!300 refactor kernel select priority scheme

Merge pull request !300 from lianliguang/refactor-kernel_select

mindspore/ccsrc/device/ascend/kernel_select_ascend.cc

@@ -31,12 +31,13 @@ namespace mindspore {
 namespace device {
 namespace ascend {
 namespace {
+const float kWegihtBaseScore = 1;
+const float kFeatureMapBaseScore = 10;
 enum MatchCountPriority : int {
   MATCH_COUNT_PRIORITY_BEGIN = 0,
   MATCH_DTYPE_COUNT = MATCH_COUNT_PRIORITY_BEGIN,
   MATCH_FORMAT_COUNT,
   MATCH_SPECIAL_FORMAT_COUNT,
-  MATCH_5D_FORMAT_COUNT,
   MATCH_OUTPUT_DTYPE_COUNT,
   MATCH_COUNT_PRIORITY_END
 };
@@ -82,13 +83,6 @@ bool IsValidKernelInfo(const std::shared_ptr<CNode> &kernel_node, const kernel::
     }
     return true;
   };
-  if (AnfAlgo::GetCNodeName(kernel_node) == "Adam") {
-    auto input_num = AnfAlgo::GetInputTensorNum(kernel_node);
-    if (AnfAlgo::GetPrevNodeOutputFormat(kernel_node, input_num - 1) !=
-        kernel_build_info.GetInputFormat(input_num - 1)) {
-      return false;
-    }
-  }
   if (AnfAlgo::GetCNodeName(kernel_node) == prim::kPrimCast->name()) {
     return AnfAlgo::GetOutputInferDataType(kernel_node, 0) == kernel_build_info.GetOutputDeviceType(0) &&
            AnfAlgo::GetPrevNodeOutputInferDataType(kernel_node, 0) == kernel_build_info.GetInputDeviceType(0);
@@ -112,21 +106,7 @@ bool MatchInferOutputDataType(const CNodePtr &cnode, const kernel::KernelBuildIn
   MS_EXCEPTION_IF_NULL(cnode);
   // Check input data type
   for (size_t input_index = 0; input_index < kernel_build_info.GetInputNum(); ++input_index) {
-    AnfNodePtr cur_input = AnfAlgo::GetInputNode(cnode, input_index);
-    MS_EXCEPTION_IF_NULL(cur_input);
-    TypeId input_origin_type;
-    if (cur_input->isa<Parameter>() && AnfAlgo::IsParameterWeight(cur_input->cast<ParameterPtr>())) {
-      // weight
-      input_origin_type = AnfAlgo::GetOutputDeviceDataType(cur_input, 0);
-    } else if (cur_input->isa<ValueNode>()) {
-      input_origin_type = AnfAlgo::GetOutputDeviceDataType(cur_input, 0);
-    } else {
-      // feature map
-      input_origin_type = AnfAlgo::GetPrevNodeOutputInferDataType(cnode, input_index);
-    }
-    if (input_origin_type == kTypeUnknown) {
-      continue;
-    }
+    TypeId input_origin_type = AnfAlgo::GetPrevNodeOutputInferDataType(cnode, input_index);
     if (kernel_build_info.GetInputDeviceType(input_index) != input_origin_type) {
       return false;
     }
@@ -140,6 +120,29 @@ bool MatchInferOutputDataType(const CNodePtr &cnode, const kernel::KernelBuildIn
   return true;
 }
 
