!2134 upgrade Ascend software package 8 Sep 22

Merge pull request !2134 from yanghaoran/master

inc/external/acl/acl_prof.h

@@ -64,7 +64,7 @@ typedef enum {
   ACL_SUBSCRIBE_OP = 0,
   ACL_SUBSCRIBE_SUBGRAPH = 1,
   ACL_SUBSCRIBE_OP_THREAD = 2,
-  ACL_SUBSCRIBE_NONE
+  ACL_SUBSCRIBE_NONE = 0xFF
 } aclprofSubscribeOpFlag;
 
 typedef enum { ACL_SUBSCRIBE_ATTRI_THREADID = 0, ACL_SUBSCRIBE_ATTRI_NONE } aclprofSubscribeOpAttri;

inc/external/acl/acl_tdt_queue.h

@@ -53,6 +53,8 @@ typedef enum {
   ACL_TDT_QUEUE_ROUTE_QUERY_DST_ID_UINT32
 } acltdtQueueRouteQueryInfoParamType;
 
+typedef enum { ACL_TDT_NORMAL_MEM = 0, ACL_TDT_DVPP_MEM } acltdtAllocBufType;
+
 /**
  * @ingroup AscendCL
  * @brief create queue
@@ -257,6 +259,53 @@ ACL_FUNC_VISIBILITY aclError acltdtFreeBuf(acltdtBuf buf);
  */
 ACL_FUNC_VISIBILITY aclError acltdtGetBufData(const acltdtBuf buf, void **dataPtr, size_t *size);
 
+/**
+ * @ingroup AscendCL
+ * @brief get private data buf address and size
+ *
+ * @param buf [IN] acltdtBuf
+ * @param dataPtr [IN/OUT] pointer to the user ptr
+ * @param size [IN] the current private data area size, less than or equal to 96B
+ * @param offset [IN] address offset, less than or equal to 96B
+ *
+ * @retval ACL_SUCCESS  The function is successfully executed.
+ * @retval OtherValues Failure
+ *
+ * @see acltdtGetBufUserData
+ */
+ACL_FUNC_VISIBILITY aclError acltdtGetBufUserData(const acltdtBuf buf, void *dataPtr, size_t size, size_t offset);
+
+/**
+ * @ingroup AscendCL
+ * @brief set private data buf address and size
+ *
+ * @param buf [OUT] acltdtBuf
+ * @param dataPtr [IN] pointer to the user ptr
+ * @param size [IN] the current private data area size, less than or equal to 96B
+ * @param offset [IN] address offset, less than or equal to 96B
+ *
+ * @retval ACL_SUCCESS  The function is successfully executed.
+ * @retval OtherValues Failure
+ *
+ * @see acltdtSetBufUserData
+ */
+ACL_FUNC_VISIBILITY aclError acltdtSetBufUserData(acltdtBuf buf, const void *dataPtr, size_t size, size_t offset);
+
+/**
+ * @ingroup AscendCL
+ * @brief copy buf ref
+ *
+ * @param buf [IN] acltdtBuf
+ * @param newBuf [OUT] Make a reference copy of the data area of buf and
+ *                     create a new buf header pointing to the same data area
+ *
+ * @retval ACL_SUCCESS  The function is successfully executed.
+ * @retval OtherValues Failure
+ *
+ * @see acltdtCopyBufRef
+ */
+ACL_FUNC_VISIBILITY aclError acltdtCopyBufRef(const acltdtBuf buf, acltdtBuf *newBuf);
+
 /**
  * @ingroup AscendCL
  * @brief Create the queue attr

inc/external/ge/ge_api_types.h

@@ -395,9 +395,6 @@ const std::string RECOMPUTE = "ge.recompute";
 
 const char_t *const GRAPH_PARALLEL_OPTION_PATH = "ge.graphParallelOptionPath";
 
