fix(mgb): remove internal for cambricon and atlas

GitOrigin-RevId: 861e349eb4

CMakeLists.txt

@@ -38,7 +38,9 @@ option(MGE_CUDA_USE_STATIC "Enable MegEngine CUDA static linking." ON)
 option(MGE_WITH_TRT "Build MegEngine with TensorRT." ON)
 option(MGE_USE_SYSTEM_LIB "Build MegEngine with system libraries." OFF)
 option(MGB_WITH_FLATBUFFERS "Build MegBrain with FlatBuffers serialization support." ON)
+option(MGE_WITH_CAMBRICON "Build MegEngine with Cambricon support" OFF)
 option(BUILD_SHARED_LIBS "Build shared libraries" ON)
+option(MGE_WITH_ATLAS "Build MegEngine with Atlas support" OFF)
 option(MGE_ENABLE_RTTI "Build with RTTI" ON)
 option(MGE_ENABLE_LOGGING "Build with logging" ON)
 option(MGE_DEBUG_UTIL "Enable debug utility" ON)
@@ -406,6 +408,51 @@ if(MGE_WITH_CUDA)
     set(MGE_CUDA_LIBS "${MGE_CUDA_LIBS}")
 endif()
 
+if(MGE_WITH_CAMBRICON)
+    include_directories("$ENV{NEUWARE_HOME}/include")
+    link_directories("$ENV{NEUWARE_HOME}/lib64")
+    include(cmake/FindBANG/FindBANG.cmake)
+    if (${MGE_MLU_ARCH} STREQUAL "MLU100")
+        set(BANG_ARCH "100")
+    elseif (${MGE_MLU_ARCH} STREQUAL "MLU1h8")
+        set(BANG_ARCH "110")
+    elseif (${MGE_MLU_ARCH} STREQUAL "MLU220")
+        set(BANG_ARCH "220")
+    elseif (${MGE_MLU_ARCH} STREQUAL "MLU270")
+        set(BANG_ARCH "270")
+    elseif (${MGE_MLU_ARCH} STREQUAL "MLU290")
+        set(BANG_ARCH "290")
+    elseif (${MGE_MLU_ARCH} STREQUAL "MLU200")
+        set(BANG_ARCH "200")
+    else()
+        message (FATAL_ERROR "Unsupported MLU arch.")
+    endif()
+    set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} --bang-mlu-arch=${MGE_MLU_ARCH}")
+    set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -std=c++11 -Werror")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__BANG_ARCH__=${BANG_ARCH}")
+    if (${CMAKE_BUILD_TYPE} STREQUAL "Debug")
+        set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -g -O0")
+    elseif (${CMAKE_BUILD_TYPE} STREQUAL "Release")
+        set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -O3")
+    elseif (${CMAKE_BUILD_TYPE} STREQUAL "RelWithDebInfo")
+        set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -g -O3")
+    elseif (${CMAKE_BUILD_TYPE} STREQUAL "MinSizeRel")
+        set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -Os")
+    endif()
+    include(cmake/cnrt.cmake)
+    include(cmake/cndev.cmake)
+    include(cmake/cnml.cmake)
+    list(APPEND MGE_CAMBRICON_LIBS libcnrt libcndev libcnml)
+    set(MGE_CAMBRICON_LIBS "${MGE_CAMBRICON_LIBS}")
+endif()
+
+
+if(MGE_WITH_ATLAS)
+    include(cmake/aclrt.cmake)
+    list(APPEND MGE_ATLAS_LIBS libascendcl)
+    set(MGE_ATLAS_LIBS "${MGE_ATLAS_LIBS}")
+    set(MGB_ATLAS ${MGE_WITH_ATLAS})
+endif()
 
 find_program(CCACHE_BIN ccache)
 if(CCACHE_BIN)
@@ -494,6 +541,11 @@ set(MGB_CUDA ${MGE_WITH_CUDA})
 set(MEGDNN_WITH_CUDA ${MGE_WITH_CUDA})
 
 
+# CAMBRICON
+set(MGB_CAMBRICON ${MGE_WITH_CAMBRICON})
+set(MEGDNN_WITH_CAMBRICON ${MGE_WITH_CAMBRICON})
+
+
 # Debug info
 if(${CMAKE_BUILD_TYPE} STREQUAL "Debug" OR ${CMAKE_BUILD_TYPE} STREQUAL "RelWithDebInfo")
     set(MGB_ASSERT_LOC 1)

cmake/aclrt.cmake

@@ -0,0 +1,35 @@
+if($ENV{LIBRARY_PATH})
+    string(REPLACE ":" ";" SYSTEM_LIBRARY_PATHS $ENV{LIBRARY_PATH})
+endif()
+
+find_library(ACLRT_LIBRARY
+    NAMES libascendcl.so
+    PATHS $ENV{LD_LIBRARY_PATH} "$ENV{ACLRT_HOME}/lib64/stub" ${CMAKE_INSTALL_PREFIX}
+    HINTS ${SYSTEM_LIBRARY_PATHS}
+    PATH_SUFFIXES stub
+    DOC "ACL library." )
+
+if(ACLRT_LIBRARY STREQUAL "ACLRT_LIBRARY-NOTFOUND")
+    message(FATAL_ERROR "Can not find ACLRT Library")
+endif()
+
+get_filename_component(__found_aclrt_root "${ACLRT_LIBRARY}/../../../" REALPATH)
+find_path(ACLRT_INCLUDE_DIR
+    NAMES acl/acl.h
+    HINTS "$ENV{ACLRT_HOME}/include" ${__found_aclrt_root}
+    PATH_SUFFIXES include
+    DOC "Path to ACLRT include directory." )
+
+if(ACLRT_INCLUDE_DIR STREQUAL "ACLRT_INCLUDE_DIR-NOTFOUND")
+    message(FATAL_ERROR "Can not find ACLRT Library")
+endif()
+
+add_library(libascendcl SHARED IMPORTED)
+
+set_target_properties(libascendcl PROPERTIES
+    IMPORTED_LOCATION ${ACLRT_LIBRARY}
+    INTERFACE_INCLUDE_DIRECTORIES ${ACLRT_INCLUDE_DIR}
+)
+
+message("-- Found ACLRT: ${__found_aclrt_root}")
+

cmake/cndev.cmake

@@ -0,0 +1,48 @@
+if($ENV{LIBRARY_PATH})
+    string(REPLACE ":" ";" SYSTEM_LIBRARY_PATHS $ENV{LIBRARY_PATH})
+endif()
+
+find_library(CNDEV_LIBRARY 
+    NAMES libcndev.so
+    PATHS $ENV{LD_LIBRARY_PATH} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
+    HINTS ${SYSTEM_LIBRARY_PATHS}
+    PATH_SUFFIXES lib lib64
+    DOC "CNDEV library." )
+
+if(CNDEV_LIBRARY STREQUAL "CNDEV_LIBRARY-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNDEV Library")
+endif()
+
+get_filename_component(__found_cndev_root "${CNDEV_LIBRARY}/../include" REALPATH)
+find_path(CNDEV_INCLUDE_DIR 
+    NAMES cndev.h
+    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cndev_root}
+    PATH_SUFFIXES include 
+    DOC "Path to CNDEV include directory." )
+
+if(CNDEV_INCLUDE_DIR STREQUAL "CNDEV_INCLUDE_DIR-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNDEV Library")
+endif()
+
+file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_1 REGEX "^#define CNDEV_VERSION_1 [0-9]+.*$")
+file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_2 REGEX "^#define CNDEV_VERSION_2 [0-9]+.*$")
+file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_3 REGEX "^#define CNDEV_VERSION_3 [0-9]+.*$")
+file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_4 REGEX "^#define CNDEV_VERSION_4 [0-9]+.*$")
+file(STRINGS "${CNDEV_INCLUDE_DIR}/cndev.h" CNDEV_5 REGEX "^#define CNDEV_VERSION_5 [0-9]+.*$")
+
+string(REGEX REPLACE "^#define CNDEV_VERSION_1 ([0-9]+).*$" "\\1" CNDEV_VERSION_1 "${CNDEV_1}")
+string(REGEX REPLACE "^#define CNDEV_VERSION_2 ([0-9]+).*$" "\\1" CNDEV_VERSION_2 "${CNDEV_2}")
+string(REGEX REPLACE "^#define CNDEV_VERSION_3 ([0-9]+).*$" "\\1" CNDEV_VERSION_3 "${CNDEV_3}")
+string(REGEX REPLACE "^#define CNDEV_VERSION_4 ([0-9]+).*$" "\\1" CNDEV_VERSION_4 "${CNDEV_4}")
+string(REGEX REPLACE "^#define CNDEV_VERSION_5 ([0-9]+).*$" "\\1" CNDEV_VERSION_5 "${CNDEV_5}")
+set(CNDEV_VERSION_STRING "${CNDEV_VERSION_1}.${CNDEV_VERSION_2}.${CNDEV_VERSION_3}.${CNDEV_VERSION_4}.${CNDEV_VERSION_5}")
+
+add_library(libcndev SHARED IMPORTED)
+
+set_target_properties(libcndev PROPERTIES
+    IMPORTED_LOCATION ${CNDEV_LIBRARY}
+    INTERFACE_INCLUDE_DIRECTORIES ${CNDEV_INCLUDE_DIR}
+)
+
+message("-- Found CNDEV: ${__found_cndev_root} (found version: ${CNDEV_VERSION_STRING})")
+

cmake/cnml.cmake

@@ -0,0 +1,44 @@
+if($ENV{LIBRARY_PATH})
+    string(REPLACE ":" ";" SYSTEM_LIBRARY_PATHS $ENV{LIBRARY_PATH})
+endif()
+
+find_library(CNML_LIBRARY 
+    NAMES libcnml.so
+    PATHS $ENV{LD_LIBRARY_PATH} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
+    HINTS ${SYSTEM_LIBRARY_PATHS}
+    PATH_SUFFIXES lib lib64
+    DOC "CNML library." )
+
+if(CNML_LIBRARY STREQUAL "CNML_LIBRARY-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNML Library")
+endif()
+
+get_filename_component(__found_cnml_root "${CNML_LIBRARY}/../include" REALPATH)
+find_path(CNML_INCLUDE_DIR 
+    NAMES cnml.h
+    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnml_root}
+    PATH_SUFFIXES include 
+    DOC "Path to CNML include directory." )
+
+if(CNML_INCLUDE_DIR STREQUAL "CNML_INCLUDE_DIR-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNML Library")
+endif()
+
+file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MAJOR REGEX "^#define CNML_MAJOR_VERSION [0-9]+.*$")
+file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_MINOR REGEX "^#define CNML_MINOR_VERSION [0-9]+.*$")
+file(STRINGS "${CNML_INCLUDE_DIR}/cnml.h" CNML_PATCH REGEX "^#define CNML_PATCH_VERSION [0-9]+.*$")
+
+string(REGEX REPLACE "^#define CNML_MAJOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MAJOR "${CNML_MAJOR}")
+string(REGEX REPLACE "^#define CNML_MINOR_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_MINOR "${CNML_MINOR}")
+string(REGEX REPLACE "^#define CNML_PATCH_VERSION ([0-9]+).*$" "\\1" CNML_VERSION_PATCH "${CNML_PATCH}")
+set(CNML_VERSION_STRING "${CNML_VERSION_MAJOR}.${CNML_VERSION_MINOR}.${CNML_VERSION_PATCH}")
+
+add_library(libcnml SHARED IMPORTED)
+
+set_target_properties(libcnml PROPERTIES
+    IMPORTED_LOCATION ${CNML_LIBRARY}
+    INTERFACE_INCLUDE_DIRECTORIES ${CNML_INCLUDE_DIR}
+)
+
+message("-- Found CNML: ${__found_cnml_root} (found version: ${CNML_VERSION_STRING})")
+

cmake/cnrt.cmake

@@ -0,0 +1,44 @@
+if($ENV{LIBRARY_PATH})
+    string(REPLACE ":" ";" SYSTEM_LIBRARY_PATHS $ENV{LIBRARY_PATH})
+endif()
+
+find_library(CNRT_LIBRARY 
+    NAMES libcnrt.so
+    PATHS $ENV{LD_LIBRARY_PATH} "$ENV{NEUWARE_HOME}/lib64" ${CMAKE_INSTALL_PREFIX}
+    HINTS ${SYSTEM_LIBRARY_PATHS}
+    PATH_SUFFIXES lib lib64
+    DOC "CNRT library." )
+
+if(CNRT_LIBRARY STREQUAL "CNRT_LIBRARY-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNRT Library")
+endif()
+
+get_filename_component(__found_cnrt_root "${CNRT_LIBRARY}/../include" REALPATH)
+find_path(CNRT_INCLUDE_DIR 
+    NAMES cnrt.h
+    HINTS "$ENV{NEUWARE_HOME}/include" ${__found_cnrt_root}
+    PATH_SUFFIXES include 
+    DOC "Path to CNRT include directory." )
+
+if(CNRT_INCLUDE_DIR STREQUAL "CNRT_INCLUDE_DIR-NOTFOUND")
+    message(FATAL_ERROR "Can not find CNRT Library")
+endif()
+
+file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MAJOR REGEX "^#define CNRT_MAJOR_VERSION [0-9]+.*$")
+file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_MINOR REGEX "^#define CNRT_MINOR_VERSION [0-9]+.*$")
+file(STRINGS "${CNRT_INCLUDE_DIR}/cnrt.h" CNRT_PATCH REGEX "^#define CNRT_PATCH_VERSION [0-9]+.*$")
+
+string(REGEX REPLACE "^#define CNRT_MAJOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MAJOR "${CNRT_MAJOR}")
+string(REGEX REPLACE "^#define CNRT_MINOR_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_MINOR "${CNRT_MINOR}")
+string(REGEX REPLACE "^#define CNRT_PATCH_VERSION ([0-9]+).*$" "\\1" CNRT_VERSION_PATCH "${CNRT_PATCH}")
+set(CNRT_VERSION_STRING "${CNRT_VERSION_MAJOR}.${CNRT_VERSION_MINOR}.${CNRT_VERSION_PATCH}")
+
+add_library(libcnrt SHARED IMPORTED)
+
+set_target_properties(libcnrt PROPERTIES
+    IMPORTED_LOCATION ${CNRT_LIBRARY}
+    INTERFACE_INCLUDE_DIRECTORIES ${CNRT_INCLUDE_DIR}
+)
+
+message("-- Found CNRT: ${__found_cnrt_root} (found version: ${CNRT_VERSION_STRING})")
+

dnn/include/megcore_atlas.h

@@ -0,0 +1,64 @@
+/**
+ * \file dnn/include/megcore_atlas.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include "megcore.h"
+
+#include <acl/acl.h>
+
+#include "megdnn/internal/visibility_prologue.h"
+
+namespace megcore {
+megcoreStatus_t createAtlasDeviceHandleWithGlobalInitStatus(
+        megcoreDeviceHandle_t* devHandle, int deviceID, unsigned int flags,
+        bool global_initialized);
+
+struct AtlasContext {
+    aclrtStream stream = nullptr;
+    AtlasContext() = default;
+    AtlasContext(aclrtStream s) : stream{s} {}
+};
+
+megcoreStatus_t createComputingHandleWithAtlasContext(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, const AtlasContext& ctx);
+
+megcoreStatus_t getAtlasContext(megcoreComputingHandle_t handle,
+                                AtlasContext* ctx);
+
+namespace atlas {
+//! convert acl error code to error string
+const char* get_error_str(aclError error);
+}  // namespace atlas
+
+}  // namespace megcore
+
+inline megcoreStatus_t megcoreCreateComputingHandleWithACLStream(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, aclrtStream stream) {
+    megcore::AtlasContext ctx{stream};
+    return megcore::createComputingHandleWithAtlasContext(compHandle, devHandle,
+                                                          flags, ctx);
+}
+
+inline megcoreStatus_t megcoreGetACLStream(megcoreComputingHandle_t handle,
+                                           aclrtStream* stream) {
+    megcore::AtlasContext ctx;
+    auto ret = megcore::getAtlasContext(handle, &ctx);
+    *stream = ctx.stream;
+    return ret;
+}
+
+#include "megdnn/internal/visibility_epilogue.h"
+
+// vim: syntax=cpp.doxygen

dnn/include/megcore_cambricon.h

@@ -0,0 +1,61 @@
+/**
+ * \file include/megcore_cambricon.h
+ *
+ * This file is part of MegDNN, a deep neural network run-time library
+ * developed by Megvii.
+ *
+ * \copyright Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ */
+
+#pragma once
+
+#include "megcore.h"
+
+#include <cndev.h>
+#include <cnml.h>
+#include <cnrt.h>
+
+#include "megdnn/internal/visibility_prologue.h"
+
+namespace megcore {
+megcoreStatus_t createDeviceHandleWithGlobalInitStatus(
+        megcoreDeviceHandle_t* devHandle, int deviceID, unsigned int flags,
+        bool global_initialized);
+
+struct CambriconContext {
+    cnrtQueue_t queue = nullptr;
+
+    CambriconContext() = default;
+
+    CambriconContext(cnrtQueue_t q) : queue{q} {}
+};
+
+megcoreStatus_t createComputingHandleWithCambriconContext(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, const CambriconContext& ctx);
+
+megcoreStatus_t getCambriconContext(megcoreComputingHandle_t handle,
+                                    CambriconContext* ctx);
+
+}  // namespace megcore
+
+static inline megcoreStatus_t megcoreCreateComputingHandleWithCNRTQueue(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, cnrtQueue_t queue) {
+    megcore::CambriconContext ctx{queue};
+    return megcore::createComputingHandleWithCambriconContext(
+            compHandle, devHandle, flags, ctx);
+}
+
+static inline megcoreStatus_t megcoreGetCNRTQueue(
+        megcoreComputingHandle_t handle, cnrtQueue_t* queue) {
+    megcore::CambriconContext ctx;
+    auto ret = megcore::getCambriconContext(handle, &ctx);
+    *queue = ctx.queue;
+    return ret;
+}
+
+#include "megdnn/internal/visibility_epilogue.h"
+
+// vim: syntax=cpp.doxygen
+

dnn/include/megcore_cdefs.h

@@ -19,6 +19,8 @@
 typedef enum {
     megcorePlatformCPU = 1,
     megcorePlatformCUDA = 4,
+    megcorePlatformCambricon = 7,
+    megcorePlatformAtlas = 8,
 } megcorePlatform_t;
 
 /**

dnn/include/megdnn/handle.h

@@ -33,6 +33,8 @@ class Handle {
             ARMV7 = 4,
             AARCH64 = 5,
             CUDA = 6,
+            ATLAS = 13,
+            CAMBRICON = 12,
         };
 
     protected:

dnn/src/CMakeLists.txt

@@ -45,6 +45,24 @@ if(MGE_WITH_CUDA)
     list(APPEND SOURCES ${CUSOURCES})
 endif()
 
+if(MGE_WITH_CAMBRICON)
+    file(GLOB_RECURSE SOURCES_ cambricon/*.cpp)
+    list(APPEND SOURCES ${SOURCES_})
+
+    file(GLOB_RECURSE BANG_SOURCES cambricon/*.mlu)
+    list(APPEND MEGDNN_INCLUDES "${PROJECT_SOURCE_DIR}/dnn/include")
+    list(APPEND MEGDNN_INCLUDES "${PROJECT_SOURCE_DIR}/dnn")
+    list(APPEND MEGDNN_INCLUDES "${PROJECT_BINARY_DIR}/genfiles")
+    bang_compile(BANG_OBJS "${BANG_SOURCES}" "${MEGDNN_INCLUDES}")
+    list(APPEND SOURCES ${BANG_OBJS})
+
+endif()
+
+if(MGE_WITH_ATLAS)
+    file(GLOB_RECURSE SOURCES_ atlas/*.cpp)
+    list(APPEND SOURCES ${SOURCES_})
+    list(APPEND LIBMEGDNN_DEF -DMEGDNN_WITH_ATLAS=1)
+endif()
 
 
 add_definitions(${LIBMEGDNN_DEF})
@@ -97,8 +115,21 @@ else()
     target_link_libraries(megdnn PRIVATE ${MGE_BLAS_LIBS})
 endif()
 
+if(MGE_WITH_ATLAS)
+    if (BUILD_SHARED_LIBS)
+        target_link_libraries(megdnn PRIVATE $<BUILD_INTERFACE:${MGE_ATLAS_LIBS}>)
+    else()
+        target_link_libraries(megdnn PRIVATE ${MGE_ATLAS_LIBS})
+    endif()
+
+endif()
+
 if(CMAKE_THREAD_LIBS_INIT)
     target_link_libraries(megdnn PRIVATE Threads::Threads)
 endif()
 
+if(MGE_WITH_CAMBRICON)
+    target_link_libraries(megdnn PRIVATE ${BANG_OBJS} ${MGE_CAMBRICON_LIBS})
+endif()
+
 install(TARGETS megdnn EXPORT ${MGE_EXPORT_TARGETS})

dnn/src/atlas/checksum/opr_impl.cpp

@@ -0,0 +1,53 @@
+/**
+ * \file dnn/src/atlas/checksum/opr_impl.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "src/atlas/checksum/opr_impl.h"
+#include "src/atlas/utils.h"
+#include "src/naive/handle.h"
+
+#include "src/common/utils.h"
+#include "src/common/opr_delegate.h"
+
+#include <cstring>
+
+using namespace megdnn;
+using namespace atlas;
+
+size_t ChecksumForwardImpl::get_workspace_in_bytes(const TensorLayout&) {
+    return 0;
+}
+
+ChecksumForward::Result ChecksumForwardImpl::exec(_megdnn_tensor_in data,
+                                                  _megdnn_workspace workspace) {
+    check_exec(data.layout, workspace.size);
+    //! FIXME currently the cce programming interface is not so stable, here i
+    //! just allocate some memory of cpu here and compute the result in cpu
+    std::vector<uint8_t> cpu_data(data.layout.span().dist_byte(), 0);
+
+    megcoreDeviceHandle_t dev_handle;
+    megcoreComputingHandle_t comp_handle = handle()->megcore_computing_handle();
+    megcoreGetDeviceHandle(comp_handle, &dev_handle);
+    megcoreMemcpy(comp_handle, cpu_data.data(), data.raw_ptr, cpu_data.size(),
+                  megcoreMemcpyDeviceToHost);
+    megcoreSynchronize(comp_handle);
+
+    auto opr = inplace_cpu_handle()->create_operator<ChecksumForward>();
+    size_t workspace_size = opr->get_workspace_in_bytes(data.layout);
+    std::vector<uint8_t> cpu_workspace_data(workspace_size, 0);
+
+    Workspace cpu_workspace(
+            reinterpret_cast<dt_byte*>(cpu_workspace_data.data()),
+            cpu_workspace_data.size());
+
+    return opr->exec(TensorND{cpu_data.data(), data.layout}, cpu_workspace);
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/checksum/opr_impl.h

@@ -0,0 +1,34 @@
+/**
+ * \file dnn/src/atlas/checksum/opr_impl.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include "megdnn/oprs.h"
+
+namespace megdnn {
+namespace atlas {
+
+class ChecksumForwardImpl final : public ChecksumForward {
+public:
+    using ChecksumForward::ChecksumForward;
+
+    bool is_thread_safe() const override { return true; }
+
+    size_t get_workspace_in_bytes(const TensorLayout& data) override;
+
+    Result exec(_megdnn_tensor_in data, _megdnn_workspace workspace) override;
+};
+
+}  // namespace naive
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/handle.cpp

@@ -0,0 +1,59 @@
+/**
+ * \file dnn/src/atlas/handle.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megcore_atlas.h"
+#include "src/common/handle_impl.h"
+#include "src/atlas/handle.h"
+#include "src/atlas/checksum/opr_impl.h"
+
+#include <acl/acl.h>
+
+namespace megdnn {
+namespace atlas {
+
+HandleImpl::HandleImpl(megcoreComputingHandle_t comp_handle)
+        : HandleImplHelper(comp_handle, HandleType::ATLAS) {
+    // Get megcore device handle
+    megcoreDeviceHandle_t dev_handle;
+    megcoreGetDeviceHandle(comp_handle, &dev_handle);
+
+    int dev_id;
+    megcoreGetDeviceID(dev_handle, &dev_id);
+    m_device_id = dev_id;
+    megcore::getAtlasContext(comp_handle, &m_megcore_context);
+}
+
+HandleImpl::~HandleImpl() noexcept = default;
+
+template <typename Opr>
+std::unique_ptr<Opr> HandleImpl::create_operator() {
+    megdnn_throw("unsupported atlas opr");
+    return nullptr;
+}
+
+size_t HandleImpl::alignment_requirement() const {
+    //! because memcpyasync api requires that the memory is 128bytes alignment
+    return 64;
+}
+
+MEGDNN_SPECIALIZE_CREATE_OPERATOR(ChecksumForward);
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wpragmas"
+#pragma GCC diagnostic ignored "-Winstantiation-after-specialization"
+MEGDNN_FOREACH_OPR_CLASS(MEGDNN_INST_CREATE_OPERATOR)
+#pragma GCC diagnostic pop
+
+}  // namespace atlas
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/handle.h

@@ -0,0 +1,63 @@
+/**
+ * \file dnn/src/atlas/handle.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+#include "megcore_atlas.h"
+#include "megdnn/basic_types.h"
+#include "megdnn/handle.h"
+#include "megdnn/oprs/general.h"
+
+#include "src/common/handle_impl.h"
+#include "src/common/megcore/common/device_context.hpp"
+#include "src/common/utils.h"
+#include "src/atlas/megcore/device_context.hpp"
+
+#include <atomic>
+#include <mutex>
+
+#include "acl/acl_rt.h"
+
+namespace megdnn {
+namespace atlas {
+
+class HandleImpl : public HandleImplHelper {
+public:
+    HandleImpl(megcoreComputingHandle_t computing_handle);
+    ~HandleImpl() noexcept;
+
+    size_t alignment_requirement() const override;
+
+    template <typename Opr>
+    std::unique_ptr<Opr> create_operator();
+
+    const megcore::AtlasContext& megcore_context() const {
+        return m_megcore_context;
+    }
+
+    int device_id() const { return m_device_id; }
+
+    aclrtStream stream() const { return megcore_context().stream; }
+
+    //! global matmul opr
+    Checksum* checksum_opr() override final {
+        return get_helper_opr<Checksum, 0>(this);
+    }
+
+private:
+    int m_device_id;
+    //! MegDNN handle does not manage the lifetime of cnrt queue.
+    megcore::AtlasContext m_megcore_context;
+};
+
+}  // namespace atlas
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/megcore/computing_context.cpp

@@ -0,0 +1,71 @@
+/**
+ * \file dnn/src/atlas/megcore/atlas_computing_context.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#include "megcore.h"
+
+#include "src/atlas//megcore/computing_context.hpp"
+#include "src/atlas/utils.h"
+#include "src/common/utils.h"
+
+using namespace megcore;
+using namespace megcore::atlas;
+
+AtlasComputingContext::AtlasComputingContext(megcoreDeviceHandle_t dev_handle,
+                                             unsigned int flags,
+                                             const AtlasContext& ctx)
+        : ComputingContext(dev_handle, flags),
+          m_own_stream{ctx.stream == nullptr},
+          m_ctx{ctx} {
+    megcorePlatform_t platform;
+    megcoreGetPlatform(dev_handle, &platform);
+    megdnn_assert(platform == megcorePlatformAtlas);
+    if (m_own_stream) {
+        acl_check(aclrtCreateStream(&m_ctx.stream));
+    }
+}
+
+AtlasComputingContext::~AtlasComputingContext() {
+    if (m_own_stream) {
+        acl_check(aclrtDestroyStream(m_ctx.stream));
+    }
+}
+
+void AtlasComputingContext::memcpy(void* dst, const void* src,
+                                   size_t size_in_bytes,
+                                   megcoreMemcpyKind_t kind) {
+    aclrtMemcpyKind atlas_kind;
+    switch (kind) {
+        case megcoreMemcpyDeviceToHost:
+            atlas_kind = ACL_MEMCPY_DEVICE_TO_HOST;
+            break;
+        case megcoreMemcpyHostToDevice:
+            atlas_kind = ACL_MEMCPY_HOST_TO_DEVICE;
+            break;
+        case megcoreMemcpyDeviceToDevice:
+            atlas_kind = ACL_MEMCPY_DEVICE_TO_DEVICE;
+            break;
+        default:
+            megdnn_throw("bad atlas memcpy kind");
+    }
+    acl_check(aclrtMemcpyAsync(dst, size_in_bytes, src, size_in_bytes,
+                               atlas_kind, m_ctx.stream));
+}
+
+void AtlasComputingContext::memset(void* dst, int value, size_t size_in_bytes) {
+    acl_check(aclrtSynchronizeStream(m_ctx.stream));
+    acl_check(aclrtMemset(dst, size_in_bytes, value, size_in_bytes));
+}
+
+void AtlasComputingContext::synchronize() {
+    acl_check(aclrtSynchronizeStream(m_ctx.stream));
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/megcore/computing_context.hpp

@@ -0,0 +1,45 @@
+/**
+ * \file dnn/src/atlas/megcore/computing_context.hpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+
+#include "megcore_atlas.h"
+#include "src/common/megcore/common/computing_context.hpp"
+
+#include <acl/acl_rt.h>
+
+namespace megcore {
+namespace atlas {
+
+class AtlasComputingContext final : public ComputingContext {
+public:
+    AtlasComputingContext(megcoreDeviceHandle_t dev_handle, unsigned int flags,
+                          const AtlasContext& ctx = {});
+    ~AtlasComputingContext();
+
+    void memcpy(void* dst, const void* src, size_t size_in_bytes,
+                megcoreMemcpyKind_t kind) override;
+    void memset(void* dst, int value, size_t size_in_bytes) override;
+    void synchronize() override;
+
+    const AtlasContext& context() const { return m_ctx; }
+
+    aclrtStream stream() const { return context().stream; }
+
+private:
+    bool m_own_stream;
+    AtlasContext m_ctx;
+};
+
+}  // namespace atlas
+}  // namespace megcore
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/megcore/device_context.cpp

@@ -0,0 +1,67 @@
+/**
+ * \file dnn/src/atlas/megcore/device_context.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "src/atlas/megcore/device_context.hpp"
+
+#include "megcore.h"
+#include "src/atlas/utils.h"
+#include "src/common/utils.h"
+
+#include "acl/acl.h"
+
+using namespace megcore;
+using namespace atlas;
+
+AtlasDeviceContext::AtlasDeviceContext(int device_id, unsigned int flags,
+                                       bool global_initialized)
+        : DeviceContext(megcorePlatformAtlas, device_id, flags) {
+    if (!global_initialized)
+        init_status.init();
+
+    int id = device_id;
+    if (id < 0) {
+        acl_check(aclrtGetDevice(&id));
+    }
+}
+
+AtlasDeviceContext::~AtlasDeviceContext() noexcept = default;
+
+size_t AtlasDeviceContext::mem_alignment_in_bytes() const noexcept {
+    return 64;
+}
+
+void AtlasDeviceContext::activate() {
+    int id = device_id();
+    if (id >= 0) {
+        acl_check(aclrtSetDevice(id));
+    }
+}
+
+void AtlasDeviceContext::deactivate() {
+    int id = device_id();
+    megdnn_assert(id >= 0);
+    acl_check(aclrtResetDevice(id));
+}
+
+void* AtlasDeviceContext::malloc(size_t size_in_bytes) {
+    void* ptr;
+    acl_check(aclrtMalloc(&ptr, size_in_bytes, ACL_MEM_MALLOC_HUGE_FIRST));
+    return ptr;
+}
+
+void AtlasDeviceContext::free(void* ptr) {
+    acl_check(aclrtFree(ptr));
+}
+
+AtlasDeviceContext::InitStatus AtlasDeviceContext::init_status;
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/megcore/device_context.hpp

@@ -0,0 +1,64 @@
+/**
+ * \file dnn/src/atlas/megcore/device_context.hpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+
+#include "src/common/megcore/common/device_context.hpp"
+#include "src/common/utils.h"
+#include "megcore_atlas.h"
+
+#include <mutex>
+#include "acl/acl.h"
+
+namespace megcore {
+namespace atlas {
+
+class AtlasDeviceContext : public DeviceContext {
+public:
+    AtlasDeviceContext(int device_id, unsigned int flags,
+                       bool global_initialized = false);
+    ~AtlasDeviceContext() noexcept;
+
+    size_t mem_alignment_in_bytes() const noexcept override;
+
+    void activate() override;
+    void deactivate() override;
+    void* malloc(size_t size_in_bytes) override;
+    void free(void* ptr) override;
+
+    struct InitStatus {
+        bool initialized;
+        std::mutex mtx;
+        InitStatus() : initialized{false} {}
+        void init() {
+            std::lock_guard<std::mutex> guard{mtx};
+            if (!initialized) {
+                auto err = aclInit(nullptr);
+                initialized = err == ACL_ERROR_NONE;
+                megdnn_assert(initialized,
+                              "aclrt initialize failed: (acl:%d): %s",
+                              static_cast<int>(err),
+                              megcore::atlas::get_error_str(err));
+            }
+        }
+        ~InitStatus() {
+            if (initialized) {
+                initialized = false;
+            }
+        }
+    };
+    static InitStatus init_status;
+};
+
+}  // namespace atlas
+}  // namespace megcore
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/megcore/public_api/computing.cpp

