;
+                } else {
+                    // Broadcast bShapeMatmul in batch dimension
+                    Shape bDimsTarget = {b, bDimsMatmul[1], bDimsMatmul[2]};
+                    auto numInput =
+                        shapeProd(bDimsTarget.begin(), bDimsTarget.end());
+                    wkspace = context->getWorkspace(numInput * dtype.getSize());
+                    checkKUNLUNError(xdnn::broadcast<float>(
+                        context->KUNLUNHandle(), (float *)bData,
+                        (float *)wkspace, bDimsMatmul, bDimsTarget));
+                    AData = aData;
+                    BData = wkspace;
+                    CData =
+                        biasTensor
+                            ? context->getWorkspace(numOutput * dtype.getSize())
+                            : outData;
+                }    // endif batchA == 1
+            } else { // batchA == batchB, no need to broadcast
+                AData = aData;
+                BData = bData;
+                CData = biasTensor
+                            ? context->getWorkspace(numOutput * dtype.getSize())
+                            : outData;
+            }
+            checkKUNLUNError((xdnn::fc_batched<float, float, float, float>(
+                context->KUNLUNHandle(), b, transA, transB, m, n, k, alpha,
+                (float *)AData, m * k, (float *)BData, n * k, beta,
+                (float *)CData, m * n, nullptr, nullptr)));
+            // Broadcast_add xw and bias if bias exists
+            if (biasTensor) {
+                auto biasShape = biasTensor->getDims();
+                broadcastShape(cDims, biasShape);
+                checkKUNLUNError(baidu::xpu::api::broadcast_add<float>(
+                    context->KUNLUNHandle(), (float *)CData,
+                    biasTensor->getRawDataPtr<float *>(), (float *)outData,
+                    cDims, biasShape));
+            }
+        } else {
+            // Matmul with no batch, call fc_fusion
+            const int lda = transA ? m : k, ldb = transB ? k : n, ldc = n;
+            auto kunlunAct = parseActType(std::move(op->getAct()));
+            checkKUNLUNError(
+                (baidu::xpu::api::fc_fusion<float, float, float, float>(
+                    context->KUNLUNHandle(), (float *)aData, (float *)bData,
+                    (float *)outData, m, n, k, transA, transB, nullptr, nullptr,
+                    nullptr, lda, ldb, ldc, alpha, 0.f,
+                    biasTensor ? biasTensor->getRawDataPtr<float *>() : nullptr,
+                    kunlunAct, nullptr)));
+        }
         return;
     }
 };

src/kernels/kunlun/pooling.cc

@@ -14,11 +14,23 @@ class AvgPooling : public KUNLUNKernelWithoutConfig {
 
         auto [n, c, h, w, kh, kw] = op->getNCHWRS();
         auto [ph, pw, sh, sw, dh, dw] = op->getPadStrideDilation();
+        auto outShape = op->getOutput()->getDims();
 
         std::vector<int> ksize = {kh, kw};
         std::vector<int> stride = {sh, sw};
         std::vector<int> pad = {ph, pw};
 
+        int yh = outShape[op->getOutput()->getRank() - 2];
+        int yw = outShape[op->getOutput()->getRank() - 1];
+
+        // If Maxpool with ceilMode true
+        // We need to change padding in order to call xdnn api
+        if (op->getCeilMode() && yh > (h + 2 * ph - kh) / sh + 1) {
+            auto padh = yh - ((h + 2 * ph - kh) / sh + 1);
+            auto padw = yw - ((w + 2 * pw - kw) / sw + 1);
+            pad = {0, padh, 0, padw};
+        }
+
         auto ret = baidu::xpu::api::avg_pool2d<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)cData, n, c, h, w,
             ksize, stride, pad, true, true, nullptr, nullptr);
@@ -38,21 +50,30 @@ class MaxPooling : public KUNLUNKernelWithoutConfig {
 
         auto [n, c, h, w, kh, kw] = op->getNCHWRS();
         auto [ph, pw, sh, sw, dh, dw] = op->getPadStrideDilation();
+        auto outShape = op->getOutput()->getDims();
 
         std::vector<int> ksize = {kh, kw};
         std::vector<int> stride = {sh, sw};
         std::vector<int> pad = {ph, pw};
 
-        int yh = (h + ph * 2 - kh) / sh + 1;
+        int yh = outShape[op->getOutput()->getRank() - 2];
-        int yw = (w + pw * 2 - kw) / sw + 1;
+        int yw = outShape[op->getOutput()->getRank() - 1];
 
-        KUNLUNPtr indices = context->getWorkspace(yh * yw * 4);
+        // If Maxpool with ceilMode true
+        // We need to change padding in order to call xdnn api
+        if (op->getCeilMode() && yh > (h + 2 * ph - kh) / sh + 1) {
+            auto padh = yh - ((h + 2 * ph - kh) / sh + 1);
+            auto padw = yw - ((w + 2 * pw - kw) / sw + 1);
+            pad = {0, padh, 0, padw};
+        }
 
-        auto ret = baidu::xpu::api::max_pool2d<float>(
+        KUNLUNPtr indices = context->getWorkspace(yh * yw * sizeof(int));
+
+        checkKUNLUNError(baidu::xpu::api::max_pool2d<float>(
             context->KUNLUNHandle(), (float *)aData, (float *)cData,
             (int *)indices, n, c, h, w, ksize, stride, pad, true, nullptr,
-            nullptr, false);
+            nullptr, false));
-        assert(ret == 0);
+
         return;
     }
 };

src/kernels/kunlun/reduce.cc

@@ -1,8 +1,9 @@
+#include "operators/reduce.h"
 #include "kunlun/kunlun_kernel_without_config.h"
 #include "kunlun/kunlun_runtime.h"
-#include "operators/reduce.h"
 
 namespace infini {
+
 class ReduceMeanXdnn : public KUNLUNKernelWithoutConfig {
     void compute(const Operator &_op,
                  const RuntimeObj *_context) const override {
@@ -26,6 +27,31 @@ class ReduceMeanXdnn : public KUNLUNKernelWithoutConfig {
     }
 };
 