-// 1: Complete graph resource evaluation(Inclusion graph fusion optimization)
-// 2: Simplified graph resource evaluation(No graph fusion optimization)
-const char_t *const EVALUATE_GRAPH_RESOURCE_MODE = "ge.evaluateGraphResourceMode";
 // Graph run mode
 enum GraphRunMode { PREDICTION = 0, TRAIN };
 // Topo sorting mode
@@ -518,8 +515,7 @@ const std::set<std::string> ir_builder_suppported_options = {INPUT_FORMAT,
                                                              SHAPE_GENERALIZED_BUILD_MODE,
                                                              MODIFY_MIXLIST,
                                                              CUSTOMIZE_DTYPES,
-                                                             BUILD_INNER_MODEL,
-                                                             EVALUATE_GRAPH_RESOURCE_MODE};
+                                                             BUILD_INNER_MODEL};
 
 // for interface: aclgrphParse
 const std::set<std::string> ir_parser_suppported_options = {

inc/external/hccl/hccl_types.h

@@ -88,6 +88,7 @@ typedef enum {
   HCCL_DATA_TYPE_UINT16 = 8, /**< uint16 */
   HCCL_DATA_TYPE_UINT32 = 9, /**< uint32 */
   HCCL_DATA_TYPE_FP64 = 10,  /**< fp64 */
+  HCCL_DATA_TYPE_BFP16 = 11, /**< bfp16 */
   HCCL_DATA_TYPE_RESERVED    /**< reserved */
 } HcclDataType;
 

inc/framework/common/debug/log.h

@@ -178,16 +178,16 @@
   }
 
 #define GE_ERRORLOG_AND_ERRORMSG(_status, errormsg)        \
-  {                                                        \
+  do {                                                     \
     GELOGE((_status), "[Check][InnerData]%s", (errormsg)); \
     REPORT_INNER_ERROR("E19999", "%s", (errormsg));        \
-  }
+  } while (false)
 
 #define GE_WARNINGLOG_AND_ERRORMSG(errormsg)                                             \
-  {                                                                                      \
+  do {                                                                                   \
     GELOGW("%s", (errormsg));                                                            \
     ErrorManager::GetInstance().ATCReportErrMessage("E10052", {"reason"}, {(errormsg)}); \
-  }
+  } while (false)
 
 #define GE_CHK_LOG_AND_ERRORMSG(expr, _status, errormsg)                                   \
   do {                                                                                     \

inc/framework/common/ge_types.h

@@ -68,6 +68,8 @@ const char_t *const GE_OPTION_EXEC_PLACEMENT = "ge.exec.placement";
 
 const std::string kTaskTypeAicore = "AI_CORE";
 const std::string kTaskTypeAicpu = "AI_CPU";
+const std::string kTaskTypeWriteBackData = "WRITE_BACK";
+const std::string kTaskTypeInvalidData = "INVALID";
 const std::string kTaskTypeInvalid = "TASK_TYPE_INVALID";
 const std::string kTaskTypeFftsPlus = "FFTS_PLUS";
 const std::string kEngineNameVectorCore = "VectorEngine";
@@ -381,5 +383,9 @@ struct ModelQueueParam {
   std::vector<uint32_t> input_queues;
   std::vector<uint32_t> output_queues;
 };
+
+// internal options
+// 1: Graph resource evaluation does not limit model memory size.
+const char_t *const EVALUATE_GRAPH_RESOURCE_MODE = "ge.evaluateGraphResourceMode";
 }  // namespace ge
 #endif  // INC_FRAMEWORK_COMMON_GE_TYPES_H_

inc/framework/common/helper/model_helper.h

@@ -49,6 +49,10 @@ class GE_FUNC_VISIBILITY ModelHelper {
     is_offline_ = val;
   }
 
