Merge branch 'Cambricon:master' into nl22222

.github/workflows/daily.yaml

@@ -12,9 +12,9 @@ jobs:
     strategy:
       matrix:
         runner: [mlu370-m8]
-        mlu_ops_version : [1.3.1]
-        cntoolkit_version : [3.14.0]
-        cnnl_version: [1.26.6]
+        mlu_ops_version : [1.4.1]
+        cntoolkit_version : [3.15.2]
+        cnnl_version: [1.28.0]
     runs-on: ${{matrix.runner}}
     steps:
       - uses: actions/checkout@v3

.github/workflows/mluops_ci.yaml

@@ -39,7 +39,7 @@ jobs:
     strategy:
       matrix:
         runner: [mlu370-m8]
-        mlu_ops_version : [v1.3.1]
+        mlu_ops_version : [v1.4.1]
     runs-on: [yellow]
     steps:
       - uses: actions/checkout@v3

CMakeLists.txt

@@ -58,6 +58,9 @@ endif()
 
 # -- BANG memcheck
 #TODO remove this option after cntoolkit upgraded to 4.0
+
+set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} --no-neuware-version-check")  # no need after toolkit 4.0
+
 if(${MLUOP_BUILD_BANG_MEMCHECK} MATCHES "ON")
   message("-- BANG memcheck enabled")
   set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS} -mllvm -enable-mlisa-sanitizer")
@@ -137,7 +140,8 @@ endif()
 if (NOT MLUOP_MLU_ARCH_LIST)
   message(STATUS "build all arch")
   set(BANG_CNCC_FLAGS "${BANG_CNCC_FLAGS}" "--bang-mlu-arch=mtp_372"
-                                           "--bang-mlu-arch=mtp_592")
+                                           "--bang-mlu-arch=mtp_592"
+                                           "--bang-mlu-arch=mtp_613")
 else()
   foreach (arch ${MLUOP_MLU_ARCH_LIST})
     set(CNCC_FLAGS_ARCH ${CNCC_FLAGS_ARCH} "--bang-mlu-arch=${arch}" )

README.md

@@ -49,10 +49,11 @@ MLU-OPS™提供了以下功能：
 ## 依赖条件
 
 - 操作系统：
-  - 支持 x86_64 架构下的 Ubuntu20.04、Centos7.6、Centos8.5、Kylin10
-  - MLU-OPS™ v1.0.0版本后将不再支持 Ubuntu18.04。Ubuntu22.04系统将在后续的版本提供支持。
+  - 支持 x86_64 架构下的 Ubuntu22.04、Centos7.6、Centos8.5、Kylin10
+  - MLU-OPS™ v1.0.0版本后将不再支持 Ubuntu18.04。
+  - MLU-OPS™ v1.4.1版本后将不再支持 Ubuntu20.04。
 - 寒武纪 MLU SDK：
-  - 编译和运行时依赖 CNToolkit v3.14.0 或更高版本，CNNL v1.26.6 或者更高版本
+  - 编译和运行时依赖 CNToolkit v3.15.2 或更高版本，CNNL v1.27.4 或者更高版本
 - 寒武纪 MLU 驱动：
   - 运行时依赖驱动 v6.0.3 或更高版本
 - 外部链接库：

build.property

@@ -1,8 +1,8 @@
 {
-  "version": "1.3.1-1",
+  "version": "1.4.1-1",
   "python": "3.6.0",
-  "build_requires": {"cntoolkit": ["release","3.15.1-1"],
-                     "cnnl":["release","1.27.4-1"],
+  "build_requires": {"cntoolkit": ["release","3.15.2-1"],
+                     "cnnl":["release","1.28.0-1"],
                      "driver": "6.0.3",
                      "eigen3": "3.4.0",
                      "libxml2": "2.9.0",

core/context.cpp

@@ -29,9 +29,9 @@
 #include "core/tool.h"
 #include "kernels/kernel.h"
 
-#define DEP_CHECK_LOG(level)                                              \
+#define DEP_CHECK_LOG(level)                                                \
   mluop::logging::LogMessage(__FILE__, __LINE__, 4, level, "MLU-OPS", true, \
-                             true, true, true)                            \
+                             true, true, true)                              \
       .stream()
 
 namespace mluop {
@@ -46,27 +46,23 @@ static struct deviceName name_list_table[] = {
     // case.
 };
 