+class ReduceSumXdnn : public KUNLUNKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<ReduceSumObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
+        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
+        void *const cData = (op->getOutput()->getRawDataPtr<void *>());
+
+        auto axes_set = op->getAxes();
+        std::vector<int> axes;
+        axes.assign(axes_set.begin(), axes_set.end());
+        auto shape = op->getInputs(0)->getDims();
+
+        auto ret = baidu::xpu::api::reduce_sum<float>(
+            context->KUNLUNHandle(), (float *)aData, (float *)cData, shape,
+            axes);
+        assert(ret == 0);
+        return;
+    }
+};
+
 REGISTER_KERNEL(Device::KUNLUN, OpType::ReduceMean, ReduceMeanXdnn,
                 "ReduceMean_xdnn_KUNLUN");
+REGISTER_KERNEL(Device::KUNLUN, OpType::ReduceSum, ReduceSumXdnn,
+                "ReduceSum_xdnn_KUNLUN");
 }; // namespace infini

src/kernels/kunlun/select.cc

@@ -1,32 +0,0 @@
-#include "kunlun/kunlun_kernel_without_config.h"
-#include "kunlun/kunlun_runtime.h"
-#include "operators/where.h"
-
-namespace infini {
-class WhereXdnn : public KUNLUNKernelWithoutConfig {
-    void compute(const Operator &_op,
-                 const RuntimeObj *_context) const override {
-        auto op = as<WhereObj>(_op);
-        IT_ASSERT(op->getDType() == DataType::Float32);
-        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
-
-        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
-        void *const bData = (op->getInputs(1)->getRawDataPtr<void *>());
-        void *const cData = (op->getInputs(2)->getRawDataPtr<void *>());
-        void *const dData = (op->getOutput()->getRawDataPtr<void *>());
-
-        auto aDim = op->getInputs(0)->getDims();
-        auto bDim = op->getInputs(1)->getDims();
-        auto cDim = op->getInputs(2)->getDims();
-        auto dDim = op->getOutput()->getDims();
-
-        auto ret = baidu::xpu::api::select<float>(
-            context->KUNLUNHandle(), (bool *)cData, (float *)aData,
-            (float *)bData, (float *)dData, cDim, aDim);
-        assert(ret == 0);
-        return;
-    }
-};
-
-REGISTER_KERNEL(Device::KUNLUN, OpType::Where, WhereXdnn, "Where_xdnn_KUNLUN");
-}; // namespace infini

src/kernels/kunlun/slice.cc

@@ -0,0 +1,39 @@
+#include "operators/slice.h"
+#include "kunlun/kunlun_kernel_without_config.h"
+#include "kunlun/kunlun_runtime.h"
+
+namespace infini {
+class SliceXdnn : public KUNLUNKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<SliceObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
+        void *inData = op->getInputs(0)->getRawDataPtr<void *>();
+        void *outData = op->getOutput()->getRawDataPtr<void *>();
+
+        // Get attributes of Slice OP
+        Shape starts = op->getStarts(), ends = op->getEnds(),
+              steps = op->getSteps();
+        Shape inShape = op->getInputs(0)->getDims();
+        // If all steps are 1, set continuous True
+        bool continuous =
+            (size_t)std::count(steps.begin(), steps.end(), 1) == steps.size();
+        if (continuous) {
+            // if continuous, call xdnn::slice
+            checkKUNLUNError(
+                xdnn::slice<float>(context->KUNLUNHandle(), (float *)inData,
+                                   (float *)outData, inShape, starts, ends));
+
+        } else {
+            // else call xdnn::strided_slice
+            checkKUNLUNError(xdnn::strided_slice<float>(
+                context->KUNLUNHandle(), (float *)inData, (float *)outData,
+                inShape, starts, ends, steps));
+        }
+    }
+};
+
+REGISTER_KERNEL(Device::KUNLUN, OpType::Slice, SliceXdnn, "Slice_xdnn_KUNLUN")
+}; // namespace infini

src/kernels/kunlun/softmax.cc

@@ -15,9 +15,9 @@ class SoftmaxXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
 
-        auto ret = baidu::xpu::api::softmax<float>(
+        checkKUNLUNError(xdnn::softmax<float>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, dim, axis);
+                                              (float *)aData, (float *)cData,
-        assert(ret == 0);
+                                              dim, axis));
         return;
     }
 };

src/kernels/kunlun/transpose.cc

@@ -16,13 +16,9 @@ class TransposeXdnn : public KUNLUNKernelWithoutConfig {
         auto dimin = op->getInputs(0)->getDims();
         auto permute = op->getPermute();
 
-        if (dimin.size() != 4) {
+        auto ret =
-            IT_TODO_HALT();
+            xdnn::transpose<float>(context->KUNLUNHandle(), (float *)aData,
-        }
+                                   (float *)cData, dimin, permute);
-
-        auto ret = baidu::xpu::api::transpose<float>(
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, dimin,
-            permute);
         assert(ret == 0);
         return;
     }
@@ -46,9 +42,9 @@ class DepthToSpaceXdnn : public KUNLUNKernelWithoutConfig {
         } else {
             permute = {0, 1, 4, 2, 5, 3};
         }
-        auto ret = baidu::xpu::api::transpose<float>(
+        auto ret =
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, reshape,
+            xdnn::transpose<float>(context->KUNLUNHandle(), (float *)aData,
-            permute);
+                                   (float *)cData, reshape, permute);
         assert(ret == 0);
         return;
     }

src/kernels/kunlun/unary.cc

@@ -14,8 +14,8 @@ class ReluXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::relu<float>(
+        auto ret = xdnn::relu<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -32,8 +32,8 @@ class SigmoidXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::sigmoid<float>(
+        auto ret = xdnn::sigmoid<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                        (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -50,8 +50,45 @@ class TanhXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::tanh<float>(
+        auto ret = xdnn::tanh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
+        assert(ret == 0);
+        return;
+    }
+};
+
+class HardSwishXdnn : public KUNLUNKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<UnaryObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
+        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
+        void *const cData = (op->getOutput()->getRawDataPtr<void *>());
+        auto len = op->getInputs(0)->size();
+
+        auto ret = xdnn::hard_swish<float>(context->KUNLUNHandle(),
+                                           (float *)aData, (float *)cData, len);
+        assert(ret == 0);
+        return;
+    }
+};
+
+class HardSigmoidXdnn : public KUNLUNKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<UnaryObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
+        void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
+        void *const cData = (op->getOutput()->getRawDataPtr<void *>());
+        auto len = op->getInputs(0)->size();
+
+        // Slop set to 0.2 as default
+        auto ret = xdnn::hard_sigmoid<float>(
+            context->KUNLUNHandle(), (float *)aData, (float *)cData, len, 0.2);
         assert(ret == 0);
         return;
     }
@@ -68,8 +105,8 @@ class SquareXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::square<float>(
+        auto ret = xdnn::square<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                       (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -86,8 +123,8 @@ class SqrtXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::sqrt<float>(
+        auto ret = xdnn::sqrt<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -104,8 +141,8 @@ class RsqrtXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::rsqrt<float>(
+        auto ret = xdnn::rsqrt<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                      (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -122,8 +159,8 @@ class ExpXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::exp<float>(
+        auto ret = xdnn::exp<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -140,8 +177,8 @@ class CeilXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::ceil<float>(
+        auto ret = xdnn::ceil<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -160,9 +197,8 @@ class ClipXdnn : public KUNLUNKernelWithoutConfig {
         float min = op->getMin().value();
         float max = op->getMax().value();
 
