!15928 unify dump path and support npy format

From: @zhangbuxue
Reviewed-by: @guoqi1024,@zhaizhiqiang
Signed-off-by: @zhaizhiqiang

mindspore/ccsrc/backend/kernel_compiler/cpu/reduce_cpu_kernel.h

@@ -37,7 +37,7 @@ class ReduceCPUKernel : public CPUKernel {
   enum ReduceType { kReduceAll, kReduceAny, kReduceMax, kReduceMin, kReduceSum, kReduceMean };
   std::vector<size_t> input_shape_;
   std::vector<int64_t> axis_;
-  ReduceType reduce_type_;
+  ReduceType reduce_type_{kReduceAll};
   std::function<void(const T *, size_t, T *)> reduce_func_;
 };
 

mindspore/ccsrc/debug/CMakeLists.txt

@@ -12,6 +12,7 @@ set(_DEBUG_SRC_LIST
     "${CMAKE_CURRENT_SOURCE_DIR}/data_dump/dump_json_parser.cc"
     "${CMAKE_CURRENT_SOURCE_DIR}/data_dump/cpu_e2e_dump.cc"
     "${CMAKE_CURRENT_SOURCE_DIR}/data_dump/dump_utils.cc"
+    "${CMAKE_CURRENT_SOURCE_DIR}/data_dump/npy_header.cc"
 )
 
 set(_OFFLINE_SRC_LIST

mindspore/ccsrc/debug/data_dump/cpu_e2e_dump.cc

@@ -19,7 +19,7 @@
 #include "backend/session/anf_runtime_algorithm.h"
 
 namespace mindspore {
-void CPUE2eDump::DumpCNodeData(const CNodePtr &node) {
+void CPUE2eDump::DumpCNodeData(const CNodePtr &node, uint32_t graph_id) {
   MS_EXCEPTION_IF_NULL(node);
   auto &dump_json_parser = DumpJsonParser::GetInstance();
   std::string kernel_name = node->fullname_with_scope();
@@ -29,7 +29,7 @@ void CPUE2eDump::DumpCNodeData(const CNodePtr &node) {
 
   MS_LOG(DEBUG) << "E2e dump CNode data start: " << kernel_name << ", current iteration is "
                 << dump_json_parser.cur_dump_iter();
-  std::string dump_path = GenerateDumpPath();
+  std::string dump_path = GenerateDumpPath(graph_id);
   if (dump_json_parser.InputNeedDump()) {
     DumpCNodeInputs(node, dump_path);
   }
@@ -129,12 +129,12 @@ void CPUE2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t outp
   DumpMemToFile(file_path, NOT_NULL(addr), int_shapes, type);
 }
 
-void CPUE2eDump::DumpParametersAndConst(const session::KernelGraph *graph) {
+void CPUE2eDump::DumpParametersAndConst(const session::KernelGraph *graph, uint32_t graph_id) {
   MS_EXCEPTION_IF_NULL(graph);
   MS_LOG(INFO) << "Start e2e dump parameters and Const values";
   std::map<std::string, size_t> const_map;
   GetConstantId(graph, &const_map);
-  const std::string &dump_path = GenerateDumpPath();
+  const std::string &dump_path = GenerateDumpPath(graph_id);
 
   // dump parameters
   const auto &parameters = graph->inputs();

mindspore/ccsrc/debug/data_dump/cpu_e2e_dump.h

@@ -29,9 +29,9 @@ class CPUE2eDump {
   CPUE2eDump() = default;
   ~CPUE2eDump() = default;
   // Dump data when task error.
-  static void DumpParametersAndConst(const session::KernelGraph *graph);
+  static void DumpParametersAndConst(const session::KernelGraph *graph, uint32_t graph_id);
 
-  static void DumpCNodeData(const CNodePtr &node);
+  static void DumpCNodeData(const CNodePtr &node, uint32_t graph_id);
 
  private:
   static void DumpCNodeInputs(const CNodePtr &node, const std::string &dump_path);

mindspore/ccsrc/debug/data_dump/dump_json_parser.cc

@@ -20,12 +20,14 @@
 #include "utils/ms_context.h"
 #include "utils/convert_utils_base.h"
 #include "backend/session/anf_runtime_algorithm.h"
+#include "debug/data_dump/npy_header.h"
 