@@ -0,0 +1,131 @@
+/**
+ * \file dnn/src/atlas/megcore/public_api/computing.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#include "megcore_atlas.h"
+
+#include "src/atlas/megcore/computing_context.hpp"
+#include "src/atlas/megcore/device_context.hpp"
+#include "src/common/megcore/public_api/computing.hpp"
+#include "src/common/megcore/public_api/device.hpp"
+#include "src/common/utils.h"
+
+using namespace megcore;
+
+megcoreStatus_t megcore::createAtlasDeviceHandleWithGlobalInitStatus(
+        megcoreDeviceHandle_t* devHandle, int deviceID, unsigned int flags,
+        bool global_initialized) {
+    auto content = megdnn::make_unique<atlas::AtlasDeviceContext>(
+            deviceID, flags, global_initialized);
+    auto& ctx = *devHandle;
+    ctx = new megcoreDeviceContext;
+    ctx->content = std::move(content);
+    return megcoreSuccess;
+}
+
+megcoreStatus_t megcore::createComputingHandleWithAtlasContext(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, const AtlasContext& ctx) {
+    MEGDNN_MARK_USED_VAR(flags);
+    megdnn_assert(flags == 0);
+    auto content = megdnn::make_unique<atlas::AtlasComputingContext>(
+            devHandle, flags, ctx);
+    auto& H = *compHandle;
+    H = new megcoreComputingContext;
+    H->content = std::move(content);
+    return megcoreSuccess;
+}
+
+megcoreStatus_t megcore::getAtlasContext(megcoreComputingHandle_t handle,
+                                         AtlasContext* ctx) {
+    auto&& H = handle;
+    megdnn_assert(H);
+    megcoreDeviceHandle_t dev_handle = H->content->dev_handle();
+    megcorePlatform_t platform;
+    megcoreGetPlatform(dev_handle, &platform);
+    megdnn_assert(platform == megcorePlatformAtlas);
+    auto context = static_cast<megcore::atlas::AtlasComputingContext*>(
+            H->content.get());
+    *ctx = context->context();
+    return megcoreSuccess;
+}
+
+const char* megcore::atlas::get_error_str(aclError error) {
+#define ERROR(_err) \
+    case _err:      \
+        return #_err;
+
+    switch (error) {
+        ERROR(ACL_ERROR_NONE);
+
+        ERROR(ACL_ERROR_INVALID_PARAM);
+        ERROR(ACL_ERROR_UNINITIALIZE);
+        ERROR(ACL_ERROR_REPEAT_INITIALIZE);
+        ERROR(ACL_ERROR_INVALID_FILE);
+        ERROR(ACL_ERROR_WRITE_FILE);
+        ERROR(ACL_ERROR_INVALID_FILE_SIZE);
+        ERROR(ACL_ERROR_PARSE_FILE);
+        ERROR(ACL_ERROR_FILE_MISSING_ATTR);
+        ERROR(ACL_ERROR_FILE_ATTR_INVALID);
+        ERROR(ACL_ERROR_INVALID_DUMP_CONFIG);
+        ERROR(ACL_ERROR_INVALID_PROFILING_CONFIG);
+        ERROR(ACL_ERROR_INVALID_MODEL_ID);
+        ERROR(ACL_ERROR_DESERIALIZE_MODEL);
+        ERROR(ACL_ERROR_PARSE_MODEL);
+        ERROR(ACL_ERROR_READ_MODEL_FAILURE);
+        ERROR(ACL_ERROR_MODEL_SIZE_INVALID);
+        ERROR(ACL_ERROR_MODEL_MISSING_ATTR);
+        ERROR(ACL_ERROR_MODEL_INPUT_NOT_MATCH);
+        ERROR(ACL_ERROR_MODEL_OUTPUT_NOT_MATCH);
+        ERROR(ACL_ERROR_MODEL_NOT_DYNAMIC);
+        ERROR(ACL_ERROR_OP_TYPE_NOT_MATCH);
+        ERROR(ACL_ERROR_OP_INPUT_NOT_MATCH);
+        ERROR(ACL_ERROR_OP_OUTPUT_NOT_MATCH);
+        ERROR(ACL_ERROR_OP_ATTR_NOT_MATCH);
+        ERROR(ACL_ERROR_OP_NOT_FOUND);
+        ERROR(ACL_ERROR_OP_LOAD_FAILED);
+        ERROR(ACL_ERROR_UNSUPPORTED_DATA_TYPE);
+        ERROR(ACL_ERROR_FORMAT_NOT_MATCH);
+        ERROR(ACL_ERROR_BIN_SELECTOR_NOT_REGISTERED);
+        ERROR(ACL_ERROR_KERNEL_NOT_FOUND);
+        ERROR(ACL_ERROR_BIN_SELECTOR_ALREADY_REGISTERED);
+        ERROR(ACL_ERROR_KERNEL_ALREADY_REGISTERED);
+        ERROR(ACL_ERROR_INVALID_QUEUE_ID);
+        ERROR(ACL_ERROR_REPEAT_SUBSCRIBE);
+        ERROR(ACL_ERROR_STREAM_NOT_SUBSCRIBE);
+        ERROR(ACL_ERROR_THREAD_NOT_SUBSCRIBE);
+        ERROR(ACL_ERROR_WAIT_CALLBACK_TIMEOUT);
+        ERROR(ACL_ERROR_REPEAT_FINALIZE);
+        ERROR(ACL_ERROR_NOT_STATIC_AIPP);
+
+        ERROR(ACL_ERROR_BAD_ALLOC);
+        ERROR(ACL_ERROR_API_NOT_SUPPORT);
+        ERROR(ACL_ERROR_INVALID_DEVICE);
+        ERROR(ACL_ERROR_MEMORY_ADDRESS_UNALIGNED);
+        ERROR(ACL_ERROR_RESOURCE_NOT_MATCH);
+        ERROR(ACL_ERROR_INVALID_RESOURCE_HANDLE);
+        ERROR(ACL_ERROR_FEATURE_UNSUPPORTED);
+
+        ERROR(ACL_ERROR_STORAGE_OVER_LIMIT);
+
+        ERROR(ACL_ERROR_INTERNAL_ERROR);
+        ERROR(ACL_ERROR_FAILURE);
+        ERROR(ACL_ERROR_GE_FAILURE);
+        ERROR(ACL_ERROR_RT_FAILURE);
+        ERROR(ACL_ERROR_DRV_FAILURE);
+        ERROR(ACL_ERROR_PROFILING_FAILURE);
+
+        default:
+            return "unknown error";
+    }
+#undef ERROR
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/utils.cpp

@@ -0,0 +1,26 @@
+/**
+ * \file dnn/src/atlas/utils.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "src/atlas/utils.h"
+#include "megcore_atlas.h"
+#include "src/common/utils.h"
+
+using namespace megdnn;
+using namespace atlas;
+
+void atlas::__throw_acl_error__(aclError err, const char* msg) {
+    auto s = ssprintf("acl return %s(%d) occurred; expr: %s",
+                      megcore::atlas::get_error_str(err), int(err), msg);
+    megdnn_throw(s.c_str());
+}
+
+// vim: syntax=cpp.doxygen

dnn/src/atlas/utils.h

@@ -0,0 +1,40 @@
+/**
+ * \file dnn/src/atlas/utils.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+
+#include "megdnn/handle.h"
+#include "src/atlas/handle.h"
+
+#include <acl/acl_base.h>
+
+#define acl_check(_x)                                        \
+    do {                                                     \
+        aclError _ret = (_x);                                \
+        if (_ret != ACL_ERROR_NONE) {                        \
+            ::megdnn::atlas::__throw_acl_error__(_ret, #_x); \
+        }                                                    \
+    } while (0)
+
+namespace megdnn {
+namespace atlas {
+
+inline HandleImpl* concrete_handle(Handle* handle) {
+    return static_cast<atlas::HandleImpl*>(handle);
+}
+
+//! Error handling funcions
+MEGDNN_NORETURN void __throw_acl_error__(aclError err, const char* msg);
+
+}  // namespace atlas
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/src/cambricon/BUILD

@@ -0,0 +1,31 @@
+load("//brain/megbrain/dnn:flags.bzl", "megdnn_opts")
+load("@megvii3//tools/build_rules:bangc.bzl", "bangc_library")
+
+package(default_visibility = ["//brain/megbrain/dnn:__subpackages__"])
+
+bangc_library(
+    name = "bangc_kernels",
+    srcs = glob([
+        "**/*.mlu",
+    ]) + [
+        "//brain/megbrain/dnn:src/common/utils.cuh",
+    ],
+    hdrs = glob([
+        "**/*.mlu.h",
+    ]),
+    deps = [
+        "//brain/megbrain/dnn:public_headers",
+    ],
+    copts = megdnn_opts + [
+        "-Ibrain/megbrain/dnn",
+    ],
+)
+
+filegroup(
+    name = "cambricon_backend_files",
+    srcs = glob([
+        "**/*.cpp",
+        "**/*.h",
+        "**/*.hpp",
+    ]),
+)

dnn/src/cambricon/checksum/checksum.mlu.h

@@ -0,0 +1,27 @@
+/**
+ * \file dnn/src/cambricon/checksum/checksum.mlu.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#pragma once
+
+#include "src/cambricon/utils.mlu.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+void checksum_kernel_union1(uint32_t* dst, const uint32_t* src, int num_elems);
+void checksum_kernel_union4(uint32_t* dst, const uint32_t* src, int num_elems);
+#ifdef __cplusplus
+}
+#endif
+
+// vim: ft=cpp syntax=cpp.doxygen
+
+

dnn/src/cambricon/checksum/checksum_kernel_union1.mlu

@@ -0,0 +1,61 @@
+/**
+ * \file dnn/src/cambricon/checksum/checksum_kernel_union1.mlu
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "checksum.mlu.h"
+#include "cnsccl.h"
+#include "mlu.h"
+
+#define CLUSTER_DIM 1
+#define CORE_DIM 4
+#define STRIDE 1024
+
+__mlu_entry__ void checksum_kernel_union1(uint32_t* dst, uint32_t* src,
+                                          int nr_elems) {
+    __nram__ uint32_t sum = 0;
+    __nram__ uint32_t val[STRIDE];
+    const uint32_t TASK_DIM = CLUSTER_DIM * CORE_DIM;
+    __mlu_shared__ uint32_t partial_sum[TASK_DIM];
+
+    int task_stride = STRIDE;
+    int start_offset = taskId * task_stride;
+    int global_stride = taskDim * task_stride;
+
+    for (int task_offset = start_offset; task_offset < nr_elems;
+         task_offset += global_stride) {
+        int end_offset = task_offset + task_stride;
+        end_offset = end_offset > nr_elems ? nr_elems : end_offset;
+        int copy_elems = end_offset - task_offset;
+        __memcpy(val, src + task_offset, copy_elems * sizeof(uint32_t),
+                 GDRAM2NRAM);
+        for (int i = 0; i < copy_elems; i++) {
+            sum = sum + val[i] * (task_offset + i + 1);
+        }
+    }
+
+    partial_sum[taskId] = sum;
+
+    __sync_cluster();
+
+    if (taskId == 0) {
+        uint32_t res = 0;
+        for (int i = 0; i < taskDim; i++) {
+            res += partial_sum[i];
+        }
+        dst[0] = res;
+    }
+}
+
+#undef CLUSTER_DIM
+#undef CORE_DIM
+#undef STRIDE
+
+// vim: ft=cpp syntax=cpp.doxygen
+

dnn/src/cambricon/checksum/checksum_kernel_union4.mlu

@@ -0,0 +1,71 @@
+/**
+ * \file dnn/src/cambricon/checksum/checksum_kernel_union4.mlu
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "checksum.mlu.h"
+#include "cnsccl.h"
+#include "mlu.h"
+
+#define CLUSTER_DIM 4
+#define CORE_DIM 4
+#define STRIDE 1024
+__mlu_entry__ void checksum_kernel_union4(uint32_t* dst, uint32_t* src,
+                                          int nr_elems) {
+    __nram__ uint32_t sum = 0;
+    __nram__ uint32_t val[STRIDE];
+    __mlu_shared__ uint32_t partial_sum_send[CORE_DIM];
+    __mlu_shared__ uint32_t partial_sum_recv[CLUSTER_DIM];
+
+    int task_stride = STRIDE;
+    int start_offset = taskId * task_stride;
+    int global_stride = taskDim * task_stride;
+
+    for (int task_offset = start_offset; task_offset < nr_elems;
+         task_offset += global_stride) {
+        int end_offset = task_offset + task_stride;
+        end_offset = end_offset > nr_elems ? nr_elems : end_offset;
+        int copy_elems = end_offset - task_offset;
+        __memcpy(val, src + task_offset, copy_elems * sizeof(uint32_t),
+                 GDRAM2NRAM);
+        for (int i = 0; i < copy_elems; i++) {
+            sum = sum + val[i] * (task_offset + i + 1);
+        }
+    }
+
+    partial_sum_send[coreId] = sum;
+
+    __sync_cluster();
+
+    if (coreId == 0) {
+        for (int i = 1; i < CORE_DIM; ++i) {
+            partial_sum_send[0] += partial_sum_send[i];
+        }
+    }
+
+    __sync_all();
+    cnscclGather((void*)&partial_sum_send, (void*)&partial_sum_recv, 1,
+                 cnscclInt, 0);
+
+    __sync_all();
+
+    if (clusterId == 0 && coreId == 0) {
+        uint32_t res = 0;
+        for (int i = 0; i < CLUSTER_DIM; ++i) {
+            res += partial_sum_recv[i];
+        }
+        dst[0] = res;
+    }
+}
+#undef CLUSTER_DIM
+#undef CORE_DIM
+#undef STRIDE
+
+// vim: ft=cpp syntax=cpp.doxygen
+

dnn/src/cambricon/checksum/opr_impl.cpp

@@ -0,0 +1,85 @@
+/**
+ * \file dnn/src/cambricon/checksum/opr_impl.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "src/cambricon/checksum/checksum.mlu.h"
+#include "src/cambricon/checksum/opr_impl.h"
+
+#include "src/cambricon/utils.h"
+
+#include <algorithm>
+
+using namespace megdnn;
+using namespace cambricon;
+
+namespace {
+void bang_c_wrapper(uint32_t* dst, const uint32_t* src, int nr_elems,
+                    cnrtQueue_t queue, cnrtCoreVersion_t core_version) {
+    cnrtKernelParamsBuffer_t params;
+    cnrt_check(cnrtGetKernelParamsBuffer(&params));
+    cnrt_check(cnrtKernelParamsBufferAddParam(params, &dst, sizeof(uint32_t*)));
+    cnrt_check(cnrtKernelParamsBufferAddParam(params, &src, sizeof(uint32_t*)));
+    cnrt_check(cnrtKernelParamsBufferAddParam(params, &nr_elems, sizeof(int)));
+    if (core_version == CNRT_MLU270) {
+        cnrtDim3_t dim;
+        dim.x = 16;
+        dim.y = 1;
+        dim.z = 1;
+        cnrtFunctionType_t c = CNRT_FUNC_TYPE_UNION4;
+        cnrt_check(cnrtInvokeKernel_V2((void*)&checksum_kernel_union4, dim,
+                                       params, c, queue));
+    } else if (core_version == CNRT_MLU220) {
+        cnrtDim3_t dim;
+        dim.x = 4;
+        dim.y = 1;
+        dim.z = 1;
+        cnrtFunctionType_t c = CNRT_FUNC_TYPE_UNION1;
+        cnrt_check(cnrtInvokeKernel_V2((void*)&checksum_kernel_union1, dim,
+                                       params, c, queue));
+    }
+    after_kernel_launch();
+    cnrt_check(cnrtDestroyKernelParamsBuffer(params));
+}
+}  // namespace
+
+size_t ChecksumForwardImpl::get_workspace_in_bytes(const TensorLayout& /* data */) {
+    size_t ws_size = sizeof(ChecksumForward::Result::checksum);
+    return ws_size;
+}
+
+ChecksumForward::Result ChecksumForwardImpl::exec(_megdnn_tensor_in data,
+                                                  _megdnn_workspace workspace) {
+    Result result;
+    memset(&result, 0, sizeof(result));
+    check_exec(data.layout, workspace.size);
+    auto queue = cnrt_queue(handle());
+
+    auto ptr = static_cast<uint8_t*>(data.raw_ptr);
+    size_t size_all = data.layout.shape[0],
+           size_ints = size_all / sizeof(uint32_t);
+    auto last_val_size = std::min<size_t>(size_all, 4);
+    cnrt_check(cnrtMemcpyAsync(&result.last_val, ptr + size_all - last_val_size,
+                               last_val_size, queue,
+                               CNRT_MEM_TRANS_DIR_DEV2HOST));
+    if (size_ints) {
+        auto&& device_info = current_device_info();
+        bang_c_wrapper(reinterpret_cast<uint32_t*>(workspace.raw_ptr),
+                       static_cast<uint32_t*>(data.raw_ptr), size_ints, queue,
+                       device_info.core_version);
+        cnrt_check(cnrtMemcpyAsync(&result.checksum, workspace.raw_ptr,
+                                   sizeof(result.checksum), queue,
+                                   CNRT_MEM_TRANS_DIR_DEV2HOST));
+    }
+    cnrt_check(cnrtSyncQueue(queue));
+    return result;
+}
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/checksum/opr_impl.h

@@ -0,0 +1,36 @@
+/**
+ * \file dnn/src/cambricon/checksum/opr_impl.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#pragma once
+
+#include "megdnn/oprs.h"
+#include "src/cambricon/utils.h"
+
+namespace megdnn {
+namespace cambricon {
+
+class ChecksumForwardImpl final : public ChecksumForward {
+public:
+    using ChecksumForward::ChecksumForward;
+
+    size_t get_workspace_in_bytes(const TensorLayout&) override;
+
+    bool is_thread_safe() const override { return true; }
+
+    Result exec(_megdnn_tensor_in data, _megdnn_workspace workspace) override;
+};
+
+}  // namespace cambricon
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen
+
+

dnn/src/cambricon/handle.cpp

@@ -0,0 +1,68 @@
+/**
+ * \file dnn/src/cambricon/handle.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "src/common/handle_impl.h"
+#include "src/common/version_symbol.h"
+
+#include "src/cambricon/handle.h"
+#include "src/cambricon/utils.h"
+
+#include "src/cambricon/checksum/opr_impl.h"
+#include <cnrt.h>
+
+namespace megdnn {
+namespace cambricon {
+
+HandleImpl::HandleImpl(megcoreComputingHandle_t comp_handle)
+        : HandleImplHelper(comp_handle, HandleType::CAMBRICON) {
+    // Get megcore device handle
+    megcoreDeviceHandle_t dev_handle;
+    megcoreGetDeviceHandle(comp_handle, &dev_handle);
+    int dev_id;
+    megcoreGetDeviceID(dev_handle, &dev_id);
+    unsigned int dev_num;
+    cnrt_check(cnrtGetDeviceCount(&dev_num));
+    MEGDNN_MARK_USED_VAR(dev_num);
+    // check validity of device_id
+    megdnn_assert(dev_id >= 0 && static_cast<unsigned int>(dev_id) < dev_num);
+    m_device_id = dev_id;
+    cnrt_check(cnrtGetDeviceInfo(&m_device_info, dev_id));
+    megcore::getCambriconContext(comp_handle, &m_megcore_context);
+}
+
+HandleImpl::~HandleImpl() noexcept = default;
+
+template <typename Opr>
+std::unique_ptr<Opr> HandleImpl::create_operator() {
+    megdnn_throw("unsupported cambricon opr");
+    return nullptr;
+}
+
+size_t HandleImpl::alignment_requirement() const {
+    return 1;
+}
+
+MEGDNN_SPECIALIZE_CREATE_OPERATOR(ChecksumForward);
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wpragmas"
+#pragma GCC diagnostic ignored "-Winstantiation-after-specialization"
+MEGDNN_FOREACH_OPR_CLASS(MEGDNN_INST_CREATE_OPERATOR)
+#pragma GCC diagnostic pop
+
+}  // namespace cambricon
+}  // namespace megdnn
+
+MEGDNN_VERSION_SYMBOL3(CNRT, CNRT_MAJOR_VERSION, CNRT_MINOR_VERSION,
+                       CNRT_PATCH_VERSION);
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/handle.h

@@ -0,0 +1,65 @@
+/**
+ * \file dnn/src/cambricon/handle.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+#include "megcore_cambricon.h"
+#include "megdnn/basic_types.h"
+#include "megdnn/handle.h"
+#include "megdnn/oprs/general.h"
+
+#include "src/common/handle_impl.h"
+#include "src/common/utils.h"
+
+#include <atomic>
+#include <mutex>
+
+#include <cnrt.h>
+
+namespace megdnn {
+namespace cambricon {
+
+class HandleImpl : public HandleImplHelper {
+public:
+    HandleImpl(megcoreComputingHandle_t computing_handle);
+    ~HandleImpl() noexcept;
+
+    size_t alignment_requirement() const override;
+
+    const cnrtDeviceInfo_t& device_info() const { return m_device_info; }
+
+    template <typename Opr>
+    std::unique_ptr<Opr> create_operator();
+
+    const megcore::CambriconContext& megcore_context() const {
+        return m_megcore_context;
+    }
+
+    int device_id() const { return m_device_id; }
+
+    cnrtQueue_t queue() const { return megcore_context().queue; }
+
+    //! global matmul opr
+    Checksum* checksum_opr() override final {
+        return get_helper_opr<Checksum, 0>(this);
+    }
+
+private:
+    int m_device_id;
+    //! MegDNN handle does not manage the lifetime of cnrt queue.
+    megcore::CambriconContext m_megcore_context;
+
+    cnrtDeviceInfo_t m_device_info;
+};
+
+}  // namespace cambricon
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/megcore/cambricon_computing_context.cpp

@@ -0,0 +1,78 @@
+/**
+ * \file dnn/src/cambricon/megcore/cambricon_computing_context.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#include "megcore.h"
+
+#include "src/cambricon/utils.h"
+#include "src/common/utils.h"
+
+#include "src/cambricon/megcore/cambricon_computing_context.hpp"
+
+using namespace megcore;
+using namespace megcore::cambricon;
+
+CambriconComputingContext::CambriconComputingContext(
+        megcoreDeviceHandle_t dev_handle, unsigned int flags,
+        const CambriconContext& ctx)
+        : ComputingContext(dev_handle, flags),
+          own_queue{ctx.queue == nullptr},
+          context_{ctx} {
+    megcorePlatform_t platform;
+    megcoreGetPlatform(dev_handle, &platform);
+    megdnn_assert(platform == megcorePlatformCambricon);
+    if (own_queue) {
+        cnrt_check(cnrtCreateQueue(&context_.queue));
+    }
+}
+
+CambriconComputingContext::~CambriconComputingContext() {
+    if (own_queue) {
+        cnrt_check(cnrtDestroyQueue(context_.queue));
+    }
+}
+
+void CambriconComputingContext::memcpy(void* dst, const void* src,
+                                       size_t size_in_bytes,
+                                       megcoreMemcpyKind_t kind) {
+    cnrtMemTransDir_t dir;
+    switch (kind) {
+        case megcoreMemcpyDeviceToHost:
+            dir = CNRT_MEM_TRANS_DIR_DEV2HOST;
+            break;
+        case megcoreMemcpyHostToDevice:
+            dir = CNRT_MEM_TRANS_DIR_HOST2DEV;
+            break;
+        case megcoreMemcpyDeviceToDevice:
+            dir = CNRT_MEM_TRANS_DIR_DEV2DEV;
+            break;
+        default:
+            megdnn_throw(megdnn_mangle("bad cnrt mem trans dir"));
+    }
+    if (kind == megcoreMemcpyDeviceToDevice) {
+        cnrt_check(cnrtSyncQueue(context_.queue));
+        cnrt_check(cnrtMemcpy(dst, const_cast<void*>(src), size_in_bytes, dir));
+        return;
+    }
+    cnrt_check(cnrtMemcpyAsync(dst, const_cast<void*>(src), size_in_bytes,
+                               context_.queue, dir));
+}
+
+void CambriconComputingContext::memset(void* dst, int value,
+                                       size_t size_in_bytes) {
+    cnrt_check(cnrtSyncQueue(context_.queue));
+    cnrt_check(cnrtMemset(dst, value, size_in_bytes));
+}
+
+void CambriconComputingContext::synchronize() {
+    cnrt_check(cnrtSyncQueue(context_.queue));
+}
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/megcore/cambricon_computing_context.hpp

@@ -0,0 +1,44 @@
+/**
+ * \file dnn/src/cambricon/megcore/cambricon_computing_context.hpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+
+#include "megcore_cambricon.h"
+#include "src/common/megcore/common/computing_context.hpp"
+
+namespace megcore {
+namespace cambricon {
+
+class CambriconComputingContext final : public ComputingContext {
+public:
+    CambriconComputingContext(megcoreDeviceHandle_t dev_handle,
+                              unsigned int flags,
+                              const CambriconContext& ctx = {});
+    ~CambriconComputingContext();
+
+    void memcpy(void* dst, const void* src, size_t size_in_bytes,
+                megcoreMemcpyKind_t kind) override;
+    void memset(void* dst, int value, size_t size_in_bytes) override;
+    void synchronize() override;
+
+    const CambriconContext& context() const { return context_; }
+
+    cnrtQueue_t queue() const { return context().queue; }
+
+private:
+    bool own_queue;
+    CambriconContext context_;
+};
+
+}  // namespace cambricon
+}  // namespace megcore
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/megcore/cambricon_device_context.cpp

@@ -0,0 +1,80 @@
+/**
+ * \file dnn/src/cambricon/megcore/cambricon_device_context.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "megcore.h"
+
+#include "src/cambricon/utils.h"
+#include "src/common/utils.h"
+
+#include "src/cambricon/megcore/cambricon_device_context.hpp"
+
+#define STR_HELPER(x) #x
+#define STR(x) STR_HELPER(x)
+
+#define CNRT_VERSION_STR    \
+    STR(CNRT_MAJOR_VERSION) \
+    "." STR(CNRT_MINOR_VERSION) "." STR(CNRT_PATCH_VERSION)
+
+#pragma message "compile with cnrt " CNRT_VERSION_STR " "
+
+#undef STR_HELPER
+#undef STR
+
+using namespace megcore;
+using namespace cambricon;
+
+CambriconDeviceContext::CambriconDeviceContext(int device_id,
+                                               unsigned int flags,
+                                               bool global_initialized)
+        : DeviceContext(megcorePlatformCambricon, device_id, flags) {
+    if (!global_initialized)
+        init_status.init();
+    unsigned int version;
+    cnrt_check(cnrtGetVersion(&version));
+    megdnn_assert(version == CNRT_VERSION,
+                  "megcore compiled with cnrt %d, get %d at runtime",
+                  CNRT_VERSION, version);
+    unsigned int dev_num;
+    cnrt_check(cnrtGetDeviceCount(&dev_num));
+    MEGDNN_MARK_USED_VAR(dev_num);
+    // check validity of device_id
+    megdnn_assert(device_id >= 0 &&
+                  static_cast<unsigned int>(device_id) < dev_num);
+    cnrt_check(cnrtGetDeviceInfo(&device_info, device_id));
+}
+
+CambriconDeviceContext::~CambriconDeviceContext() noexcept = default;
+
+size_t CambriconDeviceContext::mem_alignment_in_bytes() const noexcept {
+    return 1;
+}
+
+void CambriconDeviceContext::activate() {
+    int id = device_id();
+    cnrtDev_t dev;
+    cnrt_check(cnrtGetDeviceHandle(&dev, id));
+    cnrt_check(cnrtSetCurrentDevice(dev));
+}
+
+void* CambriconDeviceContext::malloc(size_t size_in_bytes) {
+    void* ptr;
+    cnrt_check(cnrtMalloc(&ptr, size_in_bytes));
+    return ptr;
+}
+
+void CambriconDeviceContext::free(void* ptr) {
+    cnrt_check(cnrtFree(ptr));
+}
+
+CambriconDeviceContext::InitStatus CambriconDeviceContext::init_status;
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/megcore/cambricon_device_context.hpp

@@ -0,0 +1,63 @@
+/**
+ * \file dnn/src/cambricon/megcore/cambricon_device_context.hpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+
+#include <mutex>
+#include "megcore_cambricon.h"
+#include "src/common/megcore/common/device_context.hpp"
+#include "src/common/utils.h"
+
+namespace megcore {
+namespace cambricon {
+
+class CambriconDeviceContext : public DeviceContext {
+public:
+    CambriconDeviceContext(int device_id, unsigned int flags,
+                           bool global_initialized = false);
+    ~CambriconDeviceContext() noexcept;
+
+    size_t mem_alignment_in_bytes() const noexcept override;
+
+    void activate() override;
+    void* malloc(size_t size_in_bytes) override;
+    void free(void* ptr) override;
+
+    struct InitStatus {
+        bool initialized;
+        std::mutex mtx;
+        InitStatus() : initialized{false} {}
+        void init() {
+            std::lock_guard<std::mutex> guard{mtx};
+            if (!initialized) {
+                auto cnrt_err = cnrtInit(0);
+                initialized = cnrt_err == CNRT_RET_SUCCESS;
+                megdnn_assert(initialized, "cnrt initialize failed: (cnrt:%d)",
+                              static_cast<int>(cnrt_err));
+            }
+        }
+        ~InitStatus() {
+            if (initialized) {
+                cnrtDestroy();
+                initialized = false;
+            }
+        }
+    };
+    static InitStatus init_status;
+
+private:
+    cnrtDeviceInfo_t device_info;
+};
+
+}  // namespace cambricon
+}  // namespace megcore
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/megcore/public_api/computing.cpp

@@ -0,0 +1,59 @@
+/**
+ * \file dnn/src/cambricon/megcore/public_api/computing.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#include "megcore_cambricon.h"
+
+#include "src/cambricon/megcore/cambricon_computing_context.hpp"
+#include "src/cambricon/megcore/cambricon_device_context.hpp"
+#include "src/common/megcore/public_api/computing.hpp"
+#include "src/common/megcore/public_api/device.hpp"
+#include "src/common/utils.h"
+
+using namespace megcore;
+
+megcoreStatus_t megcore::createDeviceHandleWithGlobalInitStatus(
+        megcoreDeviceHandle_t* devHandle, int deviceID, unsigned int flags,
+        bool global_initialized) {
+    auto content = megdnn::make_unique<cambricon::CambriconDeviceContext>(
+            deviceID, flags, global_initialized);
+    auto& ctx = *devHandle;
+    ctx = new megcoreDeviceContext;
+    ctx->content = std::move(content);
+    return megcoreSuccess;
+}
+
+megcoreStatus_t megcore::createComputingHandleWithCambriconContext(
+        megcoreComputingHandle_t* compHandle, megcoreDeviceHandle_t devHandle,
+        unsigned int flags, const CambriconContext& ctx) {
+    auto content = megdnn::make_unique<cambricon::CambriconComputingContext>(
+            devHandle, flags, ctx);
+    auto& H = *compHandle;
+    H = new megcoreComputingContext;
+    H->content = std::move(content);
+    return megcoreSuccess;
+}
+
+megcoreStatus_t megcore::getCambriconContext(megcoreComputingHandle_t handle,
+                                             CambriconContext* ctx) {
+    auto&& H = handle;
+    megdnn_assert(H);
+    megcoreDeviceHandle_t dev_handle = H->content->dev_handle();
+    megcorePlatform_t platform;
+    megcoreGetPlatform(dev_handle, &platform);
+    megdnn_assert(platform == megcorePlatformCambricon);
+    auto context = static_cast<megcore::cambricon::CambriconComputingContext*>(
+            H->content.get());
+    *ctx = context->context();
+    return megcoreSuccess;
+}
+
+// vim: syntax=cpp.doxygen
+
+

dnn/src/cambricon/utils.cpp

@@ -0,0 +1,75 @@
+/**
+ * \file dnn/src/cambricon/utils.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#include "src/cambricon/utils.h"
+#include "src/cambricon/utils.mlu.h"
+
+#include "src/cambricon/handle.h"
+#include "src/common/utils.h"
+
+#include <mutex>
+#include <unordered_map>
+
+using namespace megdnn;
+using namespace cambricon;
+
+namespace {
+struct DeviceInfoRecord {
+    bool init = false;
+    cnrtDeviceInfo_t device_info;
+    std::mutex mtx;
+};
+std::unordered_map<cnrtDev_t, int> dev2device_id;
+std::mutex dev2device_id_mtx;
+constexpr int MAX_NR_DEVICE = 64;
+DeviceInfoRecord device_info_rec[MAX_NR_DEVICE];
+}  // namespace
+
+void cambricon::__throw_cnrt_error__(cnrtRet_t err, const char* msg) {
+    auto s = ssprintf("cnrt return %s(%d) occurred; expr: %s",
+                      cnrtGetErrorStr(err), int(err), msg);
+    megdnn_throw(s.c_str());
+}
+
+cnrtDeviceInfo_t cambricon::current_device_info() {
+    static bool dev2device_id_init = false;
+    {
+        std::lock_guard<std::mutex> lock(dev2device_id_mtx);
+        if (!dev2device_id_init) {
+            unsigned int dev_num = 0;
+            cnrt_check(cnrtGetDeviceCount(&dev_num));
+            for (unsigned int dev_id = 0; dev_id < dev_num; ++dev_id) {
+                cnrtDev_t dev;
+                cnrt_check(cnrtGetDeviceHandle(&dev, dev_id));
+                dev2device_id[dev] = dev_id;
+            }
+            dev2device_id_init = true;
+        }
+    }
+
+    cnrtDev_t dev;
+    cnrt_check(cnrtGetCurrentDevice(&dev));
+    {
+        std::lock_guard<std::mutex> lock(dev2device_id_mtx);
+        int dev_id = dev2device_id.at(dev);
+        auto& rec = device_info_rec[dev_id];
+        {
+            std::lock_guard<std::mutex> lock(rec.mtx);
+            if (!rec.init) {
+                cnrt_check(cnrtGetDeviceInfo(&rec.device_info, dev_id));
+                rec.init = true;
+            }
+        }
+        return rec.device_info;
+    }
+}
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/utils.h