-        auto ret = baidu::xpu::api::clip<float>(context->KUNLUNHandle(),
+        auto ret = xdnn::clip<float>(context->KUNLUNHandle(), (float *)aData,
-                                                (float *)aData, (float *)cData,
+                                     (float *)cData, len, min, max);
-                                                len, min, max);
         assert(ret == 0);
         return;
     }
@@ -179,8 +215,8 @@ class FloorXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::floor<float>(
+        auto ret = xdnn::floor<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                      (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -197,8 +233,8 @@ class NegXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::neg<float>(
+        auto ret = xdnn::neg<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -214,8 +250,8 @@ class CopyXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::copy<float>(
+        auto ret = xdnn::copy<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -232,8 +268,8 @@ class ReciprocalXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::reciprocal<float>(
+        auto ret = xdnn::reciprocal<float>(context->KUNLUNHandle(),
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                           (float *)aData, (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -250,8 +286,8 @@ class AbsXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::abs<float>(
+        auto ret = xdnn::abs<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -268,8 +304,8 @@ class ATanXdnn : public KUNLUNKernelWithoutConfig {
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
 
-        auto ret = baidu::xpu::api::arctan<float>(
+        auto ret = xdnn::arctan<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                       (float *)cData, len);
         assert(ret == 0);
         return;
     }
@@ -288,35 +324,35 @@ class LogXdnn : public KUNLUNKernelWithoutConfig {
             1,
         };
         auto len = op->getInputs(0)->size();
+        auto dtype = op->getDType();
         // get ptr of tempspace
-        KUNLUNPtr temp = context->getWorkspace(len * sizeof(float));
+        KUNLUNPtr temp = context->getWorkspace(len * dtype.getSize());
         LogObj::LogType type = op->getType();
         // get output of xpu::api::loge(x)
-        auto ret = baidu::xpu::api::log<float>(
+        auto ret = xdnn::log<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)temp, len);
+                                    (float *)temp, len);
         // get ptr of divider
-        KUNLUNPtr dd =
+        KUNLUNPtr dd = context->getWorkspace(1 * dtype.getSize());
-            (float *)(context->getWorkspace((1 + len) * sizeof(float))) + len;
         // choose from logE, log2, log10
         switch (type) {
             float constant;
         case LogObj::LogE:
             // if use loge, copy from temp to cData
-            ret = baidu::xpu::api::copy<float>(
+            ret = xdnn::copy<float>(context->KUNLUNHandle(), (float *)temp,
-                context->KUNLUNHandle(), (float *)temp, (float *)cData, len);
+                                    (float *)cData, len);
             break;
         case LogObj::Log2:
             constant = std::log(2);
             context->copyBlobFromCPU(dd, &constant, sizeof(float));
-            ret = baidu::xpu::api::broadcast_div<float>(
+            ret = xdnn::broadcast_div<float>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (float *)temp, (float *)dd,
+                                             (float *)temp, (float *)dd,
                                              (float *)cData, aDim, divDim);
             break;
         case LogObj::Log10:
             constant = std::log(10);
             context->copyBlobFromCPU(dd, &constant, sizeof(float));
-            ret = baidu::xpu::api::broadcast_div<float>(
+            ret = xdnn::broadcast_div<float>(context->KUNLUNHandle(),
-                context->KUNLUNHandle(), (float *)temp, (float *)dd,
+                                             (float *)temp, (float *)dd,
                                              (float *)cData, aDim, divDim);
             break;
         default:
@@ -337,8 +373,8 @@ class CosXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::cos<float>(
+        auto ret = xdnn::cos<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -354,8 +390,8 @@ class SinXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::sin<float>(
+        auto ret = xdnn::sin<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -371,8 +407,8 @@ class TanXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::tan<float>(
+        auto ret = xdnn::tan<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -388,8 +424,8 @@ class SinhXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::sinh<float>(
+        auto ret = xdnn::sinh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -405,8 +441,8 @@ class CoshXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::cosh<float>(
+        auto ret = xdnn::cosh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                     (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -422,8 +458,8 @@ class ErfXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::erf<float>(
+        auto ret = xdnn::erf<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                    (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -439,8 +475,8 @@ class ACosXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::arccos<float>(
+        auto ret = xdnn::arccos<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                       (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -456,8 +492,8 @@ class ACoshXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::acosh<float>(
+        auto ret = xdnn::acosh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                      (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -473,8 +509,8 @@ class ASinXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::arcsin<float>(
+        auto ret = xdnn::arcsin<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                       (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -490,8 +526,8 @@ class ASinhXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::asinh<float>(
+        auto ret = xdnn::asinh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                      (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -507,8 +543,8 @@ class ATanhXdnn : public KUNLUNKernelWithoutConfig {
         void *const aData = (op->getInputs(0)->getRawDataPtr<void *>());
         void *const cData = (op->getOutput()->getRawDataPtr<void *>());
         auto len = op->getInputs(0)->size();
-        auto ret = baidu::xpu::api::atanh<float>(
+        auto ret = xdnn::atanh<float>(context->KUNLUNHandle(), (float *)aData,
-            context->KUNLUNHandle(), (float *)aData, (float *)cData, len);
+                                      (float *)cData, len);
 
         assert(ret == 0);
         return;
@@ -546,7 +582,10 @@ REGISTER_KERNEL(Device::KUNLUN, OpType::Erf, ErfXdnn, "Erf_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Acos, ACosXdnn, "ACos_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Acosh, ACoshXdnn, "ACosh_xdnn");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Asin, ASinXdnn, "ASin_xdnn");
-REGISTER_KERNEL(Device::KUNLUN, OpType::Asinh, ASinhXdnn,
+REGISTER_KERNEL(Device::KUNLUN, OpType::Asinh, ASinhXdnn, "ASinh_xdnn");
-                "ASinh_xdnn_Float3 2");
 REGISTER_KERNEL(Device::KUNLUN, OpType::Atanh, ATanhXdnn, "ATanh_xdnn");
+REGISTER_KERNEL(Device::KUNLUN, OpType::HardSwish, HardSwishXdnn,
+                "HardSwish_xdnn");
+REGISTER_KERNEL(Device::KUNLUN, OpType::HardSigmoid, HardSigmoidXdnn,
+                "HardSigmoid_xdnn");
 }; // namespace infini