 namespace {
 constexpr auto kCommonDumpSettings = "common_dump_settings";
 constexpr auto kAsyncDumpSettings = "async_dump_settings";
 constexpr auto kE2eDumpSettings = "e2e_dump_settings";
 constexpr auto kDumpMode = "dump_mode";
+constexpr auto kDumpFormat = "dump_format";
 constexpr auto kPath = "path";
 constexpr auto kNetName = "net_name";
 constexpr auto kIteration = "iteration";
@@ -42,6 +44,8 @@ constexpr auto kMindsporeDumpConfig = "MINDSPORE_DUMP_CONFIG";
 }  // namespace
 
 namespace mindspore {
+uint32_t DumpJsonParser::dump_format_ = 0;
+
 auto DumpJsonParser::CheckJsonKeyExist(const nlohmann::json &content, const std::string &key) {
   auto iter = content.find(key);
   if (iter == content.end()) {
@@ -137,13 +141,15 @@ bool DumpJsonParser::GetIterDumpFlag() {
   return e2e_dump_enabled_ && (iteration_ == 0 || cur_dump_iter_ == iteration_);
 }
 
-bool DumpJsonParser::DumpToFile(const std::string &filename, const void *data, size_t len) {
+bool DumpJsonParser::DumpToFile(const std::string &filename, const void *data, size_t len, const ShapeVector &shape,
+                                TypeId type) {
   if (filename.empty() || data == nullptr || len == 0) {
     MS_LOG(ERROR) << "Incorrect parameter.";
     return false;
   }
 
-  auto realpath = Common::GetRealPath(filename);
+  std::string file_format = dump_format_ == 1 ? ".npy" : ".bin";
+  auto realpath = Common::GetRealPath(filename + file_format);
   if (!realpath.has_value()) {
     MS_LOG(ERROR) << "Get real path failed.";
     return false;
@@ -154,6 +160,10 @@ bool DumpJsonParser::DumpToFile(const std::string &filename, const void *data, s
     MS_LOG(ERROR) << "Open file " << realpath.value() << " fail.";
     return false;
   }
+  if (dump_format_ == 1) {
+    std::string npy_header = GenerateNpyHeader(shape, type);
+    fd << npy_header;
+  }
   (void)fd.write(reinterpret_cast<const char *>(data), SizeToLong(len));
   fd.close();
   return true;
@@ -176,6 +186,7 @@ void DumpJsonParser::ParseCommonDumpSetting(const nlohmann::json &content) {
   ParseInputOutput(*input_output);
   ParseKernels(*kernels);
   ParseSupportDevice(*support_device);
+  ParseDumpFormat(*common_dump_settings);
 }
 
 void DumpJsonParser::ParseAsyncDumpSetting(const nlohmann::json &content) {
@@ -209,19 +220,19 @@ void DumpJsonParser::ParseE2eDumpSetting(const nlohmann::json &content) {
 
 void CheckJsonUnsignedType(const nlohmann::json &content, const std::string &key) {
   if (!content.is_number_unsigned()) {
-    MS_LOG(EXCEPTION) << "Dump Json Parse Failed." << key << " should be unsigned int type";
+    MS_LOG(EXCEPTION) << "Dump config parse failed, " << key << " should be unsigned int type";
   }
 }
 
 void CheckJsonStringType(const nlohmann::json &content, const std::string &key) {
   if (!content.is_string()) {
-    MS_LOG(EXCEPTION) << "Dump Json Parse Failed." << key << " should be string type";
+    MS_LOG(EXCEPTION) << "Dump config parse failed, " << key << " should be string type";
   }
 }
 
 void CheckJsonArrayType(const nlohmann::json &content, const std::string &key) {
   if (!content.is_array()) {
-    MS_LOG(EXCEPTION) << "Dump Json Parse Failed." << key << " should be array type";
+    MS_LOG(EXCEPTION) << "Dump config parse failed, " << key << " should be array type";
   }
 }
 
@@ -229,7 +240,18 @@ void DumpJsonParser::ParseDumpMode(const nlohmann::json &content) {
   CheckJsonUnsignedType(content, kDumpMode);
   dump_mode_ = content;
   if (dump_mode_ != 0 && dump_mode_ != 1) {
-    MS_LOG(EXCEPTION) << "Dump Json Parse Failed. dump_mode should be 0 or 1";
+    MS_LOG(EXCEPTION) << "Dump config parse failed, dump_mode should be 0 or 1, but got " << dump_format_;
+  }
+}
+
+void DumpJsonParser::ParseDumpFormat(const nlohmann::json &content) {
+  auto iter = content.find(kDumpFormat);
+  if (iter == content.end()) {
+    return;
+  }
+  dump_format_ = *iter;
+  if (dump_format_ != 0 && dump_format_ != 1) {
+    MS_LOG(EXCEPTION) << "Dump config parse failed, dump_format should be 0(.bin) or 1(.npy), but got " << dump_format_;
   }
 }
 