@@ -0,0 +1,40 @@
+/**
+ * \file dnn/src/cambricon/utils.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+
+#include "megcore_cdefs.h"
+#include "megdnn/handle.h"
+#include "src/cambricon/utils.mlu.h"
+#include "src/common/utils.h"
+
+#include "src/cambricon/handle.h"
+
+#include <cnrt.h>
+
+namespace megdnn {
+namespace cambricon {
+
+static inline HandleImpl* concrete_handle(Handle* handle) {
+    return static_cast<cambricon::HandleImpl*>(handle);
+}
+
+static inline cnrtQueue_t cnrt_queue(Handle* handle) {
+    return concrete_handle(handle)->queue();
+}
+
+//! get device info of current active device
+cnrtDeviceInfo_t current_device_info();
+
+}  // namespace cambricon
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen
+

dnn/src/cambricon/utils.mlu.h

@@ -0,0 +1,42 @@
+/**
+ * \file dnn/src/cambricon/utils.mlu.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+
+#include "src/common/utils.cuh"
+
+#include <stdint.h>
+
+#include <cnrt.h>
+
+#define cnrt_check(_x)                                            \
+    do {                                                          \
+        cnrtRet_t _ret = (_x);                                    \
+        if (_ret != CNRT_RET_SUCCESS) {                           \
+            ::megdnn::cambricon::__throw_cnrt_error__(_ret, #_x); \
+        }                                                         \
+    } while (0)
+
+#define after_kernel_launch()         \
+    do {                              \
+        cnrt_check(cnrtGetLastErr()); \
+    } while (0)
+
+namespace megdnn {
+namespace cambricon {
+
+//! Error handling funcions
+MEGDNN_NORETURN void __throw_cnrt_error__(cnrtRet_t err, const char* msg);
+
+}  // namespace cambricon
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen
+

dnn/src/common/handle.cpp

@@ -36,6 +36,14 @@
 #endif
 
 
+#if MEGDNN_WITH_CAMBRICON
+#include "src/cambricon/handle.h"
+#endif
+
+#ifdef MEGDNN_WITH_ATLAS
+#include "src/atlas/handle.h"
+#endif
+
 using namespace megdnn;
 
 MIDOUT_DECL(HandlePlatform);
@@ -83,6 +91,20 @@ std::unique_ptr<Handle> Handle::make(megcoreComputingHandle_t computing_handle,
             }
         }
         MIDOUT_END();
+#endif
+        }
+        else if (platform == megcorePlatformCambricon) {
+#if MEGDNN_WITH_CAMBRICON
+            return make_unique<cambricon::HandleImpl>(computing_handle);
+#else
+            return nullptr;
+#endif
+        }
+        else if (platform == megcorePlatformAtlas) {
+#if MEGDNN_WITH_ATLAS
+            return make_unique<atlas::HandleImpl>(computing_handle);
+#else
+            return nullptr;
 #endif
         }
         else {
@@ -94,6 +116,7 @@ std::unique_ptr<Handle> Handle::make(megcoreComputingHandle_t computing_handle,
             return nullptr;
 #endif
         }
+        return nullptr;
     }
 
 
@@ -166,6 +189,12 @@ std::unique_ptr<Handle> Handle::make(megcoreComputingHandle_t computing_handle,
 #endif  // !MEGDNN_NAIVE
 #if MEGDNN_WITH_CUDA
             CASE(CUDA,cuda);
+#endif
+#if MEGDNN_WITH_ATLAS
+            CASE(ATLAS, atlas);
+#endif
+#if MEGDNN_WITH_CAMBRICON
+            CASE(CAMBRICON, cambricon);
 #endif
             default:
                 megdnn_throw(megdnn_mangle("bad handle type"));

dnn/src/common/megcore/common/computing_context.cpp

@@ -18,6 +18,14 @@
 #endif
 
 
+#if MEGDNN_WITH_CAMBRICON
+#include "src/cambricon/megcore/cambricon_computing_context.hpp"
+#endif
+
+#if MEGDNN_WITH_ATLAS
+#include "src/atlas/megcore/computing_context.hpp"
+#endif
+
 using namespace megcore;
 using namespace megdnn;
 
@@ -32,6 +40,15 @@ std::unique_ptr<ComputingContext> ComputingContext::make(
 #if MEGDNN_WITH_CUDA
         case megcorePlatformCUDA:
             return make_unique<cuda::CUDAComputingContext>(dev_handle, flags);
+#endif
+#if MEGDNN_WITH_CAMBRICON
+        case megcorePlatformCambricon:
+            return make_unique<cambricon::CambriconComputingContext>(dev_handle,
+                                                                     flags);
+#endif
+#if MEGDNN_WITH_ATLAS
+        case megcorePlatformAtlas:
+            return make_unique<atlas::AtlasComputingContext>(dev_handle, flags);
 #endif
         default:
             megdnn_throw("bad platform");

dnn/src/common/megcore/common/device_context.cpp

@@ -15,6 +15,13 @@
 #if MEGDNN_WITH_CUDA
 #include "src/cuda/megcore/cuda_device_context.hpp"
 #endif
+#if MEGDNN_WITH_CAMBRICON
+#include "src/cambricon/megcore/cambricon_device_context.hpp"
+#endif
+
+#if MEGDNN_WITH_ATLAS
+#include "src/atlas/megcore/device_context.hpp"
+#endif
 
 using namespace megcore;
 using namespace megdnn;
@@ -28,6 +35,15 @@ std::unique_ptr<DeviceContext> DeviceContext::make(megcorePlatform_t platform,
 #if MEGDNN_WITH_CUDA
         case megcorePlatformCUDA:
             return make_unique<cuda::CUDADeviceContext>(deviceID, flags);
+#endif
+#if MEGDNN_WITH_CAMBRICON
+        case megcorePlatformCambricon:
+            return make_unique<cambricon::CambriconDeviceContext>(deviceID,
+                                                                  flags);
+#endif
+#if MEGDNN_WITH_ATLAS
+        case megcorePlatformAtlas:
+            return make_unique<atlas::AtlasDeviceContext>(deviceID, flags);
 #endif
         default:
             megdnn_throw("bad platform");

dnn/test/CMakeLists.txt

@@ -26,6 +26,16 @@ if(MGE_WITH_CUDA)
 endif()
 
 
+if(MGE_WITH_CAMBRICON)
+    file(GLOB_RECURSE SOURCES_ cambricon/*.cpp)
+    list(APPEND SOURCES ${SOURCES_})
+endif()
+
+if(MGE_WITH_ATLAS)
+    file(GLOB_RECURSE SOURCES_ atlas/*.cpp)
+    list(APPEND SOURCES ${SOURCES_})
+endif()
+
 
 add_executable(megdnn_test ${SOURCES})
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")

dnn/test/atlas/checksum.cpp

@@ -0,0 +1,74 @@
+/**
+ * \file dnn/test/atlas/checksum.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megdnn/oprs.h"
+#include "test/atlas/fixture.h"
+#include "test/common/checker.h"
+
+using namespace megdnn;
+using namespace test;
+
+TEST_F(ATLAS, CHECKSUM_FORWARD) {
+    auto atlas_opr = handle_atlas()->create_operator<megdnn::Checksum>(),
+         naive_opr = handle_naive()->create_operator<megdnn::Checksum>();
+    std::mt19937 rng(std::random_device{}());
+    for (size_t size :
+         {3, 8, 4 * 4 * 1024, 12345, 1024 * 1024, 1024 * 1024 * 10}) {
+        auto aligned_size = size + ((512 - size % 512) % 512);
+        auto run = [&](megdnn::Checksum* opr, void* ptr, bool log_size) {
+            TensorND tensor;
+            tensor.raw_ptr = ptr;
+            tensor.layout.init_contiguous_stride({size});
+            tensor.layout.dtype = dtype::Byte();
+            WorkspaceWrapper workspace(
+                    handle_atlas(), opr->get_workspace_in_bytes(tensor.layout));
+            if (log_size) {
+                printf("checksum(%zu): workspace=%zu\n", size,
+                       workspace.workspace().size);
+            }
+            return opr->exec(tensor, workspace.workspace());
+        };
+        std::vector<uint8_t> buf(aligned_size);
+        for (size_t i = 0; i < size; ++i)
+            buf[i] = 1;
+        auto run_offsset = [&](size_t offset) {
+            void* dev_ptr = megdnn_malloc(handle_atlas(), buf.size() + offset);
+            void* dev_buf = static_cast<char*>(dev_ptr) + offset;
+
+            Checksum::Result res_cambricon[2], res_naive[2];
+
+            for (int change_last = 0; change_last < 2; ++change_last) {
+                if (change_last)
+                    ++buf[size - 1];
+
+                megdnn_memcpy_H2D(handle_atlas(), dev_buf, buf.data(), size);
+                res_cambricon[change_last] =
+                        run(atlas_opr.get(), dev_buf, !change_last);
+                res_naive[change_last] =
+                        run(naive_opr.get(), buf.data(), false);
+            }
+
+            megdnn_free(handle_atlas(), dev_ptr);
+
+            ASSERT_EQ(res_naive[0], res_cambricon[0])
+                    << "failed for size " << size;
+            ASSERT_EQ(res_naive[1], res_cambricon[1]);
+            ASSERT_NE(res_cambricon[0], res_cambricon[1]);
+        };
+
+        for (size_t i = 0; i < 8; ++i) {
+            run_offsset(i);
+        }
+    }
+}
+
+// vim: syntax=cpp.doxygen

dnn/test/atlas/fixture.cpp

@@ -0,0 +1,55 @@
+/**
+ * \file dnn/test/atlas/fixture.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#include "test/atlas/fixture.h"
+#include "src/atlas/handle.h"
+#include "src/atlas/megcore/device_context.hpp"
+#include "src/atlas/utils.h"
+
+#include "test/common/memory_manager.h"
+#include "test/common/random_state.h"
+#include "test/common/utils.h"
+
+#include "acl/acl.h"
+
+#include <cstdlib>
+
+using namespace megdnn;
+using namespace test;
+
+void ATLAS::SetUp() {
+    RandomState::reset();
+
+    // use card 0
+    megcore_check(
+            megcoreCreateDeviceHandle(&m_dev_handle, megcorePlatformAtlas, 0));
+
+    megcoreActivate(m_dev_handle);
+    megcoreComputingHandle_t comp_handle;
+    megcore_check(megcoreCreateComputingHandle(&comp_handle, m_dev_handle));
+    m_handle_atlas = Handle::make(comp_handle);
+    megdnn_assert(m_handle_atlas);
+}
+
+Handle* ATLAS::handle_naive() {
+    if (!m_handle_naive)
+        m_handle_naive = create_cpu_handle(2);
+    return m_handle_naive.get();
+}
+
+void ATLAS::TearDown() {
+    m_handle_naive.reset();
+    m_handle_atlas.reset();
+    MemoryManagerHolder::instance()->clear();
+    megcoreDeactivate(m_dev_handle);
+}
+
+// vim: syntax=cpp.doxygen

dnn/test/atlas/fixture.h

@@ -0,0 +1,40 @@
+/**
+ * \file dnn/test/atlas/fixture.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+#pragma once
+#include <gtest/gtest.h>
+
+#include "megcore_cdefs.h"
+#include "megdnn/handle.h"
+
+#include <memory>
+
+namespace megdnn {
+namespace test {
+
+class ATLAS : public ::testing::Test {
+public:
+    void SetUp() override;
+    void TearDown() override;
+
+    Handle* handle_atlas() { return m_handle_atlas.get(); }
+    Handle* handle_naive();
+
+private:
+    std::unique_ptr<Handle> m_handle_naive;
+    std::unique_ptr<Handle> m_handle_atlas;
+    megcoreDeviceHandle_t m_dev_handle;
+};
+
+}  // namespace test
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen

dnn/test/cambricon/checksum.cpp

@@ -0,0 +1,78 @@
+/**
+ * \file dnn/test/cambricon/checksum.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "megdnn/oprs.h"
+#include "test/cambricon/fixture.h"
+#include "test/common/checker.h"
+
+using namespace megdnn;
+using namespace test;
+
+TEST_F(CAMBRICON, CHECKSUM_FORWARD) {
+    auto cambricon_opr =
+                 handle_cambricon()->create_operator<megdnn::Checksum>(),
+         naive_opr = handle_naive()->create_operator<megdnn::Checksum>();
+    std::mt19937 rng(std::random_device{}());
+    for (size_t size :
+         {3, 8, 4 * 4 * 1024, 12345, 1024 * 1024, 1024 * 1024 * 10}) {
+        auto aligned_size = size + ((512 - size % 512) % 512);
+        auto run = [&](megdnn::Checksum* opr, void* ptr, bool log_size) {
+            TensorND tensor;
+            tensor.raw_ptr = ptr;
+            tensor.layout.init_contiguous_stride({size});
+            tensor.layout.dtype = dtype::Byte();
+            WorkspaceWrapper workspace(
+                    handle_cambricon(),
+                    opr->get_workspace_in_bytes(tensor.layout));
+            if (log_size) {
+                printf("checksum(%zu): workspace=%zu\n", size,
+                       workspace.workspace().size);
+            }
+            return opr->exec(tensor, workspace.workspace());
+        };
+        std::vector<uint8_t> buf(aligned_size);
+        for (size_t i = 0; i < size; ++i)
+            buf[i] = 1;
+        auto run_offsset = [&](size_t offset) {
+            void* dev_ptr =
+                    megdnn_malloc(handle_cambricon(), buf.size() + offset);
+            void* dev_buf = static_cast<char*>(dev_ptr) + offset;
+
+            Checksum::Result res_cambricon[2], res_naive[2];
+
+            for (int change_last = 0; change_last < 2; ++change_last) {
+                if (change_last)
+                    ++buf[size - 1];
+
+                megdnn_memcpy_H2D(handle_cambricon(), dev_buf, buf.data(),
+                                  size);
+                res_cambricon[change_last] =
+                        run(cambricon_opr.get(), dev_buf, !change_last);
+                res_naive[change_last] =
+                        run(naive_opr.get(), buf.data(), false);
+            }
+
+            megdnn_free(handle_cambricon(), dev_ptr);
+
+            ASSERT_EQ(res_naive[0], res_cambricon[0])
+                    << "failed for size " << size;
+            ASSERT_EQ(res_naive[1], res_cambricon[1]);
+            ASSERT_NE(res_cambricon[0], res_cambricon[1]);
+        };
+
+        for (size_t i = 0; i < 8; ++i) {
+            run_offsset(i);
+        }
+    }
+}
+
+// vim: syntax=cpp.doxygen
+

dnn/test/cambricon/fixture.cpp

@@ -0,0 +1,51 @@
+/**
+ * \file dnn/test/cambricon/fixture.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#include "test/cambricon/fixture.h"
+#include "src/cambricon/handle.h"
+#include "src/cambricon/utils.h"
+#include "test/common/memory_manager.h"
+#include "test/common/random_state.h"
+#include "test/common/utils.h"
+
+#include <cnrt.h>
+#include <cstdlib>
+
+using namespace megdnn;
+using namespace test;
+
+void CAMBRICON::SetUp() {
+    RandomState::reset();
+
+    megcoreDeviceHandle_t dev_handle;
+    // use card 0
+    megcore_check(megcoreCreateDeviceHandle(&dev_handle,
+                                            megcorePlatformCambricon, 0));
+
+    megcoreComputingHandle_t comp_handle;
+    megcore_check(megcoreCreateComputingHandle(&comp_handle, dev_handle));
+    m_handle_cambricon = Handle::make(comp_handle);
+    megdnn_assert(m_handle_cambricon);
+}
+
+Handle* CAMBRICON::handle_naive() {
+    if (!m_handle_naive)
+        m_handle_naive = create_cpu_handle(2);
+    return m_handle_naive.get();
+}
+
+void CAMBRICON::TearDown() {
+    m_handle_naive.reset();
+    m_handle_cambricon.reset();
+    MemoryManagerHolder::instance()->clear();
+}
+
+// vim: syntax=cpp.doxygen
+

dnn/test/cambricon/fixture.h

@@ -0,0 +1,40 @@
+/**
+ * \file dnn/test/cambricon/fixture.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+#pragma once
+#include <gtest/gtest.h>
+#include "test/common/fix_gtest_on_platforms_without_exception.inl"
+
+#include "megcore_cdefs.h"
+#include "megdnn/handle.h"
+
+#include <memory>
+
+namespace megdnn {
+namespace test {
+
+class CAMBRICON : public ::testing::Test {
+public:
+    void SetUp() override;
+    void TearDown() override;
+
+    Handle* handle_cambricon() { return m_handle_cambricon.get(); }
+    Handle* handle_naive();
+
+private:
+    std::unique_ptr<Handle> m_handle_naive;
+    std::unique_ptr<Handle> m_handle_cambricon;
+};
+
+}  // namespace test
+}  // namespace megdnn
+
+// vim: syntax=cpp.doxygen
+

python_module/megengine/functional/external.py

@@ -0,0 +1,54 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# pylint: disable=too-many-lines
+from typing import List
+
+import megengine._internal as mgb
+
+from ..core import Tensor, wrap_io_tensor
+
+
+@wrap_io_tensor
+def cambricon_subgraph(
+    inputs: List[Tensor], data: bytes, symbol: str, tensor_dim_mutable: bool,
+) -> List[Tensor]:
+    """Load a serialized Cambricon subgraph (i.e. cnrtModel_t) and
+    execute the operations defined in the subgraph.
+
+    :param inputs: List of input tensors of the subgraph.
+    :param data: The serialized subgraph.
+    :param symbol: The name of the function in the subgraph.
+        The function is corresponding to a cnmlFusionOp
+        which is added to the cnmlModel_t/cnrtModel_t.
+    :param tensor_dim_mutable: Whether the input tensors' shapes are mutalbe
+        in cnrtModel_t
+    """
+    return mgb.opr.cambricon_runtime(
+        data, symbol, tuple(map(lambda x: x._symvar, inputs)), tensor_dim_mutable
+    )
+
+
+@wrap_io_tensor
+def extern_opr_subgraph(
+    inputs, output_shapes: List[tuple], dump_name: str, dump_data: bytes,
+) -> List[Tensor]:
+    """Load a serialized extern opr subgraph and fake execute the operator
+
+    :param inputs: Tensor or list of input tensors.
+    :param output_shapes: The output shapes.
+    :param dump_name: The serialized subgraph name.
+    :param dump_data: The serialized subgraph.
+
+    :return: List of tensors
+    """
+    if not isinstance(inputs, list):
+        inputs = [inputs]
+    return mgb.opr.extern_c_opr_placeholder(
+        inputs, output_shapes, dump_name=dump_name, dump_data=dump_data,
+    )

python_module/megengine/module/external.py

@@ -0,0 +1,56 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import numpy as np
+
+from ..functional.external import cambricon_subgraph, extern_opr_subgraph
+from .module import Module
+
+
+class CambriconSubgraph(Module):
+    r"""Load a serialized Cambricon subgraph.
+
+    See :func:`~.cambricon_subgraph` for more details.
+    """
+
+    def __init__(
+        self, data, symbol, tensor_dim_mutable,
+    ):
+        super(CambriconSubgraph, self).__init__()
+        self._data = data
+        self.symbol = symbol
+        self.tensor_dim_mutable = tensor_dim_mutable
+
+    @property
+    def data(self):
+        return self._data.tobytes()
+
+    @data.setter
+    def data(self, val):
+        self._data = np.frombuffer(val, dtype=np.uint8)
+
+    def forward(self, inputs):
+        outputs = cambricon_subgraph(
+            inputs, self._data, self.symbol, self.tensor_dim_mutable,
+        )
+        return outputs
+
+
+class ExternOprSubgraph(Module):
+    r"""Load a serialized extern opr subgraph.
+    """
+
+    def __init__(self, data, name, output_shapes):
+        super(ExternOprSubgraph, self).__init__()
+        self.data = data
+        self.name = name
+        self.output_shapes = output_shapes
+
+    def forward(self, inputs):
+        outputs = extern_opr_subgraph(inputs, self.output_shapes, self.name, self.data,)
+        return outputs

python_module/src/cpp/opr_defs.cpp

@@ -246,6 +246,27 @@ SymbolVarArray _Opr::tensor_rt_runtime(const SymbolVarArray& inputs,
 }
 #endif
 
+#if MGB_ATLAS
+
+#include "megbrain/opr/atlas_runtime_op.h"
+
+SymbolVarArray _Opr::atlas_runtime(const SymbolVarArray& inputs,
+                                   PyObject* data_bytes,
+                                   const OperatorNodeConfig& config) {
+    mgb_assert(PyBytes_Check(data_bytes));
+    auto size = PyBytes_Size(data_bytes);
+    mgb_assert(size, "atlas data bytes should not be empty");
+
+    return opr::AtlasRuntimeOpr::make(PyBytes_AsString(data_bytes), size,
+                                      inputs, config);
+}
+#else
+SymbolVarArray _Opr::atlas_runtime(const SymbolVarArray& inputs,
+                                   PyObject* data_bytes,
+                                   const OperatorNodeConfig& config) {
+    mgb_throw(MegBrainError, "Atlas disabled at compile time");
+}
+#endif
 
 SymbolVar _Opr::timestamp(SymbolVar input, PyObject* dest, size_t dest_off,
                            const OperatorNodeConfig& config) {
@@ -266,4 +287,27 @@ SymbolVar _Opr::virtual_dep(const SymbolVarArray& symvars,
 }
 
 
+#if MGB_CAMBRICON
+#include "megbrain/cambricon/cambricon_runtime_opr.h"
+
+SymbolVarArray _Opr::cambricon_runtime(PyObject* data_bytes, const char* symbol,
+                                       const SymbolVarArray& inputs,
+                                       bool tensor_dim_mutable,
+                                       const OperatorNodeConfig& config) {
+    mgb_assert(PyBytes_Check(data_bytes));
+    auto size = PyBytes_Size(data_bytes);
+    mgb_assert(size, "cambricon data bytes should not be empty");
+    return opr::CambriconRuntimeOpr::make(PyBytes_AsString(data_bytes), size,
+                                          symbol, inputs, tensor_dim_mutable,
+                                          config);
+}
+#else
+SymbolVarArray _Opr::cambricon_runtime(PyObject* data_bytes, const char* symbol,
+                                       const SymbolVarArray& inputs,
+                                       bool tensor_dim_mutable,
+                                       const OperatorNodeConfig& config) {
+    mgb_throw(MegBrainError, "Cambricon disabled at compile time");
+}
+#endif
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

python_module/src/cpp/opr_defs.h

@@ -130,6 +130,16 @@ static SymbolVar virtual_loss(const SymbolVarArray& ys,
 static SymbolVar virtual_dep(const SymbolVarArray& symvars,
                              const OperatorNodeConfig& config);
 
+static SymbolVarArray atlas_runtime(const SymbolVarArray& inputs,
+        PyObject* data_bytes,
+        const OperatorNodeConfig& config);
+
+
+static SymbolVarArray cambricon_runtime(PyObject* data_bytes,
+                                        const char* symbol,
+                                        const SymbolVarArray& inputs,
+                                        bool tensor_dim_mutable,
+                                        const OperatorNodeConfig& config);
 
 #ifdef SWIG
 %pythoncode {

python_module/test/unit/module/test_external.py

@@ -0,0 +1,43 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import os
+
+import numpy as np
+
+import megengine as mge
+from megengine import tensor
+from megengine.module import Module
+from megengine.module.external import CambriconSubgraph
+
+
+class MyModule(Module):
+    def __init__(self, data):
+        super().__init__()
+        self.cambricon = CambriconSubgraph(data, "subnet0", True)
+
+    def forward(self, inputs):
+        out = self.cambricon(inputs)
+        return out
+
+
+def test_cambricon_module():
+    model = "CambriconRuntimeOprTest.MutableBatchSize.mlu"
+    model = os.path.join(os.path.dirname(__file__), model)
+    with open(model, "rb") as f:
+        data = f.read()
+        m = MyModule(data)
+        inputs = []
+        inputs.append(tensor(dtype=np.float16, device="cambricon0"))
+        inputs[0].set_value(np.random.normal(size=(1, 64, 32, 32)).astype(np.float16))
+
+        def inference(inps):
+            pred = m(inps)
+            return pred
+
+        pred = inference(inputs)

src/CMakeLists.txt

@@ -33,6 +33,14 @@ if(MGE_WITH_CUDA AND MGE_WITH_TRT)
     list(APPEND SOURCES ${SOURCES_})
 endif()
 
+if(MGE_WITH_CAMBRICON)
+    list(APPEND MGB_INC ${CMAKE_CURRENT_LIST_DIR}/cambricon/include)
+    file(GLOB_RECURSE SOURCES_ cambricon/impl/*.cpp cambricon/impl/*.inl)
+    list(APPEND SOURCES ${SOURCES_})
+endif()
+set(MGB_CAMBRICON ${MGE_WITH_CAMBRICON})
+
+set(MGB_ATLAS ${MGE_WITH_ATLAS})
 
 if(MGE_WITH_CUDA)
     file(GLOB_RECURSE SOURCES_ opr/impl/standalone/*.cu)
@@ -77,6 +85,8 @@ if(MGE_WITH_DISTRIBUTED)
     target_link_libraries (megbrain PRIVATE megray)
 endif()
 target_link_libraries(megbrain PRIVATE ${MGE_CUDA_LIBS})
+target_link_libraries(megbrain PUBLIC ${MGE_CAMBRICON_LIBS})
+target_link_libraries(megbrain PUBLIC ${MGE_ATLAS_LIBS})
 if(MGE_WITH_JIT AND MGE_WITH_HALIDE)
     target_link_libraries(megbrain PRIVATE libhalide)
     target_link_libraries(megbrain PRIVATE ${HALIDE_LLVM_LIBS})