+  void SetSharedWeightFlag(const bool val) {
+    is_shared_weight_ = val;
+  }
+
   bool GetModelType() const {
     return is_unknown_shape_model_;
   }
@@ -60,6 +64,7 @@ class GE_FUNC_VISIBILITY ModelHelper {
   bool is_assign_model_ = false;
   bool is_offline_ = true;
   bool is_unknown_shape_model_ = false;
+  bool is_shared_weight_ = false;
   ModelFileHeader *file_header_ = nullptr;
   GeModelPtr model_;
   GeRootModelPtr root_model_;

inc/framework/common/profiling/ge_profiling.h

@@ -20,10 +20,8 @@
 #include "external/ge/ge_api_error_codes.h"
 #include "runtime/base.h"
 
-///
 /// @brief Output the profiling data of single operator in Pytorch, and does not support multithreading
 /// @return Status result
-///
 GE_FUNC_VISIBILITY ge::Status ProfSetStepInfo(const uint64_t index_id, const uint16_t tag_id, rtStream_t const stream);
 
 GE_FUNC_VISIBILITY ge::Status ProfGetDeviceFormGraphId(const uint32_t graph_id, uint32_t &device_id);

inc/framework/common/profiling_definitions.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2021 Huawei Technologies Co., Ltd
+ * Copyright (c) Huawei Technologies Co., Ltd. 2021. All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.

inc/framework/common/taskdown_common.h

@@ -39,12 +39,12 @@ enum class ccStatus_t {
 enum class ccKernelType {
   CCE_AI_CORE = 0, /* cce aicore */
   CCE_AI_CPU = 1,  /* cce aicpu */
-  TE = 2,          /* te operator*/
+  TE = 2,          /* te operator */
   CUSTOMIZED = 3,  /* customized operator */
-  TE_AI_CORE = 4,  /* te aicore operator*/
+  TE_AI_CORE = 4,  /* te aicore operator */
   TE_AI_CPU = 5,   /* te aicpu operator */
   AI_CPU = 6,      /* aicpu */
-  CUST_AI_CPU = 7, /* custom aicpu*/
+  CUST_AI_CPU = 7, /* custom aicpu */
   HOST_CPU = 8,    /* host cpu */
   INVALID = 10000  /* unknown kernel type */
 };

inc/framework/common/util.h

@@ -245,15 +245,12 @@ GE_FUNC_VISIBILITY std::string ToString(const std::vector<T> &v) {
   return ss.str();
 }
 
-///
 /// @ingroup: domi_common
 /// @brief: get length of file
 /// @param [in] input_file: path of file
 /// @return int64_t： File length. If the file length fails to be obtained, the value -1 is returned.
-///
 GE_FUNC_VISIBILITY extern int64_t GetFileLength(const std::string &input_file);
 
-///
 /// @ingroup domi_common
 /// @brief Reads all data from a binary file.
 /// @param [in] file_name  path of file
@@ -261,31 +258,23 @@ GE_FUNC_VISIBILITY extern int64_t GetFileLength(const std::string &input_file);
 /// @param [out] length  Output memory size
 /// @return false fail
 /// @return true success
-///
 GE_FUNC_VISIBILITY bool ReadBytesFromBinaryFile(const char_t *const file_name, char_t **const buffer, int32_t &length);
 
-///
 /// @ingroup domi_common
 /// @brief Recursively Creating a Directory
 /// @param [in] directory_path  Path, which can be a multi-level directory.
 /// @return 0 success
 /// @return -1 fail
-///
 GE_FUNC_VISIBILITY extern int32_t CreateDirectory(const std::string &directory_path);
 
-///
 /// @ingroup domi_common
 /// @brief Obtains the current time string.
 /// @return Time character string in the format ： %Y%m%d%H%M%S, eg: 20171011083555
-///
 GE_FUNC_VISIBILITY std::string CurrentTimeInStr();
 
-///
 /// @ingroup domi_common
 /// @brief Obtains the absolute time (timestamp) of the current system.
 /// @return Timestamp, in microseconds (US)
-///
-///
 GE_FUNC_VISIBILITY uint64_t GetCurrentTimestamp();
 