mindspore/ccsrc/debug/data_dump/dump_json_parser.h

@@ -33,7 +33,8 @@ class DumpJsonParser {
   }
 
   void Parse();
-  static bool DumpToFile(const std::string &filename, const void *data, size_t len);
+  static bool DumpToFile(const std::string &filename, const void *data, size_t len, const ShapeVector &shape,
+                         TypeId type);
   void CopyJsonToDir();
   bool NeedDump(const std::string &op_full_name) const;
   void MatchKernel(const std::string &kernel_name);
@@ -62,6 +63,7 @@ class DumpJsonParser {
   DISABLE_COPY_AND_ASSIGN(DumpJsonParser)
 
   std::mutex lock_;
+  static uint32_t dump_format_;
   bool async_dump_enabled_{false};
   bool e2e_dump_enabled_{false};
   uint32_t dump_mode_{0};
@@ -84,6 +86,7 @@ class DumpJsonParser {
   auto CheckJsonKeyExist(const nlohmann::json &content, const std::string &key);
 
   void ParseDumpMode(const nlohmann::json &content);
+  void ParseDumpFormat(const nlohmann::json &content);
   void ParseDumpPath(const nlohmann::json &content);
   void ParseNetName(const nlohmann::json &content);
   void ParseIteration(const nlohmann::json &content);

mindspore/ccsrc/debug/data_dump/dump_utils.cc

@@ -34,7 +34,7 @@ uint32_t ConvertPhysicalDeviceId(uint32_t device_id) {
   return kernel_runtime->device_id();
 }
 
-std::string GenerateDumpPath(const uint32_t *device_id) {
+std::string GenerateDumpPath(uint32_t graph_id, const uint32_t *device_id) {
   auto &dump_json_parser = DumpJsonParser::GetInstance();
   std::string net_name = dump_json_parser.net_name();
   std::string iterator = std::to_string(dump_json_parser.cur_dump_iter());
@@ -42,12 +42,9 @@ std::string GenerateDumpPath(const uint32_t *device_id) {
   if (dump_path.back() != '/') {
     dump_path += "/";
   }
-  if (device_id == nullptr) {
-    dump_path += (net_name + "/iteration_" + iterator);
-  } else {
-    auto physical_device = ConvertPhysicalDeviceId(*device_id);
-    dump_path += (net_name + "/device_" + std::to_string(physical_device) + "/iteration_" + iterator);
-  }
+  uint32_t physical_device = device_id == nullptr ? 0 : ConvertPhysicalDeviceId(*device_id);
+  dump_path += ("rank_" + std::to_string(physical_device) + "/" + net_name + "/graph_" + std::to_string(graph_id) +
+                "/iteration_" + iterator);
   return dump_path;
 }
 

mindspore/ccsrc/debug/data_dump/dump_utils.h

@@ -27,7 +27,7 @@ namespace mindspore {
 static const size_t PARAMETER_OUTPUT_INDEX = 0;
 static const size_t VALUE_NODE_OUTPUT_INDEX = 0;
 
-std::string GenerateDumpPath(const uint32_t *device_id = nullptr);
+std::string GenerateDumpPath(uint32_t graph_id, const uint32_t *device_id = nullptr);
 
 void GetFileKernelName(NotNull<std::string *> kernel_name);
 

mindspore/ccsrc/debug/data_dump/e2e_dump.cc

@@ -223,18 +223,19 @@ void E2eDump::DumpParametersAndConst(const session::KernelGraph *graph, const st
 bool E2eDump::DumpData(const session::KernelGraph *graph, uint32_t device_id, const Debugger *debugger) {
   MS_EXCEPTION_IF_NULL(graph);
   auto &dump_json_parser = DumpJsonParser::GetInstance();
+  uint32_t graph_id = graph->graph_id();
   if (starting_graph_id == INT32_MAX) {
-    starting_graph_id = graph->graph_id();
+    starting_graph_id = graph_id;
   }
-  if (starting_graph_id == graph->graph_id()) {
+  if (starting_graph_id == graph_id) {
     dump_json_parser.UpdateDumpIter();
   }
   if (!dump_json_parser.GetIterDumpFlag()) {
     return true;
   }
   MS_LOG(INFO) << "Start e2e dump. Current iteration is " << dump_json_parser.cur_dump_iter();
-  MS_LOG(INFO) << "Current graph id is " << graph->graph_id();
-  std::string dump_path = GenerateDumpPath(&device_id);
+  MS_LOG(INFO) << "Current graph id is " << graph_id;
+  std::string dump_path = GenerateDumpPath(graph_id, &device_id);
 