src/kernels/kunlun/where.cc

@@ -0,0 +1,67 @@
+#pragma GCC diagnostic ignored "-Wunused-variable"
+#include "operators/where.h"
+#include "kunlun/kunlun_kernel_without_config.h"
+#include "kunlun/kunlun_runtime.h"
+#include "utils/operator_utils.h"
+
+namespace infini {
+
+class WhereXdnn : public KUNLUNKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<WhereObj>(_op);
+        IT_ASSERT(op->getDType() == DataType::Float32);
+        auto context = dynamic_cast<const KUNLUNRuntimeObj *>(_context);
+
+        void *const aData =
+            (op->getInputs(0)->getRawDataPtr<void *>()); // inputX
+        void *const bData =
+            (op->getInputs(1)->getRawDataPtr<void *>()); // inputY
+        void *const cData =
+            (op->getInputs(2)->getRawDataPtr<void *>()); // condition
+        void *const dData =
+            (op->getOutput()->getRawDataPtr<void *>()); // output
+
+        auto aDim = op->getInputs(0)->getDims(); // dimX
+        auto bDim = op->getInputs(1)->getDims(); // dimY
+        auto cDim = op->getInputs(2)->getDims(); // dimCondition
+        auto dDim = op->getOutput()->getDims();  //  dimOutput
+
+        auto dtype = op->getDType();
+
+        if (aDim != bDim) {
+            // Infer broadcast for X and Y
+            Shape XYDim = infer_broadcast(aDim, bDim);
+            int XYSize = std::accumulate(XYDim.begin(), XYDim.end(), 1,
+                                         std::multiplies<int>());
+            // Align rank for XYDim and aDim or bDim
+            broadcastShape(XYDim, aDim);
+            broadcastShape(XYDim, bDim);
+            // Get workspace
+            void *wkspace = context->getWorkspace(XYSize * dtype.getSize());
+            // Broadcast X Y
+            checkKUNLUNError(xdnn::broadcast<float>(
+                context->KUNLUNHandle(),
+                (float *)(XYDim == aDim ? bData : aData), (float *)wkspace,
+                (XYDim == aDim ? bDim : aDim), XYDim));
+            // Align Rank
+            broadcastShape(dDim, XYDim);
+            broadcastShape(dDim, XYDim);
+            // Where
+            void *XData = XYDim == aDim ? aData : wkspace;
+            void *YData = XYDim == bDim ? bData : wkspace;
+            checkKUNLUNError(xdnn::select<float>(
+                context->KUNLUNHandle(), (bool *)cData, (float *)XData,
+                (float *)YData, (float *)dData, cDim, XYDim));
+        } else {
+            checkKUNLUNError(xdnn::select<float>(
+                context->KUNLUNHandle(), (bool *)cData, (float *)aData,
+                (float *)bData, (float *)dData, cDim, aDim));
+        }
+
+        return;
+    }
+};
+
+REGISTER_KERNEL(Device::KUNLUN, OpType::Where, WhereXdnn, "Where_xdnn_KUNLUN");
+}; // namespace infini

src/kunlun/kunlun_runtime.cc

@@ -19,6 +19,7 @@ void KUNLUNRuntimeObj::runWithoutSync(const Graph &graph, bool tune = false,
         auto perfData = perfEngine.getPerfData(perfKey);
         if (!perfData && !tune) {
             kernel->compute(op, this);
+            workspace->resetWorkspace();
             continue;
         }
 
@@ -52,8 +53,20 @@ void KUNLUNRuntimeObj::run(const Graph &graph, bool tune,
     sync();
 }
 
-void KUNLUNRuntimeObj::sync() const { ; }
+void KUNLUNRuntimeObj::sync() const { xpu_wait(); }
 
 string KUNLUNRuntimeObj::toString() const { return "KUNLUN Runtime"; }
 
+void KUNLUNRuntimeObj::initComm(const string &name, int worldSize, int rank) {
+    IT_ASSERT(worldSize > 0);
+    IT_ASSERT(rank >= 0);
+    IT_ASSERT(rank < worldSize);
+    IT_ASSERT(!comm) << "communicator is already initialized.";
+#ifdef INFINI_USE_XCCL
+    comm = std::make_unique<XcclCommunicatorObj>(name, worldSize, rank);
+#else
+    IT_TODO_HALT_MSG("Not compiled with XCCL");
+#endif
+}
+
 } // namespace infini

src/operators/matmul.cc

@@ -25,7 +25,7 @@ optional<vector<Shape>> MatmulObj::inferShape(const TensorVec &inputs) {
     auto A = inputs[0], B = inputs[1];
     auto shapeA = A->getDims();
     auto shapeB = B->getDims();
-    int rankA = A->getRank();
+    int rankA = A->getRank(); // Rank is the Shape of TensorDims
     int rankB = B->getRank();
     Shape shapeA1(shapeA.begin(), shapeA.begin() + (rankA - 2));
     Shape shapeB1(shapeB.begin(), shapeB.begin() + (rankB - 2));

src/operators/unary.cc

@@ -231,6 +231,8 @@ DataType CastObj::getOutputDataType() const {
         return DataType::Float32;
     case CastType::Float2BFloat16:
         return DataType::BFloat16;
+    case CastType::Float2Float:
+        return DataType::Float32;
     default:
         IT_TODO_HALT();
     }

src/utils/operator_utils.cc

@@ -114,4 +114,28 @@ std::string get_kernel_attrs_str(const KernelAttrs &kernelAttrs) {
     std::string opStr = OpType(std::get<1>(kernelAttrs)).toString();
     return deviceStr + ", " + opStr;
 }
+
+int shapeProd(std::vector<int>::iterator start,
+              std::vector<int>::iterator end) {
+    return std::accumulate(start, end, 1, std::multiplies<int>());
+}
+
+void broadcastShape(const Shape &originShape, SmallArray &modifyShape,
+                    int nDims, int size) {
+    for (int i = nDims - size - 1; i >= 0; --i) {
+        modifyShape.data[i] = 1;
+    }
+    for (int i = nDims - 1; i >= nDims - size; --i) {
+        modifyShape.data[i] = originShape[i - nDims + size];
+    }
+}
+
+void broadcastShape(const Shape &tempShape, Shape &modifyShape) {
+    // Align Rank, Add 1 in the start of smallShape
+    IT_ASSERT(tempShape.size() >= modifyShape.size());
+    modifyShape.insert(modifyShape.begin(),
+                       tempShape.size() - modifyShape.size(), 1);
+    return;
+}
+
 } // namespace infini