 ///
@@ -296,15 +285,12 @@ GE_FUNC_VISIBILITY uint64_t GetCurrentTimestamp();
 ///
 GE_FUNC_VISIBILITY uint32_t GetCurrentSecondTimestap();
 
-///
 /// @ingroup domi_common
 /// @brief Absolute path for obtaining files.
 /// @param [in] path of input file
 /// @param [out] Absolute path of a file. If the absolute path cannot be obtained, an empty string is returned
-///
 GE_FUNC_VISIBILITY std::string RealPath(const char_t *path);
 
-///
 /// @ingroup domi_common
 /// @brief Check whether the specified input file path is valid.
 /// 1.  The specified path cannot be empty.
@@ -312,23 +298,18 @@ GE_FUNC_VISIBILITY std::string RealPath(const char_t *path);
 /// 3.  The file path exists and is readable.
 /// @param [in] file_path path of input file
 /// @param [out] result
-///
 GE_FUNC_VISIBILITY bool CheckInputPathValid(const std::string &file_path, const std::string &atc_param = "");
 
-///
 /// @ingroup domi_common
 /// @brief Checks whether the specified output file path is valid.
 /// @param [in] file_path path of output file
 /// @param [out] result
-///
 GE_FUNC_VISIBILITY bool CheckOutputPathValid(const std::string &file_path, const std::string &atc_param = "");
 
-///
 /// @ingroup domi_common
 /// @brief Check whether the file path meets the whitelist verification requirements.
 /// @param [in] str file path
 /// @param [out] result
-///
 GE_FUNC_VISIBILITY bool ValidateStr(const std::string &file_path, const std::string &mode);
 
 GE_FUNC_VISIBILITY Status ConvertToInt32(const std::string &str, int32_t &val);

inc/framework/runtime/model_desc.h

@@ -27,6 +27,10 @@
 namespace gert {
 class VISIBILITY_EXPORT ShapeRange {
  public:
+  ShapeRange() = default;
+  ShapeRange(const Shape &min_shape, const Shape &max_shape);
+  bool operator==(const ShapeRange &other) const;
+
   const Shape &GetMin() const;
   const Shape &GetMax() const;
   Shape &MutableMin();
@@ -76,6 +80,9 @@ class VISIBILITY_EXPORT ModelDesc {
   const ModelIoDesc *GetOutputDesc(size_t index) const;
   const ModelIoDesc *GetAllOutputsDesc(size_t &output_num) const;
 
+  size_t GetInputNum() const;
+  size_t GetOutputNum() const;
+
   ModelIoDesc *MutableInputDesc(size_t index);
   ModelIoDesc *MutableOutputDesc(size_t index);
   ModelIoDesc *AllMutableIoDesc(size_t &input_num, size_t &output_num);

metadef

@@ -1 +1 @@
-Subproject commit 03482feb52fd7cc8544231f32891c86db3bc91a2
+Subproject commit ed43bd5a956c70e85e6f9d0a8f2277691a5ce4a2

third_party/fwkacllib/inc/cce/aicpu_engine.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ * Copyright (C) Huawei Technologies Co., Ltd. 2019-2022. All rights reserved.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -13,8 +13,8 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifndef AICPU_ENGINE_H__
-#define AICPU_ENGINE_H__
+#ifndef AICPU_ENGINE_H
+#define AICPU_ENGINE_H
 