   DumpInput(graph, dump_path, debugger);
   DumpOutput(graph, dump_path, debugger);

mindspore/ccsrc/debug/data_dump/npy_header.cc

@@ -0,0 +1,145 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "debug/data_dump/npy_header.h"
+
+#include <unordered_map>
+#include <utility>
+#include <sstream>
+#include <cmath>
+
+#include "mindspore/core/utils/log_adapter.h"
+#include "mindspore/core/ir/dtype.h"
+
+namespace mindspore {
+namespace {
+// npy file header start information
+const char kMagicPrefix[] = "\x93NUMPY";
+// magical length include kMagicPrefix length and version length
+const size_t kMagicLen = 8;
+const size_t kArrayAlign = 64;
+
+// first: header_length_type, second: encoding_type
+// header_length_type: 1 represents 2 bytes; 2 and 3 represents 4 bytes
+// encoding_type: 1 and 2 represents 'latin1'; 3 represents 'utf8'
+using version_type = std::pair<int, int>;
+
+// data type description
+// byteorder char: '<' is little endian; '>' is big endian; '|' is ignore(no change to byte order)
+// type char: 'b' represents bool; 'u' represents uint; 'i' represents int; 'f' represents float
+struct DtypeDescr {
+  char byteorder;
+  char type;
+  size_t length;
+
+  std::string str() const;
+};
+
+// npy file header description, includes data type description, fortran_order and array shape
+// fortran_order: true represents the array data Fortran-contiguous; false represents the array data C-contiguity
+struct NpyHeader {
+ public:
+  DtypeDescr dtype_descr;
+  bool fortran_order;
+  ShapeVector shape;
+
+  std::string str() const;
+
+ private:
+  std::string fortran_order_to_str() const;
+  std::string shape_to_str() const;
+};
+
+std::string DtypeDescr::str() const {
+  std::ostringstream buffer;
+  buffer << "\'" << byteorder << type << length << "\'";
+  return buffer.str();
+}
+
+std::string NpyHeader::str() const {
+  const std::string first_field = "'descr': ";
+  const std::string second_field = "'fortran_order': ";
+  const std::string third_field = "'shape': ";
+  std::ostringstream buffer;
+  buffer << "{" << first_field << dtype_descr.str() << ", " << second_field << fortran_order_to_str() << ", "
+         << third_field << shape_to_str() << ", }";
+  return buffer.str();
+}
+
+std::string NpyHeader::fortran_order_to_str() const { return fortran_order ? "True" : "False"; }
+
+std::string NpyHeader::shape_to_str() const {
+  std::ostringstream buffer;
+  buffer << "(";
+  for (const auto i : shape) {
+    buffer << std::to_string(i) << ",";
+  }
+  buffer << ")";
+  return buffer.str();
+}
+}  // namespace
+
+void int_to_byte(size_t number, char *byte, size_t length) {
+  for (size_t i = 0; i < length; i++) {
+    byte[i] = (number >> (i * 8)) & 0xff;
+  }
+}
+
+std::string GenerateNpyHeader(const ShapeVector &shape, TypeId type_id, bool fortran_order) {
+  static std::unordered_map<TypeId, DtypeDescr> type_desc_map = {
+    {kNumberTypeBool, DtypeDescr{'|', 'b', 1}},    {kNumberTypeInt8, DtypeDescr{'|', 'i', 1}},
+    {kNumberTypeInt16, DtypeDescr{'<', 'i', 2}},   {kNumberTypeInt32, DtypeDescr{'<', 'i', 4}},
+    {kNumberTypeInt64, DtypeDescr{'<', 'i', 8}},   {kNumberTypeUInt8, DtypeDescr{'|', 'u', 1}},
+    {kNumberTypeUInt16, DtypeDescr{'<', 'u', 2}},  {kNumberTypeUInt32, DtypeDescr{'<', 'u', 4}},
+    {kNumberTypeUInt64, DtypeDescr{'<', 'u', 8}},  {kNumberTypeFloat16, DtypeDescr{'<', 'f', 2}},
+    {kNumberTypeFloat32, DtypeDescr{'<', 'f', 4}}, {kNumberTypeFloat64, DtypeDescr{'<', 'f', 8}},
+  };
+  auto type_desc = type_desc_map.find(type_id);
+  if (type_desc == type_desc_map.end()) {
+    MS_LOG(EXCEPTION) << "Not support dump the " << TypeIdToType(type_id)->ToString() << " data to npy file.";
+  }
+
+  NpyHeader npy_header{type_desc->second, fortran_order, shape};
+  std::string header_str = npy_header.str();
+  size_t header_len = header_str.length();
+  version_type version{1, 0};
+  size_t total_len = kMagicLen + 2 + header_len + 1;
+  if (total_len > std::pow(2, 16)) {
+    version = {2, 0};
+    total_len = kMagicLen + 4 + header_len + 1;
+  }
+  std::ostringstream out;
+  out << kMagicPrefix;
+  out.put(version.first);
+  out.put(version.second);
+
+  size_t pad_len = kArrayAlign - total_len % kArrayAlign;
+  size_t padding_header_len = header_len + pad_len + 1;
+  if (version == version_type{1, 0}) {
+    char length_byte[2];
+    int_to_byte(padding_header_len, length_byte, 2);
+    out.write(length_byte, 2);
+  } else {
+    char length_byte[4];
+    int_to_byte(padding_header_len, length_byte, 4);
+    out.write(length_byte, 4);
+  }
+
+  std::string padding(pad_len, ' ');
+  out << header_str << padding << "\n";
+  return out.str();
+}
+}  // namespace mindspore