test/kernels/kunlun/test_kunlun_allgather.cc

@@ -0,0 +1,50 @@
+#ifdef INFINI_USE_XCCL
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/all_gather.h"
+#include "test.h"
+#include "xpu/bkcl.h"
+#include <thread>
+
+static int WORLD_SIZE = 2;
+
+namespace infini {
+
+void allGather(const string taskName, int deviceID, vector<float> data,
+               vector<vector<float>> ans) {
+    // Create Runtimes and initiate communication
+    Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    Runtime kunlunRuntime = make_ref<KUNLUNRuntimeObj>(deviceID);
+    kunlunRuntime->initComm(taskName, WORLD_SIZE, deviceID);
+    // Create Graph and insert allReduce operation
+    Graph g = make_ref<GraphObj>(kunlunRuntime);
+    auto input =
+        g->addTensor(Shape{static_cast<int>(data.size())}, DataType::Float32);
+    auto op = g->addOp<AllGatherObj>(input, std::nullopt, WORLD_SIZE);
+    // Copy data from CPU to GPU
+    g->dataMalloc();
+    input->copyin(data);
+    // Run operation
+    kunlunRuntime->run(g);
+    // Copy output from GPU to CPU
+    for (int i = 0; i < WORLD_SIZE; ++i) {
+        auto result = op->getOutputs()[i]->clone(cpuRuntime);
+        EXPECT_TRUE(result->equalData(ans[i]));
+    }
+}
+
+TEST(KUNLUN_AllGather, run) {
+    vector<float> data[2] = {{2., 3.}, {5., 6.}};
+    vector<vector<float>> ans = {{2., 3.}, {5., 6.}};
+
+    std::vector<std::thread> threads;
+    for (int gpu = 0; gpu < WORLD_SIZE; ++gpu) {
+        threads.emplace_back(allGather, "test_all_gather", gpu, data[gpu], ans);
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+}
+} // namespace infini
+#endif

test/kernels/kunlun/test_kunlun_allreduce.cc

@@ -0,0 +1,72 @@
+#ifdef INFINI_USE_XCCL
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/all_reduce.h"
+#include "test.h"
+#include "xpu/bkcl.h"
+#include <thread>
+
+static int WORLD_SIZE = 2;
+
+using namespace infini;
+
+template <typename OperatorObj>
+void allReduce(const string taskName, int deviceID, vector<float> data,
+               vector<float> ans) {
+    Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    Runtime kunlunRuntime = make_ref<KUNLUNRuntimeObj>(deviceID);
+    kunlunRuntime->initComm(taskName, WORLD_SIZE, deviceID);
+    Graph g = make_ref<GraphObj>(kunlunRuntime);
+    auto input =
+        g->addTensor(Shape{static_cast<int>(data.size())}, DataType::Float32);
+    auto op = g->addOp<OperatorObj>(input, nullptr);
+    g->dataMalloc();
+    input->copyin(data);
+    kunlunRuntime->run(g);
+    auto result = op->getOutput()->clone(cpuRuntime);
+
+    EXPECT_TRUE(result->equalData(ans));
+}
+
+TEST(KUNLUN_AllReduce, sum) {
+    vector<float> data[2] = {{2., 3.}, {5., 6.}};
+    vector<float> ans = {7., 9.};
+    std::vector<std::thread> threads;
+    for (int rank = 0; rank < WORLD_SIZE; ++rank) {
+        threads.emplace_back(allReduce<AllReduceSumObj>, "test_allreduce_sum",
+                             rank, data[rank], ans);
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+}
+
+TEST(KUNLUN_AllReduce, max) {
+    vector<float> data[2] = {{2., 3.}, {5., 6.}};
+    vector<float> ans = {5., 6.};
+
+    std::vector<std::thread> threads;
+    for (int gpu = 0; gpu < WORLD_SIZE; ++gpu) {
+        threads.emplace_back(allReduce<AllReduceMaxObj>, "test_allreduce_max",
+                             gpu, data[gpu], ans);
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+}
+
+TEST(KUNLUN_AllReduce, min) {
+    vector<float> data[2] = {{2., 3.}, {5., 6.}};
+    vector<float> ans = {2., 3.};
+
+    std::vector<std::thread> threads;
+    for (int gpu = 0; gpu < WORLD_SIZE; ++gpu) {
+        threads.emplace_back(allReduce<AllReduceMinObj>, "test_allreduce_min",
+                             gpu, data[gpu], ans);
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+}
+#endif

test/kernels/kunlun/test_kunlun_broadcast.cc

@@ -0,0 +1,56 @@
+#ifdef INFINI_USE_XCCL
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/broadcast.h"
+#include "test.h"
+#include <thread>
+#include <xpu/bkcl.h>
+
+static int WORLD_SIZE = 2;
+static int root = 0;
+
+namespace infini {
+
+void broadcast(const string taskName, int deviceID, vector<float> data,
+               vector<float> ans) {
+    // Create Runtimes and initiate communication
+    Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    Runtime kunlunRuntime = make_ref<KUNLUNRuntimeObj>(deviceID);
+    kunlunRuntime->initComm(taskName, WORLD_SIZE, deviceID);
+    // Create Graph and insert allReduce operation
+    Graph g = make_ref<GraphObj>(kunlunRuntime);
+    auto input =
+        g->addTensor(Shape{static_cast<int>(data.size())}, DataType::Float32);
+    auto op = g->addOp<BroadcastObj>(input, nullptr, root);
+    // Copy data from CPU to GPU
+    g->dataMalloc();
+    // Only rank 0 has the data
+    if (deviceID == root) {
+        input->copyin(data);
+    }
+    // Run broadcast operation
+    kunlunRuntime->run(g);
+    // Copy output from GPU to CPU
+    auto result = op->getOutput()->clone(cpuRuntime);
+
+    EXPECT_TRUE(result->equalData(ans));
+}
+
+TEST(KUNLUN_Broadcast, run) {
+    // Only 1 device gets data. Every rank should have the same data after
+    // broadcast.
+    vector<float> data = {2., 3., 5., 6.};
+    vector<float> ans = {2., 3., 5., 6.};
+
+    std::vector<std::thread> threads;
+    for (int gpu = 0; gpu < WORLD_SIZE; ++gpu) {
+        threads.emplace_back(broadcast, "test_broadcast", gpu, data, ans);
+    }
+    for (auto &thread : threads) {
+        thread.join();
+    }
+}
+} // namespace infini
+
+#endif