 #include <stdint.h>
 
@@ -23,16 +23,16 @@ extern "C" {
 #endif
 
 typedef enum {
-  AE_STATUS_SUCCESS = 0,
-  AE_STATUS_BAD_PARAM = 1,
-  AE_STATUS_OPEN_SO_FAILED = 2,
-  AE_STATUS_GET_KERNEL_NAME_FAILED = 3,
-  AE_STATUS_INNER_ERROR = 4,
-  AE_STATUS_KERNEL_API_INNER_ERROR = 5,
-  AE_STATUS_END_OF_SEQUENCE = 6,
-  AE_STATUS_DUMP_FAILED = 7,
-  AE_STATUS_TASK_WAIT = 101,
-  AE_STATUS_RESERVED
+    AE_STATUS_SUCCESS = 0,
+    AE_STATUS_BAD_PARAM = 1,
+    AE_STATUS_OPEN_SO_FAILED = 2,
+    AE_STATUS_GET_KERNEL_NAME_FAILED = 3,
+    AE_STATUS_INNER_ERROR = 4,
+    AE_STATUS_KERNEL_API_INNER_ERROR = 5,
+    AE_STATUS_END_OF_SEQUENCE = 6,
+    AE_STATUS_DUMP_FAILED = 7,
+    AE_STATUS_TASK_WAIT = 101,
+    AE_STATUS_RESERVED
 } aeStatus_t;
 
 /**
@@ -59,4 +59,4 @@ aeStatus_t aeBatchLoadKernelSo(const uint32_t loadSoNum, const char *soPaths[],
 }
 #endif
 
-#endif  // AICPU_ENGINE_H__
+#endif  // AICPU_ENGINE_H

third_party/fwkacllib/inc/ops/experiment_ops.h

@@ -293,6 +293,66 @@ REG_OP(AutoBatch)
     .ATTR(padding, Bool, false)
     .ATTR(slide_stride, Int, 0)
     .OP_END_FACTORY_REG(AutoBatch)
+
+/**
+* @brief YUVToRGB
+
+* @par Inputs:
+* @li x: A 4-D uint8 Tensor.
+*        Must set the format, supported format list ["NYUV"].
+* @li matrix: A 1-D float tensor of 2x3x3 elements
+
+* @par Outputs:
+* @li y: A 4-D uint8 Tensor.
+*        Must set the format, supported format list ["NCHW, NHWC"].
+
+* @par Attributes:
+* @li matrix_type: An Int attr, Defaults to 0.
+*                  support list [ 0: CSC_MATRIX_BT601_WIDE,
+*                                 1: CSC_MATRIX_BT601_NARROW,
+*                                 2: CSC_MATRIX_BT709_WIDE,
+*                                 3: CSC_MATRIX_BT709_NARROW,
+*                                 4: CSC_MATRIX_BT2020_WIDE,
+*                                 5: CSC_MATRIX_BT2020_NARROW,
+*                                 6: CSC_MATRIX_USR_DEFINE ]
+* @li rb_swap: An Int attr, Defaults to 0.
+*              support list [ 0: RGB, 1: BGR ]
+
+* @attention Constraints:
+* @li Only support in dvpp
+*/
+
+REG_OP(YUVToRGB)
+    .INPUT(x, TensorType({DT_UINT8}))
+    .OPTIONAL_INPUT(matrix, TensorType({DT_FLOAT}))
+    .OUTPUT(y, TensorType({DT_UINT8}))
+    .ATTR(matrix_type, Int, 0)
+    .ATTR(rb_swap, Int, 0)
+    .OP_END_FACTORY_REG(YUVToRGB)
+
+/**
+* @brief DecodeJpegPre
+
+* @par Inputs:
+* @li contents: A Tensor of type string. 0-D. The JPEG-encoded image.
+
+* @par Outputs:
+* @li dvpp_support: indicates if the dvpp support this jpeg image decode.
+
+* @par Attributes:
+* @li w_range: An required listInt contains width [min, max].
+* @li h_range: An required listInt contains height [min, max].
+
+* @attention Constraints:
+* @li Only support in dvpp
+*/
+
+REG_OP(DecodeJpegPre)
+    .INPUT(contents, TensorType({DT_STRING}))
+    .OUTPUT(dvpp_support, BOOL)
+    .REQUIRED_ATTR(w_range, ListInt)
+    .REQUIRED_ATTR(h_range, ListInt)
+    .OP_END_FACTORY_REG(DecodeJpegPre)
 }  // namespace ge
 