mindspore/ccsrc/debug/data_dump/npy_header.h

@@ -0,0 +1,28 @@
+/**
+ * Copyright 2021 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_MINDSPORE_CCSRC_DEBUG_DATA_DUMP_NPY_HEADER_H_
+#define MINDSPORE_MINDSPORE_CCSRC_DEBUG_DATA_DUMP_NPY_HEADER_H_
+
+#include <string>
+
+#include "mindspore/core/utils/shape_utils.h"
+#include "mindspore/core/ir/dtype/type_id.h"
+
+namespace mindspore {
+std::string GenerateNpyHeader(const ShapeVector &shape, TypeId type_id, bool fortran_order = false);
+}  // namespace mindspore
+#endif  // MINDSPORE_MINDSPORE_CCSRC_DEBUG_DATA_DUMP_NPY_HEADER_H_

mindspore/ccsrc/debug/tensor_load.h

@@ -172,12 +172,11 @@ class TensorLoader {
     } else {
       shape = shape + "_0";
     }
-    std::string file_extension = ".bin";
     std::string path = "";
     if (trans_flag) {
-      path = filepath + '_' + shape + '_' + TypeIdToType(host_type)->ToString() + '_' + host_fmt + file_extension;
+      path = filepath + '_' + shape + '_' + TypeIdToType(host_type)->ToString() + '_' + host_fmt;
     } else {
-      path = filepath + '_' + shape + '_' + TypeIdToType(device_type)->ToString() + '_' + addr_format + file_extension;
+      path = filepath + '_' + shape + '_' + TypeIdToType(device_type)->ToString() + '_' + addr_format;
     }
 
     MS_LOG(INFO) << "Dump path is " << path;
@@ -188,7 +187,7 @@ class TensorLoader {
       std::shared_ptr<TensorData> node = iter->second;
       size_t host_size = node->GetByteSize();
 
-      return DumpJsonParser::DumpToFile(path, node->GetDataPtr(), host_size);
+      return DumpJsonParser::DumpToFile(path, node->GetDataPtr(), host_size, host_shape, host_type);
     }
     MS_LOG(INFO) << "Tensor name:" << tensor_name << " not found in tensor_list_map";
     return true;