+string GetPriorityMatchFormat(const CNodePtr &cnode) {
+  string priority_matched_format = kOpFormat_NC1HWC0;
+  bool is_init = false;
+  bool need_change_nd = false;
+  for (size_t index = 0; index < AnfAlgo::GetInputTensorNum(cnode); ++index) {
+    auto pre_output_format = AnfAlgo::GetPrevNodeOutputFormat(cnode, index);
+    if (AnfAlgo::IsFeatureMapInput(cnode, index) &&
+        kNeedTransFormatSet.find(pre_output_format) != kNeedTransFormatSet.end()) {
+      priority_matched_format = !is_init ? priority_matched_format : pre_output_format;
+      is_init = true;
+    }
+    // feature map has two or more special format;
+    if (priority_matched_format != pre_output_format && pre_output_format != kOpFormat_DEFAULT) {
+      priority_matched_format = kOpFormat_DEFAULT;
+    }
+    auto input_shape_size = AnfAlgo::GetPrevNodeOutputInferShape(cnode, index).size();
+    need_change_nd = (need_change_nd || (input_shape_size != 4 && input_shape_size > 1));
+  }
+  if (need_change_nd) {
+    priority_matched_format = kOpFormat_DEFAULT;
+  }
+  return priority_matched_format;
+}
 /**
  * compare two vector by priority, select a better vector, like compare two num, first compare highest num location,
  * if equal then next num location
@@ -172,34 +175,18 @@ void UpdateCurMatchCounts(const kernel::KernelBuildInfo &kernel_build_info, cons
   if (cur_kernelinfo_match_counts->size() < MATCH_COUNT_PRIORITY_END) {
     MS_LOG(EXCEPTION) << "Out of range cur_kernelinfo_match_counts " << MATCH_COUNT_PRIORITY_END;
   }
+  auto pri_match_format = GetPriorityMatchFormat(kernel_node);
   for (size_t input_index = 0; input_index < AnfAlgo::GetInputTensorNum(kernel_node); ++input_index) {
-    AnfNodePtr input_anf_node = AnfAlgo::GetInputNode(kernel_node, input_index);
-    MS_EXCEPTION_IF_NULL(input_anf_node);
-    // if a input parameter is a weight with default format, the input shouldn't participate the judge
-    if (input_anf_node->isa<Parameter>()) {
-      auto para = input_anf_node->cast<ParameterPtr>();
-      if (AnfAlgo::IsParameterWeight(para) && AnfAlgo::GetOutputDeviceDataType(para, 0) == kTypeUnknown) {
-        continue;
-      }
-    }
+    auto base_score = AnfAlgo::IsFeatureMapInput(kernel_node, input_index) ? kFeatureMapBaseScore : kWegihtBaseScore;
     if (kernel_build_info.GetInputFormat(input_index) == AnfAlgo::GetPrevNodeOutputFormat(kernel_node, input_index)) {
-      if (AnfAlgo::IsFeatureMapInput(kernel_node, input_index) &&
-          kNeedTransFormatSet.find(kernel_build_info.GetInputFormat(input_index)) != kNeedTransFormatSet.end()) {
-        (*cur_kernelinfo_match_counts)[MATCH_SPECIAL_FORMAT_COUNT]++;
-      }
-      (*cur_kernelinfo_match_counts)[MATCH_FORMAT_COUNT]++;
+      (*cur_kernelinfo_match_counts)[MATCH_FORMAT_COUNT] += base_score;
     }
     if (kernel_build_info.GetInputDeviceType(input_index) ==
         AnfAlgo::GetPrevNodeOutputDeviceDataType(kernel_node, input_index)) {
-      (*cur_kernelinfo_match_counts)[MATCH_DTYPE_COUNT]++;
+      (*cur_kernelinfo_match_counts)[MATCH_DTYPE_COUNT] += base_score;
     }
-    if (kernel_build_info.GetInputFormat(input_index) == kOpFormat_NC1HWC0) {
-      // input is from a feature map & this input's shape is not 4d
-      if (AnfAlgo::IsFeatureMapInput(kernel_node, input_index) &&
-          AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, input_index).size() != kShape4dDims) {
-        continue;
-      }
-      (*cur_kernelinfo_match_counts)[MATCH_5D_FORMAT_COUNT]++;
+    if (kernel_build_info.GetInputFormat(input_index) == pri_match_format) {
+      (*cur_kernelinfo_match_counts)[MATCH_SPECIAL_FORMAT_COUNT] += base_score;
     }
   }
 
@@ -207,7 +194,7 @@ void UpdateCurMatchCounts(const kernel::KernelBuildInfo &kernel_build_info, cons
     // cal count of same output dtype between abstract and kernel info
     if (kernel_build_info.GetOutputDeviceType(output_index) ==
         AnfAlgo::GetOutputInferDataType(kernel_node, output_index)) {
-      (*cur_kernelinfo_match_counts)[MATCH_OUTPUT_DTYPE_COUNT]++;
+      (*cur_kernelinfo_match_counts)[MATCH_OUTPUT_DTYPE_COUNT] += 1;
     }
   }
 }
@@ -517,7 +504,7 @@ void SelectKernelInfo(const CNodePtr &kernel_node) {
   std::vector<std::shared_ptr<kernel::KernelBuildInfo>> kernel_info_list;
   MS_EXCEPTION_IF_NULL(kernel_node);
   kernel::KernelQuery(kernel_node, &kernel_info_list);
-  std::vector<int> most_match_counts = {-1, -1, -1, -1, -1};
+  std::vector<int> most_match_counts = {-1, -1, -1, -1};
   int selected_index = -1;
   auto context_ptr = MsContext::GetInstance();
   MS_EXCEPTION_IF_NULL(context_ptr);
@@ -527,7 +514,7 @@ void SelectKernelInfo(const CNodePtr &kernel_node) {
   std::vector<int> node_mix_precision_datatype_index;
   std::vector<TypeId> node_mix_precision_datatype;
   for (size_t info_index = 0; info_index < kernel_info_list.size(); ++info_index) {
-    std::vector<int> cur_kernel_info_match_counts = {0, 0, 0, 0, 0};
+    std::vector<int> cur_kernel_info_match_counts = {0, 0, 0, 0};
     auto kernel_build_info = *(kernel_info_list[info_index]);
     if (!IsValidKernelInfo(kernel_node, kernel_build_info)) {
       continue;