src/cambricon/impl/cambricon_runtime_opr.cpp

@@ -0,0 +1,320 @@
+/**
+ * \file src/cambricon/impl/cambricon_runtime_opr.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "megbrain/cambricon/cambricon_runtime_opr.h"
+#include "megbrain/common.h"
+
+#if MGB_CAMBRICON
+
+using namespace mgb;
+using namespace opr;
+
+namespace {
+SmallVector<int> mgb_shape_to_cnrt_shape(TensorShape mgb_shp) {
+    int ndim = mgb_shp.ndim;
+    SmallVector<int> cnrt_shp(ndim);
+    for (int i = 0; i < ndim; ++i) {
+        cnrt_shp[i] = mgb_shp[i];
+    }
+    return cnrt_shp;
+}
+TensorShape cnrt_shape_to_mgb_shape(int* dim_values, int dim_num) {
+    TensorShape ret;
+    ret.ndim = dim_num;
+    for (int i = 0; i < dim_num; ++i) {
+        ret[i] = dim_values[i];
+    }
+    return ret;
+}
+DType cnrt_dtype_to_mgb_dtype(cnrtDataType_t data_type) {
+    switch (data_type) {
+        case CNRT_FLOAT16:
+#if !MEGDNN_DISABLE_FLOAT16
+            return dtype::Float16();
+#else
+            mgb_throw(MegBrainError,
+                      "Float16 support is disabled at compile time.");
+#endif
+        case CNRT_FLOAT32:
+            return dtype::Float32();
+        case CNRT_INT8:
+            return dtype::QuantizedS8(1.f);
+        case CNRT_INT16:
+            return dtype::Int16();
+        case CNRT_INT32:
+            return dtype::Int32();
+        case CNRT_UINT8:
+            return dtype::Uint8();
+        //! TODO: check scale
+        case CNRT_QUANT8:
+            return dtype::QuantizedS8(1.f);
+        default:
+            mgb_throw(MegBrainError,
+                      "cnrtDataType %x is not supported by MegBrain.",
+                      data_type);
+    }
+}
+cnrtDataType_t mgb_dtype_to_cnrt_dtype(DType data_type) {
+    switch (data_type.enumv()) {
+#if !MEGDNN_DISABLE_FLOAT16
+        case DTypeEnum::Float16:
+            return CNRT_FLOAT16;
+#endif
+        case DTypeEnum::Float32:
+            return CNRT_FLOAT32;
+        case DTypeEnum::QuantizedS8:
+            return CNRT_QUANT8;
+        case DTypeEnum::Int32:
+            return CNRT_INT32;
+        default:
+            mgb_throw(MegBrainError,
+                      "megbrain data type %s is not supported by cnrt.",
+                      data_type.name());
+    }
+}
+};  // namespace
+
+/* ====================== CambriconRuntimeOpr ==================== */
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(CambriconRuntimeOpr);
+CambriconRuntimeOpr::CambriconRuntimeOpr(SharedBuffer buf, std::string symbol,
+                                         const VarNodeArray& inputs,
+                                         bool tensor_dim_mutable,
+                                         const OperatorNodeConfig& config)
+        : Super(inputs[0]->owner_graph(), config, "cambricon_runtime", inputs),
+          m_buffer{std::move(buf)},
+          m_symbol{std::move(symbol)},
+          m_model{nullptr},
+          m_function{nullptr},
+          m_context{nullptr},
+          m_tensor_dim_mutable{tensor_dim_mutable} {
+    mgb_assert(inputs[0]->comp_node().device_type() ==
+                       CompNode::DeviceType::CAMBRICON,
+               "CambriconRuntimeOpr can only be used on cambricon comp node; "
+               "got %s",
+               inputs[0]->comp_node().to_string().c_str());
+
+    for (auto i : inputs) {
+        add_input({i});
+    }
+    if (m_model == nullptr) {
+        m_model = {new cnrtModel_t(), cnrt_intl::ModelUnloader()};
+        MGB_CNRT_CHECK(cnrtLoadModelFromMem(
+                m_model.get(),
+                reinterpret_cast<char*>(const_cast<void*>(m_buffer.data()))));
+    }
+    if (m_function == nullptr) {
+        m_function = {new cnrtFunction_t(), cnrt_intl::FunctionDeleter()};
+        MGB_CNRT_CHECK(cnrtCreateFunction(m_function.get()));
+        MGB_CNRT_CHECK(cnrtExtractFunction(m_function.get(), *m_model,
+                                           m_symbol.c_str()));
+    }
+    int nr_inputs = 0;
+    int nr_outputs = 0;
+    int64_t* inputs_size = nullptr;
+    int64_t* outputs_size = nullptr;
+    MGB_CNRT_CHECK(cnrtGetInputDataSize(&inputs_size, &nr_inputs, *m_function));
+    mgb_assert(static_cast<size_t>(nr_inputs) == inputs.size(),
+               "inputs size mismatch: expect=%d, got=%zu", nr_inputs,
+               inputs.size());
+    MGB_CNRT_CHECK(
+            cnrtGetOutputDataSize(&outputs_size, &nr_outputs, *m_function));
+    if (nr_outputs == 1) {
+        add_output(None);
+    } else {
+        for (int i = 0; i < nr_outputs; ++i) {
+            add_output(ssprintf("o%d", i));
+        }
+    }
+    add_equivalence_component<mgb::ScalarHash<const void*>>(m_buffer.data());
+};
+
+void CambriconRuntimeOpr::scn_do_execute() {
+    mgb_assert(m_function != nullptr);
+    auto&& cnrt_env =
+            CompNodeEnv::from_comp_node(input(0)->comp_node()).cnrt_env();
+    cnrt_env.activate();
+    if (m_context == nullptr) {
+        m_context = {new cnrtRuntimeContext_t(),
+                     cnrt_intl::RuntimeContextDeleter()};
+        MGB_CNRT_CHECK(cnrtCreateRuntimeContext(m_context.get(), *m_function,
+                                                nullptr));
+        int dev_id = cnrt_env.device;
+        MGB_CNRT_CHECK(cnrtSetRuntimeContextDeviceId(*m_context, dev_id));
+        MGB_CNRT_CHECK(cnrtInitRuntimeContext(*m_context, nullptr));
+    }
+    size_t nr_inputs = input().size(), nr_outputs = output().size();
+    SmallVector<void*> params(nr_inputs + nr_outputs);
+    SmallVector<cnrtParamDesc_t> param_descs(nr_inputs + nr_outputs);
+    for (size_t i = 0; i < nr_inputs; ++i) {
+        params[i] = input(i)->dev_tensor().raw_ptr();
+        MGB_CNRT_CHECK(cnrtCreateParamDesc(&param_descs[i]));
+        MGB_CNRT_CHECK(cnrtSetDataTypeToParamDesc(
+                param_descs[i], mgb_dtype_to_cnrt_dtype(input(i)->dtype())));
+        auto dims = mgb_shape_to_cnrt_shape(input(i)->shape());
+        MGB_CNRT_CHECK(cnrtSetShapeToParamDesc(param_descs[i], dims.data(),
+                                               static_cast<int>(dims.size())));
+    }
+    for (size_t i = 0; i < nr_outputs; ++i) {
+        params[nr_inputs + i] = output(i)->dev_tensor().raw_ptr();
+        MGB_CNRT_CHECK(cnrtCreateParamDesc(&param_descs[nr_inputs + i]));
+        MGB_CNRT_CHECK(cnrtSetDataTypeToParamDesc(
+                param_descs[nr_inputs + i],
+                mgb_dtype_to_cnrt_dtype(output(i)->dtype())));
+        auto dims = mgb_shape_to_cnrt_shape(output(i)->shape());
+        MGB_CNRT_CHECK(cnrtSetShapeToParamDesc(param_descs[nr_inputs + i],
+                                               dims.data(),
+                                               static_cast<int>(dims.size())));
+    }
+    MGB_CNRT_CHECK(cnrtInvokeRuntimeContext_V2(*m_context, param_descs.data(),
+                                               params.data(), cnrt_env.queue,
+                                               nullptr));
+    for (auto& param : param_descs) {
+        MGB_CNRT_CHECK(cnrtDestroyParamDesc(param));
+    }
+}
+
+void CambriconRuntimeOpr::get_output_var_shape(
+        const TensorShapeArray& inp_shape, TensorShapeArray& out_shape) const {
+    mgb_assert(m_function != nullptr);
+    mgb_assert(input().size() == inp_shape.size());
+    if (m_tensor_dim_mutable) {
+        cnrtParamDescArray_t input_descs, output_descs;
+        int inp_param_num = input().size();
+        int out_param_num = output().size();
+        MGB_CNRT_CHECK(cnrtCreateParamDescArray(&input_descs, inp_param_num));
+        MGB_CNRT_CHECK(cnrtCreateParamDescArray(&output_descs, out_param_num));
+        for (int i = 0; i < inp_param_num; ++i) {
+            MGB_CNRT_CHECK(cnrtSetDataTypeToParamDesc(
+                    input_descs[i],
+                    mgb_dtype_to_cnrt_dtype(input(i)->dtype())));
+            auto dims = mgb_shape_to_cnrt_shape(inp_shape[i]);
+            MGB_CNRT_CHECK(
+                    cnrtSetShapeToParamDesc(input_descs[i], dims.data(),
+                                            static_cast<int>(dims.size())));
+        }
+        MGB_CNRT_CHECK(cnrtInferFunctionOutputShape(*m_function, inp_param_num,
+                                                    input_descs, out_param_num,
+                                                    output_descs));
+        for (int i = 0; i < out_param_num; ++i) {
+            int* dims = nullptr;
+            int dim_num = 0;
+            MGB_CNRT_CHECK(cnrtGetShapeFromParamDesc(output_descs[i], &dims,
+                                                     &dim_num));
+            out_shape[i] = cnrt_shape_to_mgb_shape(dims, dim_num);
+        }
+        MGB_CNRT_CHECK(cnrtDestroyParamDescArray(input_descs, inp_param_num));
+        MGB_CNRT_CHECK(cnrtDestroyParamDescArray(output_descs, out_param_num));
+    } else {
+        //! check input shape match
+        for (size_t i = 0; i < inp_shape.size(); ++i) {
+            int* dim_values = nullptr;
+            int dim_num = 0;
+            MGB_CNRT_CHECK(cnrtGetInputDataShape(
+                    &dim_values, &dim_num, static_cast<int>(i), *m_function));
+            auto shp_in_func = cnrt_shape_to_mgb_shape(dim_values, dim_num);
+            auto inpshp = inp_shape[i];
+            MGB_MARK_USED_VAR(shp_in_func);
+            mgb_assert(
+                    inpshp.eq_shape(shp_in_func),
+                    "input shape(%s) mismatch with that(%s) in cnrtFunction_t.",
+                    inpshp.to_string().c_str(),
+                    shp_in_func.to_string().c_str());
+        }
+        //! remarks: cnrt does not provide interface to let user manage
+        //! workspace
+        MGB_MARK_USED_VAR(mgb_dtype_to_cnrt_dtype);
+        for (size_t i = 0; i < out_shape.size(); ++i) {
+            int* dim_values = nullptr;
+            int dim_num = 0;
+            MGB_CNRT_CHECK(cnrtGetOutputDataShape(
+                    &dim_values, &dim_num, static_cast<int>(i), *m_function));
+            out_shape[i] = cnrt_shape_to_mgb_shape(dim_values, dim_num);
+        }
+    }
+}
+
+void CambriconRuntimeOpr::add_input_layout_constraint() {
+    //! default contiguous
+    for (auto i : input()) {
+        i->add_layout_constraint_contiguous();
+    }
+}
+
+void CambriconRuntimeOpr::init_output_dtype() {
+    cnrtDataType_t* inp_dtype_array = nullptr;
+    int inp_num;
+    MGB_CNRT_CHECK(
+            cnrtGetInputDataType(&inp_dtype_array, &inp_num, *m_function));
+    for (size_t i = 0; i < input().size(); ++i) {
+        auto dt_cnrt = cnrt_dtype_to_mgb_dtype(inp_dtype_array[i]);
+        auto dt_inp = input(i)->dtype();
+        MGB_MARK_USED_VAR(dt_cnrt);
+        MGB_MARK_USED_VAR(dt_inp);
+        mgb_assert(dt_cnrt.valid() && dt_inp.valid() &&
+                           dt_cnrt.enumv() == dt_inp.enumv(),
+                   "Input %zu's data type mismatch with that in "
+                   "cnrtFunction_t: expected %s, got %s",
+                   i, dt_cnrt.name(), dt_inp.name());
+    }
+    cnrtDataType_t* out_dtype_array = nullptr;
+    int out_num;
+    MGB_CNRT_CHECK(
+            cnrtGetOutputDataType(&out_dtype_array, &out_num, *m_function));
+    for (size_t i = 0; i < output().size(); ++i) {
+        auto dt_cnrt = cnrt_dtype_to_mgb_dtype(out_dtype_array[i]);
+        mgb_assert(dt_cnrt.valid(),
+                   "output dtype checking failed: invalid dtype returned.");
+        if (dt_cnrt.enumv() == DTypeEnum::QuantizedS8) {
+            mgb_assert(output(i)->dtype().valid(),
+                       "user should specify scale of output tensor of "
+                       "CambriconRuntimeOpr.");
+        }
+        if (!output(i)->dtype().valid())
+            output(i)->dtype(dt_cnrt);
+    }
+}
+
+SymbolVarArray CambriconRuntimeOpr::make(SharedBuffer buf, std::string symbol,
+                                         const SymbolVarArray& src,
+                                         bool tensor_dim_mutable,
+                                         const OperatorNodeConfig& config) {
+    VarNodeArray var_node_array = cg::to_var_node_array(src);
+    auto cambricon_runtime_opr = std::make_unique<CambriconRuntimeOpr>(
+            std::move(buf), std::move(symbol), var_node_array,
+            tensor_dim_mutable, config);
+    auto ret = cg::to_symbol_var_array(
+            src[0].node()
+                    ->owner_graph()
+                    ->insert_opr(std::move(cambricon_runtime_opr))
+                    ->output());
+    return ret;
+}
+
+SymbolVarArray CambriconRuntimeOpr::make(const void* buf, size_t size,
+                                         std::string symbol,
+                                         const SymbolVarArray& src,
+                                         bool tensor_dim_mutable,
+                                         const OperatorNodeConfig& config) {
+    mgb_throw_if(!CompNode::get_device_count(CompNode::DeviceType::CAMBRICON),
+                 SystemError,
+                 "can not create CambriconRuntimeOpr when Cambricon is not "
+                 "available");
+    std::shared_ptr<uint8_t> shptr{new uint8_t[size],
+                                   [](uint8_t* p) { delete[] p; }};
+    memcpy(shptr.get(), buf, size);
+    SharedBuffer buffer{std::move(shptr), size};
+    return make(std::move(buffer), std::move(symbol), src, tensor_dim_mutable,
+                config);
+}
+
+#endif  // MGB_CAMBRICON
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/cambricon/impl/cambricon_runtime_opr.oprdecl

@@ -0,0 +1,20 @@
+decl_raw_opr(
+    'cambricon_runtime',
+    desc='create an operator that could load and run cnrt offline models',
+    inputs=[
+        Doc('data_bytes', 'serialized cnrt/cnml model'),
+        Doc('symbol', 'name of cnrt/cnml function', 'str'), 
+        Doc('inputs', 'input vars', 'list of :class:`.SymbolVar`'),
+        Doc('tensor_dim_mutable', 'whether tensor shape is mutable in cnrt/cnml model', 'bool'),
+    ],
+    body=[
+        'assert isinstance(data_bytes, bytes), '
+            '"data must be bytes; got {}".format(type(data_bytes))',
+        'output = _mgb._Opr.cambricon_runtime(data_bytes, symbol, inputs, tensor_dim_mutable, config)',
+        'cvt_result_kwargs["explode_single"] = False',
+    ],
+)
+
+# vim: ft=python
+
+

src/cambricon/impl/cambricon_runtime_opr.sereg.h

@@ -0,0 +1,71 @@
+/**
+ * \file src/cambricon/impl/cambricon_runtime_opr.sereg.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "megbrain/cambricon/cambricon_runtime_opr.h"
+#include "megbrain/serialization/sereg.h"
+
+namespace mgb {
+namespace serialization {
+
+template <>
+struct OprLoadDumpImpl<opr::CambriconRuntimeOpr, 0> {
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        auto&& opr = opr_.cast_final_safe<opr::CambriconRuntimeOpr>();
+        auto&& buf = opr.buffer();
+        ctx.dump_buf_with_len(buf.data(), buf.size());
+        auto&& symbol = opr.symbol();
+        ctx.dump_buf_with_len(symbol.data(), symbol.size());
+        bool tensor_dim_mutable = opr.is_tensor_dim_mutable();
+        ctx.dump_buf_with_len(&tensor_dim_mutable, sizeof(bool));
+    }
+    static cg::OperatorNodeBase* load(OprLoadContext& ctx,
+                                      const cg::VarNodeArray& inputs,
+                                      const OperatorNodeConfig& config) {
+        inputs.at(0)->comp_node().activate();
+        auto buf = ctx.load_shared_buf_with_len();
+        auto symbol = ctx.load_buf_with_len();
+        auto tensor_dim_mutable_storage = ctx.load_buf_with_len();
+        bool tensor_dim_mutable;
+        memcpy(&tensor_dim_mutable, tensor_dim_mutable_storage.data(),
+               sizeof(bool));
+        return opr::CambriconRuntimeOpr::make(std::move(buf), std::move(symbol),
+                                              cg::to_symbol_var_array(inputs),
+                                              tensor_dim_mutable, config)
+                .at(0)
+                .node()
+                ->owner_opr();
+    }
+};
+}  // namespace serialization
+
+namespace opr {
+cg::OperatorNodeBase* opr_shallow_copy_cambricon_runtime_opr(
+        const serialization::OprShallowCopyContext& ctx,
+        const cg::OperatorNodeBase& opr_, const VarNodeArray& inputs,
+        const OperatorNodeConfig& config) {
+    auto&& opr = opr_.cast_final_safe<CambriconRuntimeOpr>();
+    return CambriconRuntimeOpr::make(opr.buffer(), opr.symbol(),
+                                     cg::to_symbol_var_array(inputs),
+                                     opr.is_tensor_dim_mutable(), config)
+            .at(0)
+            .node()
+            ->owner_opr();
+}
+
+MGB_SEREG_OPR(CambriconRuntimeOpr, 0);
+MGB_REG_OPR_SHALLOW_COPY(CambriconRuntimeOpr,
+                         opr_shallow_copy_cambricon_runtime_opr);
+}  // namespace opr
+}  // namespace mgb
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+
+

src/cambricon/include/megbrain/cambricon/cambricon_runtime_opr.h

@@ -0,0 +1,100 @@
+/**
+ * \file src/cambricon/include/megbrain/cambricon/cambricon_runtime_opr.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#pragma once
+
+#include "megbrain/comp_node_env.h"
+#include "megbrain/graph.h"
+#include "megbrain/serialization/file.h"
+
+#if MGB_CAMBRICON
+
+namespace mgb {
+namespace opr {
+namespace cnrt_intl {
+struct ModelUnloader {
+    void operator()(cnrtModel_t* model) {
+        if (model != nullptr)
+            MGB_CNRT_CHECK(cnrtUnloadModel(*model));
+    }
+};
+struct FunctionDeleter {
+    void operator()(cnrtFunction_t* function) {
+        if (function != nullptr)
+            MGB_CNRT_CHECK(cnrtDestroyFunction(*function));
+    }
+};
+struct RuntimeContextDeleter {
+    void operator()(cnrtRuntimeContext_t* context) {
+        if (context != nullptr)
+            MGB_CNRT_CHECK(cnrtDestroyRuntimeContext(*context));
+    }
+};
+using CnrtModelUniquePtr = std::unique_ptr<cnrtModel_t, ModelUnloader>;
+using CnrtFunctionUniquePtr = std::unique_ptr<cnrtFunction_t, FunctionDeleter>;
+using CnrtRuntimeContextUniquePtr =
+        std::unique_ptr<cnrtRuntimeContext_t, RuntimeContextDeleter>;
+};  // namespace cnrt_intl
+
+MGB_DEFINE_OPR_CLASS(CambriconRuntimeOpr, cg::SingleCNOutshapePureByInshapeOprBase) // {
+public:
+    using CnrtModelUniquePtr = cnrt_intl::CnrtModelUniquePtr;
+    using CnrtFunctionUniquePtr = cnrt_intl::CnrtFunctionUniquePtr;
+    using CnrtRuntimeContextUniquePtr = cnrt_intl::CnrtRuntimeContextUniquePtr;
+    using SharedBuffer = mgb::serialization::SharedBuffer;
+    void scn_do_execute() override;
+    void get_output_var_shape(const TensorShapeArray& inp_shape,
+                              TensorShapeArray& out_shape) const override;
+    void add_input_layout_constraint() override;
+    void init_output_dtype() override;
+
+    CambriconRuntimeOpr(SharedBuffer buf, std::string symbol,
+                        const VarNodeArray& inputs, bool tensor_dim_mutable,
+                        const OperatorNodeConfig& config);
+
+    const SharedBuffer& buffer() const {
+        return m_buffer;
+    }
+
+    const std::string& symbol() const {
+        return m_symbol;
+    }
+
+    bool is_tensor_dim_mutable() const {
+        return m_tensor_dim_mutable;
+    }
+
+    static SymbolVarArray make(SharedBuffer buf, std::string symbol,
+                               const SymbolVarArray& src,
+                               bool tensor_dim_mutable = false,
+                               const OperatorNodeConfig& config = {});
+
+    static SymbolVarArray make(const void* buf, size_t size, std::string symbol,
+                               const SymbolVarArray& src,
+                               bool tensor_dim_mutable = false,
+                               const OperatorNodeConfig& config = {});
+
+private:
+    SharedBuffer m_buffer; 
+    std::string m_symbol;
+    CnrtModelUniquePtr m_model;
+    CnrtFunctionUniquePtr m_function;
+    CnrtRuntimeContextUniquePtr m_context;
+    bool m_tensor_dim_mutable;
+};
+
+}  // namespace opr
+}  // namespace mgb
+
+#endif  // MGB_CAMBRICON
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+

src/cambricon/test/cambricon_runtime_opr.cpp

@@ -0,0 +1,562 @@
+/**
+ * \file src/cambricon/test/cambricon_runtime_opr.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "megbrain/comp_node_env.h"
+#include "megbrain/opr/io.h"
+#include "megbrain/plugin/profiler.h"
+#include "megbrain/serialization/serializer.h"
+#include "megbrain/test/helper.h"
+
+#if MGB_CAMBRICON
+
+#include "megbrain/cambricon/cambricon_runtime_opr.h"
+
+using namespace mgb;
+
+namespace {
+class CnmlModelContext {
+public:
+    const CompNode& cn;
+    bool batch_size_changable;
+    cnmlModel_t model;
+    cnmlTensor_t conv_input_tensor, relu_output_tensor;
+    cnmlFusionOp_t fusion_op;
+    bool built;
+    CnmlModelContext(const CompNode& cn, bool batch_size_changable = false)
+            : cn{cn},
+              batch_size_changable{batch_size_changable},
+              built{false} {}
+    ~CnmlModelContext() {
+        MGB_CNML_CHECK(cnmlDestroyTensor(&conv_input_tensor));
+        MGB_CNML_CHECK(cnmlDestroyTensor(&relu_output_tensor));
+        MGB_CNML_CHECK(cnmlDestroyFusionOp(&fusion_op));
+        MGB_CNML_CHECK(cnmlDestroyModel(model));
+    }
+
+    void build() {
+        auto&& cnrt_env = CompNodeEnv::from_comp_node(cn).cnrt_env();
+        cnrt_env.activate();
+        constexpr int core_num = 4;
+        cnrtCoreVersion_t core_version = cnrt_env.device_info.core_version;
+
+        // prepare parameter for addpad and conv
+        constexpr int dim_num = 4;
+        const int ni = 16, ci = 64, hi = 32, wi = 32;
+        const int no = 16, co = 64, ho = 32, wo = 32;
+        const int kh = 3, kw = 3;
+        const int stride_h = 1, stride_w = 1, dilation = 1;
+        const int pad_h = 2, pad_w = 2;
+
+        // count tensor nums
+        int conv_filter_count = co * kh * kw * ci;
+        int conv_bias_count = 1 * 1 * 1 * co;
+
+        // prepare cpu origin data
+        std::vector<float> conv_filter_cpu_data(conv_filter_count);
+        std::vector<float> conv_bias_cpu_data(conv_bias_count);
+
+        // prepare input data for addpad
+        unsigned int seed = time(0);
+        // prepare filter data for conv
+        for (int index = 0; index < conv_filter_count; ++index) {
+            conv_filter_cpu_data[index] =
+                    ((rand_r(&seed) % 200 / 200.0) - 0.5) / 2;
+        }
+        // prepare bias data for conv
+        for (int index = 0; index < conv_bias_count; ++index) {
+            conv_bias_cpu_data[index] = rand_r(&seed) % 100 / 100.0;
+        }
+
+        // prepare cpu data to converts to mlu memory
+        std::vector<int16_t> conv_bias_cpu(conv_bias_count);
+
+        // converts data format for mlu computing
+        // converts conv bias data
+        MGB_CNRT_CHECK(cnrtCastDataType(conv_bias_cpu_data.data(), CNRT_FLOAT32,
+                                        conv_bias_cpu.data(), CNRT_FLOAT16,
+                                        conv_bias_count, nullptr));
+
+        // u should set value depending op the data or your own needs
+        int filter_position = -6;
+        float filter_scale = 1, filter_offset = 0;
+
+        // count tensor nums
+        int conv_input_shape[] = {ni, ci, hi, wi};
+        int conv_filter_shape[] = {co, ci, kh, kw};
+        int conv_bias_shape[] = {1, co, 1, 1};
+        int conv_output_shape[] = {no, co, ho, wo};
+        int relu_output_shape[] = {no, co, ho, wo};
+
+        // setup tensors
+        // setup conv input tensor
+        conv_input_tensor = nullptr;
+        MGB_CNML_CHECK(cnmlCreateTensor_V2(&conv_input_tensor, CNML_TENSOR));
+        MGB_CNML_CHECK(cnmlSetTensorShape_V2(conv_input_tensor, dim_num,
+                                             conv_input_shape, nullptr));
+        MGB_CNML_CHECK(
+                cnmlSetTensorDataType(conv_input_tensor, CNML_DATA_FLOAT16));
+
+        // setup conv filter tensor
+        cnmlTensor_t conv_filter_tensor = nullptr;
+        MGB_CNML_CHECK(cnmlCreateTensor_V2(&conv_filter_tensor, CNML_FILTER));
+        MGB_CNML_CHECK(cnmlSetTensorShape_V2(conv_filter_tensor, dim_num,
+                                             conv_filter_shape, nullptr));
+        MGB_CNML_CHECK(
+                cnmlSetTensorDataType(conv_filter_tensor, CNML_DATA_FLOAT32));
+
+        // setup conv bias tensor
+        cnmlTensor_t conv_bias_tensor = nullptr;
+        MGB_CNML_CHECK(cnmlCreateTensor_V2(&conv_bias_tensor, CNML_CONST));
+        MGB_CNML_CHECK(cnmlSetTensorShape_V2(conv_bias_tensor, dim_num,
+                                             conv_bias_shape, nullptr));
+        MGB_CNML_CHECK(
+                cnmlSetTensorDataType(conv_bias_tensor, CNML_DATA_FLOAT16));
+
+        // setup conv output tensor
+        cnmlTensor_t conv_output_tensor = nullptr;
+        MGB_CNML_CHECK(cnmlCreateTensor_V2(&conv_output_tensor, CNML_TENSOR));
+        MGB_CNML_CHECK(cnmlSetTensorShape_V2(conv_output_tensor, dim_num,
+                                             conv_output_shape, nullptr));
+        MGB_CNML_CHECK(
+                cnmlSetTensorDataType(conv_output_tensor, CNML_DATA_FLOAT16));
+
+        // setup relu output tensor
+        relu_output_tensor = nullptr;
+        MGB_CNML_CHECK(cnmlCreateTensor_V2(&relu_output_tensor, CNML_TENSOR));
+        MGB_CNML_CHECK(cnmlSetTensorShape_V2(relu_output_tensor, dim_num,
+                                             relu_output_shape, nullptr));
+        MGB_CNML_CHECK(
+                cnmlSetTensorDataType(relu_output_tensor, CNML_DATA_FLOAT16));
+
+        // bind filters and bias to cnml const tensor
+        MGB_CNML_CHECK(cnmlBindConstData_V2(
+                conv_filter_tensor, conv_filter_cpu_data.data(), false));
+        MGB_CNML_CHECK(cnmlBindConstData_V2(conv_bias_tensor,
+                                            conv_bias_cpu.data(), false));
+
+        // create conv param and conv op
+        cnmlBaseOp_t conv_op;
+        cnmlConvOpParam_t conv_param;
+        // create relu op
+        cnmlBaseOp_t relu_op;
+
+        // setup conv param
+        MGB_CNML_CHECK(cnmlCreateConvOpParam(&conv_param, stride_h, stride_w,
+                                             dilation, dilation, pad_h, pad_w));
+        // setup conv operation
+        MGB_CNML_CHECK(cnmlCreateConvOp(&conv_op, conv_param, conv_input_tensor,
+                                        conv_output_tensor, conv_filter_tensor,
+                                        conv_bias_tensor));
+
+        // u should set value depending op the data or your own needs
+        int input_position = -6;
+        float input_scale = 1, input_offset = 0;
+        // prepare input tensor quant param for conv op
+        cnmlQuantizedParam_t input_quant_param;
+        MGB_CNML_CHECK(cnmlCreateQuantizedParam(
+                &input_quant_param, input_position, input_scale, input_offset));
+        // setup conv op computing datatype
+        MGB_CNML_CHECK(cnmlSetOperationComputingDataType(
+                conv_op, conv_input_tensor, CNML_DATA_INT8, input_quant_param));
+
+        // prepare filter tensor quant param for conv op
+        cnmlQuantizedParam_t filter_compute_quant;
+        MGB_CNML_CHECK(cnmlCreateQuantizedParam(&filter_compute_quant,
+                                                filter_position, filter_scale,
+                                                filter_offset));
+        // setup conv op computing datatype
+        MGB_CNML_CHECK(cnmlSetOperationComputingDataType(
+                conv_op, conv_filter_tensor, CNML_DATA_INT8,
+                filter_compute_quant));
+
+        // setup conv op computing layout
+        MGB_CNML_CHECK(cnmlSetOperationComputingLayout(conv_op, CNML_NCHW));
+
+        // setup active op using relu fuction
+        MGB_CNML_CHECK(cnmlCreateActiveOp(&relu_op, CNML_ACTIVE_RELU,
+                                          conv_output_tensor,
+                                          relu_output_tensor));
+
+        // setup fusion op, fuse addpad op and conv op to fusion op
+        MGB_CNML_CHECK(cnmlCreateFusionOp(&fusion_op));
+        MGB_CNML_CHECK(cnmlFuseOp(conv_op, fusion_op));
+        MGB_CNML_CHECK(cnmlFuseOp(relu_op, fusion_op));
+
+        MGB_CNML_CHECK(cnmlSetTensorDimMutable(conv_input_tensor,
+                                               &batch_size_changable, 4));
+        MGB_CNML_CHECK(cnmlSetTensorDimMutable(relu_output_tensor,
+                                               &batch_size_changable, 4));
+
+        // setup the input and output of the fusion op
+        MGB_CNML_CHECK(cnmlAddFusionInput(fusion_op, conv_input_tensor));
+        MGB_CNML_CHECK(cnmlAddFusionOutput(fusion_op, relu_output_tensor));
+
+        // set operation corenum
+        MGB_CNML_CHECK(cnmlSetFusionOpCorenum(fusion_op, core_num));
+        // set operation coreversion
+        MGB_CNML_CHECK(cnmlSetFusionOpCoreVersion(
+                fusion_op, static_cast<cnmlCoreVersion_t>(core_version)));
+        // set batch size changable
+        MGB_CNML_CHECK(cnmlSetFusionOpBatchsizeChangable(fusion_op,
+                                                         batch_size_changable));
+        // compile fusion op
+        MGB_CNML_CHECK(cnmlCompileFusionOp_V2(fusion_op));
+
+        // delete tensors
+        MGB_CNML_CHECK(cnmlDestroyTensor(&conv_filter_tensor));
+        MGB_CNML_CHECK(cnmlDestroyTensor(&conv_bias_tensor));
+        MGB_CNML_CHECK(cnmlDestroyTensor(&conv_output_tensor));
+
+        // delete quant param
+        MGB_CNML_CHECK(cnmlDestroyQuantizedParam(&input_quant_param));
+
+        // destory filter compute quant-param
+        MGB_CNML_CHECK(cnmlDestroyQuantizedParam(&filter_compute_quant));
+
+        // delete conv params
+        MGB_CNML_CHECK(cnmlDestroyConvOpParam(&conv_param));
+
+        // delete base ops and fusion op
+        MGB_CNML_CHECK(cnmlDestroyBaseOp(&conv_op));
+        MGB_CNML_CHECK(cnmlDestroyBaseOp(&relu_op));
+        built = true;
+    }
+
+    SmallVector<uint8_t> get_serialized_model() {
+        if (!built)
+            build();
+        MGB_CNML_CHECK(cnmlCreateModel(&model, "mlp"));
+        MGB_CNML_CHECK(cnmlAddFusionOpToModel(model, fusion_op, "subnet0"));
+        std::string fname =
+                ssprintf("./output/CambriconRuntimeOprTest.%s.mlu",
+                         batch_size_changable ? "MutableBatchSize"
+                                              : "ImmutableBatchSize");
+        MGB_CNML_CHECK(cnmlSaveModel(model, fname.c_str()));
+        int len = 0;
+        MGB_CNRT_CHECK(cnrtGetModelSize(fname.c_str(), &len));
+        SmallVector<uint8_t> buf(len);
+        FILE* fstream = fopen(fname.c_str(), "rb");
+        if (fstream != nullptr) {
+            auto ret = fread(buf.data(), 1, len, fstream);
+            mgb_assert(static_cast<int>(ret) == len);
+        }
+        auto fstream_close = [](FILE* fp) { fclose(fp); };
+        std::unique_ptr<FILE, decltype(fstream_close)> fstream_holder{
+                fstream, fstream_close};
+        return std::move(buf);
+    }
+
+    void do_inference(void** input_mlu_ptrs, void** output_mlu_ptrs) {
+        if (!built)
+            build();
+        auto&& cnrt_env = CompNodeEnv::from_comp_node(cn).cnrt_env();
+        cnrt_env.activate();
+        auto&& queue = cnrt_env.queue;
+        cnrtNotifier_t start, end;
+        MGB_CNRT_CHECK(cnrtCreateNotifier(&start));
+        MGB_CNRT_CHECK(cnrtCreateNotifier(&end));
+        MGB_CNRT_CHECK(cnrtPlaceNotifier(start, queue));
+        MGB_CNML_CHECK(cnmlComputeFusionOpForward_V4(
+                fusion_op, &conv_input_tensor, input_mlu_ptrs, 1,
+                &relu_output_tensor, output_mlu_ptrs, 1, queue, nullptr));
+        MGB_CNRT_CHECK(cnrtPlaceNotifier(end, queue));
+        MGB_CNRT_CHECK(cnrtSyncQueue(queue));
+        float time = 0.f;
+        MGB_CNRT_CHECK(cnrtNotifierDuration(start, end, &time));
+        printf("inference time = %.2fs\n", time * 1e-3);
+        MGB_CNRT_CHECK(cnrtDestroyNotifier(&start));
+        MGB_CNRT_CHECK(cnrtDestroyNotifier(&end));
+    }
+};
+}  // namespace
+
+TEST(TestCambriconRuntimeOpr, Basic) {
+    REQUIRE_CAMBRICON_DEVICE(1);
+    auto cn = CompNode::load("cambricon0");
+    CnmlModelContext ctx{cn, false};
+
+    // prepare parameter for addpad and conv
+    const int ni = 16, ci = 64, hi = 32, wi = 32;
+    const int no = 16, co = 64, ho = 32, wo = 32;
+
+    // count tensor nums
+    int conv_input_count = ni * hi * wi * ci;
+    int relu_output_count = no * ho * wo * co;
+
+    // prepare cpu origin data
+    std::vector<float> conv_input_cpu_data(conv_input_count);
+    std::vector<float> relu_output_cpu_data(relu_output_count);
+
+    // prepare input data for addpad
+    unsigned int seed = time(0);
+    for (int index = 0; index < conv_input_count; ++index) {
+        conv_input_cpu_data[index] = ((rand_r(&seed) % 100 / 100.0) - 0.5) / 2;
+    }
+
+    // prepare cpu data to converts to mlu memory
+    std::vector<int16_t> conv_input_cpu(conv_input_count);
+    std::vector<int16_t> relu_output_cpu(relu_output_count);
+    MGB_CNRT_CHECK(cnrtCastDataType(conv_input_cpu_data.data(), CNRT_FLOAT32,
+                                    conv_input_cpu.data(), CNRT_FLOAT16,
+                                    conv_input_count, nullptr));
+
+    auto mlu_deleter = [](void* p) { MGB_CNRT_CHECK(cnrtFree(p)); };
+    void* input_mlu_ptr;
+    void* output_mlu_ptr;
+
+    // malloc mlu mem for fusion input and output
+    MGB_CNRT_CHECK(
+            cnrtMalloc(&input_mlu_ptr, conv_input_count * sizeof(int16_t)));
+    MGB_CNRT_CHECK(
+            cnrtMalloc(&output_mlu_ptr, relu_output_count * sizeof(int16_t)));
+    // memory copy cpu->mlu
+    MGB_CNRT_CHECK(cnrtMemcpy(input_mlu_ptr, conv_input_cpu.data(),
+                              conv_input_count * sizeof(int16_t),
+                              CNRT_MEM_TRANS_DIR_HOST2DEV));
+    std::unique_ptr<void, decltype(mlu_deleter)> input_holder{input_mlu_ptr,
+                                                              mlu_deleter};
+    std::unique_ptr<void, decltype(mlu_deleter)> output_holder{output_mlu_ptr,
+                                                               mlu_deleter};
+
+    ctx.do_inference(&input_mlu_ptr, &output_mlu_ptr);
+
+    // result memory copy cnml->cpu
+    // memory copy cpu->mlu
+    MGB_CNRT_CHECK(cnrtMemcpy(relu_output_cpu.data(), output_mlu_ptr,
+                              relu_output_count * sizeof(int16_t),
+                              CNRT_MEM_TRANS_DIR_DEV2HOST));
+    MGB_CNRT_CHECK(cnrtCastDataType(relu_output_cpu.data(), CNRT_FLOAT16,
+                                    relu_output_cpu_data.data(), CNRT_FLOAT32,
+                                    relu_output_count, nullptr));
+
+    auto buf = ctx.get_serialized_model();
+    std::shared_ptr<HostTensorND> input = std::make_shared<HostTensorND>(
+            cn, TensorLayout{{ni, ci, hi, wi}, dtype::Float16()});
+    memcpy(reinterpret_cast<void*>(input->ptr<dt_float16>()),
+           conv_input_cpu.data(), conv_input_count * sizeof(int16_t));
+    auto graph = ComputingGraph::make();
+    auto x = opr::Host2DeviceCopy::make(*graph, input);
+    auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(), "subnet0",
+                                            {x}, false)[0];
+    HostTensorND output(cn, {no, co, ho, wo}, dtype::Float16());
+    auto func = graph->compile({make_callback_copy(y, output)});
+    func->execute();
+    HostTensorND out_cnml(cn, {no, co, ho, wo}, dtype::Float32()),
+            out_mgb(cn, {no, co, ho, wo}, dtype::Float32());
+    memcpy(out_cnml.ptr<float>(), relu_output_cpu_data.data(),
+           relu_output_count * sizeof(float));
+    MGB_CNRT_CHECK(cnrtCastDataType(
+            reinterpret_cast<void*>(output.ptr<dt_float16>()), CNRT_FLOAT16,
+            out_mgb.ptr<float>(), CNRT_FLOAT32, relu_output_count, nullptr));
+    MGB_ASSERT_TENSOR_NEAR(out_cnml, out_mgb, 1e-4);
+}
+
+TEST(TestCambriconRuntimeOpr, BatchSizeChangable) {
+    REQUIRE_CAMBRICON_DEVICE(1);
+    auto cn = CompNode::load("cambricon0");
+    CnmlModelContext ctx{cn, true};
+
+    // prepare parameter for addpad and conv
+    size_t ni = 16, ci = 64, hi = 32, wi = 32;
+    size_t no = 16, co = 64, ho = 32, wo = 32;
+
+    // count tensor nums
+    int conv_input_count = ni * hi * wi * ci;
+    int relu_output_count = no * ho * wo * co;
+
+    // prepare cpu origin data
+    std::vector<float> conv_input_cpu_data(conv_input_count);
+    std::vector<float> relu_output_cpu_data(relu_output_count);
+
+    // prepare input data for addpad
+    unsigned int seed = time(0);
+    for (int index = 0; index < conv_input_count; ++index) {
+        conv_input_cpu_data[index] = ((rand_r(&seed) % 100 / 100.0) - 0.5) / 2;
+    }
+
+    // prepare cpu data to converts to mlu memory
+    std::vector<int16_t> conv_input_cpu(conv_input_count);
+    std::vector<int16_t> relu_output_cpu(relu_output_count);
+    MGB_CNRT_CHECK(cnrtCastDataType(conv_input_cpu_data.data(), CNRT_FLOAT32,
+                                    conv_input_cpu.data(), CNRT_FLOAT16,
+                                    conv_input_count, nullptr));
+
+    auto mlu_deleter = [](void* p) { MGB_CNRT_CHECK(cnrtFree(p)); };
+    void* input_mlu_ptr;
+    void* output_mlu_ptr;
+
+    // malloc mlu mem for fusion input and output
+    MGB_CNRT_CHECK(
+            cnrtMalloc(&input_mlu_ptr, conv_input_count * sizeof(int16_t)));
+    MGB_CNRT_CHECK(
+            cnrtMalloc(&output_mlu_ptr, relu_output_count * sizeof(int16_t)));
+    // memory copy cpu->mlu
+    MGB_CNRT_CHECK(cnrtMemcpy(input_mlu_ptr, conv_input_cpu.data(),
+                              conv_input_count * sizeof(int16_t),
+                              CNRT_MEM_TRANS_DIR_HOST2DEV));
+    std::unique_ptr<void, decltype(mlu_deleter)> input_holder{input_mlu_ptr,
+                                                              mlu_deleter};
+    std::unique_ptr<void, decltype(mlu_deleter)> output_holder{output_mlu_ptr,
+                                                               mlu_deleter};
+
+    ctx.do_inference(&input_mlu_ptr, &output_mlu_ptr);
+
+    // result memory copy cnml->cpu
+    // memory copy cpu->mlu
+    MGB_CNRT_CHECK(cnrtMemcpy(relu_output_cpu.data(), output_mlu_ptr,
+                              relu_output_count * sizeof(int16_t),
+                              CNRT_MEM_TRANS_DIR_DEV2HOST));
+    MGB_CNRT_CHECK(cnrtCastDataType(relu_output_cpu.data(), CNRT_FLOAT16,
+                                    relu_output_cpu_data.data(), CNRT_FLOAT32,
+                                    relu_output_count, nullptr));
+    // cnml inference finished
+    {
+        // change batch size
+        ni = 32, no = 32;
+
+        auto buf = ctx.get_serialized_model();
+        std::shared_ptr<HostTensorND> input = std::make_shared<HostTensorND>(
+                cn, TensorLayout{{ni, ci, hi, wi}, dtype::Float16()});
+        memcpy(reinterpret_cast<void*>(input->ptr<dt_float16>()),
+               conv_input_cpu.data(), conv_input_count * sizeof(int16_t));
+        memcpy(reinterpret_cast<void*>(input->ptr<dt_float16>() +
+                                       conv_input_count),
+               conv_input_cpu.data(), conv_input_count * sizeof(int16_t));
+        auto graph = ComputingGraph::make();
+        auto x = opr::Host2DeviceCopy::make(*graph, input);
+        auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(),
+                                                "subnet0", {x}, true)[0];
+        HostTensorND output(cn, {no, co, ho, wo}, dtype::Float16());
+        auto func = graph->compile({make_callback_copy(y, output)});
+        func->execute();
+        HostTensorND out_cnml(cn, {no, co, ho, wo}, dtype::Float32()),
+                out_mgb(cn, {no, co, ho, wo}, dtype::Float32());
+        memcpy(out_cnml.ptr<float>(), relu_output_cpu_data.data(),
+               relu_output_count * sizeof(float));
+        memcpy(out_cnml.ptr<float>() + relu_output_count,
+               relu_output_cpu_data.data(), relu_output_count * sizeof(float));
+        MGB_CNRT_CHECK(cnrtCastDataType(
+                reinterpret_cast<void*>(output.ptr<dt_float16>()), CNRT_FLOAT16,
+                out_mgb.ptr<float>(), CNRT_FLOAT32, 2 * relu_output_count,
+                nullptr));
+        MGB_ASSERT_TENSOR_NEAR(out_cnml, out_mgb, 1e-4);
+    }
+    {
+        // change batch size
+        ni = 1, no = 1;
+        conv_input_count = ni * hi * wi * ci;
+        relu_output_count = no * ho * wo * co;
+
+        auto buf = ctx.get_serialized_model();
+        std::shared_ptr<HostTensorND> input = std::make_shared<HostTensorND>(
+                cn, TensorLayout{{ni, ci, hi, wi}, dtype::Float16()});
+        memcpy(reinterpret_cast<void*>(input->ptr<dt_float16>()),
+               conv_input_cpu.data(), conv_input_count * sizeof(int16_t));
+        auto graph = ComputingGraph::make();
+        auto x = opr::Host2DeviceCopy::make(*graph, input);
+        auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(),
+                                                "subnet0", {x}, true)[0];
+        HostTensorND output(cn, {no, co, ho, wo}, dtype::Float16());
+        auto func = graph->compile({make_callback_copy(y, output)});
+        func->execute();
+        HostTensorND out_cnml(cn, {no, co, ho, wo}, dtype::Float32()),
+                out_mgb(cn, {no, co, ho, wo}, dtype::Float32());
+        memcpy(out_cnml.ptr<float>(), relu_output_cpu_data.data(),
+               relu_output_count * sizeof(float));
+        MGB_CNRT_CHECK(cnrtCastDataType(
+                reinterpret_cast<void*>(output.ptr<dt_float16>()), CNRT_FLOAT16,
+                out_mgb.ptr<float>(), CNRT_FLOAT32, relu_output_count,
+                nullptr));
+        MGB_ASSERT_TENSOR_NEAR(out_cnml, out_mgb, 1e-4);
+    }
+}
+
+TEST(TestCambriconRuntimeOpr, Serialization) {
+    using namespace serialization;
+    REQUIRE_CAMBRICON_DEVICE(1);
+    auto cn = CompNode::load("cambricon0");
+    CnmlModelContext ctx{cn, true};
+    auto buf = ctx.get_serialized_model();
+
+    // prepare parameter for addpad and conv
+    const int ni = 1, ci = 64, hi = 32, wi = 32;
+    std::shared_ptr<HostTensorND> input = std::make_shared<HostTensorND>(
+            cn, TensorLayout{{ni, ci, hi, wi}, dtype::Float16()});
+    auto graph = ComputingGraph::make();
+    auto x = opr::Host2DeviceCopy::make(*graph, input);
+    auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(), "subnet0",
+                                            {x}, true)[0];
+    auto fname = output_file("CambriconRuntimeOprTest");
+    auto dump = [&]() {
+        auto dumper = GraphDumper::make(OutputFile::make_fs(fname.c_str()));
+        auto rst = dumper->dump({y});
+        ASSERT_EQ(rst.outputs.size(), 1u);
+    };
+    auto load = [&]() {
+        auto loader = GraphLoader::make(InputFile::make_fs(fname.c_str()));
+        auto rst = loader->load();
+        ASSERT_EQ(rst.output_var_list.size(), 1u);
+    };
+    dump();
+    load();
+}
+
+// TODO: this test will be improved later due to peer copy for cambricon is not
+// correct
+TEST(TestCambriconRuntimeOpr, MultipleDevice) {
+    REQUIRE_CAMBRICON_DEVICE(2);
+    auto cn0 = CompNode::load("cambricon0");
+    auto cn1 = CompNode::load("cambricon1");
+    CnmlModelContext ctx{cn0, true};
+    auto buf = ctx.get_serialized_model();
+
+    const int ni = 8, ci = 64, hi = 32, wi = 32;
+
+    auto graph = ComputingGraph::make();
+    auto xv = std::make_shared<DeviceTensorND>(cn0, TensorShape{ni, ci, hi, wi},
+                                               dtype::Float16());
+    auto x = opr::SharedDeviceTensor::make(*graph, xv),
+         x1 = opr::Copy::make(x, cn1);
+    auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(), "subnet0",
+                                            {x}, true)[0],
+         y1 = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(), "subnet0",
+                                             {x1}, true)[0];
+    HostTensorND host_y, host_y1;
+    auto func = graph->compile(
+            {make_callback_copy(y, host_y), make_callback_copy(y1, host_y1)});
+    func->execute();
+}
+
+TEST(TestCambriconRuntimeOpr, Profiling) {
+    REQUIRE_CAMBRICON_DEVICE(1);
+    auto cn = CompNode::load("cambricon0");
+    CnmlModelContext ctx{cn, true};
+    auto buf = ctx.get_serialized_model();
+    const int ni = 8, ci = 64, hi = 32, wi = 32;
+
+    HostTensorGenerator<dtype::Float16, RandomDistribution::GAUSSIAN> gen(
+            dt_float16(0.f), dt_float16(1.f));
+    auto input = gen({ni, ci, hi, wi}, cn);
+    auto graph = ComputingGraph::make();
+    GraphProfiler profiler{graph.get()};
+    auto x = opr::Host2DeviceCopy::make(*graph, input);
+    auto y = opr::CambriconRuntimeOpr::make(buf.data(), buf.size(), "subnet0",
+                                            {x}, true)[0];
+    HostTensorND output;
+    graph->options().var_sanity_check_first_run = false;
+    auto func = graph->compile({make_callback_copy(y, output)});
+    func->execute();
+    profiler.to_json_full(func.get())
+            ->writeto_fpath(output_file("cambricon_runtime_opr_profile.json"));
+}
+#endif
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/core/impl/comp_node/atlas/comp_node.cpp