test/kernels/kunlun/test_kunlun_gather.cc

@@ -0,0 +1,144 @@
+#include "core/graph.h"
+#include "core/kernel.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/gather.h"
+
+#include "test.h"
+
+namespace infini {
+/*
+test1:
+input = [
+      [1, 2],
+      [3, 4],
+      [5, 6],
+  ]
+  indices = [
+      [0, 1],
+      [1, 2],
+  ]
+  output = [
+      [
+          [1, 2],
+          [3, 4],
+      ],
+      [
+          [3, 4],
+          [5, 6],
+      ],
+  ]
+  axis=0
+  */
+
+/*
+test2
+input = [
+      [0, 1, 2],
+      [3, 4, 5],
+      [6, 7, 8],
+  ]
+  indices = [
+      [0, 2],
+  ]
+  axis = 1,
+  output = [
+          [[0, 2]],
+          [[3, 5]],
+          [[6, 8]],
+  ]
+*/
+/*
+test3
+input=[[[ 0,  1],
+         [ 2,  3],
+         [ 4,  5],
+         [ 6,  7]],
+
+        [[ 8,  9],
+         [10, 11],
+         [12, 13],
+         [14, 15]]]  //(2,4,2)
+indices=[[0],[3],[1]] //(3,1)
+axis=1
+output=
+
+*/
+
+TEST(Gather, KUNLUN) {
+    {
+        Runtime runtime = NativeCpuRuntimeObj::getInstance();
+        Graph gCpu = make_ref<GraphObj>(runtime);
+        auto input = gCpu->addTensor({3, 2}, DataType::Float32);
+        auto index = gCpu->addTensor({2, 2}, DataType::Int32);
+        gCpu->dataMalloc();
+        input->copyin(vector<float>{1, 2, 3, 4, 5, 6});
+        index->copyin(vector<int>{0, 1, 1, 2});
+        auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+        Graph gCuda = make_ref<GraphObj>(kunlunRuntime);
+
+        auto inputCuda = gCuda->cloneTensor(input);
+        auto indexCuda = gCuda->cloneTensor(index);
+        auto op = gCuda->addOp<GatherObj>(inputCuda, indexCuda, nullptr, 0);
+        gCuda->dataMalloc();
+        inputCuda->copyin(vector<float>{1, 2, 3, 4, 5, 6});
+        indexCuda->copyin(vector<int>{0, 1, 1, 2});
+        kunlunRuntime->run(gCuda);
+
+        // cudaPrintTensor(op->getOutput());
+        //   copy output from CUDA to CPU
+        auto oCpu = gCpu->cloneTensor(op->getOutput());
+        EXPECT_TRUE(oCpu->equalData(vector<float>{1, 2, 3, 4, 3, 4, 5, 6}));
+    }
+    {
+        Runtime runtime = NativeCpuRuntimeObj::getInstance();
+        Graph gCpu = make_ref<GraphObj>(runtime);
+        auto input = gCpu->addTensor({3, 3}, DataType::Float32);
+        auto index = gCpu->addTensor({1, 2}, DataType::Int32);
+        gCpu->dataMalloc();
+        input->setData(IncrementalGenerator());
+        index->copyin(vector<int>{0, 2});
+        auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+        Graph gCuda = make_ref<GraphObj>(kunlunRuntime);
+
+        auto inputCuda = gCuda->cloneTensor(input);
+        auto indexCuda = gCuda->cloneTensor(index);
+        auto op = gCuda->addOp<GatherObj>(inputCuda, indexCuda, nullptr, 1);
+        gCuda->dataMalloc();
+        inputCuda->setData(IncrementalGenerator());
+        indexCuda->copyin(vector<int>{0, 2});
+        kunlunRuntime->run(gCuda);
+
+        // cudaPrintTensor(op->getOutput());
+        //  copy output from CUDA to CPU
+        auto oCpu = gCpu->cloneTensor(op->getOutput());
+        EXPECT_TRUE(oCpu->equalData(vector<float>{0, 2, 3, 5, 6, 8}));
+    }
+    {
+        Runtime runtime = NativeCpuRuntimeObj::getInstance();
+        Graph gCpu = make_ref<GraphObj>(runtime);
+        auto input = gCpu->addTensor({3, 2}, DataType::Float32);
+        auto index = gCpu->addTensor({2, 2}, DataType::Int32);
+        gCpu->dataMalloc();
+        input->copyin(std::vector<float>{1.0, 1.2, 2.3, 3.4, 4.5, 5.7});
+        index->copyin(std::vector<int>{0, 1, 1, 2});
+        auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+        Graph gCuda = make_ref<GraphObj>(kunlunRuntime);
+
+        auto inputCuda = gCuda->cloneTensor(input);
+        auto indexCuda = gCuda->cloneTensor(index);
+        auto op = gCuda->addOp<GatherObj>(inputCuda, indexCuda, nullptr, 0);
+        gCuda->dataMalloc();
+        inputCuda->copyin(std::vector<float>{1.0, 1.2, 2.3, 3.4, 4.5, 5.7});
+        indexCuda->copyin(std::vector<int>{0, 1, 1, 2});
+        kunlunRuntime->run(gCuda);
+
+        // cudaPrintTensor(op->getOutput());
+        //  copy output from CUDA to CPU
+        auto oCpu = gCpu->cloneTensor(op->getOutput());
+        EXPECT_TRUE(oCpu->equalData(
+            std::vector<float>{1.0, 1.2, 2.3, 3.4, 2.3, 3.4, 4.5, 5.7}));
+    }
+}
+
+} // namespace infini

test/kernels/kunlun/test_kunlun_matmul.cc

@@ -7,52 +7,112 @@
 #include "test.h"
 
 namespace infini {
+using ExpectOutput = vector<float>;
 
-template <class T>
+void testMatmulKUNLUNWithBias(
-void testMatmul(const std::function<void(void *, size_t, DataType)> &generatorA,
+    const std::function<void(void *, size_t, DataType)> &generatorA,
     const std::function<void(void *, size_t, DataType)> &generatorB,
-                bool transA, bool transB, const Shape &shapeA,
+    const std::function<void(void *, size_t, DataType)> &generatorBias,
-                const Shape &shapeB) {
+    bool transA, bool transB, const Shape &shapeA, const Shape &shapeB,
-    // Runtime
+    const Shape &shapeBias, const ExpectOutput &ansVec) {
-    Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    auto cpuRuntime = NativeCpuRuntimeObj::getInstance();
-    auto xpuRuntime = make_ref<KUNLUNRuntimeObj>();
+    Graph gCpu = make_ref<GraphObj>(cpuRuntime);
+    auto ACpu = gCpu->addTensor(shapeA, DataType::Float32);
+    auto BCpu = gCpu->addTensor(shapeB, DataType::Float32);
+    auto BiasCpu = gCpu->addTensor(shapeBias, DataType::Float32);
+    gCpu->dataMalloc();
+    ACpu->setData(generatorA);
+    BCpu->setData(generatorB);
+    BiasCpu->setData(generatorBias);
 