mindspore/ccsrc/runtime/device/ascend/ascend_device_address.cc

@@ -664,10 +664,8 @@ bool AscendDeviceAddress::DumpMemToFile(const std::string &filepath, const std::
   } else {
     shape = shape + "_0";
   }
-  std::string file_extension = ".bin";
   if (trans_flag) {
-    std::string path =
-      filepath + '_' + shape + '_' + TypeIdToType(host_type)->ToString() + '_' + host_fmt + file_extension;
+    std::string path = filepath + '_' + shape + '_' + TypeIdToType(host_type)->ToString() + '_' + host_fmt;
     MS_LOG(INFO) << "E2E Dump path is " << path;
     mindspore::tensor::TensorPtr out_tensor = std::make_shared<tensor::Tensor>(host_type, host_shape);
     size_t host_size = out_tensor->data().nbytes();
@@ -676,17 +674,16 @@ bool AscendDeviceAddress::DumpMemToFile(const std::string &filepath, const std::
       MS_LOG(ERROR) << "Copy device mem to host failed";
       return ret;
     }
-    ret = DumpJsonParser::DumpToFile(path, out_tensor->data_c(), host_size);
+    ret = DumpJsonParser::DumpToFile(path, out_tensor->data_c(), host_size, host_shape, host_type);
   } else {
     auto host_tmp = std::vector<uint8_t>(size_);
     auto ret_rt_memcpy = rtMemcpy(host_tmp.data(), size_, ptr_, size_, RT_MEMCPY_DEVICE_TO_HOST);
     if (ret_rt_memcpy != RT_ERROR_NONE) {
       MS_LOG(ERROR) << "SyncDeviceToHost: rtMemcpy mem size[" << size_ << "] fail, ret[" << ret_rt_memcpy << "]";
     }
-    std::string path =
-      filepath + '_' + shape + '_' + TypeIdToType(type_id_)->ToString() + '_' + format_ + file_extension;
+    std::string path = filepath + '_' + shape + '_' + TypeIdToType(type_id_)->ToString() + '_' + format_;
     MS_LOG(INFO) << "E2E Dump path is " << path;
-    ret = DumpJsonParser::DumpToFile(path, host_tmp.data(), size_);
+    ret = DumpJsonParser::DumpToFile(path, host_tmp.data(), size_, host_shape_, type_id_);
   }
 
   return ret;

mindspore/ccsrc/runtime/device/cpu/cpu_device_address.cc

@@ -37,10 +37,9 @@ bool CPUDeviceAddress::DumpMemToFile(const std::string &filepath, const std::str
       shape += '_' + std::to_string(value);
     }
   }
-  std::string file_extension = ".bin";
-  std::string path = filepath + '_' + shape + '_' + TypeIdToType(type_id_)->ToString() + '_' + format_ + file_extension;
+  std::string path = filepath + '_' + shape + '_' + TypeIdToType(type_id_)->ToString() + '_' + format_;
   MS_LOG(DEBUG) << "E2E Dump path is " << path;
-  ret = DumpJsonParser::DumpToFile(path, ptr_, size_);
+  ret = DumpJsonParser::DumpToFile(path, ptr_, size_, host_shape, host_type);
   return ret;
 }
 

mindspore/ccsrc/runtime/device/cpu/cpu_kernel_runtime.cc

@@ -380,6 +380,7 @@ bool CPUKernelRuntime::Run(session::KernelGraph *kernel_graph, bool is_task_sink
   auto &dump_json_parser = DumpJsonParser::GetInstance();
   dump_json_parser.UpdateDumpIter();
   bool iter_dump_flag = dump_json_parser.GetIterDumpFlag();
+  uint32_t graph_id = kernel_graph->graph_id();
 
   for (const auto &kernel : kernels) {
 #ifdef ENABLE_PROFILE
@@ -421,7 +422,7 @@ bool CPUKernelRuntime::Run(session::KernelGraph *kernel_graph, bool is_task_sink
       MS_LOG(EXCEPTION) << e.what() << "\nTrace:" << trace::DumpSourceLines(kernel);
     }
     if (iter_dump_flag) {
-      CPUE2eDump::DumpCNodeData(kernel);
+      CPUE2eDump::DumpCNodeData(kernel, graph_id);
     }
     if (profiler_inst->GetEnableFlag()) {
       profiler_inst->OpDataProducerEnd();
@@ -439,7 +440,7 @@ bool CPUKernelRuntime::Run(session::KernelGraph *kernel_graph, bool is_task_sink
 #endif
   }
   if (iter_dump_flag) {
-    CPUE2eDump::DumpParametersAndConst(kernel_graph);
+    CPUE2eDump::DumpParametersAndConst(kernel_graph, graph_id);
   }
   return true;
 }

tests/st/dump/e2e_dump_npy.json

@@ -0,0 +1,16 @@
+{
+  "common_dump_settings": {
+    "dump_mode": 0,
+    "path": "/test",
+    "net_name": "Net",
+    "iteration": 0,
+    "input_output": 0,
+    "kernels": ["Default/Conv-op12"],
+    "support_device": [0,1,2,3,4,5,6,7],
+    "dump_format": 1
+  },
+  "e2e_dump_settings": {
+    "enable": true,
+    "trans_flag": false
+  }
+}

tests/st/dump/test_data_dump.py

@@ -14,8 +14,10 @@
 # ============================================================================
 import os
 import json
+import sys
 import time
 import shutil
+
 import numpy as np
 import pytest
 import mindspore.context as context
@@ -29,7 +31,6 @@ from mindspore.nn import Momentum
 from mindspore.nn import TrainOneStepCell
 from mindspore.nn import WithLossCell
 
