!79 Dump Failed when trans_flag set to true

Merge pull request !79 from laiyongqiang/develop

config/e2e_dump_config.json

@@ -6,17 +6,17 @@
     "net_name": "ResNet50",
     "mode": 0,
     "iteration": 0,
-    "kernels": ["TensorAdd"]
+    "kernels": ["Default/Conv2D-op2", "Default/TensorAdd-op10"]
   },
 
   "DumpSettingsSpec": {
-    "enable": "true: dump enable false: dump disable",
-    "trans_flag": "true: trans to host format,false: not trans format",
+    "enable": "true: dump enable, false: dump disable",
+    "trans_flag": "true: trans to host format, false: not trans format",
     "path": "the dump file folder",
     "net_name": "net name eg:ResNet50",
-    "mode": "0: dump all kernels 1: dump kernels in kernels list",
-    "iteration": "0: all iteration  others: specified iteration ",
-    "kernels": "kernel name list need to be dump"
+    "mode": "0: dump all kernels, 1: dump kernels in kernels list",
+    "iteration": "0: all iteration, others: specified iteration ",
+    "kernels": "op's full scope name which need to be dump"
   },
   "other": {}
 }

config/e2e_dump_config_0.json

@@ -6,17 +6,17 @@
     "net_name": "ResNet50",
     "mode": 0,
     "iteration": 0,
-    "kernels": ["AllReduce","BiasAddGrad","Conv2DBackpropFilter","SparseSoftmaxCrossEntropyWithLogits"]
+    "kernels": ["Default/Conv2D-op2", "Default/TensorAdd-op10"]
   },
 
   "DumpSettingsSpec": {
-    "enable": "true: dump enable false: dump disable",
-    "trans_flag": "true: trans to host format,false: not trans format",
+    "enable": "true: dump enable, false: dump disable",
+    "trans_flag": "true: trans to host format, false: not trans format",
     "path": "the dump file folder",
     "net_name": "net name eg:ResNet50",
-    "mode": "0: dump all kernels 1: dump kernels in kernels list",
-    "iteration": "0: all iteration  others: specified iteration ",
-    "kernels": "kernel name list need to be dump"
+    "mode": "0: dump all kernels, 1: dump kernels in kernels list",
+    "iteration": "0: all iteration, others: specified iteration ",
+    "kernels": "op's full scope name which need to be dump"
   },
   "other": {}
-}
+}

config/e2e_dump_config_1.json

@@ -6,17 +6,17 @@
     "net_name": "ResNet50",
     "mode": 0,
     "iteration": 0,
-    "kernels": ["AllReduce","BiasAddGrad","Conv2DBackpropFilter","SparseSoftmaxCrossEntropyWithLogits"]
+    "kernels": ["Default/Conv2D-op2", "Default/TensorAdd-op10"]
   },
 
   "DumpSettingsSpec": {
-    "enable": "true: dump enable false: dump disable",
-    "trans_flag": "true: trans to host format,false: not trans format",
+    "enable": "true: dump enable, false: dump disable",
+    "trans_flag": "true: trans to host format, false: not trans format",
     "path": "the dump file folder",
     "net_name": "net name eg:ResNet50",
-    "mode": "0: dump all kernels 1: dump kernels in kernels list",
-    "iteration": "0: all iteration  others: specified iteration ",
-    "kernels": "kernel name list need to be dump"
+    "mode": "0: dump all kernels, 1: dump kernels in kernels list",
+    "iteration": "0: all iteration, others: specified iteration ",
+    "kernels": "op's full scope name which need to be dump"
   },
   "other": {}
-}
+}

mindspore/ccsrc/common/trans.cc

@@ -53,6 +53,7 @@ enum DataTypeTransMode {
   FROM_INT8_TO_FLOAT,
   FROM_INT8_TO_INT32,
   FROM_INT64_TO_INT32,
+  FROM_UINT16_TO_INT32,
 };
 
 const std::map<std::pair<TypeId, TypeId>, DataTypeTransMode> mode_map{
@@ -68,7 +69,8 @@ const std::map<std::pair<TypeId, TypeId>, DataTypeTransMode> mode_map{
   {std::pair<TypeId, TypeId>(kNumberTypeUInt8, kNumberTypeInt32), FROM_UINT8_TO_INT32},
   {std::pair<TypeId, TypeId>(kNumberTypeInt8, kNumberTypeFloat32), FROM_INT8_TO_FLOAT},
   {std::pair<TypeId, TypeId>(kNumberTypeInt8, kNumberTypeInt32), FROM_INT8_TO_INT32},
-  {std::pair<TypeId, TypeId>(kNumberTypeInt64, kNumberTypeInt32), FROM_INT64_TO_INT32}};
+  {std::pair<TypeId, TypeId>(kNumberTypeInt64, kNumberTypeInt32), FROM_INT64_TO_INT32},
+  {std::pair<TypeId, TypeId>(kNumberTypeUInt16, kNumberTypeInt32), FROM_UINT16_TO_INT32}};
 