@@ -0,0 +1,508 @@
+/**
+ * \file src/core/impl/comp_node/atlas/comp_node.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "./comp_node.h"
+#include "megbrain/comp_node_env.h"
+
+#include <memory>
+#include <string>
+
+using namespace mgb;
+
+#if MGB_ATLAS
+
+#include "megbrain/common.h"
+#include "megbrain/comp_node/alloc.h"
+#include "megbrain/utils//timer.h"
+#include "megcore_atlas.h"
+
+#include <cctype>
+#include <cstdio>
+
+#include <acl/acl.h>
+#include <limits>
+
+
+using AtlasCompNodeImpl = AtlasCompNode::CompNodeImpl;
+
+/* ===================== AtlasCompNodeImpl  ===================== */
+class AtlasCompNode::CompNodeImpl final : public CompNode::Impl {
+    MGB_DYN_TYPE_OBJ_FINAL_DECL;
+
+    friend class EventImpl;
+    friend class AtlasCompNode;
+
+    struct DeviceInfo;
+    struct StaticData;
+    static StaticData* sd;
+    static Spinlock sd_mtx;
+
+    //! set to true when m_locator is assigned; set to false if async init
+    //! failed
+    bool m_initialized = false;
+    Locator m_locator, m_locator_logical;
+    DeviceInfo* m_device_info;
+
+    std::unique_ptr<Event> m_sync_event;
+    Spinlock m_sync_event_mtx;
+
+    void activate() { m_env.atlas_env().activate(); }
+
+    void init(const Locator& locator, const Locator& locator_logical);
+    void fini();
+
+    //! return whether global finalized, and print warning in such case
+    static inline bool check_global_finalized();
+
+    //! enable peer copy from dev0 to dev1
+    static void enable_peer_access(int dev0, int dev1);
+
+    static void static_free_device(ImplBase* self, void* ptr) {
+        static_cast<CompNodeImpl*>(self)->free_device(ptr);
+    }
+
+    static void static_free_host(ImplBase* self, void* ptr) {
+        static_cast<CompNodeImpl*>(self)->free_host(ptr);
+    }
+
+public:
+    CompNodeImpl() : Impl(static_free_device, static_free_host) {}
+
+    void* alloc_device(size_t size) override {
+        activate();
+        void* addr;
+        MGB_ATLAS_CHECK(aclrtMalloc(&addr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        return addr;
+    }
+
+    void free_device(void* ptr) {
+        if (check_global_finalized())
+            return;
+
+        activate();
+
+        MGB_ATLAS_CHECK(aclrtFree(ptr));
+    }
+
+    void* alloc_host(size_t size) override {
+        void* ptr;
+        MGB_ATLAS_CHECK(aclrtMallocHost(&ptr, size));
+        return ptr;
+    }
+
+    void free_host(void* ptr) { MGB_ATLAS_CHECK(aclrtFreeHost(ptr)); }
+
+    void copy_to_host(void* host_ptr, const void* device_ptr,
+                      size_t size) override {
+        activate();
+        MGB_ATLAS_CHECK(aclrtMemcpyAsync(host_ptr, size, device_ptr, size,
+                                         ACL_MEMCPY_DEVICE_TO_HOST,
+                                         m_env.atlas_env().stream));
+    }
+
+    void copy_to_device(void* device_ptr, const void* host_ptr,
+                        size_t size) override {
+        activate();
+        MGB_ATLAS_CHECK(aclrtMemcpy(device_ptr, size, host_ptr, size,
+                                    ACL_MEMCPY_HOST_TO_DEVICE));
+    }
+
+    void peer_copy_to(Impl* dest_impl, void* dest, const void* src,
+                      size_t size) override;
+
+    size_t get_mem_addr_alignment() override {
+        return m_env.property().mem_alignment;
+    }
+
+    std::unique_ptr<Event> create_event(size_t flags) override;
+
+    void sync() override;
+
+    MemNode mem_node() override;
+
+    size_t get_mem_padding() override { return 32; }
+
+    std::pair<size_t, size_t> get_mem_status_bytes() override {
+        return {std::numeric_limits<size_t>::max(),
+                std::numeric_limits<size_t>::max()};
+    }
+
+    Locator locator() override { return m_locator; }
+
+    Locator locator_logical() override { return m_locator_logical; }
+};
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(AtlasCompNode::CompNodeImpl);
+
+struct AtlasCompNodeImpl::DeviceInfo {
+    int dev_num = -1;
+
+    void init(const CompNodeEnv& env) {
+        auto&& atlas_env = env.atlas_env();
+        atlas_env.activate();
+        dev_num = atlas_env.device;
+    }
+
+    void fini() {
+        MGB_ATLAS_CHECK(aclrtResetDevice(dev_num));
+    }
+};
+
+
+struct AtlasCompNodeImpl::StaticData {
+    static constexpr int MAX_NR_COMP_NODE = 1024, MAX_NR_DEVICE = 64;
+
+    std::recursive_mutex mtx;
+
+    AtlasCompNode::CompNodeImpl node[MAX_NR_COMP_NODE];
+    DeviceInfo dev_info[MAX_NR_DEVICE];
+    int nr_node = 0,          //!< number of loaded node[]
+            nr_dev_used = 0;  //!< number of used dev_info[]
+
+    StaticData() {}
+
+    ~StaticData() {
+        for (int i = 0; i < nr_node; ++i)
+            node[i].fini();
+        for (int i = 0; i < nr_dev_used; ++i)
+            dev_info[i].fini();
+    }
+};
+AtlasCompNodeImpl::StaticData* AtlasCompNodeImpl::sd = nullptr;
+Spinlock AtlasCompNodeImpl::sd_mtx;
+
+void AtlasCompNodeImpl::init(const Locator& locator,
+                             const Locator& locator_logical) {
+    m_locator = locator;
+    m_locator_logical = locator_logical;
+    m_initialized = true;
+
+    CompNodeEnv::AtlasEnv atlas_env;
+    atlas_env.device = locator.device;
+    m_env.init_atlas(make_comp_node_from_impl(this), atlas_env);
+
+    DeviceInfo* dev_info = nullptr;
+    for (int i = 0; i < sd->nr_dev_used; ++i) {
+        if (sd->dev_info[i].dev_num == locator.device) {
+            dev_info = &sd->dev_info[i];
+            break;
+        }
+    }
+
+    if (!dev_info) {
+        dev_info = &sd->dev_info[sd->nr_dev_used];
+        dev_info->init(m_env);
+        // note: add nr_dev_used only after init succeeds
+        ++sd->nr_dev_used;
+    }
+    m_device_info = dev_info;
+
+}
+
+void AtlasCompNodeImpl::fini() {
+    if (!m_initialized)
+        return;
+
+    m_sync_event.reset();
+    m_env.fini();
+    m_initialized = false;
+    m_device_info = nullptr;
+}
+
+void AtlasCompNodeImpl::peer_copy_to(Impl* dest_impl, void* dest,
+                                     const void* src, size_t size) {
+    if (dest_impl->same_type<AtlasCompNodeImpl>()) {
+        auto&& dst_env =
+                static_cast<AtlasCompNodeImpl*>(dest_impl)->m_env.atlas_env();
+        auto&& src_env = m_env.atlas_env();
+        activate();
+        if (dst_env.device == src_env.device) {
+            MGB_ATLAS_CHECK(aclrtMemcpyAsync(dest, size, src, size,
+                                             ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                             dst_env.stream));
+        } else {
+            mgb_throw(MegBrainError,
+                      "Atlas does not support peer copy between differents "
+                      "device.");
+        }
+        return;
+
+    }
+    mgb_assert(dest_impl->env().property().type == DeviceType::CPU,
+               "cuda peer_copy_to only implemented for CPU");
+    auto copy = [this, dest, src, size]() {
+        auto stream = m_env.atlas_env().stream;
+        m_env.atlas_env().activate();
+        MGB_ATLAS_CHECK(aclrtMemcpyAsync(dest, size, src, size,
+                                         ACL_MEMCPY_DEVICE_TO_HOST,
+                                         m_env.atlas_env().stream));
+        MGB_ATLAS_CHECK(aclrtSynchronizeStream(stream));
+    };
+    dest_impl->env().cpu_env().dispatch(copy);
+
+}
+
+MemNode AtlasCompNodeImpl::mem_node() {
+    // m_device_info would be null before async init finishes; so we just return
+    // a private pointer related to device number here
+    return MemNode{sd->dev_info + m_locator.device};
+}
+
+void AtlasCompNodeImpl::sync() {
+    activate();
+
+    Event* event;
+    {
+        MGB_LOCK_GUARD(m_sync_event_mtx);
+        if (!m_sync_event)
+            m_sync_event = create_event(0);
+        event = m_sync_event.get();
+    }
+    event->record();
+    event->host_wait();
+}
+
+void AtlasCompNodeImpl::enable_peer_access(int dev0, int dev1) {
+    MGB_MARK_USED_VAR(dev0);
+    MGB_MARK_USED_VAR(dev1);
+    mgb_throw(MegBrainError,
+              "Atlas does not support peer copy between differents "
+              "device.");
+}
+
+bool AtlasCompNodeImpl::check_global_finalized() {
+    if (!sd) {
+        static std::atomic_flag warn_printed = ATOMIC_FLAG_INIT;
+        if (!warn_printed.test_and_set()) {
+            mgb_log_debug(
+                    "atlas comp node method called after global finalize");
+        }
+        return true;
+    }
+    return false;
+}
+
+/* ===================== EventImpl  ===================== */
+/**
+ * \warning Current we just use cpu timer to do record, later when the api of
+ * ddk is ready, we change to normal event.
+ */
+class AtlasCompNode::EventImpl final : public EventImplHelper {
+    AtlasCompNodeImpl* const m_comp_node_impl;
+    aclrtEvent m_atlas_event;
+    bool m_init_finished = false;
+
+    void do_record() override {
+        m_comp_node_impl->activate();
+        auto &&env = m_comp_node_impl->m_env.atlas_env();
+        MGB_ATLAS_CHECK(aclrtRecordEvent(m_atlas_event, env.stream));
+    }
+
+    bool do_finished() override {
+        m_comp_node_impl->activate();
+        aclrtEventStatus status;
+        MGB_ATLAS_CHECK(aclrtQueryEvent(m_atlas_event, &status));
+        if (status == ACL_EVENT_STATUS_COMPLETE)
+            return true;
+        if (status == ACL_EVENT_STATUS_NOT_READY)
+            return false;
+        mgb_throw(AtlasError, "invalid event status: %d", int(status));
+    }
+
+    void host_wait_cv() override {
+        MGB_ATLAS_CHECK(aclrtSynchronizeEvent(m_atlas_event));
+    }
+
+    double do_elapsed_time_until(EventImplHelper& end) override {
+        m_comp_node_impl->activate();
+        float ret = 0.0;
+        MGB_ATLAS_CHECK(aclrtEventElapsedTime(&ret, m_atlas_event,
+                    static_cast<EventImpl&>(end).m_atlas_event));
+        return static_cast<double>(ret) * 1e-3;
+
+    }
+
+    void do_device_wait_by(Impl* cn_impl) override;
+
+public:
+    EventImpl(AtlasCompNodeImpl* comp_node_impl, size_t create_flags)
+            : EventImplHelper(comp_node_impl, create_flags),
+              m_comp_node_impl{comp_node_impl} {
+        m_comp_node_impl->activate();
+        MGB_ATLAS_CHECK(aclrtCreateEvent(&m_atlas_event));
+        m_init_finished = true;
+    }
+    ~EventImpl() {
+        if (m_init_finished) {
+            MGB_TRY { MGB_ATLAS_CHECK(aclrtDestroyEvent(m_atlas_event)); }
+            MGB_CATCH(MegBrainError & exc, {
+                mgb_log_error("failed to destroy cuda event: %s", exc.what());
+            })
+        }
+    }
+};
+
+std::unique_ptr<CompNode::Event> AtlasCompNodeImpl::create_event(size_t flags) {
+    return std::make_unique<EventImpl>(this, flags);
+}
+
+void AtlasCompNode::EventImpl::do_device_wait_by(Impl* cn_impl) {
+    if (cn_impl->dyn_typeinfo() == AtlasCompNodeImpl::typeinfo()) {
+        auto imp = static_cast<AtlasCompNodeImpl*>(cn_impl);
+        auto stream = imp->m_env.atlas_env().stream;
+        imp->activate();
+        MGB_ATLAS_CHECK(aclrtStreamWaitEvent(stream, m_atlas_event));
+        return;
+    }
+    if (cn_impl->env().property().type == DeviceType::CPU) {
+        auto waiter = [this]() {
+            MGB_ATLAS_CHECK(aclrtSynchronizeEvent(m_atlas_event));
+        };
+        cn_impl->add_callback(std::move(waiter));
+        return;
+    }
+    mgb_throw(MegBrainError, "unimplemented event device_wait_by config");
+}
+
+/* ===================== AtlasCompNode static methods ===================== */
+
+bool AtlasCompNode::available() {
+    return true;
+}
+
+void AtlasCompNode::finalize() {
+    if (AtlasCompNodeImpl::sd) {
+        sync_all();
+
+        auto ptr = AtlasCompNodeImpl::sd;
+        AtlasCompNodeImpl::sd = nullptr;
+        ptr->~StaticData();
+    }
+}
+
+CompNode::Impl* AtlasCompNode::load_atlas(const Locator& locator,
+                                          const Locator& locator_logical) {
+    auto&& sdptr = AtlasCompNodeImpl::sd;
+    {
+        MGB_LOCK_GUARD(AtlasCompNodeImpl::sd_mtx);
+        if (!sdptr) {
+            // use static storage so object can be safely accessed even after
+            // global finalize
+            using T = AtlasCompNodeImpl::StaticData;
+            static std::aligned_storage_t<sizeof(T), alignof(T)> storage;
+            sdptr = new (&storage) T;
+        }
+    }
+    auto&& sd = *sdptr;
+    MGB_LOCK_GUARD(sd.mtx);
+
+    CompNodeImpl* available_node = nullptr;
+    for (int i = 0; i < sd.nr_node; ++i) {
+        auto&& cur = sd.node[i];
+        if (cur.m_initialized) {
+            if (cur.m_locator_logical == locator_logical) {
+                return &cur;
+            }
+        } else {
+            available_node = &cur;
+        }
+    }
+
+    if (!available_node) {
+        mgb_assert(sd.nr_node < sd.MAX_NR_COMP_NODE,
+                   "too many CompNode allocated");
+        mgb_assert(locator.device < sd.MAX_NR_COMP_NODE,
+                   "device number too large");
+        available_node = &sd.node[sd.nr_node++];
+    }
+
+    mgb_assert(!available_node->m_initialized);
+    available_node->init(locator, locator_logical);
+    log_comp_node_created(locator, locator_logical);
+
+    return available_node;
+}
+
+void AtlasCompNode::sync_all() {
+    auto sd = AtlasCompNodeImpl::sd;
+    if (!sd)
+        return;
+
+    for (int i = 0;; ++i) {
+        // ensure async init finished
+        CompNodeEnv* env;
+        {
+            MGB_LOCK_GUARD(sd->mtx);
+            if (i >= sd->nr_node) {
+                break;
+            }
+            env = &sd->node[i].env();
+        }
+        env->atlas_env();
+    }
+
+    MGB_LOCK_GUARD(sd->mtx);
+    MGB_ATLAS_CHECK(aclrtSynchronizeDevice());
+}
+
+void AtlasCompNode::foreach (thin_function<void(CompNode)> callback) {
+    auto sd = AtlasCompNodeImpl::sd;
+    if (!sd)
+        return;
+
+    for (int i = 0;; ++i) {
+        CompNode cur;
+        {
+            MGB_LOCK_GUARD(sd->mtx);
+            if (i >= sd->nr_node)
+                return;
+            cur = make_comp_node_from_impl(&sd->node[i]);
+        }
+        callback(cur);
+    }
+}
+
+size_t AtlasCompNode::get_device_count() {
+    static uint32_t cnt = 0;
+    static Spinlock mtx;
+    MGB_LOCK_GUARD(mtx);
+    if (cnt == 0) {
+        uint32_t dev_cnt = 0;
+        auto ret = aclrtGetDeviceCount(&dev_cnt);
+        if (ret != ACL_ERROR_NONE) {
+            mgb_log_error("aclrtGetDeviceCountfaild: %s (err %d)",
+                          ::megcore::atlas::get_error_str(ret),
+                          static_cast<int>(ret));
+            cnt = 0;
+        }
+        cnt = dev_cnt;
+    }
+    return cnt;
+}
+
+#else
+
+bool AtlasCompNode::available() {
+    return false;
+}
+void AtlasCompNode::foreach (thin_function<void(CompNode)>) {}
+void AtlasCompNode::finalize() {}
+size_t AtlasCompNode::get_device_count() {
+    return 0;
+}
+AtlasCompNode::Impl* AtlasCompNode::load_atlas(const Locator&, const Locator&) {
+    mgb_throw(MegBrainError, "atlas disabled at compile time");
+}
+void AtlasCompNode::sync_all() {}
+
+#endif  // MGB_ATLAS
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/core/impl/comp_node/atlas/comp_node.h

@@ -0,0 +1,40 @@
+/**
+ * \file src/core/impl/comp_node/atlas/comp_node.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include <mutex>
+#include "../impl_helper.h"
+
+namespace mgb {
+class AtlasCompNode final : public CompNodeImplHelper {
+public:
+    static constexpr Flag sm_flag = Flag::QUEUE_LIMITED | Flag::HAS_COPY_STREAM;
+
+    class CompNodeImpl;
+    class EventImpl;
+
+    //! whether cuda comp node is available
+    static bool available();
+
+    static void foreach (thin_function<void(CompNode)> callback);
+    static void finalize();
+    static size_t get_device_count();
+    static Impl* load_atlas(const Locator& locator,
+                            const Locator& locator_logical);
+    static void sync_all();
+
+};
+
+}  // namespace mgb
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/core/impl/comp_node/cambricon/comp_node.cpp