 #endif  // OPS_BUILT_IN_OP_PROTO_INC_EXPERIMENT_OPS_H_

third_party/fwkacllib/inc/ops/image_ops.h

@@ -2444,5 +2444,34 @@ REG_OP(ExtractGlimpseV2)
     .ATTR(uniform_noise, Bool, true)
     .ATTR(noise, String, "uniform")
     .OP_END_FACTORY_REG(ExtractGlimpseV2)
+
+/**
+* @brief NormalizeV2 \n
+
+* @par Inputs:
+* @li x: A 4-D Tensor. Must be one of the following types: uint8, float16,
+*        float. Must set the format, supported format list ["NCHW, NHWC"].
+* @li mean: A 1-D float tensor of "C(channel)" elements
+* @li variance: A 1-D float tensor of "C(channel)" elements \n
+
+* @par Outputs:
+* @li y: A 4-D Tensor. Must be one of the following types: float16, float.
+*        Must set the format, supported format list ["NCHW, NHWC"]. \n
+
+* @par Attributes:
+* @li dtype: An Type attr, support type list [DT_FLOAT16, DT_FLOAT].
+*            Defaults to DT_FLOAT. \n
+
+* @par Third-party framework compatibility
+* Compatible with pytorch normalize operator.
+*/
+
+REG_OP(NormalizeV2)
+    .INPUT(x, TensorType({DT_UINT8, DT_FLOAT16, DT_FLOAT}))
+    .INPUT(mean, TensorType({DT_FLOAT}))
+    .INPUT(variance, TensorType({DT_FLOAT}))
+    .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT}))
+    .ATTR(dtype, Type, DT_FLOAT)
+    .OP_END_FACTORY_REG(NormalizeV2)
 }  // namespace ge
 #endif  // OPS_BUILT_IN_OP_PROTO_INC_IMAGE_OPS_H_

third_party/fwkacllib/inc/ops/transformation_ops.h

@@ -418,7 +418,7 @@ REG_OP(DepthToSpace)
 *@li The size of the first dimension of input "x" must be divisible by (block_size * block_size).
 *@li "crops" is a 4Dshape [batch, height, width, depth], height = height_pad - crop_top - crop_bottom,
 *width = width_pad - crop_left - crop_right.
-*@li block_size > 2
+*@li block_size >= 2
 
 *@par Third-party framework compatibility
 * Compatible with the TensorFlow operator BatchToSpace.

third_party/fwkacllib/inc/runtime/mem.h

@@ -32,6 +32,7 @@ extern "C" {
 #define RT_MEMORY_TS_4G (0x40U)    // Used for Ts memory(only 1951)
 #define RT_MEMORY_HOST (0x81U)     // Memory on host
 #define RT_MEMORY_SVM (0x90U)      // Memory for SVM
+#define RT_MEMORY_HOST_SVM (0x90U) // Memory for host SVM
 #define RT_MEMORY_RESERVED (0x100U)
 
 #define RT_MEMORY_L1 (0x1U << 16U)

third_party/fwkacllib/inc/toolchain/prof_acl_api.h

@@ -8,6 +8,9 @@
 #ifndef MSPROFILER_API_PROF_ACL_API_H_
 #define MSPROFILER_API_PROF_ACL_API_H_
 
+#include <cstdint>
+#include <cstddef>
+
 // DataTypeConfig
 #define PROF_ACL_API                0x00000001ULL
 #define PROF_TASK_TIME              0x00000002ULL
@@ -83,9 +86,6 @@ constexpr uint64_t PROF_AICPU_MODEL_MASK = 0x4000000000000000ULL;
 #define MSVP_PROF_API __attribute__((visibility("default")))
 #endif
 
-#include <cstdint>
-#include <cstddef>
-
 namespace Msprofiler {
 namespace Api {
 /**