-    // Build input data on CPU
+    auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
-    Tensor inputCpu1 =
+    auto gKunlun = make_ref<GraphObj>(kunlunRuntime);
-        make_ref<TensorObj>(shapeA, DataType::Float32, cpuRuntime);
+    auto AKunlun = gKunlun->cloneTensor(ACpu);
-    Tensor inputCpu2 =
+    auto BKunlun = gKunlun->cloneTensor(BCpu);
-        make_ref<TensorObj>(shapeB, DataType::Float32, cpuRuntime);
+    auto BiasKunlun = gKunlun->cloneTensor(BiasCpu);
+    auto matmul = gKunlun->addOp<MatmulObj>(AKunlun, BKunlun, nullptr, transA,
+                                            transB, BiasKunlun);
 
-    // MLU
+    // allocate Kunlun memory
-    Graph xpuGraph = make_ref<GraphObj>(xpuRuntime);
+    gKunlun->dataMalloc();
-    auto inputMlu1 = xpuGraph->cloneTensor(inputCpu1);
+    AKunlun->setData(generatorA);
-    auto inputMlu2 = xpuGraph->cloneTensor(inputCpu2);
+    BKunlun->setData(generatorB);
-    auto mluOp = xpuGraph->addOp<T>(inputMlu1, inputMlu2, nullptr);
+    BiasKunlun->setData(generatorBias);
-    xpuGraph->dataMalloc();
+    kunlunRuntime->run(gKunlun);
-    inputMlu1->setData(generatorA);
+
-    inputMlu2->setData(generatorB);
+    auto CCpu = gCpu->cloneTensor(matmul->getOutput());
-    xpuRuntime->run(xpuGraph);
+    // CCpu->printData();
-    auto outputMlu = mluOp->getOutput();
+    //  check results on CPU
-    auto outputMlu2Cpu = outputMlu->clone(cpuRuntime);
+    EXPECT_TRUE(CCpu->equalData(ansVec));
-    // CPU
+    // print a tensor/operator/graph by print()
-    Graph cpuGraph = make_ref<GraphObj>(cpuRuntime);
+    // gKunlun->print();
-    auto cpuOp = cpuGraph->addOp<T>(inputCpu1, inputCpu2, nullptr);
-    cpuGraph->addTensor(inputCpu1);
-    cpuGraph->addTensor(inputCpu2);
-    cpuGraph->dataMalloc();
-    inputCpu1->setData(generatorA);
-    inputCpu2->setData(generatorB);
-    cpuRuntime->run(cpuGraph);
-    auto outputCpu = cpuOp->getOutput();
-    outputCpu->print();
-    outputMlu2Cpu->print();
-    // Check
-    EXPECT_TRUE(outputCpu->equalData(outputMlu2Cpu));
 }
 
-TEST(xpu_Matmul, run) {
+void testMatmulKUNLUN(
-    testMatmul<MatmulObj>(IncrementalGenerator(), IncrementalGenerator(), false,
+    const std::function<void(void *, size_t, DataType)> &generatorA,
-                          false, Shape{2, 3}, Shape{3, 4});
+    const std::function<void(void *, size_t, DataType)> &generatorB,
+    bool transA, bool transB, const Shape &shapeA, const Shape &shapeB,
+    const ExpectOutput &ansVec) {
+    auto cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(cpuRuntime);
+    auto ACpu = gCpu->addTensor(shapeA, DataType::Float32);
+    auto BCpu = gCpu->addTensor(shapeB, DataType::Float32);
+    gCpu->dataMalloc();
+    ACpu->setData(generatorA);
+    BCpu->setData(generatorB);
+
+    auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+    auto gKunlun = make_ref<GraphObj>(kunlunRuntime);
+    auto AKunlun = gKunlun->cloneTensor(ACpu);
+    auto BKunlun = gKunlun->cloneTensor(BCpu);
+    auto matmul = gKunlun->addOp<MatmulObj>(AKunlun, BKunlun, nullptr, transA,
+                                            transB, nullptr);
+
+    // allocate Kunlun memory
+    gKunlun->dataMalloc();
+    AKunlun->setData(generatorA);
+    BKunlun->setData(generatorB);
+    kunlunRuntime->run(gKunlun);
+
+    auto CCpu = gCpu->cloneTensor(matmul->getOutput());
+    // CCpu->printData();
+    //  check results on CPU
+    EXPECT_TRUE(CCpu->equalData(ansVec));
+    // print a tensor/operator/graph by print()
+    // gKunlun->print();
+}
+
+TEST(XDNN_Matmul, run) {
+    testMatmulKUNLUN(IncrementalGenerator(), OneGenerator(), false, false,
+                     Shape{1, 3, 5}, Shape{1, 5, 2},
+                     ExpectOutput{10, 10, 35, 35, 60, 60});
+    testMatmulKUNLUN(IncrementalGenerator(), IncrementalGenerator(), true,
+                     false, Shape{2, 3, 4}, Shape{2, 3, 2},
+                     ExpectOutput{40, 52, 46, 61, 52, 70, 58, 79, 400, 448, 424,
+                                  475, 448, 502, 472, 529});
+    testMatmulKUNLUN(
+        IncrementalGenerator(), IncrementalGenerator(), false, false,
+        Shape{2, 3, 5}, Shape{5, 2},
+        ExpectOutput{60, 70, 160, 195, 260, 320, 360, 445, 460, 570, 560, 695});
+    testMatmulKUNLUN(IncrementalGenerator(), IncrementalGenerator(), true,
+                     false, Shape{2, 5, 3}, Shape{5, 2},
+                     ExpectOutput{180, 210, 200, 235, 220, 260, 480, 585, 500,
+                                  610, 520, 635});
+    testMatmulKUNLUN(IncrementalGenerator(), IncrementalGenerator(), false,
+                     false, Shape{3, 5}, Shape{5, 2},
+                     ExpectOutput{60, 70, 160, 195, 260, 320});
+}
+
+TEST(XDNN_Matmul_With_Bias, run) {
+    testMatmulKUNLUNWithBias(IncrementalGenerator(), OneGenerator(),
+                             OneGenerator(), false, false, Shape{1, 3, 5},
+                             Shape{1, 5, 2}, Shape{2},
+                             ExpectOutput{11, 11, 36, 36, 61, 61});
+    testMatmulKUNLUNWithBias(IncrementalGenerator(), IncrementalGenerator(),
+                             OneGenerator(), true, false, Shape{2, 3, 4},
+                             Shape{2, 3, 2}, Shape{4, 2},
+                             ExpectOutput{41, 53, 47, 62, 53, 71, 59, 80, 401,
+                                          449, 425, 476, 449, 503, 473, 530});
 }
 