@@ -0,0 +1,759 @@
+/**
+ * \file src/core/impl/comp_node/cambricon/comp_node.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#include "./comp_node.h"
+#include "megbrain/comp_node_env.h"
+#include "megbrain/utils/thread.h"
+
+#include <string>
+
+using namespace mgb;
+
+#if MGB_CAMBRICON
+
+#include "megbrain/comp_node/alloc.h"
+
+#include <cctype>
+#include <cstdio>
+
+#include <thread>
+
+#include <cndev.h>
+#include <cnrt.h>
+
+using CambriconCompNodeImpl = CambriconCompNode::CompNodeImpl;
+
+namespace {
+size_t get_min_system_memory(size_t available) {
+    // taken from src/core/impl/cuda/comp_node.cpp
+    if (available < (1u << 31)) {
+        // 225MiB
+        return 225 * 1024 * 1024;
+    } else {
+        // max(300 MiB, 0.05 * available)
+        return std::max<size_t>(300 * 1024 * 1024, available / 20);
+    }
+}
+}  // anonymous namespace
+
+/* ======================= CambriconRawAlloctor ======================*/
+namespace mgb {
+namespace mem_alloc {
+class CambriconRawAlloctor final : public RawAllocator {
+public:
+    void* alloc(size_t size) override {
+        void* addr;
+        cnrtRet_t ret = cnrtMalloc(&addr, size);
+        if (ret == CNRT_RET_SUCCESS) {
+            mgb_assert(addr);
+            return addr;
+        }
+        auto msg = mgb_ssprintf_log(
+                "cnrtMalloc failed while requesting %zd bytes (%.3fMiB) of "
+                "memory; error: %s",
+                size, size / (1024.0 * 1024), cnrtGetErrorStr(ret));
+        msg.append(CnrtError::get_cnrt_extra_info());
+        mgb_throw_raw(MemAllocError{msg});
+    }
+
+    void free(void* ptr) override {
+        cnrtRet_t ret = cnrtFree(ptr);
+        if (ret == CNRT_RET_SUCCESS)
+            return;
+        auto msg = ssprintf("cnrtFree failed for %p: %s", ptr,
+                            cnrtGetErrorStr(ret));
+        msg.append(CnrtError::get_cnrt_extra_info());
+        mgb_throw_raw(MemAllocError{msg});
+    }
+
+    void get_mem_info(size_t& free, size_t& tot) override;
+};
+
+class CambriconDeviceRuntimePolicy : public DeviceRuntimePolicy {
+public:
+    CompNode::DeviceType device_type() override {
+        return CompNode::DeviceType::CAMBRICON;
+    }
+    void set_device(int device) override {
+        cnrtDev_t dev;
+        MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev, device));
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev));
+    }
+    void device_synchronize(int device) override {
+        cnrtDev_t dev;
+        MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev, device));
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev));
+        MGB_CNRT_CHECK(cnrtSyncDevice());
+    }
+};
+
+/* ====================== DevMemAlloc ================================*/
+std::unique_ptr<DevMemAlloc> DevMemAlloc::make_cambricon_alloc() {
+    return std::make_unique<FwdDevMemAlloc>(
+            std::make_shared<CambriconRawAlloctor>());
+}
+}  // namespace mem_alloc
+}  // namespace mgb
+
+/* ====================== CambriconCompNodeImpl ======================*/
+class CambriconCompNode::CompNodeImpl final : public CompNode::Impl {
+    MGB_DYN_TYPE_OBJ_FINAL_DECL;
+
+    friend class EventImpl;
+    friend class CambriconCompNode;
+    friend class mgb::mem_alloc::CambriconRawAlloctor;
+
+    struct DeviceInfo;
+    struct StaticData;
+    static StaticData* sd;
+    static Spinlock sd_mtx;
+
+    //! set to true when m_locator is assigned; set to false if init
+    //! failed
+    bool m_initialized = false;
+    Locator m_locator, m_locator_logical;
+    mem_alloc::StreamMemAlloc* m_mem_alloc;
+    DeviceInfo* m_device_info;
+    cnrtDev_t m_dev;
+
+    void activate() { m_env.cnrt_env().activate(); }
+
+    void init(const Locator& locator, const Locator& locator_logical);
+    void fini();
+
+    static inline bool check_global_finalized();
+
+    //! enable peer copy from dev0 to dev1
+    static bool enable_peer_access(int dev0, int dev1);
+
+    static void static_free_device(ImplBase* self, void* ptr) {
+        static_cast<CompNodeImpl*>(self)->free_device(ptr);
+    }
+
+    static void static_free_host(ImplBase* self, void* ptr) {
+        static_cast<CompNodeImpl*>(self)->free_host(ptr);
+    }
+
+public:
+    CompNodeImpl() : Impl(static_free_device, static_free_host) {}
+
+    void* alloc_device(size_t size) override {
+        activate();
+        return m_mem_alloc->alloc(size);
+    }
+
+    void free_device(void* ptr);
+
+    void* alloc_host(size_t size) override {
+        activate();
+        void* ptr;
+        MGB_CNRT_CHECK(cnrtMallocHost(&ptr, size, CNRT_MEMTYPE_DEFAULT));
+        return ptr;
+    }
+
+    void free_host(void* ptr) {
+        if (!check_global_finalized()) {
+            activate();
+        }
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(m_dev));
+        MGB_CNRT_CHECK(cnrtFreeHost(ptr));
+    }
+
+    void copy_to_host(void* host_ptr, const void* device_ptr,
+                      size_t size) override {
+        activate();
+        MGB_CNRT_CHECK(cnrtMemcpyAsync(host_ptr, const_cast<void*>(device_ptr),
+                                       size, m_env.cnrt_env().queue,
+                                       CNRT_MEM_TRANS_DIR_DEV2HOST));
+    }
+
+    void copy_to_device(void* device_ptr, const void* host_ptr,
+                        size_t size) override {
+        activate();
+        MGB_CNRT_CHECK(cnrtMemcpyAsync(device_ptr, const_cast<void*>(host_ptr),
+                                       size, m_env.cnrt_env().queue,
+                                       CNRT_MEM_TRANS_DIR_HOST2DEV));
+    }
+
+    void peer_copy_to(Impl* dest_impl, void* dest, const void* src,
+                      size_t size) override;
+
+    size_t get_mem_addr_alignment() override {
+        return m_env.property().mem_alignment;
+    }
+
+    std::unique_ptr<Event> create_event(size_t flags) override;
+
+    void sync() override;
+
+    MemNode mem_node() override;
+
+    std::pair<size_t, size_t> get_mem_status_bytes() override {
+        m_env.cnrt_env().activate();
+        cndevMemoryInfo_t mem_info;
+        MGB_CNDEV_CHECK(
+                cndevGetMemoryUsage(&mem_info, m_env.cnrt_env().device));
+        size_t tot, used, free;
+        constexpr size_t mb2size = 1024 * 1024;
+        tot = static_cast<size_t>(mem_info.PhysicalMemoryTotal) * mb2size;
+        used = static_cast<size_t>(mem_info.PhysicalMemoryUsed) * mb2size;
+        free = tot - used + m_mem_alloc->get_free_memory_dev().tot;
+        return {tot, free};
+    }
+
+    Locator locator() override { return m_locator; }
+
+    Locator locator_logical() override { return m_locator_logical; }
+};
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(CambriconCompNode::CompNodeImpl);
+
+struct CambriconCompNodeImpl::DeviceInfo {
+    int dev_num = -1;
+    cnrtDev_t dev;
+    std::unique_ptr<mem_alloc::DevMemAlloc> mem_alloc;
+
+    bool init_done() const { return mem_alloc.get(); }
+
+    void init(const CompNodeEnv& env);
+
+    // unlike cuda, we have to set device first, then release device memory
+    void fini() {
+        cnrtSetCurrentDevice(dev);
+        return mem_alloc.reset();
+    }
+
+    size_t get_mem_reserve_size();
+};
+
+struct CambriconCompNodeImpl::StaticData {
+    static constexpr int MAX_NR_COMP_NODE = 4096, MAX_NR_DEVICE = 64;
+
+    std::recursive_mutex mtx;
+
+    mem_alloc::DevMemAlloc::PreAllocConfig prealloc_config;
+
+    CambriconCompNode::CompNodeImpl node[MAX_NR_COMP_NODE];
+    DeviceInfo dev_info[MAX_NR_DEVICE];
+    int nr_node = 0, nr_dev_used = 0;
+
+    StaticData() {
+        prealloc_config.max_overhead = 0;
+        prealloc_config.alignment = 1;
+    }
+
+    ~StaticData() {
+        for (int i = 0; i < nr_node; ++i)
+            node[i].fini();
+        for (int i = 0; i < nr_dev_used; ++i)
+            dev_info[i].fini();
+    }
+};
+CambriconCompNodeImpl::StaticData* CambriconCompNodeImpl::sd = nullptr;
+Spinlock CambriconCompNodeImpl::sd_mtx;
+
+void CambriconCompNodeImpl::init(const Locator& locator,
+                                 const Locator& locator_logical) {
+    m_locator = locator;
+    m_locator_logical = locator_logical;
+    m_initialized = true;
+
+    auto on_succ = [this](cnrtQueue_t queue) {
+        auto locator = m_locator;
+        log_comp_node_created(locator, m_locator_logical);
+
+        MGB_LOCK_GUARD(sd->mtx);
+        DeviceInfo* dev_info = nullptr;
+        for (int i = 0; i < sd->nr_dev_used; ++i) {
+            if (sd->dev_info[i].dev_num == locator.device) {
+                dev_info = &sd->dev_info[i];
+                break;
+            }
+        }
+
+        if (!dev_info) {
+            dev_info = &sd->dev_info[sd->nr_dev_used];
+            dev_info->init(m_env);
+            ++sd->nr_dev_used;
+        }
+        m_device_info = dev_info;
+        m_mem_alloc =
+                dev_info->mem_alloc->add_stream(static_cast<void*>(queue));
+        m_dev = m_device_info->dev;
+    };
+
+    auto on_error = [this](std::exception&) {
+        MGB_LOCK_GUARD(sd->mtx);
+        m_initialized = false;
+    };
+
+    m_env.init_cnrt(locator.device, make_comp_node_from_impl(this),
+                    {on_succ, on_error});
+}
+
+void CambriconCompNodeImpl::fini() {
+    if (!m_initialized)
+        return;
+
+    m_env.fini();
+    m_mem_alloc = nullptr;
+    m_device_info = nullptr;
+    m_initialized = false;
+}
+
+void CambriconCompNodeImpl::free_device(void* ptr) {
+    if (check_global_finalized())
+        return;
+
+    activate();
+    m_mem_alloc->free(ptr);
+}
+
+void CambriconCompNodeImpl::peer_copy_to(Impl* dest_impl, void* dest,
+                                         const void* src, size_t size) {
+    if (dest_impl->same_type<CambriconCompNodeImpl>()) {
+        auto&& dst_env = static_cast<CambriconCompNodeImpl*>(dest_impl)
+                                 ->m_env.cnrt_env();
+        auto&& src_env = m_env.cnrt_env();
+        activate();
+        if (dst_env.device == src_env.device) {
+            // remark: transfering data from device to device does not
+            // support async
+            sync();
+            dest_impl->sync();
+            MGB_CNRT_CHECK(cnrtMemcpy(dest, const_cast<void*>(src), size,
+                                      CNRT_MEM_TRANS_DIR_DEV2DEV));
+        } else {
+            mgb_throw_if(
+                    !enable_peer_access(src_env.device, dst_env.device) ||
+                            !enable_peer_access(dst_env.device, src_env.device),
+                    CnrtError,
+                    "directly memory access is not available for "
+                    "src=%d,dst=%d",
+                    src_env.device, dst_env.device);
+            sync();
+            dest_impl->sync();
+            MGB_CNRT_CHECK(cnrtMemcpyPeer(dest, dst_env.device,
+                                          const_cast<void*>(src),
+                                          src_env.device, size));
+        }
+        return;
+    }
+    mgb_assert(dest_impl->env().property().type == DeviceType::CPU,
+               "cnrt peer_copy_to only implemented for CPU");
+    auto copy = [this, dest, src, size]() {
+        m_env.cnrt_env().activate();
+        auto queue = m_env.cnrt_env().queue;
+        MGB_CNRT_CHECK(cnrtMemcpyAsync(dest, const_cast<void*>(src), size,
+                                       queue, CNRT_MEM_TRANS_DIR_DEV2HOST));
+        MGB_CNRT_CHECK(cnrtSyncQueue(queue));
+    };
+    dest_impl->env().cpu_env().dispatch(copy);
+}
+
+MemNode CambriconCompNodeImpl::mem_node() {
+    return MemNode{sd->dev_info + m_locator.device};
+}
+
+void CambriconCompNodeImpl::sync() {
+    activate();
+
+    // remark: CNRT does not provide interface like cudaEventQuery to test
+    // whether an event is finished. so we just call the cnrtSyncQueue
+    MGB_CNRT_CHECK(cnrtSyncQueue(m_env.cnrt_env().queue));
+}
+
+bool CambriconCompNodeImpl::enable_peer_access(int dev0, int dev1) {
+    static bool queried_enabled[StaticData::MAX_NR_DEVICE]
+                               [StaticData::MAX_NR_DEVICE];
+    if (queried_enabled[dev0][dev1])
+        return queried_enabled[dev0][dev1];
+
+    static std::mutex global_lock;
+    MGB_LOCK_GUARD(global_lock);
+    unsigned int can = 0;
+    MGB_CNRT_CHECK(cnrtGetPeerAccessibility(&can, dev0, dev1));
+    if (can)
+        mgb_log("device(%d) can directly access memories on device(%d)", dev0,
+                dev1);
+    queried_enabled[dev0][dev1] = can;
+    return can;
+}
+
+/* ================== CambriconCompNodeImpl::DeviceInfo ===============*/
+
+void CambriconCompNodeImpl::DeviceInfo::init(const CompNodeEnv& env) {
+    mgb_assert(!mem_alloc);
+    auto&& cnenv = env.cnrt_env();
+    cnenv.activate();
+    dev_num = cnenv.device;
+    MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev, dev_num));
+    // remark: Because free_device will be called after global finalize, so the
+    // implementation of mem_alloc should handle the deallocation of memories
+    // allocated by the mem_alloc. As a result, we should use the DevMemAlloc
+    // instead of FwdDevMemAlloc.
+#if 0
+    // forward cnrtMalloc
+    mem_alloc = mem_alloc::DevMemAlloc::make_cambricon_alloc();
+#else
+    auto reserve_size = get_mem_reserve_size();
+    mem_alloc = mem_alloc::DevMemAlloc::make(
+            dev_num, reserve_size,
+            std::make_shared<mem_alloc::CambriconRawAlloctor>(),
+            std::make_shared<mem_alloc::CambriconDeviceRuntimePolicy>());
+    mem_alloc->prealloc_config(sd->prealloc_config);
+    auto align = env.property().mem_alignment;
+    mem_alloc->alignment(align);
+    cnrtDeviceInfo_t device_info;
+    MGB_CNRT_CHECK(cnrtGetDeviceInfo(&device_info, dev_num));
+    mgb_log("cambricon: card%d: name=`%s' dyn_mem_reserve=%.2fMiB "
+            "alignment=0x%zx",
+            dev_num, device_info.device_name, reserve_size / 1024.0 / 1024,
+            align);
+#endif
+}
+
+size_t CambriconCompNodeImpl::DeviceInfo::get_mem_reserve_size() {
+    if (auto setting = MGB_GETENV("MGB_CAMBRICON_RESERVE_MEMORY")) {
+        if (!strncmp(setting, "b:", 2)) {
+            return std::stoull(setting + 2);
+        }
+        size_t tot, free;
+        cndevMemoryInfo_t mem_info;
+        MGB_CNDEV_CHECK(cndevGetMemoryUsage(&mem_info, dev_num));
+        constexpr size_t mb2size = 1024 * 1024;
+        tot = static_cast<size_t>(mem_info.PhysicalMemoryTotal) * mb2size;
+        size_t used =
+                static_cast<size_t>(mem_info.PhysicalMemoryUsed) * mb2size;
+        free = tot - used;
+        return free - get_min_system_memory(free);
+    } else {
+        return 0;
+    }
+}
+
+bool CambriconCompNodeImpl::check_global_finalized() {
+    if (!sd) {
+        static std::atomic_flag warn_printed = ATOMIC_FLAG_INIT;
+        if (!warn_printed.test_and_set()) {
+            mgb_log_warn(
+                    "cambricon comp node method called after global finalize");
+        }
+        return true;
+    }
+    return false;
+}
+
+/* ================== CambriconCompNodeImpl::EventImpl ================*/
+
+class CambriconCompNode::EventImpl final : public EventImplHelper {
+    bool m_placed_notifier = false;
+    bool m_sync_queue_called = false;
+    bool m_init_finished = false;
+    cnrtNotifier_t m_cnrt_notifier;
+
+    CambriconCompNodeImpl* cambricon_comp_node_impl() const {
+        return static_cast<CambriconCompNodeImpl*>(m_comp_node_impl);
+    }
+
+    void do_record() override {
+        m_sync_queue_called = false;
+        cambricon_comp_node_impl()->activate();
+        auto&& env = cambricon_comp_node_impl()->m_env.cnrt_env();
+        if (!m_placed_notifier) {
+            MGB_CNRT_CHECK(cnrtPlaceNotifier(m_cnrt_notifier, env.queue));
+            m_placed_notifier = true;
+        }
+    }
+
+    void call_sync_queue() {
+        mgb_assert(m_placed_notifier);
+        if (!m_sync_queue_called) {
+            cambricon_comp_node_impl()->activate();
+            auto&& env = cambricon_comp_node_impl()->m_env.cnrt_env();
+            MGB_CNRT_CHECK(cnrtSyncQueue(env.queue));
+            m_sync_queue_called = true;
+        }
+    }
+
+    bool do_finished() override {
+        call_sync_queue();
+        return true;
+    }
+
+    void host_wait_cv() override {
+        mgb_assert(m_placed_notifier);
+        cambricon_comp_node_impl()->activate();
+        auto&& env = cambricon_comp_node_impl()->m_env.cnrt_env();
+        MGB_CNRT_CHECK(cnrtSyncQueue(env.queue));
+    }
+
+    double do_elapsed_time_until(EventImplHelper& end) override {
+        cambricon_comp_node_impl()->activate();
+        auto&& env = cambricon_comp_node_impl()->m_env.cnrt_env();
+        MGB_CNRT_CHECK(cnrtSyncQueue(env.queue));
+        float ret = 0.f;
+        MGB_CNRT_CHECK(cnrtNotifierDuration(
+                m_cnrt_notifier, static_cast<EventImpl&>(end).m_cnrt_notifier,
+                &ret));
+        return static_cast<double>(ret) * 1e-3;
+    }
+
+    void do_device_wait_by(Impl* cn_impl) override;
+
+public:
+    EventImpl(CambriconCompNodeImpl* comp_node_impl, size_t create_flags)
+            : EventImplHelper(comp_node_impl, create_flags) {
+        cambricon_comp_node_impl()->activate();
+        MGB_CNRT_CHECK(cnrtCreateNotifier(&m_cnrt_notifier));
+        m_init_finished = true;
+    }
+
+    ~EventImpl() {
+        if (m_init_finished) {
+            MGB_TRY { MGB_CNRT_CHECK(cnrtDestroyNotifier(&m_cnrt_notifier)); }
+            MGB_CATCH(MegBrainError & exc, {
+                mgb_log_error("failed to destroy cnrt notifier: %s",
+                              exc.what());
+            })
+        }
+    }
+};
+
+std::unique_ptr<CompNode::Event> CambriconCompNodeImpl::create_event(
+        size_t flags) {
+    return std::make_unique<EventImpl>(this, flags);
+}
+
+void CambriconCompNode::EventImpl::do_device_wait_by(Impl* cn_impl) {
+    if (cn_impl->env().property().type == DeviceType::CAMBRICON) {
+        auto imp = static_cast<CambriconCompNodeImpl*>(cn_impl);
+        auto queue = imp->m_env.cnrt_env().queue;
+        imp->activate();
+        MGB_CNRT_CHECK(cnrtSyncQueue(queue));
+        return;
+    }
+    if (cn_impl->env().property().type == DeviceType::CPU) {
+        auto waiter = [this]() {
+            cambricon_comp_node_impl()->activate();
+            auto queue = cambricon_comp_node_impl()->m_env.cnrt_env().queue;
+            MGB_CNRT_CHECK(cnrtSyncQueue(queue));
+        };
+        cn_impl->add_callback(std::move(waiter));
+        return;
+    }
+    mgb_throw(MegBrainError, "unimplemented event device_wait_by config");
+}
+
+/* ================== CambriconCompNode static methods ================*/
+
+bool CambriconCompNode::available() {
+    CompNodeEnv::CnrtEnv::init();
+    static int result = -1;
+    static Spinlock mtx;
+    MGB_LOCK_GUARD(mtx);
+    if (result == -1) {
+        unsigned int dev_num = 0;
+        auto err = cnrtGetDeviceCount(&dev_num);
+        result = err == CNRT_RET_SUCCESS && dev_num >= 1;
+        if (!result) {
+            mgb_log_warn("cambricon unavailable: %d(%s) dev_num=%u",
+                         static_cast<int>(err), cnrtGetErrorStr(err), dev_num);
+        }
+    }
+    return result;
+}
+
+void CambriconCompNode::finalize() {
+    if (CambriconCompNodeImpl::sd) {
+        sync_all();
+
+        auto ptr = CambriconCompNodeImpl::sd;
+        CambriconCompNodeImpl::sd = nullptr;
+        ptr->~StaticData();
+    }
+}
+
+CompNode::Impl* CambriconCompNode::load_cambricon(
+        const Locator& locator, const Locator& locator_logical) {
+    int nr_devs = get_device_count();
+    mgb_assert(locator.device >= 0 && locator.device < nr_devs,
+               "request device%d out of range [0, %d)", locator.device,
+               nr_devs);
+
+    auto&& sdptr = CambriconCompNodeImpl::sd;
+    {
+        MGB_LOCK_GUARD(CambriconCompNodeImpl::sd_mtx);
+        if (!sdptr) {
+            using T = CambriconCompNodeImpl::StaticData;
+            static std::aligned_storage_t<sizeof(T), alignof(T)> storage;
+            sdptr = new(&storage)T;
+        }
+    }
+    auto&& sd = *sdptr;
+    MGB_LOCK_GUARD(sd.mtx);
+
+    CompNodeImpl* available_node = nullptr;
+    for (int i = 0; i < sd.nr_node; ++i) {
+        auto&& cur = sd.node[i];
+        if (cur.m_initialized) {
+            if (cur.m_locator_logical == locator_logical) {
+                return &cur;
+            }
+        } else {
+            available_node = &cur;
+        }
+    }
+
+    if (!available_node) {
+        mgb_assert(sd.nr_node < sd.MAX_NR_COMP_NODE,
+                   "too many CompNode allocated");
+        mgb_assert(locator.device < sd.MAX_NR_COMP_NODE,
+                   "device number too large");
+        available_node = &sd.node[sd.nr_node++];
+    }
+
+    mgb_assert(!available_node->m_initialized);
+    available_node->init(locator, locator_logical);
+
+    return available_node;
+}
+
+void CambriconCompNode::try_coalesce_all_free_memory() {
+    auto sd = CambriconCompNodeImpl::sd;
+    if (!sd)
+        return;
+
+    size_t size = 0;
+    for (int i = 0; i < sd->nr_dev_used; ++i) {
+        size += sd->dev_info[i]
+                        .mem_alloc->gather_stream_free_blk_and_release_full();
+    }
+    if (size) {
+        mgb_log_debug("%zu bytes freed by try_coalesce_all_free_memory()",
+                      size);
+    }
+}
+
+void CambriconCompNode::sync_all() {
+    auto sd = CambriconCompNodeImpl::sd;
+    if (!sd)
+        return;
+    for (int i = 0;; ++i) {
+        CompNodeEnv* env;
+        {
+            MGB_LOCK_GUARD(sd->mtx);
+            if (i >= sd->nr_node) {
+                break;
+            }
+            env = &sd->node[i].env();
+        }
+        env->cnrt_env();
+    }
+
+    MGB_LOCK_GUARD(sd->mtx);
+    for (int i = 0; i < sd->nr_dev_used; ++i) {
+        cnrtDev_t dev;
+        MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev, sd->dev_info[i].dev_num));
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev));
+        MGB_CNRT_CHECK(cnrtSyncDevice());
+    }
+}
+
+void CambriconCompNode::foreach (thin_function<void(CompNode)> callback) {
+    auto sd = CambriconCompNodeImpl::sd;
+    if (!sd)
+        return;
+
+    for (int i = 0;; ++i) {
+        CompNode cur;
+        {
+            MGB_LOCK_GUARD(sd->mtx);
+            if (i >= sd->nr_node)
+                return;
+            cur = make_comp_node_from_impl(&sd->node[i]);
+        }
+        callback(cur);
+    }
+}
+
+size_t CambriconCompNode::get_device_count() {
+    CompNodeEnv::CnrtEnv::init();
+    static int cnt = -1;
+    static Spinlock mtx;
+    MGB_LOCK_GUARD(mtx);
+    if (cnt == -1) {
+        unsigned int dev_cnt = 0;
+        auto ret = cnrtGetDeviceCount(&dev_cnt);
+        if (ret != CNRT_RET_SUCCESS) {
+            mgb_log_error("cnrtGetDeviceCount faild: %s (err %d)",
+                          cnrtGetErrorStr(ret), int(ret));
+            cnt = 0;
+        }
+        cnt = dev_cnt;
+        mgb_assert(cnt >= 0);
+    }
+    return cnt;
+}
+
+void mgb::mem_alloc::CambriconRawAlloctor::get_mem_info(size_t& free,
+                                                        size_t& tot) {
+    auto sd = CambriconCompNodeImpl::sd;
+    int device = -1;
+    {
+        cnrtDev_t dev;
+        MGB_CNRT_CHECK(cnrtGetCurrentDevice(&dev));
+        for (int i = 0; i < sd->nr_dev_used; ++i) {
+            if (sd->dev_info[i].dev == dev) {
+                device = sd->dev_info[i].dev_num;
+                break;
+            }
+        }
+    }
+    mgb_assert(device >= 0,
+               "current device has not been initialized in static data");
+    cndevMemoryInfo_t mem_info;
+    auto ret = cndevGetMemoryUsage(&mem_info, device);
+    if (ret == CNDEV_SUCCESS) {
+        constexpr size_t mb2size = 1024 * 1024;
+        tot = static_cast<size_t>(mem_info.PhysicalMemoryTotal) * mb2size;
+        size_t used =
+                static_cast<size_t>(mem_info.PhysicalMemoryUsed) * mb2size;
+        free = tot - used;
+        return;
+    }
+    auto msg =
+            ssprintf("cndevGetMemoryUsage faild %s", cndevGetErrorString(ret));
+    mgb_throw_raw(MemAllocError{msg});
+}
+
+#else
+
+bool CambriconCompNode::available() {
+    return false;
+}
+void CambriconCompNode::try_coalesce_all_free_memory() {}
+void CambriconCompNode::foreach (thin_function<void(CompNode)>) {}
+void CambriconCompNode::finalize() {}
+size_t CambriconCompNode::get_device_count() {
+    return 0;
+}
+CambriconCompNode::Impl* CambriconCompNode::load_cambricon(const Locator&,
+                                                           const Locator&) {
+    mgb_throw(MegBrainError, "cambricon disabled at compile time");
+}
+void CambriconCompNode::sync_all() {}
+
+#undef err
+
+#endif  // MGB_CAMBRICON
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+

src/core/impl/comp_node/cambricon/comp_node.h

@@ -0,0 +1,39 @@
+/**
+ * \file src/core/impl/comp_node/cambricon/comp_node.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ */
+
+#pragma once
+
+#include "../impl_helper.h"
+
+namespace mgb {
+    class CambriconCompNode final: public CompNodeImplHelper {
+        public:
+            static constexpr Flag sm_flag =
+                    Flag::QUEUE_LIMITED | Flag::HAS_COPY_STREAM;
+
+            class CompNodeImpl;
+            class EventImpl;
+
+            //! whether cambricon comp node is available
+            static bool available();
+
+            static void try_coalesce_all_free_memory();
+            static void foreach(thin_function<void(CompNode)> callback);
+            static void finalize();
+            static size_t get_device_count();
+            static Impl* load_cambricon(
+                    const Locator &locator, const Locator &locator_logical);
+            static void sync_all();
+    };
+}
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
+

src/core/impl/comp_node/comp_node.cpp

@@ -15,6 +15,8 @@
 
 #include "./cuda/comp_node.h"
 #include "./cpu/comp_node.h"
+#include "./cambricon/comp_node.h"
+#include "./atlas/comp_node.h"
 
 #include <cstring>
 #include <atomic>
@@ -40,6 +42,10 @@ namespace {
                 return "gpu";
             case DT::CPU:
                 return "cpu";
+            case DT::ATLAS:
+                return "atlas";
+            case DT::CAMBRICON:
+                return "cambricon";
             case DT::MULTITHREAD:
                 return "multithread";
             default:
@@ -145,7 +151,20 @@ CompNode::Locator CompNode::Locator::parse(const std::string &id) {
     DeviceType dev_type;
 
     // parse dev_type
-    if (ptr[0] == 'm') {
+    if (ptr[0] == 'a') {
+        if (strncmp(ptr, "atlas", 5)) {
+            err();
+        }
+        dev_type = DeviceType::ATLAS;
+        ptr += 5;
+    }
+    else if (ptr[2] == 'm') {
+        if (strncmp(ptr, "cambricon", 9)) {
+            err();
+        }
+        dev_type = DeviceType::CAMBRICON;
+        ptr += 9;
+    } else if (ptr[0] == 'm') {
         if (strncmp(ptr, "multithread", 11)) {
             err();
         }
@@ -478,6 +497,13 @@ CompNode CompNode::load(const Locator& locator_physical,
         case DeviceType::CPU:
             ret = CpuCompNode::load_cpu(locator_physical, locator_logical);
             break;
+        case DeviceType::ATLAS:
+            ret = AtlasCompNode::load_atlas(locator_physical, locator_logical);
+            break;
+        case DeviceType::CAMBRICON:
+            ret = CambriconCompNode::load_cambricon(locator_physical,
+                                                    locator_logical);
+            break;
         default:
             mgb_throw(MegBrainError, "bad device type");
     }
@@ -496,20 +522,27 @@ void CompNode::finalize() {
     comp_node_detail::DepedentObjList::invoke_callback_and_clean();
     CudaCompNode::finalize();
     CpuCompNode::finalize();
+    CambriconCompNode::finalize();
+    AtlasCompNode::finalize();
 }
 
 void CompNode::try_coalesce_all_free_memory() {
     CudaCompNode::try_coalesce_all_free_memory();
+    CambriconCompNode::try_coalesce_all_free_memory();
 }
 
 void CompNode::sync_all() {
     CudaCompNode::sync_all();
     CpuCompNode::sync_all();
+    CambriconCompNode::sync_all();
+    AtlasCompNode::sync_all();
 }
 
 void CompNode::foreach(thin_function<void(CompNode)> callback) {
     CudaCompNode::foreach(callback);
     CpuCompNode::foreach(callback);
+    CambriconCompNode::foreach(callback);
+    AtlasCompNode::foreach(callback);
 }
 
 size_t CompNode::get_device_count(DeviceType type, bool warn) {
@@ -519,6 +552,10 @@ size_t CompNode::get_device_count(DeviceType type, bool warn) {
         case DeviceType::MULTITHREAD:
         case DeviceType::CPU:
             return CpuCompNode::get_device_count();
+        case DeviceType::CAMBRICON:
+            return CambriconCompNode::get_device_count();
+        case DeviceType::ATLAS:
+            return AtlasCompNode::get_device_count();
         default:
             mgb_throw(MegBrainError, "bad device type");
     }
@@ -534,6 +571,12 @@ bool CompNode::contain_flag(DeviceType device_type, Flag flag) {
         case DeviceType::CPU:
             cn_flag = CpuCompNode::sm_flag;
             break;
+        case DeviceType::CAMBRICON:
+            cn_flag = CambriconCompNode::sm_flag;
+            break;
+        case DeviceType::ATLAS:
+            cn_flag = AtlasCompNode::sm_flag;
+            break;
         default:
             mgb_throw(MegBrainError, "unexpected device type");
     }

src/core/impl/comp_node/cpu/comp_node.cpp

@@ -528,9 +528,23 @@ class CpuCompNode::CompNodeImpl final: public CpuDispatchableBase {
                 Impl *dest_impl, void *dest,
                 const void *src, size_t size) override {
             if (!dest_impl->same_type<CpuCompNode::CompNodeImpl>()) {
+                if (dest_impl->env().property().type == DeviceType::ATLAS) {
+#if MGB_ATLAS
+                    dest_impl->copy_to_device(dest, src, size);
+                    return;
+#else
+                    mgb_throw(MegBrainError,
+                              "Atlas comp_node used but "
+                              "MGB_ATLAS not enabled");
+#endif
+
+
+                } else {
                     mgb_assert(locator().device == Locator::DEVICE_CPU_DEFAULT,
-                            "currently only peer copy from default cpu comp nodes "
-                            "is implemented");
+                               "currently only peer copy from default cpu comp "
+                               "nodes "
+                               "is implemented");
+                }
             }
             dest_impl->copy_to_device(dest, src, size);
         }
@@ -841,12 +855,22 @@ void CpuCompNode::CpuDispatchableBase::EventImpl::do_device_wait_by(
         auto type = cn_impl->env().property().type;
         mgb_throw_if(type != CompNode::DeviceType::CPU
                              && type != CompNode::DeviceType::CUDA
+                             && type != CompNode::DeviceType::ATLAS
                              ,
                      MegBrainError,
-                     "currently CPU can only wait for CPU, CUDA"
+                     "currently CPU can only wait for CPU, CUDA, ATLAS"
         );
     }
 
+    if (cn_impl->env().property().type == CompNode::DeviceType::ATLAS) {
+#if MGB_ATLAS
+        return m_comp_node_impl->sync();
+#else
+        mgb_throw(MegBrainError,
+                  "Atlas comp_node used but MGB_ATLAS not enabled");
+#endif
+
+    }
 
     auto version = m_record_nr_req.load(std::memory_order_relaxed);
     mgb_assert(version, "device wait on non-recorded event");

src/core/impl/comp_node_env.cpp

@@ -22,6 +22,15 @@
 #endif
 #endif
 
+#if MGB_CAMBRICON
+#include "megcore_cambricon.h"
+#endif
+
+#if MGB_ATLAS
+#include "acl/acl.h"
+#include "megcore_atlas.h"
+#endif
+
 using namespace mgb;
 
 /* =================== MegDNNHandle =================== */
@@ -54,6 +63,28 @@ MegDNNHandle::MegDNNHandle(const CompNodeEnv& env) {
         init = true;
     }
 #endif
+#if MGB_CAMBRICON
+    if (env.property().type == CompNode::DeviceType::CAMBRICON) {
+        CompNodeEnv::CnrtEnv::init_status.init();
+        megcore::createDeviceHandleWithGlobalInitStatus(
+                &m_dev_hdl, env.cnrt_env().device, 0, true);
+        megcore::createComputingHandleWithCambriconContext(
+                &m_comp_hdl, m_dev_hdl, 0, {env.cnrt_env().queue});
+        init = true;
+    }
+#endif
+
+#if MGB_ATLAS
+    if (env.property().type == CompNode::DeviceType::ATLAS) {
+        CompNodeEnv::AtlasEnv::init_status.init();
+        megcore::createAtlasDeviceHandleWithGlobalInitStatus(
+                &m_dev_hdl, env.atlas_env().device, 0, true);
+        megcore::createComputingHandleWithAtlasContext(
+                &m_comp_hdl, m_dev_hdl, 0, {env.atlas_env().stream});
+
+        init = true;
+    }
+#endif
 
     if (env.property().type == CompNode::DeviceType::CPU) {
         megcoreCreateDeviceHandle(&m_dev_hdl, megcorePlatformCPU);
@@ -175,6 +206,73 @@ void CompNodeEnv::init_cuda_async(int dev, CompNode comp_node,
 }
 #endif
 