-context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
 
 class Net(nn.Cell):
     def __init__(self):
@@ -39,8 +40,10 @@ class Net(nn.Cell):
     def construct(self, x_, y_):
         return self.add(x_, y_)
 
-x = np.random.randn(1, 3, 3, 4).astype(np.float32)
-y = np.random.randn(1, 3, 3, 4).astype(np.float32)
+
+x = np.array([[1, 2, 3], [4, 5, 6]]).astype(np.float32)
+y = np.array([[7, 8, 9], [10, 11, 12]]).astype(np.float32)
+
 
 def change_current_dump_json(file_name, dump_path):
     with open(file_name, 'r+') as f:
@@ -50,6 +53,7 @@ def change_current_dump_json(file_name, dump_path):
     with open(file_name, 'w') as f:
         json.dump(data, f)
 
+
 @pytest.mark.level0
 @pytest.mark.platform_arm_ascend_training
 @pytest.mark.platform_x86_ascend_training
@@ -61,7 +65,7 @@ def test_async_dump():
     change_current_dump_json('async_dump.json', dump_path)
     os.environ['MINDSPORE_DUMP_CONFIG'] = pwd + "/async_dump.json"
     device_id = context.get_context("device_id")
-    dump_file_path = pwd + '/async_dump/device_{}/Net_graph_0/0/0/'.format(device_id)
+    dump_file_path = dump_path + '/device_{}/Net_graph_0/0/0/'.format(device_id)
     if os.path.isdir(dump_path):
         shutil.rmtree(dump_path)
     add = Net()
@@ -69,24 +73,90 @@ def test_async_dump():
     time.sleep(5)
     assert len(os.listdir(dump_file_path)) == 1
 
-@pytest.mark.level0
-@pytest.mark.platform_arm_ascend_training
-@pytest.mark.platform_x86_ascend_training
-@pytest.mark.env_onecard
-def test_e2e_dump():
-    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+
+def run_e2e_dump_bin():
+    if sys.platform != 'linux':
+        return
     pwd = os.getcwd()
-    dump_path = pwd + "/e2e_dump"
-    change_current_dump_json('e2e_dump.json', dump_path)
-    os.environ['MINDSPORE_DUMP_CONFIG'] = pwd + "/e2e_dump.json"
+    dump_path = pwd + '/e2e_dump'
+    change_current_dump_json('e2e_dump_bin.json', dump_path)
+    os.environ['MINDSPORE_DUMP_CONFIG'] = pwd + '/e2e_dump_bin.json'
     device_id = context.get_context("device_id")
-    dump_file_path = pwd + '/e2e_dump/Net/device_{}/iteration_1/'.format(device_id)
+    dump_file_path = dump_path + '/rank_{}/Net/graph_0/iteration_1/'.format(device_id)
     if os.path.isdir(dump_path):
         shutil.rmtree(dump_path)
     add = Net()
     add(Tensor(x), Tensor(y))
-    time.sleep(5)
-    assert len(os.listdir(dump_file_path)) == 5
+    if context.get_context("device_target") == "Ascend":
+        output_name = "Default--Add-op1_output_0_shape_2_3_Float32_DefaultFormat.bin"
+    else:
+        output_name = "Default--Add-op3_output_0_shape_2_3_Float32_DefaultFormat.bin"
+    output_path = dump_file_path + output_name
+    real_path = os.path.realpath(output_path)
+    output = np.fromfile(real_path, dtype=np.float32)
+    expect = np.array([8, 10, 12, 14, 16, 18], np.float32)
+    assert output.dtype == expect.dtype
+    assert np.array_equal(output, expect)
+
+
+def run_e2e_dump_npy():
+    if sys.platform != 'linux':
+        return
+    pwd = os.getcwd()
+    dump_path = pwd + '/e2e_dump'
+    change_current_dump_json('e2e_dump_npy.json', dump_path)
+    os.environ['MINDSPORE_DUMP_CONFIG'] = pwd + '/e2e_dump_npy.json'
+    device_id = context.get_context("device_id")
+    dump_file_path = dump_path + '/rank_{}/Net/graph_0/iteration_1/'.format(device_id)
+    if os.path.isdir(dump_path):
+        shutil.rmtree(dump_path)
+    add = Net()
+    add(Tensor(x), Tensor(y))
+    if context.get_context("device_target") == "Ascend":
+        output_name = "Default--Add-op1_output_0_shape_2_3_Float32_DefaultFormat.npy"
+    else:
+        output_name = "Default--Add-op3_output_0_shape_2_3_Float32_DefaultFormat.npy"
+    output_path = dump_file_path + output_name
+    real_path = os.path.realpath(output_path)
+    output = np.load(real_path)
+    expect = np.array([[8, 10, 12], [14, 16, 18]], np.float32)
+    assert output.dtype == expect.dtype
+    assert np.array_equal(output, expect)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_e2e_dump_bin():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+    run_e2e_dump_bin()
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_e2e_dump_npy():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+    run_e2e_dump_npy()
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_cpu_e2e_dump_bin():
+    context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
+    run_e2e_dump_bin()
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_cpu_e2e_dump_npy():
+    context.set_context(mode=context.GRAPH_MODE, save_graphs=True, device_target="CPU")
+    run_e2e_dump_npy()
+
 