 template <typename SrcT, typename DstT>
 void TransDataSrc2Dst(const TypeIdArgs &args, void *dst, const size_t data_size) {
@@ -116,6 +118,9 @@ bool CastKernel(const TypeIdArgs &args, void *dst, const size_t data_size, const
     case FROM_INT64_TO_INT32:
       TransDataSrc2Dst<int64_t, int32_t>(args, dst, data_size);
       break;
+    case FROM_UINT16_TO_INT32:
+      TransDataSrc2Dst<uint16_t, int32_t>(args, dst, data_size);
+      break;
     default:
       MS_LOG(ERROR) << "unsupported datatype trans";
       return false;

mindspore/ccsrc/device/ascend/ascend_device_address.cc

@@ -106,13 +106,13 @@ bool AscendDeviceAddress::SyncDeviceToHost(const std::vector<int> &shape, size_t
     } else {
       auto shape_size = trans::ShapeSize(host_shape);
       auto host = std::vector<uint8_t>(size_);
-      const trans::TypeIdArgs type_args{ptr_, shape_size, type_id_, type};
-      sync_ok = trans::TransDataType(type_args, host.data());
+      SyncMemory(host.data(), ptr_, size_, RT_MEMCPY_DEVICE_TO_HOST);
+      const trans::TypeIdArgs type_args{host.data(), shape_size, type_id_, type};
+      sync_ok = trans::TransDataType(type_args, host_ptr);
       if (!sync_ok) {
         MS_LOG(ERROR) << "trans data type failed.";
         return false;
       }
-      SyncMemory(host_ptr, host.data(), size, RT_MEMCPY_DEVICE_TO_HOST);
     }
   } else if (format_ == kOpFormat_NC1HWC0 || format_ == kOpFormat_FRAC_Z || format_ == kOpFormat_FRAC_NZ) {
     sync_ok = SyncDeviceToHostAndConvertFormat(shape, size, type, host_ptr);

mindspore/ccsrc/device/ascend/ascend_kernel_runtime.cc

@@ -150,9 +150,9 @@ void DumpOutput(mindspore::session::KernelGraph *graph, const string &dump_path,
     auto output_size = AnfAlgo::GetOutputTensorNum(node);
     for (size_t j = 0; j < output_size; ++j) {
       auto addr = AnfAlgo::GetOutputAddr(node, j);
-      auto shape = AnfAlgo::GetOutputDeviceShape(node, j);
-      auto type = AnfAlgo::GetOutputDeviceDataType(node, j);
-      auto format = AnfAlgo::GetOutputFormat(node, j);
+      auto shape = AnfAlgo::GetOutputInferShape(node, j);
+      auto type = AnfAlgo::GetOutputInferDataType(node, j);
+      auto format = kOpFormat_DEFAULT;
       string filepath = dump_path + '/' + kernel_name + '_' + "output_" + std::to_string(j);
       auto ascend_addr = dynamic_cast<const mindspore::device::ascend::AscendDeviceAddress *>(addr);
       std::vector<int> int_shapes;
@@ -181,9 +181,9 @@ void DumpParameters(mindspore::session::KernelGraph *graph, const string &dump_p
       continue;
     }
     auto addr = AnfAlgo::GetOutputAddr(item, PRAMATER_OUTPUT_INDEX);
-    auto shape = AnfAlgo::GetOutputDeviceShape(item, PRAMATER_OUTPUT_INDEX);
-    auto type = AnfAlgo::GetOutputDeviceDataType(item, PRAMATER_OUTPUT_INDEX);
-    auto format = AnfAlgo::GetOutputFormat(item, PRAMATER_OUTPUT_INDEX);
+    auto shape = AnfAlgo::GetOutputInferShape(item, PRAMATER_OUTPUT_INDEX);
+    auto type = AnfAlgo::GetOutputInferDataType(item, PRAMATER_OUTPUT_INDEX);
+    auto format = kOpFormat_DEFAULT;
     string filepath = dump_path + '/' + parameter_name + '_' + "output_0";
     auto ascend_addr = dynamic_cast<const mindspore::device::ascend::AscendDeviceAddress *>(addr);
     std::vector<int> int_shapes;