-// once cnrtGetDeviceProperties() update and not use
-// device_ordinal, update this funciton.
-mluOpDevType_t convertDeviceName(char *name) {
-  struct deviceName *pName = NULL;
-  int num = sizeof(name_list_table) / sizeof(struct deviceName);
-  if (CONTEXT_DEVICENAME_LEAST_SIZE > strlen(name)) {
-    LOG(ERROR)
-        << "get device name failed. device name too short. device name = "
-        << name << "\n";
-    return MLUOP_UNKNOWN_DEVICE;
-  }
-  for (int i = 0; i < num; i++) {
-    pName = &name_list_table[i];
-    if (0 == strncmp(pName->name, name, strlen(pName->name)) ||
-        (i == num - 1 &&
-         0 >= strncmp(pName->name, name, CONTEXT_DEVICENAME_LEAST_SIZE))) {
-      return pName->type;
+mluOpDevType_t convertDeviceNameFromInt(int device_code) {
+  switch (device_code) {
+    case 372: {
+      return MLUOP_MLU370;
+      break;
+    }
+    case 592: {
+      return MLUOP_MLU590;
+      break;
+    }
+    case 613: {
+      return MLUOP_MTP613;
+      break;
     }
+    default:
+      break;
   }
-  LOG(ERROR) << "get device name failed. return unknown device. device name = "
-             << name << "\n";
   return MLUOP_UNKNOWN_DEVICE;
 }
 }  // namespace mluop
@@ -179,6 +175,7 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
   int32_t persisting_l2cache_maxsize = 0;
   double memory_band_width = 0;
   char device_name[CONTEXT_DEVICENAME_BUFFER_SIZE] = "";
+  int device_code = 0;
   mluOpContext *ctx = new (std::nothrow) mluOpContext();
   CNcontext drv_ctx;
   CNctxConfigParam ctx_conf_param;
@@ -246,6 +243,11 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
           cnDeviceGetAttribute(&persisting_l2cache_maxsize,
                                CN_DEVICE_ATTRIBUTE_MAX_PERSISTING_L2_CACHE_SIZE,
                                mlu_dev));
+  INTERNAL_CHECK(
+      "[mluOpCreate]",
+      CN_SUCCESS == cnDeviceGetAttribute(&device_code,
+                                         CN_DEVICE_ATTRIBUTE_MLU_ISA_VERSION,
+                                         mlu_dev));
   INTERNAL_CHECK(
       "[mluOpCreate]",
       CN_SUCCESS == cnDeviceGetName(device_name, CONTEXT_DEVICENAME_BUFFER_SIZE,
@@ -266,8 +268,7 @@ mluOpStatus_t MLUOP_WIN_API mluOpCreate(mluOpHandle_t *handle) {
   }
 
   ctx->capability_job_limit = (int32_t)ctx_conf_param.unionLimit;
-  ctx->arch = mluop::convertDeviceName(
-      device_name);  // warning: possible return unknown.
+  ctx->arch = mluop::convertDeviceNameFromInt(device_code);
   ctx->sram_size = sram_size - REM_FOR_STACK;
 
   strncpy(ctx->device_name, device_name, sizeof(device_name));

core/context.h

@@ -57,6 +57,7 @@ typedef enum {
   MLUOP_MLU270 = 270,
   MLUOP_MLU370 = 372,
   MLUOP_MLU590 = 592,
+  MLUOP_MTP613 = 613,
   MLUOP_MLU290 = 290,
 } mluOpDevType_t;
 

core/gen_case.cpp

@@ -761,7 +761,7 @@ std::string descToString(mluOpTensorDescriptor_t desc, char delimiter) {
   tensor_info << "  layout: " << mluop::getNameOfTensorLayout(layout)
               << delimiter;
   tensor_info << "  dtype: " << mluop::getNameOfDataType(dtype) << delimiter;
-  if (desc->pointer_mode == MLUOP_POINTER_MODE_HOST) {
+  if (desc->getPointerMode() == MLUOP_POINTER_MODE_HOST) {
     tensor_info << "  pointer_mode: POINTER_MODE_HOST" << delimiter;
     if ((total_element_num != 1) || (dim != 0)) {
       LOG(WARNING) << "[gen_case] Tensor has been set to POINTER_MODE_HOST, "

core/gen_case.h

@@ -492,7 +492,7 @@ class PbNode {
     uint64_t data_size = total_num * mluop::getSizeOfDataType(dtype);
     void *data = malloc(data_size);
     auto memcpy_dir =
-        (tensors[index].desc->pointer_mode == MLUOP_POINTER_MODE_HOST
+        (tensors[index].desc->getPointerMode() == MLUOP_POINTER_MODE_HOST
              ? cnrtMemcpyHostToHost
              : cnrtMemcpyDevToHost);
     if (cnrtSuccess == cnrtMemcpy(data,