+#if MGB_ATLAS
+
+void mgb::_on_atlas_error(const char* expr, int err, const char* file,
+                          const char* func, int line) {
+    mgb_throw(AtlasError, "atlas error %d: %s (%s at %s:%s:%d)", int(err),
+              megcore::atlas::get_error_str(err), expr, file, func, line);
+}
+
+CompNodeEnv::AtlasEnv::InitStatus CompNodeEnv::AtlasEnv::init_status;
+void CompNodeEnv::init_atlas(CompNode comp_node, const AtlasEnv& env) {
+    m_comp_node = comp_node;
+    m_atlas_env = env;
+    m_property.type = DeviceType::ATLAS;
+    m_property.mem_alignment = 64;
+
+    m_atlas_env.activate();
+    MGB_ATLAS_CHECK(aclrtCreateStream(&m_atlas_env.stream));
+    m_user_data_container = std::make_unique<UserDataContainer>();
+    mgb_assert(m_property.mem_alignment ==
+               MegDNNHandle::get(*this).handle()->alignment_requirement());
+}
+#endif
+
+
+#if MGB_CAMBRICON
+const char* mgb::cnml_get_error_string(cnmlStatus_t err) {
+    switch (err) {
+#define cb(_err) \
+    case _err:   \
+        return #_err
+        cb(CNML_STATUS_SUCCESS);
+        cb(CNML_STATUS_NODEVICE);
+        cb(CNML_STATUS_DOMAINERR);
+        cb(CNML_STATUS_INVALIDARG);
+        cb(CNML_STATUS_LENGTHERR);
+        cb(CNML_STATUS_OUTOFRANGE);
+        cb(CNML_STATUS_RANGEERR);
+        cb(CNML_STATUS_OVERFLOWERR);
+        cb(CNML_STATUS_UNDERFLOWERR);
+        cb(CNML_STATUS_INVALIDPARAM);
+        cb(CNML_STATUS_BADALLOC);
+        cb(CNML_STATUS_BADTYPEID);
+        cb(CNML_STATUS_BADCAST);
+        cb(CNML_STATUS_UNSUPPORT);
+#undef cb
+    }
+    return "Unknown CNML error";
+}
+
+void mgb::_on_cnrt_error(const char* expr, cnrtRet_t err, const char* file,
+                         const char* func, int line) {
+    mgb_throw(CnrtError, "cnrt error %d: %s (%s at %s:%s:%d)", int(err),
+              cnrtGetErrorStr(err), expr, file, func, line);
+}
+
+void mgb::_on_cndev_error(const char* expr, cndevRet_t err, const char* file,
+                         const char* func, int line) {
+   mgb_throw(CndevError, "cndev error %d: %s (%s at %s:%s:%d)", int(err),
+              cndevGetErrorString(err), expr, file, func, line);
+}
+
+void mgb::_on_cnml_error(const char* expr, cnmlStatus_t err, const char* file,
+                         const char* func, int line) {
+    mgb_throw(CnmlError, "cnml error %d: %s (%s at %s:%s:%d)", int(err),
+              cnml_get_error_string(err), expr, file, func, line);
+}
+#endif
 
 void CompNodeEnv::init_cpu(const CpuEnv& env, CompNode comp_node) {
     m_comp_node = comp_node;
@@ -188,6 +286,41 @@ void CompNodeEnv::init_cpu(const CpuEnv& env, CompNode comp_node) {
 }
 
 
+#if MGB_CAMBRICON
+void CompNodeEnv::init_cnrt(int dev, CompNode comp_node,
+                            const ContinuationCtx<cnrtQueue_t>& cont) {
+    m_comp_node = comp_node;
+    m_cnrt_env.device = dev;
+    m_property.type = DeviceType::CAMBRICON;
+    MGB_CNRT_CHECK(cnrtGetDeviceInfo(&m_cnrt_env.device_info, dev));
+    // FIXME: doc doesn't describe the aligment requirement for device memory
+    // address
+    m_property.mem_alignment = 1u;
+    // ensure exception safe
+    bool queue_created = false;
+    MGB_MARK_USED_VAR(queue_created);
+    MGB_TRY {
+        m_cnrt_env.activate();
+        MGB_CNRT_CHECK(cnrtCreateQueue(&m_cnrt_env.queue));
+        queue_created = true;
+        m_user_data_container = std::make_unique<UserDataContainer>();
+        cont.next(m_cnrt_env.queue);
+        // TODO: initialize megdnn handle
+        mgb_assert(m_property.mem_alignment ==
+                   MegDNNHandle::get(*this).handle()->alignment_requirement());
+    }
+    MGB_CATCH(std::exception & exc, {
+        mgb_log_error("cnrt init failed: %s", exc.what());
+        if (queue_created) {
+            MGB_CNRT_CHECK(cnrtDestroyQueue(m_cnrt_env.queue));
+        }
+        cont.err(exc);
+        throw;
+    })
+}
+CompNodeEnv::CnrtEnv::InitStatus CompNodeEnv::CnrtEnv::init_status;
+#endif
+
 void CompNodeEnv::fini() {
     ensure_async_init_finished();
     m_user_data_container.reset();
@@ -197,6 +330,19 @@ void CompNodeEnv::fini() {
         MGB_CUDA_CHECK(cudaStreamDestroy(m_cuda_env.stream));
     }
 #endif
+#if MGB_CAMBRICON
+    if (m_property.type == DeviceType::CAMBRICON) {
+        m_cnrt_env.activate();
+        MGB_CNRT_CHECK(cnrtDestroyQueue(m_cnrt_env.queue));
+    }
+#endif
+
+#if MGB_ATLAS
+    if (m_property.type == DeviceType::ATLAS) {
+        m_atlas_env.activate();
+        MGB_ATLAS_CHECK(aclrtDestroyStream(m_atlas_env.stream));
+    }
+#endif
 }
 
 #if MGB_ENABLE_COMP_NODE_ASYNC_INIT

src/core/impl/exception.cpp

@@ -71,6 +71,29 @@ std::string CudaError::get_cuda_extra_info() {
 #endif
 }
 
+AtlasError::AtlasError(const std::string &msg):
+    SystemError(msg)
+{
+}
+
+
+CnrtError::CnrtError(const std::string& msg) : SystemError(msg) {
+    m_msg.append(get_cnrt_extra_info());
+}
+
+std::string CnrtError::get_cnrt_extra_info() {
+#if MGB_CAMBRICON
+    // get last error
+    auto err = cnrtGetLastErr();
+    return ssprintf("(last_err=%d(%s))", err, cnrtGetErrorStr(err));
+#else
+    return "cnrt disabled at compile time";
+#endif
+}
+
+CndevError::CndevError(const std::string& msg) : SystemError(msg) {}
+
+CnmlError::CnmlError(const std::string& msg) : SystemError(msg) {}
 
 bool mgb::has_uncaught_exception() {
 #if MGB_ENABLE_EXCEPTION

src/core/impl/graph/var_node_mem_mgr.cpp

@@ -124,7 +124,7 @@ StaticDeviceMemoryManager::make_default_impl() {
 #endif  // MGB_THREAD_SAFE
 
 /* ==================== AsyncVarReleaser ==================== */
-#if MGB_CUDA 
+#if MGB_CUDA || MGB_ATLAS
 class VarNodeMemManager::AsyncVarReleaser {
     struct WaiterParam {
         CompNode cn;
@@ -247,7 +247,7 @@ bool VarNodeMemManager::ImpureMemPlanManager::check_need_realloc() {
 VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
     m_owner_graph(graph),
     m_seq_mem_opt(graph)
-#if MGB_CUDA 
+#if MGB_CUDA || MGB_ATLAS
     ,m_asyn_var_releaser(new AsyncVarReleaser)
 #endif
 {
@@ -255,7 +255,7 @@ VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
         MGB_MARK_USED_VAR(ev);
         // async release is only used for sync between multiple comp nodes, and
         // does not wait for device to finish
-#if MGB_CUDA 
+#if MGB_CUDA || MGB_ATLAS
         m_asyn_var_releaser->wait_release_finish();
 #endif
         m_cpu_async_release_barrier.wait_zero();
@@ -296,8 +296,7 @@ VarNodeMemManager::VarNodeMemManager(ComputingGraphImpl *graph):
     graph->event().register_receiver_permanent<event::CompSeqExecError>(
             on_comp_seq_error);
 
-#if MGB_ENABLE_VAR_DEV_MEM_DEFRAGMENTER && (MGB_CUDA \
-                )
+#if MGB_ENABLE_VAR_DEV_MEM_DEFRAGMENTER && (MGB_CUDA || MGB_ATLAS)
     auto on_mem_defrag_start = [this](const event::BeforeMemDefrag&) {
         m_asyn_var_releaser->wait_release_finish();
     };
@@ -1350,6 +1349,13 @@ void VarNodeMemManager::decr_var_mem_refcnt(
         case DT::CUDA:
             m_asyn_var_releaser->add(dispatch_cn, var);
             break;
+#endif
+#if MGB_ATLAS
+        case DT::ATLAS:
+            {
+                m_asyn_var_releaser->add(dispatch_cn, var);
+                break;
+            }
 #endif
         default:
             mgb_throw(MegBrainError,

src/core/impl/graph/var_node_mem_mgr.h

@@ -437,7 +437,7 @@ class VarNodeMemManager {
         SyncableCounter m_cpu_async_release_barrier;
 
 
-#if MGB_CUDA 
+#if MGB_CUDA || MGB_ATLAS
         //! release dynamic var on after compnode event finishes
         class AsyncVarReleaser;
         std::unique_ptr<AsyncVarReleaser> m_asyn_var_releaser;

src/core/impl/tensor.cpp

@@ -611,6 +611,12 @@ void mgb::dev_tensor_memset(const DeviceTensorND& tensor, int val) {
             MGB_CUDA_CHECK(
                     cudaMemsetAsync(ptr, val, size, env.cuda_env().stream));
             break;
+#endif
+#if MGB_ATLAS
+       case CompNode::DeviceType::ATLAS:
+           MGB_ATLAS_CHECK(aclrtMemsetAsync(ptr, -1, val, size,
+                                            env.atlas_env().stream));
+           break;
 #endif
         case CompNode::DeviceType::CPU: {
             auto fill = [ptr, size, val]() { std::memset(ptr, val, size); };

src/core/include/megbrain/comp_node.h

@@ -112,6 +112,8 @@ class CompNode {
 
             CUDA = 1,
             CPU = 2,
+            CAMBRICON = 3,
+            ATLAS = 9,
             MULTITHREAD,
             MAX_DEVICE_ID,
         };

src/core/include/megbrain/comp_node_env.h

@@ -44,6 +44,31 @@
 #endif //MGB_ENABLE_LOGGING
 #endif //MGB_CUDA
 
+#if MGB_ATLAS
+#include "acl/acl.h"
+#include <atomic>
+
+#if MGB_ENABLE_LOGGING
+#define MGB_ATLAS_CHECK(expr)                                         \
+    do {                                                              \
+        aclError __acl_check_code = (expr);                           \
+        if (!mgb_likely(__acl_check_code == ACL_ERROR_NONE)) {        \
+            ::mgb::_on_atlas_error(#expr, __acl_check_code, __FILE__, \
+                                  __func__, __LINE__);                \
+        }                                                             \
+    } while (0)
+#else
+#define MGB_ATLAS_CHECK(expr)                                           \
+    do {                                                                \
+        aclError __acl_check_code = (expr);                             \
+        if (!mgb_likely(__acl_check_code == ACL_ERROR_NONE)) {          \
+            ::mgb::_on_atlas_error(#expr, __acl_check_code, "", "", 1); \
+        }                                                               \
+    } while (0)
+
+#endif //MGB_ENABLE_LOGGING
+
+#endif // MGB_ATLAS
 
 //! whether to enable asynchronous initialization for CompNode and CompNodeEnv
 #define MGB_ENABLE_COMP_NODE_ASYNC_INIT (MGB_CUDA)
@@ -61,6 +86,10 @@
 #include "megbrain/utils/thin/function.h"
 
 namespace mgb {
+#if MGB_ATLAS
+[[noreturn]] void _on_atlas_error(const char* expr, aclError err,
+                                  const char* file, const char* func, int line);
+#endif
 
 #if MGB_CUDA
 [[noreturn]] void _on_cuda_error(const char* expr, cudaError_t err,
@@ -68,6 +97,16 @@ namespace mgb {
 #endif
 
 
+#if MGB_CAMBRICON
+const char* cnml_get_error_string(cnmlStatus_t err);
+[[noreturn]] void _on_cnrt_error(const char* expr, cnrtRet_t err,
+                                 const char* file, const char* func, int line);
+[[noreturn]] void _on_cndev_error(const char* expr, cndevRet_t err,
+                                  const char* file, const char* func, int line);
+[[noreturn]] void _on_cnml_error(const char* expr, cnmlStatus_t err,
+                                 const char* file, const char* func, int line);
+#endif
+
 class CPUDispatcher : public MegcoreCPUDispatcher {
 public:
     using AffinityCallBack = thin_function<void(size_t)>;
@@ -80,6 +119,7 @@ public:
     }
 };
 
+using AtlasDispatcher = CPUDispatcher;
 
 /*!
  * \brief CompNode environment
@@ -158,6 +198,17 @@ public:
             m_cuda_env.activate();
         }
 #endif
+#if MGB_CAMBRICON
+        if (m_property.type == DeviceType::CAMBRICON) {
+            m_cnrt_env.activate();
+        }
+#endif
+#if MGB_ATLAS
+        if (m_property.type == DeviceType::ATLAS) {
+            m_atlas_env.activate();
+        }
+#endif
+
     }
 
     /*!
@@ -199,6 +250,113 @@ public:
                          const ContinuationCtx<cudaStream_t>& cont);
 #endif
 
+#if MGB_ATLAS
+    struct AtlasEnv {
+        int device = -1;
+        aclrtStream stream = 0;
+
+        struct InitStatus {
+            bool initialized;
+            Spinlock mtx;
+            InitStatus() : initialized{false} {}
+            void init() {
+                MGB_LOCK_GUARD(mtx);
+                if (!initialized) {
+                    auto acl_err = aclInit(nullptr);
+                    initialized = acl_err == ACL_ERROR_NONE;
+                    mgb_throw_if(!initialized, AtlasError,
+                                 "acl initialize failed: (acl: %d)",
+                                 static_cast<int>(acl_err));
+                }
+            }
+            ~InitStatus() {
+                MGB_LOCK_GUARD(mtx);
+                if (initialized) {
+                    initialized = false;
+                }
+            }
+        };
+        static InitStatus init_status;
+
+        static void init() {
+            init_status.init();
+        }
+
+        void activate() const {
+            init();
+            MGB_ATLAS_CHECK(aclrtSetDevice(device));
+        }
+    };
+
+    const AtlasEnv& atlas_env() const {
+        if (mgb_unlikely(m_property.type != DeviceType::ATLAS))
+            on_bad_device_type(DeviceType::ATLAS);
+        ensure_async_init_finished();
+        return m_atlas_env;
+    }
+
+    //! init this as a atlas env synchronously
+    void init_atlas(CompNode comp_node, const AtlasEnv& env);
+#endif
+
+#if MGB_CAMBRICON
+    struct CnrtEnv {
+        int device = -1;
+        cnrtQueue_t queue = nullptr;
+        cnrtDeviceInfo_t device_info;
+        struct InitStatus {
+            bool initialized;
+            Spinlock mtx;
+            InitStatus() : initialized{false} {}
+            void init() {
+                MGB_LOCK_GUARD(mtx);
+                if (!initialized) {
+                    auto cnrt_err = cnrtInit(0);
+                    initialized = cnrt_err == CNRT_RET_SUCCESS;
+                    auto cndev_err = cndevInit(0);
+                    initialized &= cndev_err == CNDEV_SUCCESS;
+                    auto cnml_err = cnmlInit(0);
+                    initialized &= cnml_err == CNML_STATUS_SUCCESS;
+                    mgb_throw_if(!initialized, CnrtError,
+                                 "cnrt/cndev/cnml initialize failed: (cnrt:%d, "
+                                 "cndev:%d, cnml: %d)",
+                                 static_cast<int>(cnrt_err),
+                                 static_cast<int>(cndev_err),
+                                 static_cast<int>(cnml_err));
+                }
+            }
+            ~InitStatus() {
+                if (initialized) {
+                    MGB_CNML_CHECK(cnmlExit());
+                    MGB_CNDEV_CHECK(cndevRelease());
+                    cnrtDestroy();
+                    initialized = false;
+                }
+            }
+        };
+        static InitStatus init_status;
+
+        static void init() {
+            init_status.init();
+        }
+
+        void activate() const {
+            init();
+            cnrtDev_t dev;
+            MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev, device));
+            MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev));
+        }
+    };
+
+    const CnrtEnv& cnrt_env() const {
+        if (mgb_unlikely(m_property.type != DeviceType::CAMBRICON))
+            on_bad_device_type(DeviceType::CAMBRICON);
+        return m_cnrt_env;
+    }
+
+    void init_cnrt(int dev, CompNode comp_node,
+                   const ContinuationCtx<cnrtQueue_t>& cont);
+#endif
 
     struct CpuEnv {
         using Task = CPUDispatcher::Task;
@@ -238,6 +396,12 @@ private:
 
 #if MGB_CUDA
     CudaEnv m_cuda_env;
+#endif
+#if MGB_ATLAS
+    AtlasEnv m_atlas_env;
+#endif
+#if MGB_CAMBRICON
+    CnrtEnv m_cnrt_env;
 #endif
     CpuEnv m_cpu_env;
 

src/core/include/megbrain/exception.h

@@ -139,6 +139,32 @@ public:
     CudaError(const std::string& msg);
 };
 
+class AtlasError final: public SystemError {
+public:
+    AtlasError(const std::string& msg);
+};
+
+
+class CnrtError final : public SystemError {
+public:
+    /*!
+     * \brief get extra info for current cnrt status, to be appended in
+     * error message
+     */
+    static std::string get_cnrt_extra_info();
+
+    CnrtError(const std::string& msg);
+};
+
+class CndevError final : public SystemError {
+public:
+    CndevError(const std::string& msg);
+};
+
+class CnmlError final : public SystemError {
+public:
+    CnmlError(const std::string& msg);
+};
 
 class AssertionError final : public MegBrainError {
 public:

src/core/test/comp_node.cpp

@@ -40,6 +40,13 @@ TEST(TestCompNode, Parse) {
     ASSERT_EQ(L::parse("cpu2:23"), make_lc(D::CPU, 2, 23));
     ASSERT_EQ(L::parse("cpu21:23"), make_lc(D::CPU, 21, 23));
 
+    ASSERT_EQ(L::parse("cambriconx"), make_lc(D::CAMBRICON, -1, 0));
+    ASSERT_EQ(L::parse("cambricon2"), make_lc(D::CAMBRICON, 2, 0));
+    ASSERT_EQ(L::parse("cambricon2:3"), make_lc(D::CAMBRICON, 2, 3));
+
+    ASSERT_EQ(L::parse("atlasx"), make_lc(D::ATLAS, -1, 0));
+    ASSERT_EQ(L::parse("atlas2"), make_lc(D::ATLAS, 2, 0));
+    ASSERT_EQ(L::parse("atlas2:3"), make_lc(D::ATLAS, 2, 3));
     ASSERT_EQ(L::parse("xpu"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpux"), make_lc(D::UNSPEC, -1, 0));
     ASSERT_EQ(L::parse("xpu23"), make_lc(D::UNSPEC, 23, 0));
@@ -59,6 +66,8 @@ TEST(TestCompNode, Parse) {
     ASSERT_THROW(L::parse("cpu0:"), MegBrainError);
     ASSERT_THROW(L::parse("cpu0:x"), MegBrainError);
     ASSERT_THROW(L::parse("cpu2:23x"), MegBrainError);
+    ASSERT_THROW(L::parse("cmabricon0"), MegBrainError);
+    ASSERT_THROW(L::parse("atlast0"), MegBrainError);
     ASSERT_THROW(L::parse("multithread"), MegBrainError);
     ASSERT_THROW(L::parse("multithread1:"), MegBrainError);
     ASSERT_THROW(L::parse("multithread1:default"), MegBrainError);
@@ -129,13 +138,20 @@ TEST(TestCompNode, Load) {
         ASSERT_EQ(CompNode::load("gpux"), cn2);
         ASSERT_EQ(CompNode::load("gpu1"), cn3);
     }
+
+#if MGB_ATLAS
+    auto atlas0 = CompNode::load("atlas0");
+    auto atlas1 = CompNode::load("atlas1");
+    ASSERT_NE(atlas0, atlas1);
+#endif
 }
 
 TEST(TestCompNode, FreeAfterFinalize) {
     CompNode::finalize();
     for (size_t i = 0; i < CompNode::NR_DEVICE_TYPE; ++i) {
         auto type = static_cast<CompNode::DeviceType>(i);
-        if (!CompNode::get_device_count(type))
+        if (!check_device_type_avaiable(type) ||
+            !CompNode::get_device_count(type))
             continue;
         auto cn = CompNode::load(CompNode::Locator{type, -1, {0}});
         auto ptr = cn.alloc_device(123);
@@ -275,6 +291,30 @@ TEST(TestCompNodeCuda, Uid) {
 }
 
 
+#if MGB_CAMBRICON
+TEST(TestCompNodeCambricon, MemNode) {
+    REQUIRE_CAMBRICON_DEVICE(2);
+    auto cn00 = CompNode::load("cambricon0"),
+         cn1 = CompNode::load("cambricon1"),
+         cn01 = CompNode::load("cambricon0:1");
+    ASSERT_EQ(cn00, CompNode::load("cambricon0"));
+    ASSERT_EQ(cn00.mem_node(), cn01.mem_node());
+    ASSERT_NE(cn00.mem_node(), cn1.mem_node());
+}
+#endif
+
+#if MGB_ATLAS
+TEST(TestCompNodeAtlas, MemNode) {
+    auto cn00 = CompNode::load("atlas0"),
+         cn1 = CompNode::load("atlas1"),
+         cn01 = CompNode::load("atlas0:1");
+    ASSERT_EQ(cn00, CompNode::load("atlas0"));
+    ASSERT_EQ(cn00.mem_node(), cn01.mem_node());
+    ASSERT_NE(cn00.mem_node(), cn1.mem_node());
+}
+#endif
+
+
 TEST(TestCompNodeCPU, PhysicalDispatch) {
     constexpr int ID = 0x2a6453e0;
     using L = CompNode::Locator;
@@ -421,6 +461,41 @@ TEST(TestCompNodeCPU, PeerCopyFromCUDA) {
 }
 
 
+#if MGB_CAMBRICON
+TEST(TestCompNodeCPU, PeerCopyFromCambricon) {
+    REQUIRE_CAMBRICON_DEVICE(1);
+    REQUIRE_THREAD();
+    auto cn_gpu = CompNode::load("cambriconx");
+    auto cn_cpu = CompNode::load("cpux");
+
+    HostTensorGenerator<> gen;
+    auto a = gen({20, 3, 112, 112});
+    auto b = gen({20, 3, 112, 112});
+    auto c = gen({20, 3, 112, 112});
+    DeviceTensorND dev_a{cn_gpu}, dev_b{cn_cpu}, dev_c{cn_gpu};
+    dev_a.copy_from(*a).sync();
+    dev_b.copy_from(*b).sync();
+    dev_c.copy_from(*c).sync();
+
+    auto wait_event = cn_gpu.create_event();
+
+    dev_a.copy_from(dev_c);
+    wait_event->record();
+
+    cn_cpu.device_wait_event(*wait_event);
+    dev_b.copy_from(dev_a);
+
+    dev_b.sync();
+
+    HostTensorND result;
+    result.copy_from(dev_b);
+
+    CompNode::sync_all();
+
+    MGB_ASSERT_TENSOR_EQ(result, *c);
+}
+#endif
+
 TEST(TestCompNodeSyncManager, HostWait) {
     REQUIRE_THREAD();
     CompNodeSyncManager mgr(CompNode::load("xpu0"));
@@ -542,6 +617,8 @@ TEST(TestCompNode, MultipleLoad) {
     };
     for (size_t i = 1; i < CompNode::NR_DEVICE_TYPE; ++i) {
         auto dt = static_cast<CompNode::DeviceType>(i);
+        if (!check_device_type_avaiable(dt))
+            continue;
         if (CompNode::get_device_count(dt)) {
             auto cn = CompNode::load({dt, 0, {0}});
             mgb_log("comp node %s is available", cn.to_string().c_str());
@@ -552,6 +629,111 @@ TEST(TestCompNode, MultipleLoad) {
     }
 }
 
+#if MGB_CAMBRICON
+TEST(TestCompNodeCambricon, D2DCopy) {
+auto run = [](CompNode cn) {
+        constexpr size_t size = 100 * 1024 * 1024;
+        HostTensorND a(cn, {size}, dtype::Int32{}), b;
+        auto pa = a.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            pa[i] = i;
+        }
+        DeviceTensorND tmp, tmp1;
+        tmp.copy_from(a);
+        tmp1.copy_from(tmp);
+        b.copy_from(tmp1).sync();
+        auto pb = b.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            ASSERT_EQ(static_cast<int>(i), pb[i]);
+        }
+        CompNode::finalize();
+    };
+    REQUIRE_CAMBRICON_DEVICE(1);
+    auto cn = CompNode::load("cambricon0");
+    run(cn);
+    cn = CompNode::load("cambricon1");
+    run(cn);
+}
+
+// peer copy for cambricon between different devices is not correct now, so
+// disable this testcase
+#if 0
+TEST(TestCompNodeCambricon, P2PCopy) {
+    auto run_raw = []() {
+        int v0 = 0, v1 = 1;
+        cnrtDev_t dev0, dev1;
+        MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev0, 0));
+        MGB_CNRT_CHECK(cnrtGetDeviceHandle(&dev1, 1));
+        int *dp0, *dp1;
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev0));
+        MGB_CNRT_CHECK(cnrtMalloc((void**)(&dp0), sizeof(int)));
+        MGB_CNRT_CHECK(
+                cnrtMemcpy(dp0, &v0, sizeof(int), CNRT_MEM_TRANS_DIR_HOST2DEV));
+        MGB_CNRT_CHECK(cnrtSetCurrentDevice(dev1));
+        MGB_CNRT_CHECK(cnrtMalloc((void**)(&dp1), sizeof(int)));
+        MGB_CNRT_CHECK(
+                cnrtMemcpy(dp1, &v1, sizeof(int), CNRT_MEM_TRANS_DIR_HOST2DEV));
+        unsigned int can = 0;
+        MGB_CNRT_CHECK(cnrtGetPeerAccessibility(&can, 0, 1));
+        printf("can = %s\n", can ? "TRUE" : "FALSE");
+        if (can) {
+            MGB_CNRT_CHECK(cnrtMemcpyPeer(dp1, 1, dp0, 0, sizeof(int)));
+            int get;
+            MGB_CNRT_CHECK(cnrtMemcpy(&get, dp1, sizeof(int),
+                                      CNRT_MEM_TRANS_DIR_DEV2HOST));
+            ASSERT_EQ(0, get);
+        }
+    };
+    auto run = [](CompNode cn0, CompNode cn1) {
+        constexpr size_t size = 100;
+        HostTensorND a(cn0, {size}, dtype::Int32{}), b;
+        auto pa = a.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            pa[i] = i;
+        }
+        DeviceTensorND tmp(cn0, {size}, dtype::Int32{}),
+                tmp1(cn1, {size}, dtype::Int32{});
+        tmp.copy_from(a);
+        tmp1.copy_from(tmp);
+        b.copy_from(tmp1).sync();
+        auto pb = b.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            ASSERT_EQ(static_cast<int>(i), pb[i]);
+        }
+        CompNode::finalize();
+    };
+    REQUIRE_CAMBRICON_DEVICE(2);
+    auto cn0 = CompNode::load("cambricon0"), cn1 = CompNode::load("cambricon1");
+    run_raw();
+    run(cn0, cn1);
+}
+#endif
+#endif // MGB_CAMBRICON
+
+#if MGB_ATLAS
+
+TEST(TestCompNodeAtlas, D2DCopy) {
+    auto run = [](CompNode cn) {
+        constexpr size_t size = 10 * 1024 * 1024;
+        HostTensorND a(cn, {size}, dtype::Int32{}), b;
+        auto pa = a.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            pa[i] = i;
+        }
+        DeviceTensorND tmp, tmp1;
+        tmp.copy_from(a);
+        tmp1.copy_from(tmp);
+        b.copy_from(tmp1).sync();
+        auto pb = b.ptr<int>();
+        for (size_t i = 0; i < size; ++i) {
+            ASSERT_EQ(static_cast<int>(i), pb[i]);
+        }
+        CompNode::finalize();
+    };
+    auto cn = CompNode::load("atlas0");
+    run(cn);
+}
+#endif
 
 namespace {
 class CompNodeDepedentObjectInst final : public CompNodeDepedentObject {

src/megbrain_build_config.h.in

@@ -13,6 +13,8 @@
 #define _HEADER_MGB_BUILD_CONFIG
 
 #cmakedefine01 MGB_CUDA
+#cmakedefine01 MGB_CAMBRICON
+#cmakedefine01 MGB_ATLAS
 #cmakedefine01 MGB_ASSERT_LOC
 #cmakedefine01 MGB_ENABLE_DEBUG_UTIL
 #cmakedefine01 MGB_ENABLE_LOGGING
@@ -54,6 +56,10 @@
 #cmakedefine01 MEGDNN_THREADS_512
 #cmakedefine01 MEGDNN_ENABLE_MULTI_THREADS
 