 } // namespace infini

test/kernels/kunlun/test_kunlun_slice.cc

@@ -0,0 +1,39 @@
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/slice.h"
+#include "test.h"
+
+namespace infini {
+TEST(KUNLUN_Slice, run) {
+    Runtime cpuRuntime = NativeCpuRuntimeObj::getInstance();
+    auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+
+    // Build input data on CPU
+    Tensor icpu =
+        make_ref<TensorObj>(Shape{3, 2, 1, 5}, DataType::Float32, cpuRuntime);
+    icpu->dataMalloc();
+    icpu->setData(IncrementalGenerator());
+
+    // Build CUDA graph;
+    Graph g = make_ref<GraphObj>(kunlunRuntime);
+    auto i = g->cloneTensor(icpu);
+    auto op =
+        g->addOp<SliceObj>(i, nullptr, vector<int>{1, 1}, vector<int>{2, 5},
+                           vector<int>{0, 3}, std::nullopt);
+
+    // allocate CUDA memory
+    g->dataMalloc();
+    i->setData(IncrementalGenerator());
+
+    // Execute on CUDA
+    kunlunRuntime->run(g);
+
+    // clone CUDA output to CPU
+    auto o = op->getOutput();
+    auto cpuo = o->clone(cpuRuntime);
+    // cudaPrintTensor(o);
+    //  check results on CPU
+    EXPECT_TRUE(cpuo->equalData(vector<float>{11, 12, 13, 14, 16, 17, 18, 19}));
+}
+} // namespace infini

test/kernels/kunlun/test_kunlun_where.cc

@@ -0,0 +1,77 @@
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "kunlun/kunlun_runtime.h"
+#include "operators/where.h"
+
+#include "test.h"
+
+namespace infini {
+
+void test_where(const Shape &inputXShape, const vector<float> &inputXData,
+                const Shape &inputYShape, const vector<float> &inputYData,
+                const Shape &conditionShape,
+                const vector<int8_t> &conditionData,
+                const vector<float> &ExpectData) {
+    Runtime runtime = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(runtime);
+    auto condition = gCpu->addTensor(conditionShape, DataType::Bool);
+    auto inputX = gCpu->addTensor(inputXShape, DataType::Float32);
+    auto inputY = gCpu->addTensor(inputYShape, DataType::Float32);
+
+    gCpu->dataMalloc();
+    condition->copyin(conditionData); //
+    inputX->copyin(inputXData);
+    inputY->copyin(inputYData); //
+
+    auto kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+    Graph gCuda = make_ref<GraphObj>(kunlunRuntime);
+
+    auto conditionGpu = gCuda->cloneTensor(condition);
+    auto inputXGpu = gCuda->cloneTensor(inputX);
+    auto inputYGpu = gCuda->cloneTensor(inputY);
+
+    auto op = gCuda->addOp<WhereObj>(inputXGpu, inputYGpu, conditionGpu,
+                                     nullptr); // WhereObj
+    gCuda->dataMalloc();
+    conditionGpu->copyin(conditionData);
+    inputXGpu->copyin(inputXData);
+    inputYGpu->copyin(inputYData);
+    kunlunRuntime->run(gCuda);
+
+    auto oCpu = gCpu->cloneTensor(op->getOutput()); // move Data from gpu to cpu
+    oCpu->printData();                              //->printData
+    EXPECT_TRUE(oCpu->equalData(ExpectData));
+}
+
+TEST(KUNLUN_Where, run) {
+    test_where(
+        Shape{2, 2, 3, 1}, vector<float>{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11},
+        Shape{2, 2, 3, 1}, vector<float>{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
+        Shape{2, 2, 3, 1}, vector<int8_t>{0, 1, 1, 0, 0, 0, 1, 1, 0, 1, 1, 1},
+        vector<float>{0., 1., 2., 0., 0., 0., 6., 7., 0., 9., 10., 11.});
+
+    test_where(Shape{2, 2, 1, 3}, // inputx
+               vector<float>{0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5},
+               Shape{2, 2, 1, 3}, // inputy
+               vector<float>{1, 1, 3, 2, 5, 1, 5, 2, 3, 5, 6, 7},
+               Shape{2, 2, 1, 3}, // condition
+               vector<int8_t>{0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 0, 0},
+               vector<float>{1, 1, 2, 2, 5, 1, 0, 2, 2, 3, 6, 7});
+
+    test_where(Shape{2, 2, 1, 3},
+               vector<float>{0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5}, // inputX
+               Shape{2, 2, 1, 3},
+               vector<float>{1, 1, 3, 2, 5, 1, 5, 2, 3, 5, 6, 7},   // inputY
+               Shape{2, 1, 1, 3}, vector<int8_t>{1, 1, 0, 1, 1, 1}, // condition
+               vector<float>{0, 1, 3, 3, 4, 1, 0, 1, 2, 3, 4, 5});  // result
+
+    test_where(Shape{2, 2, 1, 3},
+               vector<float>{0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5}, // inputX
+               Shape{2, 2, 1, 3},
+               vector<float>{1, 1, 3, 2, 5, 1, 5, 2, 3, 5, 6, 7}, // inputY
+               Shape{2, 1, 1, 3},
+               vector<int8_t>{1, 1, 0, 1, 1,
+                              1}, // condition               } // python output
+               vector<float>{0, 1, 3, 3, 4, 1, 0, 1, 2, 3, 4, 5}); // result
+}
+} // namespace infini

test/kunlun/test_kunlun_workspace.cc

@@ -0,0 +1,20 @@
+#include "core/runtime.h"
+#include "core/workspace.h"
+#include "kunlun/kunlun_runtime.h"
+
+#include "test.h"
+
+namespace infini {
+TEST(KunlunWorkspace, test) {
+    Ref<KUNLUNRuntimeObj> kunlunRuntime = make_ref<KUNLUNRuntimeObj>();
+    auto wkspace = kunlunRuntime->getWorkspaceObj();
+    KUNLUNPtr space1 = kunlunRuntime->getWorkspace(1024 * 1024 * sizeof(float));
+    IT_ASSERT(wkspace->getWorkspaceAlloc() == 1024 * 1024 * sizeof(float));
+    KUNLUNPtr space2 = kunlunRuntime->getWorkspace(1024 * 1024 * sizeof(float));
+    IT_ASSERT(wkspace->getWorkspaceAlloc() == 1024 * 1024 * sizeof(float) * 2);
+    IT_ASSERT((void *)(static_cast<uint8_t *>(space1) +
+                       1024 * 1024 * sizeof(float)) == (void *)space2);
+    wkspace->resetWorkspace();
+    IT_ASSERT(wkspace->getWorkspaceAlloc() == 0);
+}
+} // namespace infini