 class ReluReduceMeanDenseRelu(Cell):
     def __init__(self, kernel, bias, in_channel, num_class):
@@ -116,11 +186,13 @@ def search_path(path, keyword):
             search_path(each_path, keyword)
     return None
 
+
 @pytest.mark.level0
 @pytest.mark.platform_arm_ascend_training
 @pytest.mark.platform_x86_ascend_training
 @pytest.mark.env_onecard
 def test_async_dump_net_multi_layer_mode1():
+    context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
     test_name = "test_async_dump_net_multi_layer_mode1"
     json_file = os.path.join(os.getcwd(), "{}.json".format(test_name))
     device_id = context.get_context("device_id")
@@ -131,7 +203,8 @@ def test_async_dump_net_multi_layer_mode1():
     bias = Tensor(np.ones((1000,)).astype(np.float32))
     net = ReluReduceMeanDenseRelu(weight, bias, 2048, 1000)
     criterion = SoftmaxCrossEntropyWithLogits(sparse=False)
-    optimizer = Momentum(learning_rate=0.1, momentum=0.1, params=filter(lambda  x: x.requires_grad, net.get_parameters()))
+    optimizer = Momentum(learning_rate=0.1, momentum=0.1,
+                         params=filter(lambda x: x.requires_grad, net.get_parameters()))
     net_with_criterion = WithLossCell(net, criterion)
     train_network = TrainOneStepCell(net_with_criterion, optimizer)
     train_network.set_train()

tests/ut/cpp/debug/memory_dumper_test.cc

@@ -37,11 +37,11 @@ TEST_F(TestMemoryDumper, test_DumpToFileAbsPath) {
   }
 
   int ret;
-  char filename[] = "/tmp/dumpToFileTestFile";
-  ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int));
+  const std::string filename = "/tmp/dumpToFileTestFile";
+  ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int), ShapeVector {10, 100}, kNumberTypeInt32);
   ASSERT_EQ(ret, true);
 
-  int fd = open(filename, O_RDONLY);
+  int fd = open((filename + ".bin").c_str(), O_RDONLY);
   int readBack[1000] = {0};
   int readSize = read(fd, readBack, len * sizeof(int));
   (void)close(fd);
@@ -69,11 +69,11 @@ TEST_F(TestMemoryDumper, test_DumpToFileRelativePath) {
   }
 
   int ret;
-  char filename[] = "../../dumpToFileTestFile";
-  ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int));
+   const std::string filename = "../../dumpToFileTestFile";
+  ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int), ShapeVector{100, 10}, kNumberTypeInt32);
   ASSERT_EQ(ret, true);
 
-  int fd = open(filename, O_RDONLY);
+  int fd = open((filename + ".bin").c_str(), O_RDONLY);
   int readBack[1000] = {0};
   int readSize = read(fd, readBack, len * sizeof(int));
   (void)close(fd);
@@ -101,11 +101,11 @@ TEST_F(TestMemoryDumper, test_DumpToFileNotExistDir) {
     data[i] = i % 10;
   }
 
-  char filename[] = "./tmp/dumpToFileTestFile";
-  int ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int));
+  const std::string filename = "./tmp/dumpToFileTestFile";
+  int ret = DumpJsonParser::DumpToFile(filename, data, len * sizeof(int), ShapeVector {1,}, kNumberTypeInt32);
   ASSERT_EQ(ret, true);
 
-  int fd = open(filename, O_RDONLY);
+  int fd = open((filename + ".bin").c_str(), O_RDONLY);
   int readBack[1000] = {0};
   int readSize = read(fd, readBack, len * sizeof(int));
   (void)close(fd);