+// whether atlas is available
+#ifndef MGB_ATLAS
+#define MGB_ATLAS  0
+#endif
 
 // whether cuda is available
 #ifndef MGB_CUDA
@@ -135,6 +141,15 @@
 #endif
 
 
+#ifndef MEGDNN_WITH_CAMBRICON
+#define MEGDNN_WITH_CAMBRICON 0
+#endif
+
+#ifndef MGB_CAMBRICON
+#define MGB_CAMBRICON MEGDNN_WITH_CAMBRICON
+#endif
+
+
 // whether to enable TensorRT support
 #ifndef MGB_ENABLE_TENSOR_RT
 #define MGB_ENABLE_TENSOR_RT    MGB_CUDA

src/opr/impl/atlas_runtime_op.cpp

@@ -0,0 +1,486 @@
+/**
+ * \file src/opr/impl/atlas_runtime_op.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megbrain/opr/atlas_runtime_op.h"
+#include <memory>
+#include "megbrain/common.h"
+#include "megbrain/graph/operator_node.h"
+#include "megdnn/basic_types.h"
+#include "megdnn/dtype.h"
+
+#if MGB_ATLAS
+#include "acl/acl_mdl.h"
+
+using namespace mgb;
+using namespace opr;
+
+namespace {
+/**
+ * \brief get mgb shape from acl shape, batch from mgb
+ */
+TensorShape acl_shape_to_mgb_shape_for_output(aclmdlIODims acl_shape,
+                                              size_t batch) {
+    TensorShape ret;
+    ret.ndim = acl_shape.dimCount;
+    for (size_t i = 0; i < ret.ndim; ++i) {
+        ret[i] = acl_shape.dims[i];
+    }
+    ret[0] = batch;
+    return ret;
+}
+
+/**
+ * \brief deduce the input shape from aclFormat and aipp config.
+ *
+ * \param acl_shape shape from om file
+ * \param batch batchsize from mgb
+ * \param enable_dynamic_batch True if set dynamic batch size
+ * \param om_format layout format from om file
+ * \param aipp_input_fmt input_format in static aipp config of om file
+ */
+TensorShape acl_shape_to_mgb_shape_for_input(
+        aclmdlIODims acl_shape, size_t batch, bool enable_dynamic_batch,
+        aclFormat om_format, AtlasRuntimeOpr::AippInputFormat aipp_input_fmt) {
+    TensorShape ret;
+    ret.ndim = acl_shape.dimCount;
+    mgb_assert(ret.ndim == 4,
+               "Unexpected ndim form aclmdlIODims expected 4, but got %zu",
+               ret.ndim);
+    for (size_t i = 0; i < ret.ndim; ++i) {
+        ret[i] = acl_shape.dims[i];
+    }
+    if (enable_dynamic_batch) {
+        mgb_assert(ret[0] == static_cast<size_t>(-1),
+                   "batch size expected to be -1 when enable dynamic "
+                   "batchsize, got: %zu\n",
+                   ret[0]);
+        ret[0] = batch;
+    } else {
+        mgb_assert(ret[0] == batch,
+                   "batchsize mismatch if no dynamic batchsize enabled, "
+                   "expected: %zu got: %zu\n",
+                   ret[0], batch);
+    }
+
+    if (aipp_input_fmt != AtlasRuntimeOpr::AippInputFormat::NO_AIPP) {
+        mgb_assert(om_format == ACL_FORMAT_NHWC,
+                   "om format should be NHWC if enable aipp");
+    }
+
+    return ret;
+}
+
+DType acl_dtype_to_mgb_dtype(aclDataType data_type) {
+    switch (data_type) {
+        case ACL_UINT8:
+            return dtype::Uint8();
+        case ACL_FLOAT16:
+#if !MEGDNN_DISABLE_FLOAT16
+            return dtype::Float16();
+#else
+            mgb_throw(MegBrainError,
+                      "Float16 support is disabled at compile time.");
+#endif
+        case ACL_FLOAT:
+            return dtype::Float32();
+        case ACL_INT8:
+            return dtype::Int8();
+        case ACL_INT16:
+            return dtype::Int16();
+        case ACL_INT32:
+            return dtype::Int32();
+        default:
+            mgb_throw(MegBrainError,
+                      "aclDataType %x is not supported by MegBrain.",
+                      static_cast<int>(data_type));
+    }
+}
+
+/**
+ * \brief generate batch size which match the batch_choice
+ */
+SmallVector<size_t> gen_batch_vec(size_t origin_batch,
+                                  const SmallVector<size_t>& batch_choices) {
+    SmallVector<size_t> ret;
+    size_t idx = 0;
+    size_t nr_batch_choices = batch_choices.size();
+    size_t batch = origin_batch;
+    while (idx < nr_batch_choices) {
+        size_t val = batch_choices[idx];
+        while (batch >= batch_choices[idx]) {
+            ret.push_back(val);
+            batch -= val;
+        }
+        idx++;
+    }
+    mgb_assert(batch == 0,
+               "Invalid batch size %zu, can not be generate by batch choices",
+               origin_batch);
+
+    return ret;
+}
+
+class PtrGetter {
+public:
+    PtrGetter(const VarNodeArray& vars) {
+        for (auto&& var : vars) {
+            m_ptrs.push_back(var->dev_tensor().raw_ptr());
+            m_batch_in_bytes.push_back(var->layout().stride[0] *
+                                       var->layout().dtype.size());
+        }
+    }
+
+    std::pair<void*, size_t> get(size_t batch, size_t idx) {
+        std::pair<void*, size_t> ret;
+        ret.first = m_ptrs[idx];
+        ret.second = batch * m_batch_in_bytes[idx];
+        m_ptrs[idx] = reinterpret_cast<void*>(
+                reinterpret_cast<uintptr_t>(ret.first) + ret.second);
+        return ret;
+    }
+
+private:
+    SmallVector<void*> m_ptrs;
+    SmallVector<size_t> m_batch_in_bytes;
+};
+
+};  // namespace
+
+/* ====================== AtlasRuntimeOpr ==================== */
+MGB_DYN_TYPE_OBJ_FINAL_IMPL(AtlasRuntimeOpr);
+AtlasRuntimeOpr::AtlasRuntimeOpr(SharedBuffer buf,
+                                 const std::pair<uint32_t, aclmdlDesc*>& model,
+                                 const VarNodeArray& inputs,
+                                 const OperatorNodeConfig& config)
+        : Super(inputs[0]->owner_graph(), config, "atlas_runtime", inputs),
+          m_buffer{std::move(buf)},
+          m_model_id{model.first},
+          m_model_desc{model.second} {
+    mgb_assert(
+            inputs[0]->comp_node().device_type() == CompNode::DeviceType::ATLAS,
+            "AtlasRuntimeOpr can only be used on atlas comp node; "
+            "got %s",
+            inputs[0]->comp_node().to_string().c_str());
+    mgb_assert(m_buffer.data() != nullptr ||
+               (m_model_id != INVALID_MODEL_ID && m_model_desc != nullptr));
+
+    for (auto i : inputs) {
+        add_input({i});
+    }
+    if (m_model_id == INVALID_MODEL_ID && m_model_desc == nullptr) {
+        MGB_ATLAS_CHECK(aclmdlLoadFromMem(m_buffer.data(), m_buffer.size(),
+                                          &m_model_id));
+        m_model_desc = aclmdlCreateDesc();
+        MGB_ATLAS_CHECK(aclmdlGetDesc(m_model_desc, m_model_id));
+        m_is_model_holder = true;
+    }
+
+    //! aipp input format
+    m_aipp_input_format = SmallVector<AippInputFormat>(inputs.size());
+    aclAippInfo aipp_info;
+    for (size_t i = 0; i < inputs.size(); ++i) {
+        aclError acl_err = aclmdlGetFirstAippInfo(m_model_id, i, &aipp_info);
+        if (ACL_ERROR_NONE == acl_err) {
+            switch (aipp_info.inputFormat) {
+                case ACL_YUV420SP_U8:
+                    m_aipp_input_format[i] = AippInputFormat::YUV420SP_U8;
+                    break;
+                case ACL_RGB888_U8:
+                    m_aipp_input_format[i] = AippInputFormat::RGB888_U8;
+                    break;
+                default:
+                    mgb_throw(MegBrainError,
+                              "Unsupported aclAippInputFormat for input %zu. ",
+                              i);
+            }
+        } else if (ACL_ERROR_NOT_STATIC_AIPP == acl_err) {
+            m_aipp_input_format[i] = AippInputFormat::NO_AIPP;
+        } else {
+            MGB_ATLAS_CHECK(acl_err);
+        }
+    }
+
+    size_t dynamic_index;
+    auto errcode = aclmdlGetInputIndexByName(
+            m_model_desc, ACL_DYNAMIC_TENSOR_NAME, &dynamic_index);
+    if (errcode == ACL_ERROR_NONE) {
+        aclmdlHW hw_info;
+        MGB_ATLAS_CHECK(
+                aclmdlGetDynamicHW(m_model_desc, dynamic_index, &hw_info));
+        mgb_assert(hw_info.hwCount == 0, "Currently not support dynamic HW");
+    }
+
+    //! dynamic batch size
+    aclmdlBatch acl_batch;
+    MGB_ATLAS_CHECK(aclmdlGetDynamicBatch(m_model_desc, &acl_batch));
+    if (acl_batch.batchCount) {
+        size_t dynamic_data_size;
+        dynamic_data_size =
+                aclmdlGetInputSizeByIndex(m_model_desc, dynamic_index);
+        m_dyn_batch_tensor = DeviceTensorND(
+                inputs[0]->comp_node(), {{dynamic_data_size}, dtype::Uint8()});
+
+        for (size_t i = 0; i < acl_batch.batchCount; ++i) {
+            m_dyn_batch_choices.push_back(
+                    static_cast<size_t>(acl_batch.batch[i]));
+        }
+        std::sort(m_dyn_batch_choices.begin(), m_dyn_batch_choices.end(),
+                  std::greater<>());
+    }
+
+    //! add output
+    size_t nr_outputs = aclmdlGetNumOutputs(m_model_desc);
+    using F = VarNode::Flag;
+    if (nr_outputs == 1) {
+        add_output(None);
+    } else {
+        for (size_t i = 0; i < nr_outputs; ++i) {
+            add_output(ssprintf("o%zu", i));
+        }
+    }
+    if (!m_dyn_batch_choices.empty()) {
+        /**
+         * \warning If enable dynamic batchsize, the memory of output
+         * should be the largest be the size with the largest batch_size, so we
+         * set the flag to SYS_MEM_ALLOC.
+         */
+        for (size_t i = 0; i < nr_outputs; ++i) {
+            output(i)->add_flag(F::NO_SYS_MEM_ALLOC);
+        }
+    }
+    add_equivalence_component<mgb::ScalarHash<const void*>>(m_buffer.data());
+};
+
+AtlasRuntimeOpr::~AtlasRuntimeOpr() {
+    if (m_is_model_holder) {
+        MGB_ATLAS_CHECK(aclmdlUnload(m_model_id));
+        MGB_ATLAS_CHECK(aclmdlDestroyDesc(m_model_desc));
+    }
+}
+
+void AtlasRuntimeOpr::scn_do_execute() {
+    auto&& acl_env =
+            CompNodeEnv::from_comp_node(input(0)->comp_node()).atlas_env();
+    acl_env.activate();
+
+    if (!m_dyn_batch_choices.empty()) {
+        for (size_t i = 0; i < output().size(); i++) {
+            auto output_size = aclmdlGetOutputSizeByIndex(m_model_desc, i);
+            auto ovar = output(i);
+            ovar->shape_alloc(ovar->shape(), output_size);
+        }
+    }
+
+    PtrGetter input_getter(input());
+    PtrGetter output_getter(output());
+
+    bool enable_dynamic_batch = !m_dyn_batch_choices.empty();
+    size_t nr_inputs = aclmdlGetNumInputs(m_model_desc);
+    size_t nr_outputs = aclmdlGetNumOutputs(m_model_desc);
+    size_t input_batch = input(0)->layout()[0];
+
+    if (enable_dynamic_batch) {
+        mgb_assert(nr_inputs == input().size() + 1,
+                   "nr inputs got from om model should be one more than got "
+                   "from megbrain");
+    }
+    SmallVector<size_t> batches_each_run;
+    if (enable_dynamic_batch) {
+        batches_each_run = gen_batch_vec(input_batch, m_dyn_batch_choices);
+    } else {
+        batches_each_run.push_back(input_batch);
+    }
+
+    for (auto&& batch : batches_each_run) {
+        //! prepare input
+        auto model_inputs = aclmdlCreateDataset();
+        mgb_assert(model_inputs != nullptr,
+                   "failed to create atlas input dataset.");
+        for (size_t i = 0; i < input().size(); i++) {
+            auto value_pair = input_getter.get(batch, i);
+            auto input_size = aclmdlGetInputSizeByIndex(m_model_desc, i);
+            if (enable_dynamic_batch) {
+                mgb_assert(input_size == value_pair.second / batch *
+                                                 m_dyn_batch_choices[0],
+                           "input %zu size mismatch, expected: %zu got: %zu", i,
+                           input_size,
+                           value_pair.second / batch *
+                                   m_dyn_batch_choices[0]);
+            }
+            aclDataBuffer* input_db =
+                    aclCreateDataBuffer(value_pair.first, value_pair.second);
+            mgb_assert(input_db != nullptr,
+                       "failed to create atlas input data buffer for input "
+                       "%zu:%s.",
+                       i, input(i)->cname());
+            aclmdlAddDatasetBuffer(model_inputs, input_db);
+        }
+        //! append unit tensor for dynamic batch
+        if (enable_dynamic_batch) {
+            aclDataBuffer* input_db = aclCreateDataBuffer(
+                    reinterpret_cast<void*>(m_dyn_batch_tensor.raw_ptr()),
+                    m_dyn_batch_tensor.layout().span().dist_byte());
+            mgb_assert(input_db != nullptr,
+                       "failed to create atlas input data buffer for dynamic "
+                       "batch tensor.");
+            MGB_ATLAS_CHECK(aclmdlAddDatasetBuffer(model_inputs, input_db));
+
+            MGB_ATLAS_CHECK(aclmdlSetDynamicBatchSize(
+                    m_model_id, model_inputs, input().size(),
+                    static_cast<uint64_t>(batch)));
+        }
+
+        //! prepare output
+        auto model_outputs = aclmdlCreateDataset();
+        mgb_assert(model_outputs != nullptr,
+                   "failed to create atlas output dataset.");
+        for (size_t i = 0; i < nr_outputs; i++) {
+            auto value_pair = output_getter.get(batch, i);
+            aclDataBuffer* output_db =
+                    aclCreateDataBuffer(value_pair.first, value_pair.second);
+            mgb_assert(output_db != nullptr,
+                       "failed to create atlas output data buffer for output "
+                       "%zu:%s.",
+                       i, output(i)->cname());
+            aclmdlAddDatasetBuffer(model_outputs, output_db);
+        }
+        MGB_ATLAS_CHECK(aclmdlExecute(m_model_id, model_inputs, model_outputs));
+
+        for (size_t i = 0; i < nr_inputs; ++i) {
+            aclDataBuffer* db_ptr = aclmdlGetDatasetBuffer(model_inputs, i);
+            MGB_ATLAS_CHECK(aclDestroyDataBuffer(db_ptr));
+        }
+        for (size_t i = 0; i < nr_outputs; ++i) {
+            aclDataBuffer* db_ptr = aclmdlGetDatasetBuffer(model_outputs, i);
+            MGB_ATLAS_CHECK(aclDestroyDataBuffer(db_ptr));
+        }
+        MGB_ATLAS_CHECK(aclmdlDestroyDataset(model_inputs));
+        MGB_ATLAS_CHECK(aclmdlDestroyDataset(model_outputs));
+    }
+}
+
+void AtlasRuntimeOpr::get_output_var_shape(const TensorShapeArray& inp_shape,
+                                           TensorShapeArray& out_shape) const {
+    size_t nr_inputs = aclmdlGetNumInputs(m_model_desc);
+    size_t batch_size = inp_shape[0][0];
+    //! enable dynamic batchsize
+    if (!m_dyn_batch_choices.empty()) {
+        mgb_assert(!gen_batch_vec(batch_size, m_dyn_batch_choices).empty());
+        mgb_assert(nr_inputs == inp_shape.size() + 1,
+                   "nr inputs got from om model should be one more than got "
+                   "from megbrain");
+    }
+    for (size_t i = 0; i < inp_shape.size(); ++i) {
+        aclmdlIODims input_dims;
+        MGB_ATLAS_CHECK(aclmdlGetInputDimsV2(m_model_desc, i, &input_dims));
+        auto om_format = aclmdlGetInputFormat(m_model_desc, i);
+        TensorShape shape_from_om = acl_shape_to_mgb_shape_for_input(
+                input_dims, batch_size, !m_dyn_batch_choices.empty(), om_format,
+                m_aipp_input_format[i]);
+        mgb_assert(shape_from_om.eq_shape(inp_shape[i]),
+                   "shape mismatch of input %zu, expected: %s got: %s", i,
+                   shape_from_om.to_string().c_str(),
+                   inp_shape[i].to_string().c_str());
+    }
+
+    for (size_t i = 0; i < out_shape.size(); ++i) {
+        aclmdlIODims output_dims;
+        MGB_ATLAS_CHECK(aclmdlGetOutputDims(m_model_desc, i, &output_dims));
+        out_shape[i] =
+                acl_shape_to_mgb_shape_for_output(output_dims, batch_size);
+    }
+}
+
+void AtlasRuntimeOpr::add_input_layout_constraint() {
+    //! default contiguous
+    for (auto i : input()) {
+        i->add_layout_constraint_contiguous();
+    }
+}
+
+void AtlasRuntimeOpr::init_output_dtype() {
+    DType dt_acl, dt_input;
+    for (size_t i = 0; i < input().size(); ++i) {
+        dt_acl =
+                acl_dtype_to_mgb_dtype(aclmdlGetInputDataType(m_model_desc, i));
+        dt_input = input(i)->dtype();
+        mgb_assert(dt_acl.valid() && dt_input.valid() &&
+                           dt_acl.enumv() == dt_input.enumv(),
+                   "dtype mismatch of input %zu: expected %s, "
+                   "got %s",
+                   i, dt_acl.name(), dt_input.name());
+    }
+
+    for (size_t i = 0; i < output().size(); ++i) {
+        dt_acl = acl_dtype_to_mgb_dtype(
+                aclmdlGetOutputDataType(m_model_desc, i));
+        mgb_assert(dt_acl.valid(),
+                   "output dtype checking failed: invalid dtype returned.");
+        if (dt_acl.enumv() == DTypeEnum::QuantizedS8) {
+            mgb_assert(output(i)->dtype().valid(),
+                       "user should specify scale of output tensor of "
+                       "AtlasRuntimeOpr.");
+        }
+        if (!output(i)->dtype().valid())
+            output(i)->dtype(dt_acl);
+    }
+}
+
+SymbolVarArray AtlasRuntimeOpr::make(SharedBuffer buf,
+                                     const SymbolVarArray& src,
+                                     const OperatorNodeConfig& config) {
+    VarNodeArray var_node_array = cg::to_var_node_array(src);
+    auto atlas_runtime_opr = std::make_unique<AtlasRuntimeOpr>(
+            std::move(buf),
+            std::pair<uint32_t, aclmdlDesc*>{INVALID_MODEL_ID, nullptr},
+            var_node_array, config);
+    auto ret = cg::to_symbol_var_array(
+            src[0].node()
+                    ->owner_graph()
+                    ->insert_opr(std::move(atlas_runtime_opr))
+                    ->output());
+    return ret;
+}
+
+SymbolVarArray AtlasRuntimeOpr::make(const void* buf, size_t size,
+                                     const SymbolVarArray& src,
+                                     const OperatorNodeConfig& config) {
+    mgb_throw_if(!CompNode::get_device_count(CompNode::DeviceType::ATLAS),
+                 SystemError,
+                 "can not create AtlasRuntimeOpr when atlas is not "
+                 "available");
+    std::shared_ptr<uint8_t> shptr{new uint8_t[size],
+                                   [](uint8_t* p) { delete[] p; }};
+    memcpy(shptr.get(), buf, size);
+    SharedBuffer buffer{std::move(shptr), size};
+    return make(std::move(buffer), src, config);
+}
+
+SymbolVarArray AtlasRuntimeOpr::make(
+        const SharedBuffer buf, const std::pair<uint32_t, aclmdlDesc*>& model,
+        const SymbolVarArray& src, const OperatorNodeConfig& config) {
+    VarNodeArray var_node_array = cg::to_var_node_array(src);
+    auto atlas_runtime_opr = std::make_unique<AtlasRuntimeOpr>(
+            buf, model, var_node_array, config);
+    auto ret = cg::to_symbol_var_array(
+            src[0].node()
+                    ->owner_graph()
+                    ->insert_opr(std::move(atlas_runtime_opr))
+                    ->output());
+    return ret;
+}
+
+constexpr uint32_t AtlasRuntimeOpr::INVALID_MODEL_ID;
+
+#endif  // MGB_atlas
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/atlas_runtime_op.oprdecl

@@ -0,0 +1,17 @@
+decl_raw_opr(
+    'atlas_runtime',
+    desc='create an operator that could load and run acl offline model',
+    inputs=[
+        Doc('inputs', 'input vars', 'list of :class:`.SymbolVar`'),
+        Doc('data_bytes', 'serialized acl model'),
+    ],
+    body=[
+        'assert isinstance(data_bytes, bytes), '
+            '"data must be bytes; got {}".format(type(data_bytes))',
+        'output = _mgb._Opr.atlas_runtime(inputs, data_bytes, config)',
+        'cvt_result_kwargs["explode_single"] = False',
+    ],
+)
+
+# vim: ft=python
+

src/opr/impl/atlas_runtime_op.sereg.h

@@ -0,0 +1,68 @@
+/**
+ * \file src/opr/impl/atlas_runtime_opr.sereg.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megbrain/opr/atlas_runtime_op.h"
+#include "megbrain/serialization/sereg.h"
+
+#if MGB_ATLAS
+namespace mgb {
+
+namespace serialization {
+
+template <>
+struct OprLoadDumpImpl<opr::AtlasRuntimeOpr, 0> {
+
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        auto&& opr = opr_.cast_final_safe<opr::AtlasRuntimeOpr>();
+        auto&& buf = opr.buffer();
+        ctx.dump_buf_with_len(buf.data(), buf.size());
+    }
+    static cg::OperatorNodeBase* load(OprLoadContext& ctx,
+                                      const cg::VarNodeArray& inputs,
+                                      const OperatorNodeConfig& config) {
+        inputs.at(0)->comp_node().activate();
+        auto buf = ctx.load_shared_buf_with_len();
+        return opr::AtlasRuntimeOpr::make(
+                       std::move(buf), cg::to_symbol_var_array(inputs),
+                       config)
+                .at(0)
+                .node()
+                ->owner_opr();
+    }
+};
+}  // namespace serialization
+
+namespace opr {
+
+cg::OperatorNodeBase* opr_shallow_copy_atlas_runtime_opr(
+        const serialization::OprShallowCopyContext& ctx,
+        const cg::OperatorNodeBase& opr_, const VarNodeArray& inputs,
+        const OperatorNodeConfig& config) {
+    MGB_MARK_USED_VAR(ctx);
+    auto&& opr = opr_.cast_final_safe<AtlasRuntimeOpr>();
+    return AtlasRuntimeOpr::make(opr.buffer(), opr.model(),
+                                 cg::to_symbol_var_array(inputs),
+                                 config)
+            .at(0)
+            .node()
+            ->owner_opr();
+}
+
+MGB_SEREG_OPR(AtlasRuntimeOpr, 0);
+MGB_REG_OPR_SHALLOW_COPY(AtlasRuntimeOpr, opr_shallow_copy_atlas_runtime_opr);
+
+}  // namespace opr
+}  // namespace mgb
+
+#endif
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/include/megbrain/opr/atlas_runtime_op.h

@@ -0,0 +1,87 @@
+/**
+ * \file src/opr/include/megbrain/opr/atlas_runtime_op.h
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#pragma once
+
+#include <memory>
+#include "megbrain/comp_node_env.h"
+#include "megbrain/graph.h"
+#include "megbrain/serialization/file.h"
+
+#if MGB_ATLAS
+#include "acl/acl.h"
+
+namespace mgb {
+namespace opr {
+
+MGB_DEFINE_OPR_CLASS(AtlasRuntimeOpr,
+                     cg::SingleCNOutshapePureByInshapeOprBase) // {
+public:
+    using SharedBuffer = mgb::serialization::SharedBuffer;
+
+    enum AippInputFormat {NO_AIPP, YUV420SP_U8, RGB888_U8};
+
+    void scn_do_execute() override;
+    void get_output_var_shape(const TensorShapeArray& inp_shape,
+                              TensorShapeArray& out_shape) const override;
+    void add_input_layout_constraint() override;
+    void init_output_dtype() override;
+
+    /**
+     * \brief create AtlasRuntimeOpr with buf or set model with
+     * a existance model.
+     *
+     * \brief Neither buf is set or model_id&model_desc is set
+     */
+    AtlasRuntimeOpr(SharedBuffer buf,
+                    const std::pair<uint32_t, aclmdlDesc*>& model,
+                    const VarNodeArray& inputs,
+                    const OperatorNodeConfig& config);
+    ~AtlasRuntimeOpr();
+
+    const SharedBuffer& buffer() const { return m_buffer; }
+
+    std::pair<uint32_t, aclmdlDesc*> model() const {
+        return {m_model_id, m_model_desc};
+    }
+
+    static SymbolVarArray make(SharedBuffer buf, const SymbolVarArray& src,
+                               const OperatorNodeConfig& config = {});
+
+    static SymbolVarArray make(SharedBuffer buf,
+                               const std::pair<uint32_t, aclmdlDesc*>& model,
+                               const SymbolVarArray& src,
+                               const OperatorNodeConfig& config = {});
+
+    static SymbolVarArray make(const void* buf, size_t size,
+                               const SymbolVarArray& src,
+                               const OperatorNodeConfig& config = {});
+
+private:
+    SharedBuffer m_buffer;
+    constexpr static uint32_t INVALID_MODEL_ID = -1;
+    uint32_t m_model_id = INVALID_MODEL_ID;
+    aclmdlDesc* m_model_desc = nullptr;
+    //! if set true, it will release model
+    bool m_is_model_holder = false;
+    SmallVector<AippInputFormat> m_aipp_input_format;
+    //! Atlas need a 64bit device tensor to hold dynamic batch state
+    DeviceTensorND m_dyn_batch_tensor;
+    SmallVector<size_t> m_dyn_batch_choices;
+};
+
+}  // namespace opr
+}  // namespace mgb
+
+#endif  // MGB_ATLAS
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/test/atlas_runtime_op.cpp

@@ -0,0 +1,202 @@
+/**
+ * \file src/opr/test/atlas_runtime_op.cpp
+ * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+ *
+ * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
+ * implied.
+ */
+
+#include "megbrain/opr/tensor_manip.h"
+#include "megdnn/dtype.h"
+#if MGB_ATLAS
+
+#include "megbrain/comp_node_env.h"
+#include "megbrain/opr/io.h"
+#include "megbrain/opr/basic_arith.h"
+#include "megbrain/test/helper.h"
+
+#include "megbrain/opr/atlas_runtime_op.h"
+#include "megbrain/serialization/serializer.h"
+#include "megbrain/plugin/profiler.h"
+
+#include <random>
+#include <vector>
+#include <stdio.h>
+
+#include "./atlas_models.h"
+
+using namespace mgb;
+using namespace opr;
+using namespace serialization;
+
+TEST(TestOprAtlas, Basic) {
+    HostTensorGenerator<> gen;
+    const auto& graph = ComputingGraph::make();
+    const auto& host_x = gen({4, 3, 16, 16});
+
+    //! run om model
+    const auto& om_buffer = ATLAS_MODEL.at("model_om");
+    auto cn = CompNode::load("atlas0");
+    auto x = Host2DeviceCopy::make(*graph, host_x, cn);
+    auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                        {x})[0];
+    HostTensorND host_om;
+    auto om_func = graph->compile({make_callback_copy(y, host_om, true)});
+    om_func->execute().wait();
+
+    //! run mdl model
+    const auto& mdl_buffer = ATLAS_MODEL.at("model_mdl");
+    auto loader = GraphLoader::make(
+            InputFile::make_mem_proxy(mdl_buffer.first, mdl_buffer.second));
+    auto rst = loader->load();
+    auto input = rst.tensor_map.at("d");
+    input->copy_from(*host_x).sync();
+    HostTensorND host_mdl;
+    auto mgb_func = rst.graph_compile(
+            {make_callback_copy(rst.output_var_list[0], host_mdl)});
+    mgb_func->execute().wait();
+
+    //! In atlas, the inner compute is fp16
+    MGB_ASSERT_TENSOR_NEAR(host_mdl, host_om, 1e-3);
+}
+
+TEST(TestOprAtlas, DynamicBatch) {
+    for (size_t batch : {1, 6}) {
+        HostTensorGenerator<> gen;
+        const auto& graph = ComputingGraph::make();
+        const auto& host_x = gen({batch, 3, 16, 16});
+
+        //! run om model
+        const auto& om_buffer = ATLAS_MODEL.at("model_dyn_om");
+        auto cn = CompNode::load("atlas0");
+        auto x = Host2DeviceCopy::make(*graph, host_x, cn);
+        auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                            {x})[0];
+        HostTensorND host_om;
+        auto om_func = graph->compile({make_callback_copy(y, host_om, true)});
+        om_func->execute().wait();
+
+        //! run mdl model
+        const auto& mdl_buffer = ATLAS_MODEL.at("model_mdl");
+        auto loader = GraphLoader::make(
+                InputFile::make_mem_proxy(mdl_buffer.first, mdl_buffer.second));
+        auto rst = loader->load();
+        auto input = rst.tensor_map.at("d");
+        input->copy_from(*host_x).sync();
+        HostTensorND host_mdl;
+        auto mgb_func = rst.graph_compile(
+                {make_callback_copy(rst.output_var_list[0], host_mdl)});
+        mgb_func->execute().wait();
+
+        //! In atlas, the inner compute is fp16
+        MGB_ASSERT_TENSOR_NEAR(host_mdl, host_om, 1e-3);
+    }
+}
+
+TEST(TestOprAtlas, Rgb888) {
+    HostTensorGenerator<dtype::Uint8, RandomDistribution::UNIFORM> gen;
+    const auto& graph = ComputingGraph::make();
+    const auto &host_x = gen({1, 3, 16, 16});
+
+    //! run om model
+    const auto& om_buffer = ATLAS_MODEL.at("model_rgb_om");
+    auto x = Host2DeviceCopy::make(*graph, host_x);
+    x = opr::Dimshuffle::make(x, {0, 2, 3, 1});
+    auto cn = CompNode::load("atlas0");
+    auto atlas_x = Copy::make(x, {cn});
+    auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                        {atlas_x})[0];
+    HostTensorND host_om;
+    auto om_func = graph->compile({make_callback_copy(y, host_om, true)});
+    om_func->execute().wait();
+
+    //! run mdl model
+    const auto& mdl_buffer = ATLAS_MODEL.at("model_aipp_mdl");
+    auto loader = GraphLoader::make(
+            InputFile::make_mem_proxy(mdl_buffer.first, mdl_buffer.second));
+    auto rst = loader->load();
+    auto input = rst.tensor_map.at("d");
+    input->copy_from(*host_x).sync();
+    HostTensorND host_mdl;
+    auto mgb_func = rst.graph_compile(
+            {make_callback_copy(rst.output_var_list[0], host_mdl)});
+    mgb_func->execute().wait();
+
+    //! In atlas, the inner compute is fp16
+    MGB_ASSERT_TENSOR_NEAR(host_mdl,
+                           host_om, 1e-3);
+}
+
+TEST(TestOprAtlas, Yuv) {
+    //! As YUV420SP depends on the input processed by AIPP, so here we just
+    //! check if the shape satisfy.
+    HostTensorGenerator<dtype::Uint8, RandomDistribution::UNIFORM> gen;
+    const auto& graph = ComputingGraph::make();
+    const auto &host_x = gen({1, 24, 16, 1});
+
+    //! run om model
+    const auto& om_buffer = ATLAS_MODEL.at("model_yuv_om");
+    auto cn = CompNode::load("atlas0");
+    auto x = Host2DeviceCopy::make(*graph, host_x, cn);
+    auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                        {x})[0];
+    HostTensorND host_om;
+    auto om_func = graph->compile({make_callback_copy(y, host_om, true)});
+    om_func->execute().wait();
+}
+
+TEST(TestOprAtlas, Serialization) {
+    using namespace serialization;
+
+    HostTensorGenerator<> gen;
+    const auto& graph = ComputingGraph::make();
+    const auto& host_x = gen({4, 3, 16, 16});
+
+    const auto& om_buffer = ATLAS_MODEL.at("model_om");
+    auto cn = CompNode::load("atlas0");
+    auto x = Host2DeviceCopy::make(*graph, host_x, cn);
+    auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                        {x})[0];
+
+    auto fname = output_file("AtlasRuntimeOprTest");
+    auto dump = [&]() {
+        auto dumper = GraphDumper::make(OutputFile::make_fs(fname.c_str()));
+        auto rst = dumper->dump({y});
+        ASSERT_EQ(rst.outputs.size(), 1u);
+    };
+    auto load = [&]() {
+        auto loader = GraphLoader::make(InputFile::make_fs(fname.c_str()));
+        auto rst = loader->load();
+        ASSERT_EQ(rst.output_var_list.size(), 1u);
+    };
+    dump();
+    load();
+}
+
+TEST(TestOprAtlas, Profiling) {
+    HostTensorGenerator<> gen;
+    const auto& graph = ComputingGraph::make();
+    GraphProfiler profiler{graph.get()};
+    const auto& host_x = gen({1, 3, 16, 16});
+
+    //! run om model
+    const auto& om_buffer = ATLAS_MODEL.at("model_dyn_om");
+    auto cn = CompNode::load("atlas0");
+    auto x = Host2DeviceCopy::make(*graph, host_x, cn);
+    auto y = opr::AtlasRuntimeOpr::make(om_buffer.first, om_buffer.second,
+                                        {x})[0];
+    HostTensorND host_om;
+    auto om_func = graph->compile({make_callback_copy(y, host_om, true)});
+    om_func->execute().wait();
+
+    profiler.to_json_full(om_func.get())
+            ->writeto_fpath(output_file("atlas_runtime_opr_profile.json"));
+}
+
+#endif  // MGB_ATLAS
+
+// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/plugin/test/opr_io_dump.cpp

@@ -118,6 +118,8 @@ void run_test(const PluginMaker& plugin_maker,
               const ResultChecker& result_checker) {
     for (size_t i = 1; i < CompNode::NR_DEVICE_TYPE; ++i) {
         auto type = static_cast<CompNode::DeviceType>(i);
+        if (!check_device_type_avaiable(type))
+            continue;
         if (CompNode::get_device_count(type)) {
             auto cn = CompNode::load({type, -1, 0});
             if (cn.contain_flag(CompNode::Flag::SUPPORT_RECORDER)) {
@@ -188,4 +190,3 @@ TEST(TestOprIODump, Binary) {
 }
 
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}
-

src/serialization/impl/sereg_caller.cpp

@@ -32,6 +32,12 @@ namespace mgb{void call_sereg(){}}
 #if MGB_ENABLE_TENSOR_RT
 #include "../../tensorrt/impl/tensorrt_opr.sereg.h"
 #endif
+#if MGB_ATLAS
+#include "../../opr/impl/atlas_runtime_op.sereg.h"
+#endif
 #if MGB_JIT
 #include "../../jit/impl/jit.sereg.h"
 #endif
+#if MGB_CAMBRICON
+#include "../../cambricon/impl/cambricon_runtime_opr.sereg.h"
+#endif

test/src/helper.cpp

@@ -310,6 +310,28 @@ bool mgb::check_gpu_available(size_t num) {
 }
 
 
+bool mgb::check_cambricon_device_available(size_t num) {
+    if (CompNode::get_device_count(CompNode::DeviceType::CAMBRICON) < num) {
+        mgb_log_warn("skip test case that requires %zu cambricon device(s)",
+                     num);
+        return false;
+    }
+    return true;
+}
+
+bool mgb::check_device_type_avaiable(CompNode::DeviceType device_type) {
+    switch (device_type) {
+        case mgb::CompNode::DeviceType::CUDA:
+        case mgb::CompNode::DeviceType::CPU:
+        case mgb::CompNode::DeviceType::CAMBRICON:
+        case mgb::CompNode::DeviceType::ATLAS:
+        case mgb::CompNode::DeviceType::MULTITHREAD:
+            return true;
+        default:
+            return false;
+    }
+    return false;
+}
 
 bool mgb::check_compute_capability(int major, int minor) {
 #if MGB_CUDA

test/src/include/megbrain/test/helper.h

@@ -437,8 +437,6 @@ std::vector<CompNode> load_multiple_xpus(size_t num);
 //! check whether given number of GPUs is available
 bool check_gpu_available(size_t num);
 
-//! check whether given number of AMD GPUs is available
-bool check_amd_gpu_available(size_t num);
 
 //! check whether given number of cambricon devices is available
 bool check_cambricon_device_available(size_t num);
@@ -446,6 +444,9 @@ bool check_cambricon_device_available(size_t num);
 //! check current capability >= major.minor
 bool check_compute_capability(int major, int minor);
 
+//! check compnode avaiable
+bool check_device_type_avaiable(CompNode::DeviceType device_type);
+
 //! hook persistent cache get calls during the lifetime
 class PersistentCacheHook {
     class HookedImpl;
@@ -479,6 +480,12 @@ public:
         return; \
 } while(0)
 
+//! skip a testcase if cambricon device not available
+#define REQUIRE_CAMBRICON_DEVICE(n)               \
+    do {                                          \
+        if (!check_cambricon_device_available(n)) \
+            return;                               \
+    } while (0)
 
 #if MGB_HAVE_THREAD
 #define REQUIRE_THREAD()